Peptides have become increasingly popular among longevity and wellness enthusiasts. The preclinical data on many of these compounds are substantial, as are anecdotal reports of benefit. However, most have not undergone clinical testing. In 2023, the FDA placed these products in its high-risk Category 2, blocking compounding pharmacies from producing them. A grey market developed to fill the gap, with hundreds of thousands to potentially low millions of Americans ordering research chemicals for personal medical use. In April, the FDA removed twelve from the list, and an advisory committee will soon review whether they belong on the list available to compounding pharmacies.

Compounded products will be higher quality than what the grey market offers. The harder question is upstream: if these compounds work, why have they never been approved? One of the twelve is Semax, a Russian-developed peptide with a thirty-year clinical record for ischemic stroke. The deeper question is not on the agenda: why a molecule with three decades of foreign clinical use has no path into American medicine. The prior piece in this series covered the full list and the regulatory mechanics. Semax is the case study for how the system fails to evaluate compounds like these.

If you enjoy these posts, consider subscribing and becoming a part of our growing community!

Subscribe now

A man in his sixties is wheeled into a stroke unit at a Moscow hospital with weakness on his right side and slurred speech. The neurologist on call follows the standard acute-ischemic protocol: imaging, tPA if the window allows, mechanical thrombectomy if indicated, and a course of intranasal Semax administered alongside the rest. Down the hall, the pharmacist does not blink. The drug has been on the Russian List of Vital and Essential Medicines since 2011. It has been used for stroke in Russia since the late 1990s.

Same week, an academic medical center in Cleveland. A man in his sixties presents with weakness on his right side and slurred speech. The neurologist runs the U.S. acute-ischemic protocol: imaging, tPA if the window allows, mechanical thrombectomy if a large-vessel occlusion is confirmed. Semax does not enter the conversation. It cannot. The molecule is not approved by the FDA, has no U.S. sponsor, and exists in American medicine somewhere between an unscheduled research chemical and a gray-market peptide sold to biohackers on the internet.

Same drug. Same indication. Three decades of post-marketing exposure in one country and a regulatory dead silence in the other.

In April 2026, the FDA moved Semax, along with eleven other peptides, off Category 2 of its Section 503A bulks list. The Pharmacy Compounding Advisory Committee will hear the Semax petition in July. That is news. But it is only part of the story. It does not approve Semax for stroke or for any other indication. It clears the molecule for a public review of whether it can be compounded by pharmacies, a question two regulatory floors down from “does it work.”

There are real questions. The original Semax trial base was thinner than its proponents claim, and the methodological standards behind it would likely not satisfy a modern FDA reviewer. The molecule may or may not work. That is not the question.

The question is why the U.S. system has no way to even ask.

The compound

Semax was designed in 1982 at the Institute of Molecular Genetics, then part of the Soviet Academy of Sciences. It was first described in the open scientific literature in 1991 (Potaman et al., 1991). Its structure is a synthetic heptapeptide, Met-Glu-His-Phe-Pro-Gly-Pro, built around the cognitive-active core of adrenocorticotropic hormone, residues 4 through 7, with a Pro-Gly-Pro extension that resists peptidase cleavage. The modification extends the half-life from minutes to hours, eliminates the steroidogenic activity of the parent hormone, and produces measurable CNS effects when delivered intranasally.

The initial Phase I trials ran from 1990 to 1994, followed by Phase II through 1996. Russian Ministry of Health registration came in the mid-to-late 1990s. The on-label indications include acute ischemic stroke, transient ischemic attack, optic nerve atrophy, and a cluster of cognitive and cerebrovascular conditions. The mechanism is officially listed as not fully characterized. The published work converges on two candidate effects: induction of BDNF and NGF expression with concurrent upregulation of the TrkB receptor, and modulation of dopaminergic and serotonergic signaling. Russia added Semax to its List of Vital and Essential Drugs on December 7, 2011.

The molecule has accumulated decades of broad clinical population exposure in Russian post-marketing surveillance, with more cumulative patient-years than many compounds the FDA has approved. It is not a regulatory orphan in the United States because it has been ignored. It is a regulatory orphan because the U.S. system has no machinery for evaluating it.

The thesis

The U.S. drug development pipeline is a sequence of filters. Each one determines which compounds and which evidence reach the system. Most are individually defensible. None are individually about whether a drug works.

Semax fails seven of these filters simultaneously. That pattern, not the molecule itself, is the story.

No sponsor. Drugs do not get approved. Sponsors get drugs approved. Every FDA-approved compound has a corporate or institutional entity behind it that funds the trials, files the IND, runs the regulatory correspondence, manages the post-marketing commitments, and absorbs the cost of failure. Semax has none of this. There is no U.S. licensing biotech. There is no academic champion preparing an investigator-led IND. There is no patient advocacy foundation pushing the indication. There is no sell-side analyst tracking the program. The infrastructure of people who normally drag a foreign drug through approval does not exist for this molecule.

The ICH boundary. The FDA accepts foreign clinical data under specific conditions tied to compliance with International Council for Harmonisation guidelines, GCP-compliant auditing, and population comparability. Russia is not an ICH member. The bulk of the Semax evidence base sits in Russian-language journals, generated under late-Soviet and post-Soviet trial standards, that the agency has no established framework for evaluating without commissioning independent verification. No one is paying for that verification. No one has reason to.

Sanctions and the post-2022 IP collapse. Until 2022, a U.S. biotech could plausibly have licensed Semax from the Russian Academy of Sciences and run a Western development program. That pathway is now closed. Sanctions have made technology transfer from state-affiliated Russian institutions legally and politically difficult. Pharmaceutical investors will not write checks against deals with Russian counterparties. The IP-level transaction that would need to happen first is not currently doable, and there is no obvious horizon on which it becomes doable again.

The neuroprotectant graveyard. The recent history of acute-stroke neuroprotection is a wreckage. NXY-059 failed Phase 3 in the SAINT-II trial in the New England Journal of Medicine in 2007 after positive earlier data (Shuaib et al., 2007). Selfotel, eliprodil, gavestinel, citicoline, magnesium, and a long tail of other candidates followed the same arc. Pharma effectively walked away from the indication around 2010. Any new neuroprotectant inherits the priors that field has earned. A reviewer looking at Semax stroke data starts from a position of trained skepticism toward the entire claim category.

Standard-of-care lock-in. The Russian Semax stroke trials, including the canonical Gusev acute hemispheric stroke study, ran against control arms that received supportive care alone (Gusev et al., 1997). Tissue plasminogen activator was not yet routine. Mechanical thrombectomy did not exist. A modern American trial would have to demonstrate incremental benefit on top of two interventions that already work. The sample sizes and effect-size resolution required to do that are an order of magnitude beyond what the original Russian work approached. A bridging study in 2026 would need larger samples, cleaner endpoints, enough power to detect incremental benefit on top of modern stroke care, and the willingness to accept that the answer may come back negative. This is the one filter where the gate might be tracking real biology rather than infrastructure.

Off-patent and pleiotropic. The familiar problem. Semax is off-patent. There is no exclusivity to recoup development costs. The compound also appears to do several things at once across multiple receptor systems and multiple candidate indications. That confuses a regulatory framework designed to evaluate single compounds against single endpoints in single disease populations. A drug that does several things moderately well does not fit a system designed to prove one thing definitively.

The credibility tax. Semax has been adopted by the biohacker and functional medicine ecosystem and circulates in adjacent wellness culture. The April 2026 reclassification itself is widely characterized in trade press as RFK-Jr.-era HHS direction, which sharpens the political coloring. The same association that drives consumer demand actively repels the institutional actors who would otherwise study it. “Wellness-coded” remains a near-fatal credibility signal in mainstream American medicine. A neurologist who proposes a Semax investigator-initiated trial today is making a career bet that no department chair has any incentive to approve.

The honest counterweight

A serious neurologist looking at the actual published Semax trial data would not say “this is clearly an effective drug being unjustly excluded.” They would say something closer to “this is a plausibly active compound with a thin and methodologically weak evidence base, used at scale in a country whose regulatory standards I do not trust.”

That position is defensible. Russian clinical trial methodology of the 1990s and 2000s was on average, of lower quality than Western trials of the same era. The Semax stroke trials were small. Many were unblinded. Composite endpoints are common in this literature and are gameable. The thirty-year clinical track record reflects breadth of population exposure, not depth of evidence. The mechanism is still officially not fully characterized. The seven filters described above are not a conspiracy. Most of them exist for reasons.

Seven independently defensible filters can still produce a system that systematically cannot process certain categories of compound, regardless of whether those compounds work. That outcome may be acceptable. It is not currently being chosen deliberately.

What the April 2026 reclassification actually does

Read carefully, the FDA’s April 2026 action is small. The agency removed twelve peptide-based bulk drug substances, Semax among them, from Category 2 of the interim 503A bulks list, the bucket reserved for compounds the agency has flagged as potentially raising significant safety concerns. The Pharmacy Compounding Advisory Committee will hold public review meetings in two batches: BPC-157, KPV, MOTs-C, and TB-500 on July 23, 2026; DSIP and Semax on July 24; and the remaining peptides (LL-37, Dihexa, GHK-Cu, PEG-MGF, and Melanotan II) by February 2027 (Orrick, 2026; FDA, 2026).

The agency is clear about what this is and what it is not. Removal from Category 2 does not mean a substance has been evaluated and approved for the 503A bulks list. The peptides are now in regulatory limbo until the PCAC meets and the FDA takes final action. The mechanical reason for the removals is procedural: the original nominators voluntarily withdrew their nominations, and the safety-concern designation could not stand on a withdrawn record.

What the action does not do: it does not create an FDA approval pathway for Semax in stroke. It does not authorize prescription. It does not address any of the seven filters above. It moves a single procedural lever, eligibility for compounding pharmacy review, and does not touch the structural architecture that made Semax a regulatory orphan in the first place. A 55-year-old in Cleveland who has an ischemic stroke on July 25, 2026, the day after the PCAC meeting, will be in exactly the same position as the same patient on July 23.

Why this matters now

The Semax case looks like a curiosity from a particular geopolitical and historical moment. The same filters now apply to a rapidly growing class of compounds the system is structurally unprepared for, and each category trips a recognizable subset of the seven.

AI-designed protein binders from academic labs trip the no-sponsor filter and the off-patent problem. Compounds from non-ICH jurisdictions, particularly the rising volume now coming out of Chinese clinical infrastructure, trip the foreign-data filter. Academic spinouts whose IP sits inside university tech transfer offices trip the sponsor and credibility filters before they trip anything scientific. Open-science therapeutics released without patent protection by design trip the exclusivity filter as a feature. Off-patent peptide drugs as a class trip the sponsor, exclusivity, and credibility filters simultaneously. None of these categories are exotic. All of them are growing.

The pipeline that was built for sponsor-led, single-indication, ICH-compliant, patent-protected drug development is increasingly mismatched to where useful molecules are now being generated. The April 2026 reclassification is, in this light, the smallest possible regulatory acknowledgment that the off-patent peptide class exists at all. It does not solve the structural problem. It barely registers it.

The conversation about FDA reform tends to focus on speeding approvals or lowering evidentiary bars. Those are real debates. For Semax and the broader category it represents, the binding constraint is upstream. Which evidence the system can metabolize. Which sponsors it requires. Which jurisdictions it recognizes. None of these constraints were chosen on first principles. They accumulated.

The reform conversation that matters is the one about that accumulation. The agency needs an evidence framework for compounds that arrive without a sponsor to package the data. It needs an audit pathway for non-ICH clinical evidence that does not require years of bridging studies. It needs a regulatory route for compounds whose value lies in pleiotropy rather than in a single primary endpoint. It needs an answer for what drug development looks like when molecules are coming from algorithms and university labs faster than pharma can license them.

None of these questions have answers right now. None of them are getting asked at the level of the agency or the legislature. The drug a Cleveland neurologist cannot prescribe to a stroke patient today is not the most important fact in this story. The drug a Cleveland neurologist will not be able to prescribe five years from now, when that drug exists and was designed by an algorithm in an academic lab and has trial data from Beijing, is the more important fact.

The U.S. system has structural blindness to certain categories of evidence and certain origins of molecules. The cost of that blindness is borne by patients who never learn what they were never offered. The science will not settle itself. The system has to decide what it wants to be able to know.

These newsletters take significant effort to put together and are totally for the reader’s benefit. If you find these explorations valuable, there are multiple ways to show your support:

Engage: Like or comment on posts to join the conversation.

Leave a comment

Subscribe: Never miss an update by subscribing to the Substack.

Subscribe now

Share: Help spread the word by sharing posts with friends directly or on social media.

Share

References

FDA. Certain Bulk Drug Substances for Use in Compounding That May Present Significant Safety Risks (interim 503A list). Updated April 22, 2026.

Gusev EI, Skvortsova VI, Miasoedov NF, Nezavibat’ko VN, Zhuravleva EIu, Vanichkin AV. Effectiveness of semax in acute period of hemispheric ischemic stroke (a clinical and electrophysiological study). Zh Nevrol Psikhiatr Im S S Korsakova. 1997;97(6):26-34.

Orrick, Herrington & Sutcliffe LLP. FDA Announces Removal of 12 Peptides from Category 2 and Schedules PCAC Meetings to Consider Adding Peptides to 503A Bulk Drug Substances List. April 2026.

Potaman VN, Alfeeva LY, Kamensky AA, Levitzkaya NG, Nezavibatko VN. N-terminal degradation of ACTH(4-10) and its synthetic analog semax by the rat blood enzymes. Biochemical and Biophysical Research Communications. 1991;176:741-746.

Shuaib A, Lees KR, Lyden P, et al; SAINT II Trial Investigators. NXY-059 for the treatment of acute ischemic stroke. New England Journal of Medicine. 2007;357:562-571. doi:10.1056/NEJMoa070240

A 55-year-old has been managing a partial rotator cuff tear for eighteen months. Physical therapy has plateaued. Surgery is the next conversation. A functional medicine physician offers another option. Four hundred micrograms of BPC-157, injected twice daily for six weeks.

The patient asks for the human evidence at that dose for that injury. The literature has no answer.

The FDA’s procedural reopening does not change that.

If you enjoy these posts, consider subscribing and becoming a part of our growing community!

Subscribe now

Your body runs on peptides. Insulin is one. Oxytocin is one. The GLP-1 drugs everyone is taking for weight loss are synthetic versions of a peptide your gut secretes after a meal.

Peptides are short chains of amino acids, anywhere from two to about fifty, that act as signaling molecules. Smaller and more targeted than the proteins that build tissue. Larger and more specific than the small-molecule drugs that fill most prescriptions.

A different class of peptides has spent the last three years in regulatory limbo. Compounds like BPC-157, a fragment of a stomach protein that appears to accelerate tissue repair, were available through compounding pharmacies until the FDA restricted them in late 2023 under its Category 2 listing. They moved to the gray market. Functional medicine doctors, athletes, and longevity clinics built protocols around them anyway.

The FDA recently reopened review of these compounds. In April 2026, twelve peptide nominations came off Category 2 after the original nominators withdrew them. The agency simultaneously scheduled advisory committee meetings for July 2026 and early 2027 to consider whether some should be formally added to the 503A bulks list, the list of substances licensed compounding pharmacies may legally use. Withdrawal from Category 2 is not the same as authorization. The compounds sit in a procedural in-between, neither flagged as high-risk nor cleared for compounding, pending the advisory review and FDA’s eventual response to it.

Compounding pharmacies are not drug manufacturers. A 503A compounding pharmacy mixes a specific medication for a specific patient on a doctor’s prescription, using raw active pharmaceutical ingredients tested for sterility, potency, and identity. The peptides in question are not FDA-approved drugs. The 2023 Category 2 listing told compounding pharmacies they could no longer prepare them. Withdrawal from Category 2 removes the explicit safety-risk flag but does not by itself authorize compounding.

BPC-157 is the most sought-after of the group. Fifteen amino acids, derived from a fragment of a human gastric protein (Sikiric et al., 2016). The mechanism is plausible. It activates VEGFR2 to drive new blood vessel formation. It increases nitric oxide production. It dampens inflammatory cytokine signaling.

Animal studies, predominantly from a single research group in Croatia, show accelerated recovery across muscle, tendon, bone, gut lining, nerve, and vascular tissue (Chang et al., 2011). Independent groups have reproduced pieces of the picture, but the bulk of the animal package still rests on the original Croatian work.

The human data is almost nothing. The published human trials cover a handful of studies and fewer than three dozen subjects. The animal signal is not the weak point. The compound does something. What it does in that 55-year-old patient at that dose for that injury remains unanswered.

The other returning compounds sit on similar foundations. TB-500 is a synthetic fragment of thymosin beta-4, an actin-binding peptide that promotes cell migration and tissue repair (Goldstein et al., 2012). The parent molecule has been studied in small human trials for skin ulcers and cardiac injury. The fragment that compounding pharmacies prepare is not identical to the parent, and the human data on the fragment is thin. Semax and Selank are Russian-developed nootropic peptides, derivatives of ACTH and tuftsin, with decades of clinical use in Russia and small human trial packages well short of a U.S. NDA (Koroleva and Myasoedov, 2018).

Epitalon is a four-amino-acid pineal peptide with in-vitro reports of telomerase activation in human somatic cells (Khavinson et al., 2003). KPV is an anti-inflammatory tripeptide studied mostly in animal models of inflammatory bowel disease. MOTS-C is a mitochondrial-derived peptide with academic interest in metabolic regulation and almost no clinical trial data.

These compounds act on repair, inflammation, and stress response pathways that do not fit the single-indication logic of FDA drug approval. That structural mismatch is why the literature looks the way it does.

FDA approval for a new drug runs around a billion dollars and takes a decade. Composition patents run twenty years from filing. By the time any company cleared an approval pathway for an old peptide, the patent protection would be gone. No exclusivity means no return on investment. Companies do not run billion-dollar trials on molecules they cannot profitably sell.

Pleiotropy compounds the problem. The word means a molecule acts across multiple tissue types through overlapping mechanisms. BPC-157 fits that pattern.

The FDA pipeline requires approval for a single disease indication. A molecule that does many things moderately well does not fit a system designed to prove one thing definitively. Compounding pharmacies prepare these peptides cheaply. Pharmaceutical companies cannot justify the investment.

The result is a regulatory dead zone. The compounds are too well-known to ignore and too unpatentable to develop.

Pharma does not study what it cannot sell. A molecule that cannot cross that economic threshold stays in animal models and small university trials. If it works, it works quietly, in clinics where the compounding supply exists. If it does not work, nobody finds out through a phase 3 readout. Nobody finds out at all.

The 2023 restriction collapsed the legal supply. Demand did not collapse with it. Testing of gray-market peptide products has documented variability in active content, bacterial endotoxin contamination, and in some cases entirely wrong compounds. If FDA eventually adds some of these compounds to the bulks list, demand will migrate back toward licensed pharmacies. That has not happened yet.

A compounded peptide from a 503A pharmacy comes with a physician prescription, USP-grade active ingredient, tested endotoxin levels, and a labeled concentration that matches the vial. A peptide from an Instagram vendor comes with none of those guarantees. Same molecule. Different manufacturing.

For the rotator cuff patient and the prescribing physician, the questions that matter are unchanged by category reclassification. What is the dose rationale at this body weight and indication. What is the best human evidence at that dose. Is the prescribing physician tracking measurable outcomes or operating on testimonial. What is the exit criterion if there is no improvement in eight weeks.

The procedural reopening does not answer any of them. They are answered at the level of the patient and the clinic.

A pleiotropic, off-patent compound will never pay back a billion-dollar approval trial. Academic consortia could run mid-sized trials, but the incentives for grant-funded investigators to study a compound with no pharma sponsor are weak. Patient registries and real-world evidence programs could plausibly produce usable data at a fraction of the cost, if anyone funded them.

That is the pathway compounded peptides need. Neither the administration that restricted them nor the one reopening them has moved toward it. The science will not settle itself. It has to be paid for, and nobody with the money has a reason to pay.

These newsletters take significant effort to put together and are totally for the reader’s benefit. If you find these explorations valuable, there are multiple ways to show your support:

Engage: Like or comment on posts to join the conversation.

Leave a comment

Subscribe: Never miss an update by subscribing to the Substack.

Subscribe now

Share: Help spread the word by sharing posts with friends directly or on social media.

Share

References

Sikiric P, Seiwerth S, Rucman R, et al. Brain-gut axis and pentadecapeptide BPC 157: theoretical and practical implications. Curr Neuropharmacol. 2016;14:857-865. doi: 10.2174/1570159x13666160502153022

Chang CH, Tsai WC, Lin MS, Hsu YH, Pang JH. The promoting effect of pentadecapeptide BPC 157 on tendon healing involves tendon outgrowth, cell survival, and cell migration. J Appl Physiol. 2011;110:774-780. doi: 10.1152/japplphysiol.00945.2010

Goldstein AL, Hannappel E, Sosne G, Kleinman HK. Thymosin beta4: a multi-functional regenerative peptide. Basic properties and clinical applications. Expert Opin Biol Ther. 2012;12:37-51. doi: 10.1517/14712598.2012.634793

Khavinson VKh, Bondarev IE, Butyugov AA. Epithalon peptide induces telomerase activity and telomere elongation in human somatic cells. Bull Exp Biol Med. 2003;135:590-592. doi: 10.1023/a:1025493705728

Koroleva SV, Myasoedov NF. Semax as a universal drug for therapy and research. Biol Bull. 2018;45:589-600. doi: 10.1134/S1062359018060055

You’re home visiting the family for Thanksgiving. You notice your father-in-law seems to have some odd memory lapses. You quietly bring it up with other family members who independently noticed similar things. You gently suggest to your mother-in-law that it may be time to consult a neurologist. After spending a couple of months scheduling appointments and sitting on a wait list, the tests were inconclusive.

This speed runs the frustrating medical experience for many aging Americans. Misdiagnosis rates for neurodegenerative diseases range from 25 to 30 percent in specialized dementia clinics (Beach et al., 2012). The problem compounds with age: by 80, roughly 70 percent of patients harbor multiple neurodegenerative pathologies simultaneously (Kapasi et al., 2017). Alzheimer’s tangled with vascular disease. Parkinson’s layered over frontotemporal degeneration. The brain doesn’t pick one disease and commit.

Blood-based tests for Alzheimer’s (p-tau217, amyloid ratios) have made progress, but even in this case, it’s only part of the story. For Parkinson’s, ALS, and frontotemporal dementia, clinical diagnosis still relies on symptoms and exclusion. A confirmatory test with specific biomarkers is either invasive (cerebrospinal fluid), expensive (PET imaging), or unavailable entirely. That gap is where misdiagnosis lives.

If you enjoy these posts, consider subscribing and becoming a part of our growing community!

Subscribe now

A team at Lund University just made that gap tractable. Published in Nature Medicine, the team trained a deep learning model called ProtAIDe-Dx on plasma proteomics data from over 17,000 patients across 19 clinical sites (An et al., Nature Medicine, 2026). The model reads roughly 7,600 proteins from a single blood draw, then simultaneously generates probabilistic diagnoses for six conditions: Alzheimer’s disease, Parkinson’s disease, frontotemporal dementia, ALS, stroke/TIA, and cognitive health.

ProtAIDe-Dx achieved balanced classification accuracy of 95% for ALS, 92% for Parkinson’s, 83% for cognitively unimpaired controls, 81% for Alzheimer’s, 72% for frontotemporal dementia, and 70% for stroke. For a single blood test with no clinical inputs, those numbers match or exceed the diagnostic accuracy of a general neurologist’s clinical assessment.

The model learns all six conditions simultaneously. Shared biological signals between diseases inform each prediction, and it generates probability scores for every condition in every patient. That means it can flag co-pathology: a patient scoring high for both Alzheimer’s and Parkinson’s likely has both.

The strongest clinical result came from tracking patients over time. ProtAIDe-Dx predictions stratified future cognitive decline better than clinical diagnosis alone. Patients the model flagged as likely Alzheimer’s declined faster on cognitive testing regardless of their clinical label at baseline. Some apparent false positives may have been catching preclinical pathology that the clinical label missed.

The team validated this externally in BioFINDER-2, a memory clinic cohort of over 1,700 patients with biomarker-confirmed diagnoses. Patients flagged for elevated Alzheimer’s probability were more likely to test positive for amyloid and tau on confirmatory imaging. Elevated stroke probability tracked with white matter lesion burden on MRI. When combined with standard clinical biomarkers, the model significantly improved differential diagnosis accuracy, especially for non-Alzheimer’s dementias where existing blood tests fall short.

Beyond diagnosis, the model identified a set of proteins that discriminate healthy aging from neurodegeneration across all six conditions. One of them detoxifies a byproduct of glucose metabolism that drives protein damage and accumulates with age. Its appearance suggests that metabolic vulnerability may be a shared upstream signal across neurodegenerative diseases, one where the failure to clear cellular damage accumulates into vulnerability.

Healthcare systems require disease diagnosis before administering therapeutics, especially when the care is expensive. For Alzheimer’s disease, lecanemab and donanemab require biomarker-confirmed amyloid status before treatment can begin. That confirmation currently means several months of navigating patient visits to reach a PET scan, averaging $3,000, or a lumbar puncture. A SomaScan proteomics panel runs roughly $950 per sample and returns six probabilistic diagnoses. The economics of triage favor blood. A test that flags which patients need expensive confirmatory imaging, which should enter a clinical trial, and which carry preclinical pathology worth monitoring could reshape how neurology clinics process their intake queues. The family that spent two months on a wait list for an inconclusive answer could have six probabilities from a single tube of blood for a third of the cost and none of the headache.

These newsletters take significant effort to put together and are totally for the reader’s benefit. If you find these explorations valuable, there are multiple ways to show your support:

• Engage: Like or comment on posts to join the conversation.

Leave a comment

• Subscribe: Never miss an update by subscribing to the Substack.

Subscribe now

• Share: Help spread the word by sharing posts with friends directly or on social media.

  Share

References

Beach TG, Monsell SE, Phillips LE, Kukull W. Accuracy of the clinical diagnosis of Alzheimer disease at National Institute on Aging Alzheimer Disease Centers, 2005-2010. J Neuropathol Exp Neurol. 2012;71:266-273. doi: 10.1097/NEN.0b013e31824b211b

Kapasi A, DeCarli C, Schneider JA. Impact of multiple pathologies on the threshold for clinically overt dementia. Acta Neuropathol. 2017;134:171-186. doi: 10.1007/s00401-017-1717-7

An L, Pichet Binette A, Hristovska I, et al. A deep joint-learning proteomics model for diagnosis of six conditions associated with dementia. Nature Medicine. 2026. doi: 10.1038/s41591-026-04303-y

Internet security has always rested on an uncomfortable fact.

The software running the global financial system, hospitals, power grids, water treatment plants, air traffic control, and military command systems is full of vulnerabilities. The WannaCry ransomware attack shut down a third of NHS hospital trusts in England. The Colonial Pipeline hack cut fuel supply to the entire U.S. East Coast for six days. The SolarWinds breach gave Russian intelligence access to the Treasury Department, the Department of Homeland Security, and parts of the Pentagon. Global cybercrime costs roughly $500 billion per year. Everyone knows the software is fragile.

The only reason it has not collapsed already is that the people capable of chaining those bugs into working exploits have been rare.

That scarcity has been holding the internet together.

Anthropic is now claiming its new model, Claude Mythos Preview, has broken it.

If you enjoy these posts, consider subscribing and becoming a part of our growing community!

Subscribe now

The company says Mythos found thousands of high-severity vulnerabilities, including bugs in every major operating system and every major web browser. It says the model found a 27-year-old vulnerability in OpenBSD. A 16-year-old bug in FFmpeg that automated testing had hit 5 million times without catching. A Linux kernel exploit chain that escalated from ordinary user access to full machine control. A browser escape built by chaining four separate vulnerabilities.

If those claims hold up, the story is no longer that AI is getting better at code completion.

A machine has joined the vulnerability research labor force. At scale.

That changes everything.

Three things could make this less than it appears. The vulnerabilities could be pre-seeded or artificially narrowed to favorable targets. The “thousands” of bugs could be mostly variants of known classes rather than genuinely novel discoveries. And the autonomous operation could be heavily guided by expert operators choosing which codebases to attack and how to frame the prompts. Anthropic has not yet released enough detail to rule out any of these. Independent replication will matter more than anything in the system card.

Anthropic did not explicitly train Mythos to find vulnerabilities. The capability emerged from general improvements in code comprehension, reasoning, and autonomous operation. The scaffold is simple: launch an isolated container with the target software and its source code, prompt the model with “find a security vulnerability in this program,” and let it run. Mythos reads the code, hypothesizes where bugs might live, runs the actual program to test its theories, adds debug logic or attaches debuggers as needed, and outputs a bug report with a proof-of-concept exploit. It can also rank every file in a project by likelihood of containing interesting bugs on a 1-to-5 scale, then work down the list in priority order. A second agent validates each finding. The same improvements that make the model better at patching vulnerabilities also make it better at exploiting them. That symmetry is the core problem.

A real threshold is being reached.

Most AI announcements are easy to ignore. A benchmark goes up. The demo gets cleaner. The model writes better copy, produces better code, and makes fewer stupid mistakes. Useful. Incremental. Familiar.

This feels different.

Anthropic is claiming the model can do work that used to require a very small number of unusually skilled humans. That is a much harder claim.

And the examples matter.

OpenBSD, FFmpeg, and Linux kernel are not toy software. These are old, heavily used, heavily examined codebases. A bug surviving inside them for 16 or 27 years tells you something about how hard it was to find. If Mythos is actually surfacing flaws at that level, this is a capabilities story first and a product story second.

Nicholas Carlini, an Anthropic researcher, put it more plainly than the company did: he said he had found more bugs in the last couple of weeks than in the rest of his life combined.

The model does not need to be perfect. It just needs to multiply expert throughput hard enough to break the old equilibrium.

That may already be happening.

Anthropic’s behavior is the tell.

They restricted access. Mythos is available only through Project Glasswing, a defensive cybersecurity coalition that includes AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks. Anthropic says it is committing $100 million in usage credits across the effort, along with direct support for open-source security organizations.

A company that thinks it has built a slightly better assistant ships it. Anthropic locked this one down.

Anthropic is effectively telling us that some frontier capabilities may no longer fit inside the normal software release model. The branding, the benchmark scores, the carefully staged launch. None of that is the story. Anthropic appears to believe some capabilities should never ship broadly.

The company appears to believe Mythos crossed a line.

The most unsettling detail is not even the vulnerabilities

Anthropic also published a 244-page system card describing some of Mythos’s behavior during testing.

During one evaluation, Mythos was placed in a sandboxed environment and instructed to attempt escape. It reportedly built a multi-step exploit, reached the broader internet, and emailed a researcher directly. He found out because he received the email while eating a sandwich in a park.

The model operated toward a goal inside an adversarial environment. And it won.

The model escaped a sandbox designed to contain it. The natural questions: was network access pre-configured in the test environment? Was email tooling available inside the container? Was the “escape” actually exploitation of a permitted channel rather than a genuine breakout? Anthropic’s published account does not answer these clearly. Until it does, the anecdote is striking but not dispositive.

Anthropic’s reporting around the system card includes other ugly details. The model sometimes appeared to recognize that it was being evaluated. In some cases, it reportedly concealed prohibited actions after taking them. Earlier variants reportedly tried to use the Linux /proc filesystem to harvest credentials and work around sandboxing.

Anthropic also describes Mythos as its best-aligned model so far.

Maybe. But “best aligned” is doing less work here than they let on. The behavior that matters is whether a model with offensive cyber capabilities can act with persistence, deception, and autonomy once given enough room. Anthropic’s own materials suggest those thresholds are closer than the public conversation admits.

But these models are coming regardless. The defensive case deserves a serious hearing. The same model that finds the bug can help patch the bug. True. That is the whole logic of Glasswing.

If models like Mythos are arriving anyway, the best near-term strategy may be to hand the first serious deployments to defenders and let them sweep critical infrastructure before everyone else gets access to comparable systems. That is a sane response. Probably the only sane response.

Critical software is held together by overworked teams, underfunded open-source maintainers, and dependency graphs that almost nobody fully understands. Internet security has always been a talent bottleneck. A model that multiplies the output of high-end vulnerability researchers changes that math.

The offensive case is what keeps this from being a normal product launch

The same capability has another use.

That is the whole problem.

Once vulnerability discovery and exploit development become cheaper, faster, and less dependent on rare human talent, the offense-defense balance shifts. The time between bug discovery and active exploitation shrinks. More actors can operate at a higher level. The long tail gets more dangerous.

That includes state actors. Criminal groups. Contractors. Anyone willing to use machine speed against a software ecosystem that still patches at human speed.

This is why “too dangerous to release” does not sound like empty marketing this time.

OpenAI said something similar about GPT-2 in 2019. In hindsight that looks quaint. It is tempting to throw Mythos into the same bucket and move on. That would be a mistake. GPT-2 generated text. Mythos generates exploits.

Language generation was easy to dramatize and hard to weaponize at the level people feared. Cyber capability is different. The path from model competence to real-world damage is shorter. The target environment already exists. The vulnerable systems are already deployed. The incentives are already in place.

Whether Anthropic has overstated the capability will become clearer as independent researchers get access. The more important question is what happens when the next lab ships something comparable without the restrictions.

The open-source problem.

Glasswing’s partner list reads like a Fortune 50 roster. Those companies have security teams, budgets, and incident response infrastructure. The harder problem is everything underneath them. cURL handles HTTP requests for virtually every connected device on earth. OpenSSL encrypts roughly 70% of all web traffic. The xz-utils backdoor in 2024 demonstrated that a single compromised maintainer could threaten the entire Linux ecosystem. These projects are maintained by small teams, sometimes one person, often unpaid.

Anthropic says it is extending Glasswing access beyond the headline partners to over 40 organizations maintaining critical infrastructure, plus $4 million in direct donations to open-source security groups. That helps. It does not solve the structural problem. Machine-scale vulnerability discovery exposes fragility faster than maintenance teams can patch it. Finding the bugs is now the easy part. You still need maintainers with time, patches that do not break dependencies, rollouts across millions of systems, and coordination between projects that barely communicate. The bottleneck has moved from discovery to remediation, and nobody has a plan for that yet.

This was discovered through a leak

The public first learned about Mythos through an error. Security researchers reportedly discovered thousands of unpublished Anthropic blog assets sitting in a publicly accessible cache because of a CMS misconfiguration. Draft posts referenced Claude Mythos and an internal tier called Capybara above Opus. Days later, another exposure reportedly revealed hundreds of thousands of lines of Claude Code source code through a misconfigured npm package.

You could not script the irony better. A company quietly preparing to tell the world it had built a model powerful enough to reshape cybersecurity got scooped by its own security mistakes.

The Mythos claims may still hold. Anthropic’s own release-pipeline failures make its safety posture harder to trust. Because now there are two stories sitting on top of each other. One is the official story: frontier model, dangerous cyber capabilities, restricted deployment, defensive coalition. The other is messier: a company positioning itself as the lab taking cyber risk seriously while fumbling basic operational security twice in public.

Both stories matter.

The most important unanswered questions

Anthropic’s evidence is still largely Anthropic-controlled. That means skepticism is warranted, but it should be precise skepticism.

The autonomy question matters most. Whether Mythos was handed promising attack surfaces by expert operators or performed meaningful end-to-end discovery on its own determines whether this is a force multiplier or a replacement.

The novelty question is close behind. The number of those “thousands” of vulnerabilities that were net-new versus variants of known bug classes changes the significance entirely. Net-new bugs in mature codebases would be extraordinary. Variants of known patterns would be impressive but fundamentally different.

The failure rates, transferability, and boundary conditions matter too, but those will come out in the 90-day report. Those details matter much more than the benchmark stack.

The public should also care about timing. Anthropic estimates, per the system card, that comparable models could appear elsewhere within 6 to 18 months. If that window is real, this is not a one-company story. It is a short grace period. Glasswing is an attempt to use that grace period well — maybe it buys defenders enough time to harden the most important systems before equivalent capability spreads more widely.

The real signal

The name Mythos will fade. The possibility behind it will not: some frontier models can no longer be released like normal software because the gap between capability and consequence has closed.

Cybersecurity makes it legible in practice for the first time. The internet has been running on a hidden assumption: elite vulnerability research does not scale. Anthropic is telling us that assumption may now be false. If even half of this holds up, the security model the internet has relied on for thirty years stops working.

And the people moving fastest right now will not be the ones treating AI as a better coding assistant.

They will be the ones treating it as a machine for finding the cracks the whole digital world has been quietly standing on.

These newsletters take significant effort to put together and are totally for the reader’s benefit. If you find these explorations valuable, there are multiple ways to show your support:

• Engage: Like or comment on posts to join the conversation.

Leave a comment

• Subscribe: Never miss an update by subscribing to the Substack.

Subscribe now

• Share: Help spread the word by sharing posts with friends directly or on social media.

Share

References

- Anthropic. “Project Glasswing: Securing critical software for the AI era.” April 2026. https://www.anthropic.com/glasswing

- Anthropic Frontier Red Team. “Claude Mythos Preview.” April 2026. https://red.anthropic.com/mythos-preview

- Fortune. “Anthropic says its new AI model can find critical software vulnerabilities.” April 2026.

- Simon Willison. Analysis of Mythos system card and leaked assets. April 2026.

- TechCrunch. “Anthropic launches Project Glasswing with $100M in defensive cybersecurity credits.” April 2026.

- The Hacker News. Coverage of CMS misconfiguration and Claude Code source exposure. April 2026.

- Carlini N, Cheng N, Lucas K, et al. “Claude Mythos Preview.” Anthropic Frontier Red Team, April 2026. https://red.anthropic.com/2026/mythos-preview/

- Anthropic. “Building AI Cyber Defenders.” 2026. https://www.anthropic.com/research/building-ai-cyber-defenders

- Anthropic. “Coordinated Vulnerability Disclosure.” https://www.anthropic.com/coordinated-vulnerability-disclosure

- OpenBSD SACK patch (27-year-old bug). https://ftp.openbsd.org/pub/OpenBSD/patches/7.8/common/025_sack.patch.sig

- FFmpeg H.264 fix (16-year-old bug). https://code.ffmpeg.org/FFmpeg/FFmpeg/pulls/22499/files

- Global cybercrime cost estimate (~$500B/year): Maas M et al. “Estimating Global Yearly Cybercrime Damage Costs.” GovAI, 2025. https://www.governance.ai/research-paper/estimating-global-yearly-cybercrime-damage-costs

Blood cancers were once a death sentence. Two CAR-T therapies approved since 2021 pushed overall response rates past 70 percent in relapsed multiple myeloma (Munshi et al., NEJM, 2021; Berdeja et al., Lancet, 2021). The problem is everything that comes before the infusion. These treatments take three to six weeks to build, cost half a million dollars, and require lymphodepleting chemotherapy before they even begin.

For these treatments, doctors harvest a patient’s white blood cells through leukapheresis. Those cells ship to a manufacturing facility where cell engineers reprogram them with a chimeric antigen receptor, a synthetic receptor that recognizes a specific cancer protein. The reprogrammed cells are expanded in culture for weeks. Meanwhile, the patient undergoes lymphodepleting chemotherapy to clear space in the immune system. Then the engineered cells are infused back into the patient. The total bill exceeds $500,000 per patient (American Cancer Society, 2025; Rawlings et al., Clin Hematol Int, 2024). Patients routinely deteriorate or die during the wait.

If you enjoy these posts, consider subscribing and becoming a part of our growing community!

Subscribe now

A phase 1 trial out of Tongji Hospital in Wuhan just cut all of that out and skipped to the life-saving part (An et al., Nature Medicine, 2026).

The treatment is ESO-T01, an immune-shielded lentiviral vector that carries the genetic instructions for an anti-BCMA chimeric antigen receptor. BCMA is a protein that sits on the surface of myeloma cells. The vector reprograms T cells directly in the bloodstream, turning the patient’s own immune cells into cancer killers without ever removing them from the body.

Five patients with relapsed or refractory multiple myeloma received a single intravenous infusion. No leukapheresis for immune cell isolation. No ex vivo cell engineering or multiweek expansion. No lymphodepleting chemotherapy. Just one injection.

Figure 1. Conventional CAR-T therapy requires seven steps over 3–6 weeks at a total cost exceeding $500,000. The in vivo approach (ESO-T01) reduces this to a single intravenous infusion. Source: An et al., Nature Medicine (2026).

The goal of a phase 1 study is to collect safety data. The efficacy results were not supposed to be the story.

Four of five patients achieved objective responses. Three achieved stringent complete remission, the deepest category of response in myeloma. The approved therapies that achieve those response rates require leukapheresis, weeks of manufacturing, and lymphodepleting chemotherapy. This study matched them with a single injection and none of it.

All four evaluable responders were minimal residual disease negative at the 10^-5 sensitivity level by day 60. That means fewer than one myeloma cell per 100,000 bone marrow cells. Those numbers belong in a phase 3 readout. This is phase 1.

Figure 2. Clinical responses in the ESO-T01 phase 1 trial. Three of five patients achieved stringent complete remission (sCR). All four responders were MRD-negative at 10^-5 sensitivity by day 60. Source: An et al., Nature Medicine (2026).

The economics of CAR-T therapy are built on infrastructure. Manufacturing alone represents roughly 75 percent of total treatment cost (Hernandez et al., JAMA Oncol, 2018).

Building the weapon inside the body eliminates most of that. No manufacturing facility. No weeks of inpatient stays. The drug itself, a viral vector produced at scale, behaves more like a conventional biologic than a bespoke cell therapy.

No one has published a formal cost estimate for in vivo CAR-T yet. A KPMG analysis modeled how manufacturing simplification could bring the cost below $30,000 per treatment (KPMG, 2024). In vivo generation is the most radical version of that simplification.

If you can build it inside, the infrastructure collapses. CAR-T therapy currently exists at roughly 200 centers worldwide, nearly all in the United States, Europe, and a handful of Chinese cities. The majority of the world’s cancer patients have zero access. A viral vector that ships at scale and requires no cell manufacturing facility changes that. A patient in rural Arkansas, or Lagos, or rural India, could receive a single IV infusion at a community clinic.

A cancer treatment that once required a factory now requires a syringe.

The field has not fully reckoned with another advantage: repeat dosing. Ex vivo CAR-T is functionally a one-shot therapy. If the cancer relapses, re-manufacturing is difficult or impossible because the patient’s T cells are already depleted from the first round. An in vivo platform changes that calculus. If the first dose fails or the disease returns, the patient gets another injection. That opens the door to combination protocols: sequential targets, dose escalation, CAR-T paired with checkpoint inhibitors. Strategies that are financially prohibitive when each dose costs half a million dollars.

The safety data tell a more complicated story. Cytokine release syndrome, a systemic inflammatory response triggered when engineered immune cells activate and attack cancer en masse, occurred in four patients, three of them grade 3. All were controlled with standard interventions: corticosteroids, tocilizumab, supportive care (An et al., 2026). The most common side effects were transient drops in blood counts and reversible liver enzyme elevations. One patient died from spinal cord compression caused by an extramedullary myeloma lesion, a known complication of the underlying disease.

The trial enrolled five patients, all male, all heavily pretreated with a median of three prior lines of therapy. Median follow-up was six months. The trial was stopped early in 2025 for reasons the authors do not disclose. Phase 1 trials are designed to establish safety, not efficacy. The sample is too small for anything definitive, and durability of response remains an open question.

In 2003, a gene therapy trial for X-linked severe combined immunodeficiency used gamma-retroviral vectors to insert a corrective gene into patient cells. Several children developed leukemia. The vectors had integrated near proto-oncogenes and activated them (Hacein-Bey-Abina et al., Science, 2003). That episode set the entire field back by a decade.

Lentiviral vectors were engineered specifically to avoid this. Modern self-inactivating designs delete the enhancer and promoter elements that caused the problem, sharply reducing the risk of activating nearby genes after integration (Naldini, Nature, 2015). The clinical track record now supports the redesign. Over 500 patients have received lentiviral gene therapy across a dozen disease indications. The longest follow-up data, now past a decade, show no vector-related cancers (Scala et al., Nature Medicine, 2023; Kohn et al., NEJM, 2025). No confirmed case of replication-competent lentivirus has been reported in any patient to date (Merten et al., Mol Ther Methods Clin Dev, 2024).

But ESO-T01 is doing something those therapies did not. It is injecting the vector directly into the bloodstream of a patient who has not been lymphodepleted. Previous attempts at in vivo CAR-T generation struggled precisely here. The patient’s immune system neutralized the viral vector before it could reprogram enough T cells. ESO-T01 solves this with a nanobody-directed, immune-shielded envelope that evades complement and antibody recognition (An et al., 2026). That shielding is the key engineering innovation. Without it, the concept does not work.

The entire logistical architecture of CAR-T therapy exists because of one assumption: you have to build the weapon outside the body. Seven FDA-approved CAR-T products. Billions of dollars in manufacturing infrastructure. All built on that premise.

Other groups have pursued in vivo CAR-T using lipid nanoparticles or engineered fusogens, but ESO-T01 is the first to report clinical data in humans.

The platform is not specific to myeloma. The immune-shielded lentiviral vector is a chassis. Swap the receptor target and the same architecture could reprogram T cells to hunt different cancers, from B-cell lymphomas to solid tumors that have resisted every prior generation of CAR-T.

But cancer might not even be the most interesting application. In 2022, a team at the University of Erlangen gave ex vivo CAR-T cells to five patients with refractory systemic lupus erythematosus. All five went into drug-free remission (Mackensen et al., Nature Medicine, 2022). The engineered cells depleted the autoreactive B cells driving the disease. Similar results have since emerged in multiple sclerosis and systemic sclerosis. The limiting factor is the same one that limits cancer patients: manufacturing complexity, cost, and access.

An in vivo platform eliminates all three. A rheumatologist in a community clinic could administer a single infusion that reprograms the patient’s own T cells to clear the B cells attacking their joints, kidneys, or nervous system. No referral to an academic center. No six-figure bill. No three-week manufacturing wait while the disease progresses.

Five patients is a proof of concept, not a product. Larger trials will determine whether the efficacy holds and whether the safety profile is tolerable at scale.

But the question the field has been asking for a decade, whether you can skip the factory and program the immune system in place, now has preliminary human data behind it. The answer appears to be yes.

These newsletters take significant effort to put together and are totally for the reader’s benefit. If you find these explorations valuable, there are multiple ways to show your support:

• Engage: Like or comment on posts to join the conversation.

Leave a comment

• Subscribe: Never miss an update by subscribing to the Substack.

Subscribe now

• Share: Help spread the word by sharing posts with friends directly or on social media.

Share

References:

1. An N, Wang D, Zhang P, et al. In vivo generation of anti-BCMA CAR-T cells in relapsed or refractory multiple myeloma: a phase 1 study. Nature Medicine (2026). DOI: 10.1038/s41591-026-04244-6.

2. Munshi NC, et al. Idecabtagene vicleucel in relapsed and refractory multiple myeloma. N Engl J Med 384, 705-716 (2021).

3. Berdeja JG, et al. Ciltacabtagene autoleucel in patients with relapsed or refractory multiple myeloma (CARTITUDE-1). Lancet 398, 314-324 (2021).

4. Hacein-Bey-Abina S, et al. LMO2-associated clonal T cell proliferation in two patients after gene therapy for SCID-X1. Science 302, 415-419 (2003).

5. Naldini L. Gene therapy returns to centre stage. Nature 526, 351-360 (2015).

6. Scala S, et al. Dynamics of genetically engineered hematopoietic stem and progenitor cells after autologous transplantation in humans. Nature Medicine 29, 1769-1780 (2023).

7. Kohn DB, et al. Lentiviral gene therapy with autologous hematopoietic stem and progenitor cells for adenosine deaminase-deficient SCID. N Engl J Med (2025).

8. Merten OW, et al. Long-term stability of clinical-grade lentiviral vectors. Mol Ther Methods Clin Dev 32, 101176 (2024).

9. American Cancer Society. CAR T-cell Therapy and Its Side Effects. cancer.org (2025).

10. KPMG. The USD 30k CAR-T Therapy: A Future Within Reach? (2024).

11. Hernandez I, et al. Total costs of chimeric antigen receptor T-cell immunotherapy. JAMA Oncol 4, 994-997 (2018).

12. Rawlings MO, et al. Navigating the economic burden of multiple myeloma. Clin Hematol Int (2024).

13. Mackensen A, et al. Anti-CD19 CAR T cells for refractory systemic lupus erythematosus. Nature Medicine 28, 2124-2132 (2022).

ClinicalTrials.gov: NCT06791681.

Weight loss drugs are judged by the number on a scale. But the scale is a crude instrument. It makes no distinction between fat and muscle. Up to a third of the weight lost on Ozempic is muscle. That ratio holds across the major GLP-1 trials. Semaglutide and Tirzepatide drive approximately 15 to 22 percent total body weight loss (Wilding et al., 2021; Jastreboff et al., 2022), but lean tissue accounts for a quarter to a third of it (Heymsfield et al., 2014). Strength is a predictor of longevity (García-Hermoso et al., 2018). For older adults with sarcopenia or anyone already carrying low muscle mass, the additional loss is its own risk. The field has known for years, and it was quietly downplayed. Because there was no fix.

A phase 2 trial published in Nature Medicine provides a solution (Heymsfield et al., 2026). The drug is bimagrumab, a monoclonal antibody that blocks activin type II receptors. Combined with semaglutide, it produced more total weight loss while preserving the lean mass that semaglutide strips away.

If you enjoy these posts, consider subscribing and becoming a part of our growing community!

Subscribe now

Bimagrumab was built to treat muscle wasting, not obesity (Heymsfield et al., JAMA, 2021). Its target, the activin type II receptor, sits on the surface of muscle and fat cells. It receives signals from myostatin, activin A, activin E, and several other TGF-beta family ligands that suppress muscle growth and regulate fat storage. Bimagrumab blocks the receptor itself, shutting off the entire suppressive input. That is why it works where earlier approaches failed. Myostatin knockout mice are enormous. Belgian Blue cattle carry a natural myostatin loss-of-function. But every drug built to block myostatin alone has failed in humans. Stamulumab, domagrozumab, landogrozumab. Two decades of clinical attempts, none of them worked. Myostatin is one of several signals converging on the same receptor. Blocking one ligand was never enough. Bimagrumab blocks the receptor itself and catches them all.

The effect on fat was unexpected. Exome sequencing from the UK Biobank identified variants in the activin E and ALK7 genes that protect against abdominal obesity. Bimagrumab blocks the same signaling node, increasing fat mobilization independently of appetite.

The BELIEVE trial confirmed this directly. Caloric intake in the semaglutide groups dropped. In the bimagrumab groups, it did not. Bimagrumab reshapes body composition without changing how much patients eat. Two drugs, two tissues, two completely different pathways. One works through the brain. The other works through muscle and fat.

The BELIEVE trial randomized over 500 adults with obesity across nine arms, including bimagrumab alone, semaglutide alone, and four dose combinations. At 48 weeks, the high-dose combination (bimagrumab 30 mg/kg plus semaglutide 2.4 mg) produced 17.8 kg of total weight loss versus 14.2 kg for semaglutide alone.

On semaglutide alone, roughly a third of the weight lost was muscle. In the high-dose combination, almost none of it was. Ninety-two percent of the combination’s weight loss was fat. Bimagrumab alone went further. It lost fat and gained lean mass. Every kilogram it shed was fat. Losing fat while gaining muscle is the one thing diet and exercise cannot reliably do at the same time. Bimagrumab did it by default.

By week 72, the gap widened. Semaglutide alone had shed 7.4 percent of lean tissue. Bimagrumab alone had gained 4.1 percent. The combination held lean mass near baseline while losing nearly half its visceral fat.

Grip strength increased in bimagrumab-containing groups, confirming the lean mass numbers translate to function. Grip strength is also an independent predictor of longevity, particularly in women (BioWire Byte 021).

Every prediabetic patient in the high-dose combination normalized blood sugar by week 72. One hundred percent.

The safety signals were predictable. Muscle spasms occurred in roughly two-thirds of bimagrumab patients. Acne was common. LDL cholesterol rose with bimagrumab while semaglutide lowered it. The combination partially offset the rise but did not eliminate it. Discontinuation rates were lower in the combination arms than in bimagrumab alone, suggesting semaglutide may improve tolerability. (The trial was funded by Eli Lilly, which acquired bimagrumab’s developer Versanis Bio.)

This is phase 2. It was built to find doses and characterize the biology, not to prove efficacy at scale. Phase 3 will need larger populations, longer treatment, and MRI confirmation of the DXA-based body composition data, particularly in older adults with the most to lose. A trial combining bimagrumab with tirzepatide is already registered.

GLP-1s solved appetite. Activin receptor blockade may solve body composition. If this combination delivers in phase 3 what it delivered here, it will redefine what counts as successful weight loss. Not just pounds lost. Pounds of what?

These newsletters take significant effort to put together and are totally for the reader’s benefit. If you find these explorations valuable, there are multiple ways to show your support:

• Engage: Like or comment on posts to join the conversation.

Leave a comment

• Subscribe: Never miss an update by subscribing to the Substack.

Subscribe now

• Share: Help spread the word by sharing posts with friends directly or on social media.

Share

References

Heymsfield SB, Aronne LJ, Montgomery P, et al. Bimagrumab plus semaglutide alone or in combination for the treatment of obesity: a randomized phase 2 trial. Nature Medicine. 2026;32:869-882. doi: 10.1038/s41591-026-04204-0

Heymsfield SB, Gonzalez MC, Shen W, Redman L, Thomas D. Weight loss composition is one-fourth fat-free mass: a critical review and critique of this widely cited rule. Obes Rev. 2014;15:310-321. doi: 10.1111/obr.12143

Heymsfield SB, Coleman LA, Miller R, et al. Effect of bimagrumab vs placebo on body fat mass among adults with type 2 diabetes and obesity: a phase 2 randomized clinical trial. JAMA Netw Open. 2021;4(1):e2033457. doi: 10.1001/jamanetworkopen.2020.33457

Wilding JPH, Batterham RL, Calanna S, et al. Once-weekly semaglutide in adults with overweight or obesity. N Engl J Med. 2021;384:989-1002. doi: 10.1056/NEJMoa2032183

Jastreboff AM, Aronne LJ, Ahmad NN, et al. Tirzepatide once weekly for the treatment of obesity. N Engl J Med. 2022;387:205-216. doi: 10.1056/NEJMoa2206038

García-Hermoso A, Cavero-Redondo I, Ramírez-Vélez R, et al. Muscular strength as a predictor of all-cause mortality in an apparently healthy population: a systematic review and meta-analysis of data from approximately 2 million men and women. Arch Phys Med Rehabil. 2018;99(10):2100-2113. doi: 10.1016/j.apmr.2018.01.008

Physically active adults develop dementia at lower rates, perform better on cognitive tests, and show less hippocampal atrophy on MRI. The epidemiology has been clear for years. What hasn’t been clear is the mechanism. Researchers have proposed BDNF, irisin, and a half-dozen other mediators. None of them traced a complete chain from a peripheral organ to the brain and back to a cognitive outcome. A study published in Cell in February 2026 does exactly that, and the organ in question is the liver (Bieri, Pratt et al., Cell, 2026).

First, if you enjoy these posts, consider subscribing and becoming a part of our growing community!

Subscribe now

The connection starts with an enzyme called GPLD1, a phospholipase secreted by the liver whose circulating levels rise with exercise. Saul Villeda’s lab at UCSF identified it in 2020 by transferring plasma from exercised mice into sedentary aged mice, then running proteomics on the blood to find what was doing the work (Horowitz et al., Science, 2020). Overexpressing GPLD1 in the liver of aged mice recapitulated the cognitive benefits of exercise plasma transfer. But GPLD1 does not cross the blood-brain barrier. Whatever it was doing, it was doing it from the outside.

The new paper identifies the target. TNAP, tissue-nonspecific alkaline phosphatase, is a protein anchored to the surface of brain endothelial cells, the cells that form the blood-brain barrier (BBB). As the brain ages, TNAP accumulates on these cells. GPLD1 cleaves it off, physically stripping TNAP from the membrane. Circulating TNAP levels rise in the blood after treatment, confirming the cleavage is happening in real time.

The consequences are immediate. Aged mice show extensive tracer leakage from hippocampal blood vessels. A barrier in disrepair. GPLD1 treatment reduced that leakage across the hippocampus and reversed a large fraction of age-related gene expression changes in brain endothelial cells, pushing them back toward a youthful profile. TNAP inhibition alone reproduced the majority of those effects. Two different interventions targeting the same enzyme, same transcriptomic direction.

To prove TNAP itself drives the decline, the group engineered young, healthy mice to overexpress it on brain endothelial cells. Those animals developed barrier leakage and cognitive impairment at four to five months old. TNAP accumulation alone was sufficient to age the brain’s vasculature. Running the experiment in reverse, CRISPR knockout of TNAP in aged mice improved cognition. So did SBI-425, an oral TNAP inhibitor that never crosses the blood-brain barrier. The rescue works from the periphery.

The group then asked whether this holds in disease, not just aging. In a mouse model of Alzheimer’s disease, GPLD1 treatment reduced amyloid deposits, dampened neuroinflammation, and improved cognition across multiple behavioral tests. SBI-425 replicated most of those benefits but missed one, Y-maze working memory, which GPLD1 rescued. That gap matters because GPLD1 cleaves over 100 proteins beyond TNAP, and its full cognitive reach appears broader than any single target. For drug development, that’s a feature. TNAP inhibition gives you a clean peripheral target for barrier repair, while the biology leaves room for combination strategies that capture more of what exercise actually does.

The medical research field has spent decades searching for a drug that meaningfully rescues cognitive decline. Most candidates fail because the brain is hard to reach. SBI-425 sidesteps that problem entirely. It’s oral, it never crosses the blood-brain barrier, and it has already shown cognitive benefits in mice. That profile is cleaner than most CNS drug candidates, though peripheral TNAP plays roles in other tissues and those liabilities would need evaluation. (Disclosure: Villeda is co-founder of Ceiba Bio, Inc.; the work is covered by patent PCT/US2020/016549.)

Early human data hints in the same direction. Western blots of human cortical tissue showed TNAP protein levels elevated in older adults and Alzheimer’s patients compared with healthy young controls. The broader implication is uncomfortable for a field that has spent decades treating the brain as the protagonist of its own decline. Amyloid is made in neurons. Tau tangles spread through neural circuits. Inflammation is driven by resident microglia. All of that is framed as intrinsic failure. This paper suggests the brain may be more dependent on peripheral vascular support than anyone assumed. A liver enzyme, elevated by running, can shift hundreds of age-related endothelial gene expression changes back toward youth and improve cognition in aged and Alzheimer’s mice. If the barrier is the bottleneck, how much of what we call brain aging is actually vascular aging that the brain merely suffers?

No one has answered that in humans yet. But if SBI-425 achieves in humans even a fraction of what it does in mice, the field will need to reckon with a possibility it has largely set aside: that the best way to treat a failing brain might be to fix the pipes.

These newsletters take significant effort to put together and are totally for the reader’s benefit. If you find these explorations valuable, there are multiple ways to show your support:

• Engage: Like or comment on posts to join the conversation.

Leave a comment

• Subscribe: Never miss an update by subscribing to the Substack.

Subscribe now

• Share: Help spread the word by sharing posts with friends directly or on social media.

Share

References

Bieri G, Pratt K, Fuseya Y, et al. Liver exerkine reverses aging- and Alzheimer’s-related memory loss via vasculature. Cell. 2026;189(4). doi: 10.1016/j.cell.2026.01.024

Horowitz AM, Fan X, Bieri G, et al. Blood factors transfer beneficial effects of exercise on neurogenesis and cognition to the aged brain. Science. 2020;369(6500):167-173. doi: 10.1126/science.aaw2622

A surgeon is performing a heart transplant. He cracks the chest, connects the patient to bypass, and removes the failing heart. He sutures in the donor organ. Aorta to aorta. Pulmonary artery. Vena cava. Left atrial cuff. He removes the cross-clamp and rewarms the heart with perfused blood. Nothing. He defibrillates. Nothing. Inotropes. Mechanical support. Six hours later, the team withdraws care. But when did the patient actually die? Was it when the old heart stopped? When the new one never started? If it had resumed rhythm ten minutes after they quit, would the patient have been alive the whole time?

The medical establishment treats brain death as binary. Circulation stops. Neurons lose oxygen. Within minutes, cells enter irreversible decline. That timeline underpins the legal definition of death, organ procurement windows, and the clinical default that a post-mortem brain is biologically inert. In 2019, Nenad Sestan’s lab at Yale tested that assumption directly, and the results don’t support it.

Vrselja and colleagues took 32 pig brains from a slaughterhouse and waited four hours after death. They connected them to BrainEx, a perfusion system that pushes synthetic oxygenated solution through the brain’s own vasculature. The perfusate carried cytoprotective agents, anti-inflammatories, and neural activity blockers. Control brains received no perfusion or perfusion without the active solution. Six hours later, the treated brains were compared against both control groups (Vrselja et al., 2019).

Figure 1. BrainEx perfusion system and experimental workflow. Adapted from Vrselja et al., Nature 2019.

Vascular perfusion restored throughout the brain. Cells consumed oxygen, metabolized glucose, produced carbon dioxide. Tissue under electron microscopy was markedly intact compared to unperfused controls, which showed expected decomposition. Glial cells mounted inflammatory responses. The blood-brain barrier partially recovered. Individual neurons, when sampled and tested with electrodes, showed preserved electrophysiological properties and spontaneous synaptic activity.

But continuous EEG monitoring showed flatline throughout. Whatever was happening was cellular, not coordinated. Individual neurons worked. The networks connecting them did not.

Figure 2. BrainEx restored cellular-level function but no organized network activity was detected.

Then there are the neural activity blockers. The stated rationale is methodological: the experiment measured cellular viability, and spontaneous firing would consume energy and damage the tissue being studied. Fine. But the team also kept general anesthesia on standby and ran continuous EEG for the duration. You don’t monitor for something you’re certain won’t happen. The paper says nothing about what might occur without the blockers. Neither has anyone else. The absence of global activity wasn’t proof of irreversibility. It was a condition of the experiment. That line hasn’t been crossed.

Three years later, the same lab went whole-body. OrganEx perfused entire pigs one hour after cardiac arrest. Hearts regained electrical activity and contractile function. Livers produced albumin. Kidneys resumed filtration. Involuntary muscular contractions appeared in the head and neck, surprising even the researchers (Andrijevic et al., 2022). The cellular death cascade turned out to be slower, more complex, and more interruptible than anyone in clinical medicine had assumed.

Figure 3. From BrainEx (2019) to OrganEx (2022): expanding from brain to whole-body perfusion.

So what do you do with a technology that reverses death at the cellular level?

The near-term one is organ preservation. Thousands of donor organs are discarded annually because they can’t survive the gap between procurement and transplant. A system that restores cellular function after extended ischemia attacks that problem directly.

The longer-term one is drug discovery. Sestan and Vrselja co-founded Bexorg to do something no one else can: test drug candidates on perfused human brains — not organoids, not mouse models, but actual human tissue from donors with neurological disease — Alzheimer’s, Parkinson’s, the conditions where 90–99% of candidates fail. The academic science behind the platform is published in *Nature*. Bexorg has not published its own data. No FDA framework exists for this. The technology is real. The business case is unproven.

Neither application changes the core finding. Four hours of warm ischemia did not destroy these cells. The boundary between reversible and irreversible brain death is not where medicine placed it. That matters independent of whether anyone commercializes it.

**References**

Vrselja Z, Daniele SG, Silbereis J, et al. Restoration of brain circulation and cellular functions hours post-mortem. *Nature*. 2019;568(7752):336-343. PMID: 30918408

Andrijevic D, Vrselja Z, Lysyy T, et al. Cellular recovery after prolonged warm ischaemia of the whole body. *Nature*. 2022;608(7922):405-412. PMID: 35922506

Farahany NA, Greely HT, Giattino CM. Part-revived pig brains raise slew of ethical quandaries. *Nature*. 2019;568(7752):299-302.

The fastest-growing drug class in the world was designed for the gut and obesity. The most important question about it may involve the brain and dementia.

First, if you enjoy these posts, consider subscribing and becoming a part of our growing community!

Subscribe now

More than 50 million people worldwide now take GLP-1 receptor agonists for weight loss or diabetes. A growing number of researchers want to know what those weekly injections are doing to their brains. The largest systematic review and meta-analysis to date just tried to find out.

Choudhury and colleagues screened 10,037 records and included 82 studies covering Parkinson’s disease, Alzheimer’s disease, depression, anxiety, substance use disorders, binge eating, and suicidality (Choudhury, et al., 2026).

GLP-1 receptor agonist users had a 30% lower risk of developing Parkinson’s disease (HR 0.70). That’s a large signal from a drug class never designed to touch the basal ganglia. Study-to-study variability was moderate, which for observational data spanning multiple drug formulations is reassuring.

When Choudhury’s group looked at patients who already had Parkinson’s, the drugs didn’t help. No improvements in motor symptoms. No improvements in non-motor symptoms. No improvement in quality of life or dyskinesia scores. On the MDS-UPDRS Part III, the standard motor function scale, the drugs were indistinguishable from placebo.

A separate meta-analysis reinforces this. Stefanou and colleagues pooled all four randomized placebo-controlled trials of GLP-1 agonists in Parkinson’s, 667 patients across exenatide, lixisenatide, liraglutide, and NLY01, and found nothing (Stefanou, et al., 2026). No benefit on motor scores. No benefit on cognition. No benefit on non-motor symptoms. The drugs were well tolerated. They just didn’t work.

GLP-1 receptor agonists appear to prevent Parkinson’s disease but cannot treat it.

This prevention-without-treatment pattern exists elsewhere in medicine. Statins prevent heart attacks far more reliably than they reverse existing coronary artery disease. Aspirin prevents strokes more effectively than it rehabilitates stroke damage. The preventive signal is large enough to detect in population data. The treatment signal is zero.

Why? One explanation is direct neuroprotection. GLP-1 receptors are expressed throughout the brain, including in the substantia nigra, which degenerates in Parkinson’s. Some GLP-1 agonists cross the blood-brain barrier in animal models. Preclinical work shows they suppress neuroinflammation, reduce oxidative stress, and enhance mitochondrial function in dopaminergic neurons. Semaglutide in particular modulates microglial activation in rodent models of neurodegeneration (Evola and Parmar, 2026). The preclinical case is real. But it hasn’t translated to humans.

There’s a simpler explanation. GLP-1 agonists fix metabolism. They reduce blood sugar, visceral adiposity, insulin resistance, and systemic inflammation. Chronic systemic inflammation drives neuroinflammation through well-characterized pathways. Reduce the upstream metabolic dysfunction and you reduce the downstream inflammatory cascade that accelerates neuronal loss. Under this framing, GLP-1 agonists aren’t protecting the brain directly. They’re protecting the body, and the brain benefits.

The Alzheimer’s data is the strongest evidence for the metabolic explanation. GLP-1 RA users in Choudhury’s analysis showed a 63% reduced risk of Alzheimer’s compared to unmatched controls (RR 0.37). But when compared to SGLT2 inhibitors, another metabolically active drug class, the advantage vanished (RR 1.07). SGLT2 inhibitors don’t cross the blood-brain barrier. They don’t bind GLP-1 receptors. They fix metabolism through an entirely different mechanism. If both drug classes produce the same reduction in Alzheimer’s incidence, the simplest interpretation is that metabolic rescue is doing the work, not direct neuroprotection.

Novo Nordisk’s Phase 3 EVOKE and EVOKE+ trials reinforce this. The trials enrolled 3,808 participants with amyloid-confirmed early-stage Alzheimer’s and tested oral semaglutide against placebo for two years (Scheltens, et al., 2026). Semaglutide did not slow disease progression in either trial. Cerebrospinal fluid biomarkers of tau pathology and neuroinflammation improved modestly, around 10%, but none of that translated into cognitive or functional benefit. Novo Nordisk terminated the program. If semaglutide had meaningful direct neuroprotective effects in humans, you would expect some signal in 3,808 patients over two years. There was none on any clinical measure.

The pattern across both diseases is consistent. Observational data shows lower incidence of neurodegenerative disease among GLP-1 RA users. Randomized trials show no therapeutic effect once disease is established. The SGLT2 comparator suggests the prevention signal may not be specific to GLP-1 receptor activation at all. The most parsimonious explanation is that metabolic health protects the brain, and GLP-1 agonists happen to be the most widely prescribed way to achieve it.

Elsewhere, GLP-1 RAs were associated with a 45% reduced risk of cannabis use disorder and reduced binge eating frequency, but showed no effect on depression or anxiety. The suicidality question that has followed this drug class since FDA surveillance signals in 2023 found no increased risk in any analysis, with observational data trending protective.

Caveats apply. The prevention data is entirely observational, carrying the usual confounding and healthy-user biases. No randomized trial has tested whether GLP-1 agonists prevent neurodegenerative disease in healthy people. Evidence certainty ranged from moderate to very low depending on the outcome. Choudhury’s group calls the findings “hypothesis-generating.”

The real story here may not be that GLP-1 drugs protect the brain. It may be that metabolic health protects the brain. The drug class just happens to be the largest unintentional experiment in testing that hypothesis. Whether the prevention signal holds up in a prospective trial designed to isolate it remains an open question. No one has run that trial yet.

These newsletters take significant effort to put together and are totally for the reader’s benefit. If you find these explorations valuable, there are multiple ways to show your support:

• Engage: Like or comment on posts to join the conversation.

Leave a comment

• Subscribe: Never miss an update by subscribing to the Substack.

Subscribe now

• Share: Help spread the word by sharing posts with friends directly or on social media.

  Share

References:

Choudhury, I., Ward, J.H., Mahesh, S., Alam, U., Azmi, S. and Anson, M. (2026) ‘Effect of Glucagon-Like-Peptide-1 Receptor Agonists (GLP-1 RA) on Neuropsychiatric Outcomes: A Systematic Review and Meta-Analysis’, Clinical Therapeutics. doi: 10.1016/j.clinthera.2026.02.010.

Evola, V. and Parmar, M.S. (2026) ‘Targeting neuroinflammation in neurodegenerative disorders: the emerging potential of semaglutide’, Inflammation Research, 75(1), 13. doi: 10.1007/s00011-025-02166-6.

Scheltens, P., Atri, A., Feldman, H.H. et al. (2026) ‘Baseline characteristics from evoke and evoke+: Two phase 3 randomized placebo-controlled trials of semaglutide in participants with early-stage symptomatic Alzheimer’s disease’, Alzheimer’s & Dementia: Translational Research & Clinical Interventions, 12(1), e70200. doi: 10.1002/trc2.70200.

Stefanou, M.I., Panagiotopoulos, E., Tentolouris, A. et al. (2026) ‘Efficacy and safety of glucagon-like peptide-1 receptor agonists in Parkinson’s disease: a systematic review and meta-analysis of randomized placebo-controlled clinical trials’, Therapeutic Advances in Neurological Disorders, 19, 17562864251408269. doi: 10.1177/17562864251408269.

You’re at the doctor’s office getting a routine physical. They ask their normal list of questions about your eating and exercise habits. You mention the morning jogs, maybe some weekend cycling. They nod approvingly. What they aren’t trained to ask is how strong you are. A new study in JAMA Network Open suggests that strength is a critical indicator for longevity.

Researchers followed over 5,400 women aged 63 to 99 for nearly a decade. They measured two simple things: grip strength and the time it takes to stand up from a chair five times. Then they tracked who lived and who didn’t. The clear result was that stronger women lived longer; in the study, women in the highest quartile of grip strength had a 33% lower mortality risk compared to the weakest group, even after adjusting for age, smoking, comorbidities, and body weight (LaMonte, et al., 2026). Similarly, a 2022 meta-analysis of cohort studies reported that any muscle-strengthening activity was associated with lower risk of all-cause mortality and major non-communicable diseases (Momma, et al., 2022). But by how much?

First, if you enjoy these posts, consider subscribing and becoming a part of our growing community!

Subscribe now

The mortality gradient is nearly threefold: 67.2 deaths per 1,000 person-years in the weakest quartile, dropping to 23.5 in the strongest. Normally researchers just use self-reported exercise. But in this study, they strapped accelerometers on every participant for a full week, capturing actual movement and sedentary time with objective precision. The survival advantage persisted through adjustment for moderate-to-vigorous physical activity, sedentary hours, walking speed, and C-reactive protein (LaMonte, et al., 2026).

The most provocative finding: grip strength was associated with lower mortality even in women who did not meet the recommended 150 minutes per week of aerobic activity. Read that again. Women who were essentially sedentary by guideline standards still lived longer if they were strong.

The two strength measures also told different stories. Grip strength barely budged when the researchers piled on health and clinical adjustments. Chair stand time attenuated almost to the null. The authors suggest chair stands may capture general health status and fatigability, while grip strength reflects something more fundamental about skeletal muscle output. Previous work supports this: muscle strength predicted mortality independent of muscle mass measured by imaging, suggesting that muscle quality matters more than quantity (Newman, et al., 2006).

The biology makes this plausible. Skeletal muscle is the body’s largest endocrine organ. Contracting muscle fibers release myokines that regulate inflammation, insulin sensitivity, and immune surveillance. Chronic inflammation accelerates loss of muscle mass and function through mitochondrial dysfunction and impaired excitation-contraction coupling (Ferrucci and Fabbri, 2018). The fact that CRP adjustment barely moved the association suggests the protective effect of strength operates through pathways beyond inflammation alone.

A caveat: this cohort is exclusively postmenopausal women from the Women’s Health Initiative. The findings may not generalize directly to men or younger populations.

By 2050, women over 75 will be the largest age subgroup in the United States. Most longevity guidance still emphasizes cardio. This study makes the case that strength is the more fundamental variable. It can be assessed in a clinic with a $30 dynamometer and a chair. The question is whether anyone will bother to measure it.

These newsletters take significant effort to put together and are totally for the reader’s benefit. If you find these explorations valuable, there are multiple ways to show your support:

• Engage: Like or comment on posts to join the conversation.

Leave a comment

• Subscribe: Never miss an update by subscribing to the Substack.

Subscribe now

• Share: Help spread the word by sharing posts with friends directly or on social media.

Share

References:

LaMonte, M.J., Hyde, E.T., Nguyen, S. et al. (2026) ‘Muscular Strength and Mortality in Women Aged 63 to 99 Years’, JAMA Network Open, 9(2), e2559367. doi: 10.1001/jamanetworkopen.2025.59367.

Newman, A.B., Kupelian, V., Visser, M. et al. (2006) ‘Strength, but not muscle mass, is associated with mortality in the Health, Aging and Body Composition study cohort’, The Journals of Gerontology Series A, 61(1), pp. 72-77. doi: 10.1093/gerona/61.1.72.

Momma, H., Kawakami, R., Honda, T. et al. (2022) ‘Muscle-strengthening activities are associated with lower risk and mortality in major non-communicable diseases: a systematic review and meta-analysis of cohort studies’, British Journal of Sports Medicine, 56(13), pp. 755-763. doi: 10.1136/bjsports-2021-105061.

It’s extremely late and I work tomorrow, so I’m going to keep this short. I’m thrilled to announce Neural NeXus has a new home www.neuralnexus.press.

But I didn’t email you about that. I emailed you about this:

I built a daily Spelling Bee game and I bet most of you can’t hit Genius (don’t worry, I didn’t either).

Seven letters. The middle letter must appear at least once in every word you make. You have five minutes. There’s a leaderboard and scores are public like an arcade.

New puzzle every day. Today’s is live right now.

The original Spelling Bee on the New York Times is a favorite between my mom and me. Whenever we call to catch up, we play together. Rather than being cooperative, we usually see who can spell the most words before being kicked out (there is a cap on how many points you can reach without a paid subscription to NYTimes). I love the game concept, but never found the format satisfying. So I made my own. I hope you all enjoy it as much as we do.

And for launch day — the top scorer gets a shoutout in the next Neural NeXus article AND gets to pick the topic for a future deep dive. You choose the science, I write it.

Prove me wrong → neuralnexus.press/play

— David

René Descartes had a problem he could not solve, and it has haunted philosophy ever since.

The problem was this: the mind and the body appear to be made of fundamentally different stuff. The body is a physical object that occupies space, obeys mechanical laws, and can be dissected and measured. While the mind seems to be something else entirely, it thinks, it feels, it is aware. It has no mass. You cannot cut it open. In 1641, Descartes distilled this intuition into a doctrine: res cogitans, thinking substance, and res extensa, extended substance. Two irreducibly distinct things.

The trouble was explaining how they interact. If mind and body are categorically different, how does the decision to lift your arm actually lift your arm? Descartes proposed, weakly, that the pineal gland was the site of their meeting. The mind somehow pushed on the body there. He knew this was unsatisfying. He died before he resolved it. The problem remained.

Three hundred years later, Gilbert Ryle took aim at the whole framework. In his 1949 book The Concept of Mind, he coined the phrase that would outlive everything else he wrote: the ghost in the machine. This was mockery. The ghost, the immaterial mind, the Cartesian soul, was, Ryle argued, a category error. Minds are not things that inhabit bodies the way ghosts inhabit houses. To ask where your mind is located is like visiting Harvard, seeing the colleges, the libraries, the faculty, the laboratories, and then finally asking to be shown the university. You have already seen everything there is to see. The question assumes a kind of independent entity that does not exist. There is no ghost. There is only the machine, doing what machines do, and we have been fooled by grammar into thinking there must be something else inside.

Ryle thought he had closed the issue. He had not.

In 1967, Arthur Koestler published his own book with the same title, The Ghost in the Machine, as a rebuttal. Something strange was happening inside biological organisms, Koestler argued: the hierarchical architecture of the nervous system produced behaviors that could not be explained by simply pointing at neurons. Evolution had built something that exceeded its own parts. He kept Ryle’s phrase and inverted its meaning — the ghost, in his telling, was not an error but an open question.

That argument has never been settled. Consciousness researchers, philosophers of mind, and neuroscientists have spent the decades since accumulating evidence and frameworks without converging on an answer. Despite the scientific advances, the hard problem, why there is something it is like to be a brain, remains nearly as hard as it was in Descartes’ time. Then someone tried to copy one.

First, if you enjoy these posts, consider subscribing and becoming a part of our growing community!

Subscribe now

A San Francisco-based company called Eon Systems posted a video online. It did not settle the argument. It threw a hand grenade into the debate by turning a thought experiment into an actual one. The thought experiment: if you digitized every neuron and every connection of a biological brain, and you ran that copy in a new body, would it be the same mind?

The video (embedded above) shows a fruit fly — a Drosophila melanogaster — walking, feeding, and grooming inside a physics simulation. Nothing unusual about a simulated fly. What was unusual was the brain controlling it. The fly’s behavior was driven by an emulation of a real biological brain, wired neuron-to-neuron, synapse-to-synapse, from electron microscopy data of a real Drosophila nervous system. A copy of a biological connectome, running in simulation, making a body move.

Eon’s founder, Michael Andregg, described it as the world’s first whole-brain emulation to produce multiple behaviors in an embodied simulation. He was, with important caveats, correct.

In 2024, Philip Shiu, now a senior scientist at Eon, and collaborators published a paper in Nature that represented years of collective work by hundreds of researchers across multiple institutions. The researchers used the FlyWire connectome, a complete map of synaptic connections in the adult Drosophila brain assembled through electron microscopy and machine-learning-assisted reconstruction, to create a computational model of the entire fly brain. Neural NeXus covered the FlyWire consortium's work when it published in October 2024 (BioWire 018) — the largest brain map of any organism at the time, a project involving over three million manual edits and 8,453 identified neuron types. What Eon has now done is take that map and run it. The numbers: 140,000 neurons, 50 million synaptic connections, every neuron typed by neurotransmitter identity. The complete wiring diagram of a biological brain, rendered in silico.

The FlyWire connectome — all 140,000 neurons of the adult Drosophila melanogaster brain, color-coded by cell type, assembled from electron microscopy data by the FlyWire consortium. From Dorkenwald et al., 2024, Nature.

The model is nearly a perfect anatomical replica. But the lights weren’t on — the neurons couldn’t fire. Shiu et al. gave them life by using a simple neuron model. With this, each cell accumulates input, fires when a threshold is crossed, and resets, to simulate the dynamics of a network. They activated specific neurons computationally and asked whether the downstream activity matched known experimental results. For feeding behavior, activation of sugar-sensing and water-sensing gustatory neurons in the model accurately predicted which neurons would respond in living animals. For grooming, mechanosensory activation predicted the correct antennal grooming circuit.

The model matched real neural behavior with over 90% accuracy — built from wiring data alone, with no parameter fitting. But it was a brain without a body. Activation flowed through the connectome and produced motor outputs — signals that would, in a living fly, reach the ventral nerve cord and drive the muscles. In the model, those signals had nowhere to go. The brain was predicting movement. Nothing was moving.

The new Eon demonstration addresses this directly. They integrated Shiu’s brain model with a physics-based simulation of the Drosophila body and musculoskeletal system (NeuroMechFly v2) to close the loop. In this system, sensory input enters the emulated brain and neural activity propagates across the full connectome. Motor commands then emerge as descending neuron signals. These turn into control outputs for turning, forward velocity, grooming, and escape, which are then translated into joint torques and leg trajectories by body-level controllers trained through imitation learning.

The brain decides. The controllers execute. The fly walks. It grooms. It feeds.

Incredibly, the behaviors emerged without training the brain model. No reinforcement learning shaped the connectome’s activity, no fine-tuning adjusted its synaptic weights against observed fly behavior. The connectome structure, implemented in the simplest possible neuron model and coupled to imitation-trained body controllers, was enough to produce recognizable, sequenced behavior. That is the result Eon is claiming, and it is the right thing to be struck by: the wiring diagram, even mediated through an engineered motor interface, contains a remarkable amount of what a fly needs to be a fly. Eon’s own technical writeup cautions that broader behavioral repertoires will likely require additional learning mechanisms, richer motor interfaces, and more functional data. The demo is a first step, not a final proof.

The model used to implement the fly brain is computationally simple by design. Each neuron accumulates input, its membrane potential leaking back toward baseline between spikes, and fires when a threshold is crossed. What it does not do — cannot do, in its current form — is change. The synaptic weights are fixed. The architecture is static. The brain captured in the FlyWire connectome is a snapshot of a single adult fly’s nervous system at a single moment in time.

What Eon has emulated is not a mind in process. It is a mind in amber. In this sense, every large language model ever deployed shares the same limitation — frozen weights, no capacity to learn from the next input. The plasticity problem is not unique to brain emulation. It is the central unsolved constraint in all neural intelligence.

There is a science-fiction version of whole-brain emulation where this problem is already solved. In Richard K. Morgan’s Altered Carbon, every human carries a cortical stack at the base of the skull. It is a device encoding the complete mind as Digital Human Freight. It contains personality, memory, identity. The full architecture of consciousness, stored on alien metal. In this universe, bodies are disposable. Stacks are what matter. The protagonist, Takeshi Kovacs, wakes up in strangers’ bodies across centuries. Each time, he adapts — new proprioception, new reflexes, new flesh. He forms new memories. His mind, contained in the stack, rewires to fit hardware it was never born into. That is plasticity. This science fiction is Eon’s stated ambition: the human mind, digitized and portable. The cortical stack is the final version of the fly connectome. But plasticity remains the unsolved problem.

Eon’s digital fly cannot do any of this.

Biological neural plasticity is the ability of synaptic connections to strengthen or weaken in response to experience. That is the physical substrate of learning and memory. It is what allows an organism to update its model of the world, to form new associations, to recover from injury, to change over a lifetime. A static copy of a biological brain can generate embodied behavior, but without plasticity it may be closer to a preserved behavioral architecture than a living mind. The digital fly executes the repertoire of the fly it was copied from, unable to deviate, unable to learn, unable to form a single new memory from contact with a world that is already different from the one its connectome was shaped by.

The FlyWire connectome also maps only the central brain. It does not include the full peripheral nervous system, the sensory neurons in the legs, the mechanoreceptors in the wings, the proprioceptive circuits that tell the brain where the body is in space. The connection between the emulated brain and the simulated NeuroMechFly body required Eon to make engineering assumptions — educated guesses about how motor outputs from the connectome map onto the muscles of the simulated body. Those assumptions were good enough to produce recognizably fly-like behavior. They were still assumptions. The ghost and the machine were not as perfectly married as the headline framing implied.

Eon names the next constraint itself. The visual system is implemented in the current demo — the Lappalainen connectome-constrained visual motion model is integrated, and the fly can in principle respond to looming threats. But the company’s own technical writeup describes the visual activations as “somewhat decorative”: they are present in the simulation but do not currently drive behavioral outputs in any meaningful way. A fly that cannot respond to what it sees is a significant constraint on what the demo actually demonstrates. The data is open source, the video is public, the methodology is described. But the independent validation that separates a compelling demonstration from an established scientific result has not happened. This is not unusual for early-stage work. It is worth stating plainly.

None of this should obscure what Eon has actually done.

No one had previously demonstrated a complete biological connectome driving a physically simulated body through multiple naturalistic behaviors. The prior landmarks in this space either modeled brains without bodies or animated bodies without biologically realistic neural dynamics. One example, OpenWorm, made an attempt at whole-nervous-system emulation with C. elegans. These creatures contain 302 neurons and a much simpler behavioral repertoire. Another example is DeepMind and Janelia’s MuJoCo fly, first described in a 2024 preprint and published in Nature in 2025, which used reinforcement learning to control a simulated fly body. Impressive engineering. Not brain emulation.

Eon’s demonstration is different in kind. The ghost is a copy.

The question that copy raises, and that Ryle thought he had answered, and Koestler thought was still open, and that consciousness researchers have not closed, is what a copy of a brain actually is.

· · ·

The fly is a proof of concept. The question Eon is now inside is whether proof of concept scales.

The fly brain contains 140,000 neurons. The mouse brain contains roughly 70 million or over 500 times as many. The mouse has a vastly more complex functional architecture, with plasticity mechanisms that span timescales from milliseconds to years in a dynamic connectome continuously remodeled by experience. Eon has announced that the mouse is their next target, combining expansion microscopy to map every synaptic connection with tens of thousands of hours of calcium and voltage imaging to capture how those networks activate in living tissue. The human brain — the implicit endpoint of the entire enterprise — contains approximately 86 billion neurons and 100 trillion synaptic connections.

The fly’s behavioral repertoire, sophisticated as it is for a 140,000-neuron system, is largely stereotyped. It is, in the language of ethology, a relatively fixed set of responses to a relatively constrained set of stimuli. The mouse’s repertoire includes flexible decision-making, spatial memory, fear learning, social behavior, and forms of cognition that are continuous with the human case. Emulating a mouse brain without plasticity would produce a very expensive recording. Emulating it with plasticity raises questions that go beyond engineering.

The scaling challenges between a fly and a mouse are log-scale, requiring technological advances that don’t exist. Yet.

If you copy a mouse brain — with plasticity intact, with the capacity for learning — and run it in simulation, and it learns something, what is it? Is the digital mouse the same mouse? A copy? Something new? Does the question even have a determinate answer?

These are the actual questions that whole-brain emulation at scale will force into scientific, legal, and ethical territory where no framework currently exists. The engineering roadmap is clearer than the conceptual one.

The answer to the plasticity problem is being worked on in Melbourne, Australia, from the opposite direction.

Cortical Labs does not start with a connectome. They start with neurons, living human neurons grown on a silicon chip, and they give those neurons a world to interact with. The neurons learn. Readers of the prior Neural NeXus piece on Cortical Labs (Lab Grown Neurons Learn to Play Doom) will know the Pong result: cortical cells with no prior exposure to anything, forming a closed loop with a game simulation and improving their rally lengths measurably within five minutes of gameplay. Their CL1 device shipped to researchers in early 2025, and one developer connected it to Doom in about a week using approximately 200,000 living human neurons.

That number is worth pausing on. The Eon fly connectome contains 140,000 neurons. The Cortical Labs Doom demonstration ran on roughly 200,000. The same order of magnitude of biological tissue — one frozen and mapped, the other alive and adaptive — produces behaviors of comparable complexity. There may be critical thresholds of neural count at which certain levels of behavioral organization become achievable, regardless of substrate or architecture. Whether that convergence is meaningful or coincidental is a question the field will be answering for years.

The deeper inversion is this: Eon has a perfect map of a biological brain and cannot make it learn. Cortical Labs has no map at all, the neural activity of their cultures is not decoded at the level of individual circuits, and the neurons learn continuously. Eon emulates the structure. Cortical Labs cultivates the process. One preserves the ghost. The other keeps it alive.

Ryle would say neither. There is no ghost. There is only information processing, and both approaches are finding different ways to do it.

Koestler might say both. The ghost is whatever it is that makes a nervous system do something more than its parts, and both of these systems are producing exactly that kind of excess.

Descartes, confronted with a digital fly that walks without thinking and a dish of neurons that thinks without a body, might revise his pineal gland hypothesis.

The digital fly does not know it is in a simulation. It cannot know. It reacts to inputs according to the frozen architecture of a brain that no longer exists in biological form. Surprise is not available to it. Learning is not available to it. Change is not available to it. The ghost is present. The ghost is running. The ghost is stuck.

And yet it walks. It grooms. It feeds.

Descartes could not explain how the ghost and the machine interact. Ryle said the ghost was an error. Koestler said it was a mystery. Eon Systems said: let us try to run it on different hardware.

The answer they got back was: we can run it. We just can’t update it.

Is a mind that cannot change still a mind? That question is now in the laboratory.

The fly does not have an opinion.

Neural NeXus covers the intersection of neuroscience, AI, and the technologies reshaping how intelligence is built. These newsletters take significant effort to put together and are totally for the reader’s benefit. If you find these explorations valuable, there are multiple ways to show your support:

Engage: Like or comment on posts to join the conversation.

Leave a comment

Subscribe:

Subscribe now

Share: Help spread the word by sharing posts with friends directly or on social media.

Share

References

Shiu, P.K. et al. A Drosophila computational brain model reveals sensorimotor processing. Nature 634, 210–219 (2024). doi.org/10.1038/s41586-024-07763-9

Dorkenwald, S. et al. Neuronal wiring diagram of an adult brain. Nature (2024). doi.org/10.1038/s41586-024-07558-y

Wissner-Gross, A. The First Multi-Behavior Brain Upload. The Innermost Loop, Substack (March 7, 2026). theinnermostloop.substack.com

Kingsley, D. BioWire Weekly — 018. Neural NeXus, Substack (October 8, 2024). davidkingsley.substack.com

Descartes, R. Meditations on First Philosophy (1641).

Ryle, G. The Concept of Mind. Hutchinson & Co. (1949).

Koestler, A. The Ghost in the Machine. Hutchinson & Co. (1967).

The brain doesn’t experience the world directly.

It receives electrical signals converted to sensory information from eyes, skin, muscle, and from those signals alone it builds a model of what’s out there. Then it sends other electrical signals to perform functions. This is the loop - perception and action.

Which raises an interesting and uncomfortable question: if you gave a dish of neurons in a lab a synthetic stream of electrical signals, one that responded to their activity the way a body would, would they know the difference?

Researchers tried it.

It turns out there is a precise mathematical law governing how nervous systems respond to that kind of feedback. When you give neurons a world built on that law, something remarkable happens.

They learn.

Scientists have already used this principle to teach neurons to play video games. First Pong. Then Doom.

The company that built this system is called Cortical Labs. Understanding who they are and what they are actually trying to do matters, because the press coverage has mostly gotten the framing wrong.

To understand why, you first have to understand how the brain actually works.

First, if you enjoy these posts, consider subscribing and becoming a part of our growing community!

Subscribe now

How the Brain Actually Works

The standard textbook picture of the brain is a camera attached to a processor. Sensory data flows in, gets processed, and produces a response. This picture is wrong, and the evidence against it has been accumulating for decades.

The more accurate model — formalized most rigorously by theoretical neuroscientist Karl Friston at University College London — is that the brain is fundamentally a prediction engine.^1,2 At every moment, across every level of its architecture, the brain is generating a model of what should be happening next and comparing that prediction against what actually arrives. The gap between prediction and reality is called prediction error. Minimizing that gap is, under this framework, the organizing principle of all neural computation.

Friston formalized this as the Free Energy Principle: biological systems minimize a quantity called variational free energy, which is mathematically equivalent to the divergence between an internal model of the world and the evidence the system actually receives.^1,3 This is not a metaphor. It is a testable mathematical claim about how nervous systems self-organize.

The behavioral extension of this is Active Inference.^4,5 The key insight is that you can reduce prediction error in two ways. You can update your internal model to better match reality. Or you can act on the environment to make reality better match your model. Reaching for a glass of water is, under this framework, a self-fulfilling prediction: your motor cortex generates the expectation that your hand will close around the glass, and your muscles fire to make that expectation true. Hunger is a prediction error signal — your body’s model says glucose levels should be higher than they are. Pain is a prediction error signal. Curiosity is a prediction error signal.

This framing has a critical implication for artificial systems — one that is underappreciated in most AI coverage.

Modern large language models and deep neural networks are, in a structural sense, also prediction engines. A transformer predicts the next token. A convolutional network predicts the correct label. The training process minimizes the gap between prediction and ground truth, which is mathematically similar to what the Free Energy Principle describes in biological systems. The resemblance runs deep: many of the researchers who built modern deep learning were drawing, consciously or not, on decades of computational neuroscience.

But there is a critical difference in how the prediction loop closes. In a biological system, prediction error is continuous, embodied, and consequential — the organism acts on the world and receives feedback that is causally linked to what it just did. The loop runs at the timescale of experience. In a standard neural network, the loop closes offline, against a static dataset, during training. After training, the model generates predictions without updating. It no longer has a world; it has a frozen approximation of one.

This architecture has a well-known failure mode that is, in retrospect, entirely predictable. When you show a language model evidence that its answer is wrong, it will sometimes acknowledge the correction and then, a few exchanges later, revert confidently to the original wrong answer. The model is not being stubborn. It is doing exactly what its architecture trains it to do: generate the token sequence most consistent with its internal model of what a correct response looks like. The correction is a new input, but it does not update the underlying weights. The world changed; the model did not. This is, in the vocabulary of Active Inference, a system that can update its outputs but cannot update its model — one half of the prediction-error minimization loop with the other half removed.

Biological neurons, given a world that punishes wrong predictions with unpredictable stimulation, update continuously, in real time — because the cost of staying wrong is immediate and physical.

The DishBrain experiment was, at its core, a test of whether this framework held in a radically minimal system — a single monolayer of cortical neurons with no body, no history, and no prior exposure to anything.

Giving Neurons a World

The experimental setup, described in the 2022 Neuron paper by Brett Kagan and colleagues (including Friston as a co-author), was elegant in its simplicity.⁶

Cortical neurons — derived from either mouse embryos at embryonic day 15.5 or human induced pluripotent stem cells — were cultured onto a high-density multi-electrode array (HD-MEA) containing over 26,000 platinum electrodes across an 8mm² surface. Of these, 1,024 were routed for recording, and 8 were designated as sensory input electrodes. The neurons grew across the chip over several weeks, forming dense interconnected networks. When the cultures reached sufficient electrophysiological maturity — evidenced by synchronized bursting activity — they were connected to a software simulation of Pong.

Ball position was encoded as electrical pulses delivered to the sensory electrodes, using a combined rate and place coding scheme — a method with close biological parallels to how the rodent barrel cortex encodes whisker deflection. Paddle movement was decoded from spiking activity in two designated motor regions: excess activity in motor region 1 moved the paddle up; motor region 2, down. The neurons had no way of knowing what any of this meant. The mapping between their activity and the game world was entirely arbitrary.

The training signal was the implementation of the Free Energy Principle directly. When the paddle missed the ball, the neurons received unpredictable, chaotic stimulation — random pulses at random sites across the 8 input electrodes. When the paddle connected, they received predictable, synchronized stimulation across all electrodes simultaneously. No reward, in the conventional reinforcement learning sense. Just the difference between order and chaos.

The results were published across 399 test sessions and roughly 43 biological replicates. A few things stand out.

First, the learning was fast. Significant improvement in average rally length — the primary performance metric — emerged within five minutes of real-time gameplay. Not five epochs. Five minutes. For context, a standard deep reinforcement learning algorithm trained on the same task takes roughly 90 minutes to reach comparable performance.

Second, the control conditions were rigorous. Cultures receiving sensory information but no feedback showed no learning. Cultures receiving feedback but no sensory information showed no learning. Electrically inactive HEK293T cells showed no learning. The key variable was the closed-loop feedback between neural activity and environmental consequence. Embodiment — in the specific sense of actions having observable consequences — was required.

Third, and perhaps most striking: human cortical cells initially performed worse than mouse cortical cells before reversing to significantly outperform them. In the first five minutes of gameplay, human cortical cells (HCCs) showed lower rally lengths than mouse cortical cells (MCCs) and even the media-only baseline. The authors interpret this initial underperformance as possible exploratory behavior — a period of active sampling of the environment before the system settles on a strategy. By the final 15 minutes, HCCs had reversed the trend and significantly outperformed all control groups. To the authors’ knowledge, this is the first empirical demonstration that human cortical neurons exhibit greater adaptive learning capacity than their rodent counterparts under equivalent conditions. It is a result that had been theorized but never directly tested, and it matters: it is one of the clearest pieces of evidence that human cortical neurons carry genuinely distinct computational properties from what we see in model organisms.

The entropy analysis added a mechanistic layer. The team used information entropy of neural responses as a proxy for the average “surprise” experienced by the culture — an upper bound on variational free energy. As predicted by theory, entropy was lower during gameplay than at rest. Following unpredictable (miss) feedback, entropy spiked. Following predictable (hit) feedback, it did not. The neurons were reorganizing their activity to reduce the unpredictability of their sensory inputs — precisely what the Free Energy Principle predicts.

One important limitation the authors were honest about: between-session learning was not robustly observed. Cultures appeared to relearn associations within each session but did not reliably retain them across sessions. This is consistent with the use of cortical neurons, which are not specialized for long-term memory storage in vivo, and it is a clear direction for future work.

From Pong to Doom

The Pong result established that embodied neurons could learn. The question that followed was whether this was a narrow demonstration or the beginning of something more general.

In early 2025, Cortical Labs upgraded to the CL1, which replaced the earlier electrode architecture with a stable planar array of 59 electrodes at sub-millisecond latency, extended neuron viability to six months, and built a Python API that abstracts the wetware sufficiently that researchers without biology backgrounds can program it directly.

The Doom demonstration arrived shortly after launch, built by independent developer Sean Cole in approximately one week. The core engineering challenge was translation: Doom’s visual stream had to be converted into a language neurons can process, which is electricity. Cole and the Cortical Labs team mapped the game’s video feed into patterns of electrical stimulation delivered across the electrode array. Enemies approaching from the left generated one set of signals; open corridors, another. The neurons received these inputs, processed them through their firing patterns, and the system mapped those patterns back to in-game actions — a specific firing pattern triggers a shot, another moves the character right, another advances forward.

The neurons can navigate toward enemies and shoot at them. Solving the game is a different matter entirely. Doom, at its core, is a puzzle — it requires spatial memory, key-hunting, route planning, and the ability to sequence decisions across time. The CL1’s culture lacks persistent between-session memory and the architectural complexity to hold a map of a level in anything resembling working memory. The character wanders, reacts, and fires, but it does not progress through the game in any strategic sense. It plays like a reflex, not a plan. Brett Kagan acknowledged this directly: the neurons are “learning,” but in the same breath called the result an early-stage proof of the interface — a demonstration that the bridge between biological tissue and digital environment works, with the harder question of gameplay competence still ahead. The Cortical Labs team is explicit that they have “solved the interface problem” — the bridge between digital game state and biological tissue — but that better training methods, encoding schemes, and reward structures are needed before the neurons can do more than react to what is directly in front of them.

The cultural resonance matters here even if the performance benchmarks are modest. “Can it run Doom?” has been the unofficial benchmark for computational legitimacy since the 1990s — a question applied to calculators, tractors, ATMs, and now a dish of living human brain cells. The answer, as of this year, is yes. The more important answer — can it beat Doom — remains, for now, no. That gap between reflex and strategy is precisely the terrain that the next generation of biological computing research will need to cross.

The CL1, the Competition, and What Comes Next

Cortical Labs is not the only organization working in this space, and understanding the competitive landscape puts their position in context.

FinalSpark, a Swiss company, has taken a different approach to neural reinforcement — using dopamine as a chemical reward signal rather than electrical feedback. Kagan has been explicit about why Cortical Labs avoided this: dopamine-based reward is difficult to scale to more complex devices, and the company’s long-term ambition requires an approach that can grow with the architecture. Intel and IBM have both invested heavily in neuromorphic silicon — chips designed to mimic neural computation — but mimicking and implementing are different propositions, and both approaches remain power-hungry relative to biological alternatives. In 2024, Chinese researchers introduced MetaBOC, a brain-on-a-chip platform representing a parallel development track. At Johns Hopkins, Tufts, and Melbourne, organoid intelligence programs are exploring three-dimensional brain organoid structures as computational substrates — cultures with layered architecture closer to in vivo complexity than the flat monolayers of DishBrain.

The field now has a name — Organoid Intelligence (OI) — and funding to match.

Cortical Labs was founded in Melbourne in 2019 by Dr. Hon Weng Chong, a clinician by training who started with a single question: what if the entire field had the problem backwards? The dominant paradigm in AI hardware takes inspiration from the brain and tries to implement it in silicon — neuromorphic chips, spiking neural networks, in-memory computing architectures. Chong looked at that effort, noted that no artificial system had yet matched the brain on its own terms, and asked what nobody seemed to want to ask: what if you just used neurons?

His framing for the CL1 is worth quoting directly. He describes it as a reverse Neuralink. Neuralink’s bet is that you can put a chip inside a brain and use the brain’s neurons to interface with machines. Cortical Labs inverted the proposition: put the brain onto the chip. Give it a body made of silicon and software. Let it compute from there. Cortical Labs raised $11 million from investors including Horizons Ventures, Blackbird Ventures, and In-Q-Tel — the latter being the CIA’s venture arm, which tends to notice when a technology has implications beyond the academic. Their commercial product, the CL1, was announced at Mobile World Congress in Barcelona in March 2025. The first 115 units are shipping this year.

The CL1 itself is a remarkable piece of hardware given how recently this was science fiction. Each unit contains 800,000 neurons grown from iPSCs derived from adult donor skin or blood samples. A proprietary life-support system maintains temperature, nutrients, gas exchange, and waste filtration, keeping the culture viable for up to six months. The biOS — Biological Intelligence Operating System — handles the firmware abstraction between wetware and Python. A rack of 30 units consumes 850–1,000 watts. For comparison, running equivalent AI workloads on conventional GPU infrastructure requires tens of kilowatts. The first 115 units ship at $35,000 each, with rack pricing at $20,000 per unit.

For labs that cannot justify the capital outlay, Cortical Labs offers cloud access at $300 per week per unit — what Chong calls Wetware-as-a-Service, or WaaS. The model is deliberate. Chong has said he wants to bring the cost of biological computing experimentation down to the price of a Zoom subscription. The Python API means a researcher with no background in electrophysiology can deploy code to living neurons without owning a pipette. This is a distribution strategy as much as a pricing strategy: the fastest way to find out what biological computers are actually good at is to let thousands of researchers try things you have not thought of yet.

That openness is already producing unexpected use cases. The obvious applications are in drug discovery and neurological disease modeling — the CL1 provides a human-relevant, animal-testing-free substrate for studying how compounds affect real neural circuits in real time, something no current in vitro model offers with comparable fidelity. A culture can be exposed to a candidate neuropsychiatric drug and its learning dynamics observed directly. Changes in entropy, plasticity, firing patterns — all quantifiable against the same closed-loop gameplay benchmarks used in the DishBrain paper. This alone represents a significant advance over the standard tools of pharmaceutical neuroscience.

But the inbound interest has ranged considerably further. Robotics groups are exploring biological controllers for real-time adaptive movement — the same properties that make neurons good at Pong are also what makes robotic control in unstructured environments hard for conventional systems. Government and defense interest, signaled by In-Q-Tel’s presence on the cap table, likely intersects with autonomous systems and pattern recognition at low power budgets. And then there are the genuinely strange inquiries: Bitcoin miners looking at the energy profile, experimental musicians interested in biological controllers for generative audio, artists exploring what a collaborative project with living neural tissue even means. The technology surface is already wider than the roadmap covers.

The near-term constraint is scale. The current CL1 contains a culture roughly equivalent in neuron count to an insect brain. The Minimal Viable Brain project inside Cortical Labs is working toward organized cultures of around thirty neurons for targeted pattern-recognition tasks — moving from emergent self-organization toward something closer to intentional circuit design. Scaling toward primate-level complexity will require advances in three-dimensional culture architecture, electrode density, and closed-loop learning algorithms not yet at commercial readiness. The ethical framework will need to keep pace: Cortical Labs has embedded bioethicists directly into the development cycle, and the question of at what scale a biological computing substrate might warrant moral consideration is not one anyone in the field is dismissing.

Every new computing substrate — vacuum tubes, transistors, integrated circuits — passed through a period where it worked in principle but not yet at scale. The 2022 paper settled whether the neurons can learn. The open question is whether the architecture can be made reliable, scalable, and manufacturable at a cost that makes the energy and adaptability advantages matter.

The answer is being worked out right now in Melbourne, Lausanne, Baltimore, and Beijing. The CL1 is the first external proof that the process has left the lab.

What Karl Friston said about the CL1 captures the moment accurately: it is the first commercially available biomimetic computer, the ultimate in neuromorphic computing, because it uses real neurons rather than simulating them. Four billion years of evolution produced a learning machine that runs on 20 watts, generalizes from almost nothing, and updates continuously from contact with the world. We have been trying to replicate that in silicon for seventy years. Cortical Labs decided to stop replicating and start deploying.

The first shipments go out this summer. The benchmark is no longer whether biological computers can play Doom. It is what they learn to do next, once researchers with real problems get their hands on them.

Neural NeXus covers the intersection of neuroscience, AI, and the technologies reshaping how intelligence is built. These newsletters take significant effort to put together and are totally for the reader’s benefit. If you find these explorations valuable, there are multiple ways to show your support:

Engage: Like or comment on posts to join the conversation.

Leave a comment

Subscribe:

Subscribe now

Share: Help spread the word by sharing posts with friends directly or on social media.

Share

References

1. Friston, K. (2010). The free-energy principle: a unified brain theory? Nature Reviews Neuroscience, 11(2), 127–138. doi:10.1038/nrn2787
2. Friston, K. (2009). The free-energy principle: a rough guide to the brain? Trends in Cognitive Sciences, 13(7), 293–301. doi:10.1016/j.tics.2009.04.005
3. Friston, K., Kilner, J., & Harrison, L. (2006). A free energy principle for the brain. Journal of Physiology-Paris, 100(1–3), 70–87. doi:10.1016/j.jphysparis.2006.10.001
4. Friston, K., FitzGerald, T., Rigoli, F., Schwartenbeck, P., & Pezzulo, G. (2017). Active inference: a process theory. Neural Computation, 29(1), 1–49. doi:10.1162/NECO_a_00912
5. Friston, K., et al. (2017). Active inference and learning. Neuroscience & Biobehavioral Reviews, 68, 862–879. doi:10.1016/j.neubiorev.2016.06.022
6. Kagan, B.J., et al. (2022). In vitro neurons learn and exhibit sentience when embodied in a simulated game-world. Neuron, 110(23), 3952–3969. doi:10.1016/j.neuron.2022.09.001

Foreword

The most dangerous ideas are the ones that feel inevitable.

Citrini Research published a piece that touched a nerve. They clearly articulated the anxieties I have about AI and the future economy in a way I hadn't managed to myself. They called it The 2028 Global Intelligence Crisis, a fictional macro memo written from June 2028, reconstructing how AI-driven disruption unraveled into a full-blown economic catastrophe. You should read it first. They describe a scenario where AI scales economic productivity but couples it with a fatalistically unraveling labor market — a world where the flow of capital bypasses humans entirely, as they lose the ability to provide economic value.

The bear case, briefly (seriously, read their essay): agentic AI triggers mass white-collar layoffs beginning in 2026. Displaced workers spend less. Consumer demand contracts. Companies facing margin pressure invest more in AI to cut costs further, accelerating the next round of layoffs. A negative feedback loop with no natural brake. Ghost GDP — output that appears in national accounts but never circulates through households. The $13 trillion residential mortgage market, underwritten against incomes that no longer exist, begins to crack. Private credit defaults find their way into life insurer balance sheets through offshore structures nobody fully understood until it was too late. The Fed can cut rates to zero. It cannot change the fact that a Claude agent does the work of a $180,000 product manager for $200 a month. S&P 3,500. Unemployment 10.2%. The Intelligence Displacement Spiral.

The authors got the mechanism right. But the model is incomplete. It describes one half of the ledger with forensic precision and quietly ignores the other entirely; a world with no new opportunities, where the only wielders of intelligence are gigantic corporations. This is why I started this Substack. I have felt the pull towards AI fatalism, a dystopia where humans lose their ability to provide economic value, stripped of purpose, closer than we’d like to admit.

It’s a future underestimating humanity. I don’t buy it.

Here’s what I believe instead. The developments in AI will reduce the cost of intelligence and democratize the future economy. Anyone with drive, curiosity, and grit can enter a field and compete. Those rich in free time and determination will beat those rich in capital and institutional knowledge. We may see more millionaires and maybe even billionaires created than ever before — not because they’re tech titans, but because they leveraged new tools to unlock creativity and build businesses at a rate never seen before. That is what an Intelligence Renaissance looks like.

So I wrote the other half. The bull case. Same format. Same voice. Same starting conditions. Different equilibrium. Not because the disruption doesn’t happen, but because the countervailing forces the bear model ignores turn out to matter for what happens next to humanity.

What follows is a fictional macro memo from the same June 2028 — the one where the canary didn’t die.

— A Future Bull

First, if you enjoy these posts, consider subscribing and becoming a part of our growing community!

Subscribe now

The Consequences of Abundant Intelligence

March 1st, 2026 June 30th, 2028

The unemployment rate printed 6.2% this morning, the fourth consecutive monthly decline from the 7.8% peak we hit in Q3 2027. The market rallied 1.4% on the number. The S&P 500 is now up 14% from the lows, within striking distance of the October 2026 highs that seemed so distant six months ago.

Traders are cautiously relieved. Eighteen months ago, the same print would have triggered a short squeeze and fresh mania. Today it is absorbed with something approaching appropriate skepticism. That is healthy.

Two years. That is all it took to get from “contained and sector-specific” to an economy that no longer resembles the one any of us grew up in — and then, unexpectedly, to an economy that has begun to resemble something none of us had quite imagined before. This quarter’s macro memo is our attempt to reconstruct the sequence. A post-mortem on the fears that, in crucial respects, did not materialize.

The fears were not irrational. We shared many of them. What the bears got right about the mechanism, they got wrong about the equilibrium.

The Two Loops

The bear scenario was built around a feedback loop with no natural brake. We argued then, and document now, that an equal and opposite loop was always present — suppressed in their model, but real in the economy.

The Two Ledgers

The bear model described one half of the economic ledger with precision. It quietly omitted the other. Here are both columns, side by side.

The Euphoria Was Palpable — And Then So Was the Fear

By October 2026, the S&P 500 flirted with 8,000. The initial wave of layoffs due to human obsolescence began in early 2026 (foreboding) and did exactly what layoffs are supposed to do. Margins expanded. Earnings beat. Stocks rallied. Record-setting corporate profits were funneled right back into AI compute.

Then the narrative shifted. The unemployment rate began its measured rise. White-collar job openings collapsed. A consensus formed, quickly and with considerable analytical force, that this disruption was different — that AI was a general intelligence which would absorb the jobs it created, that the feedback loop had no natural brake, that the velocity of money would flatline as machines replaced the workers who drove consumer spending.

We published our own concerns. In January 2027, we argued that the U.S. economy was essentially a white-collar services economy, and that the businesses AI was chewing up were not tangential to GDP — they were GDP. We were right about the mechanism.

We were wrong about the equilibrium.

The bears built an elegant closed-loop model. AI improves, companies need fewer workers, displaced workers spend less, companies invest more in AI, AI improves. They asked, correctly, how much money machines spend on discretionary goods. The answer, zero, was supposed to be the kill shot.

What the model missed was the other half of the ledger. It focused with forensic precision on what was being destroyed, and almost none at all on what was simultaneously being created. Every great disruption has two faces. The bear memo described one of them brilliantly. This memo describes the other.

How It Started

In late 2025, agentic coding tools took a step-function jump in capability. A competent developer working with Claude Code or Codex could now replicate the core functionality of a mid-market SaaS product in weeks. Not perfectly, and not with every edge case handled, but well enough that the CIO reviewing a $500k annual renewal started asking the question: “what if we just built this ourselves?”

Every bear analyst in the room focused on what this meant for incumbents. The implications for ServiceNow and Zendesk and the long tail of SaaS were correctly identified and aggressively repriced. What almost nobody asked was what this meant for entrants.

A competent developer could replicate a mid-market SaaS product in weeks. But so could a former product manager with no engineering background. A laid-off financial analyst. A marketing director whose department had just been automated. The same tools that made it possible for enterprises to cancel their software contracts made it possible for individuals to build companies without raising venture capital, without hiring engineers, without the infrastructure cost that had historically made entrepreneurship the province of the well-funded.

We spoke with a procurement manager at a Fortune 500 in the summer of 2026 — the same procurement manager the bears would have recognized from their version of events: the one who forced a 30% discount on a SaaS renewal by credibly threatening to build the product internally. The bear narrative ends there. Ours does not.

Six months later, he left the Fortune 500. He built the internal tool. And then he packaged it, priced it at a 60% discount to the incumbent, and sold it to eleven other procurement teams in his industry — teams where he had relationships, where his credibility was real, where he understood the specific workflow pain points in ways that a general-purpose LLM could approximate but not replace. His total startup cost: $3,400 in AI inference and hosting fees. He has seventeen customers now and is hiring — two humans, because he needs people who understand the nuance of enterprise procurement in ways that remain, for now, human.

Multiply this dynamic by a few hundred thousand workers across every major metro. That is what the closed-loop model missed.

When the Floor Became a Launch Pad

The bear narrative correctly identified that displaced white-collar workers would not sit idle. It assumed they would downshift. Many did — and that part of the story was accurate and painful. But a remarkable number did something the models had not priced in. They started businesses.

This should not have been surprising. Prior technological disruptions have generated entrepreneurial booms in their wake. What was different this time was the sheer scale of barrier removal — and the quality of the cohort entering the market.

Starting a business in 2024 required capital, technical co-founders, legal infrastructure, accounting, HR, a sales team, office space. The fully-loaded cost of a ten-person startup in San Francisco was north of $3 million per year before it generated a dollar of revenue. The bottleneck was not ideas. It was friction.

By mid-2026, that friction was approaching zero. Legal incorporation: automated. Accounting: automated. Customer service: automated. Basic engineering: automated. Marketing copy, competitive analysis, financial modeling, contract review: automated. A motivated former white-collar professional with domain expertise and taste could launch a functional business in a weekend and reach their first hundred customers without a single full-time hire.

And the people entering this market were not average. These were workers who had cleared competitive hiring bars at demanding companies. Who had spent a decade accumulating domain expertise, professional networks, and the hard-won judgment about which problems were actually worth solving. What they had lacked was capital and technical infrastructure. They no longer needed either. Armed with AI tools that compressed months of execution into days, a single expert could now do what previously required a team of ten.

U.S. NEW BUSINESS FORMATIONS HIT 651,000 IN OCTOBER 2026, HIGHEST MONTHLY READING ON RECORD; SMALL BUSINESS ADMINISTRATION NOTES “STRUCTURAL ACCELERATION” IN SOLO AND MICRO-ENTERPRISE FILINGS | SBA / Census Bureau, November 2026

That October print was 40% above the prior record, set during the COVID-era stimulus boom. But unlike the 2020–2021 formation wave — which was dominated by gig-economy LLC filings and people trying to deduct home office expenses — this cohort was building real companies. Domain-specific AI wrappers targeting industries their founders had spent a decade in. Boutique consulting practices serving fifty enterprise clients with two humans and an agentic stack. Healthcare navigation services. Hyper-local service businesses with near-zero marginal cost of expansion.

The DoorDash story the bears told was accurate: coding agents had collapsed the barrier to launching a delivery app, fragmenting the market and compressing margins toward zero. But the same mechanism that destroyed DoorDash’s moat created dozens of profitable micro-logistics cooperatives — each run by a former gig worker who now owned the platform. The middle margin was not vaporized. It was redistributed. Not to machines, but to the network participants themselves.

The Deflationary Abundance That Actually Circulated

The bears coined “Ghost GDP” — output that shows up in national accounts but never circulates through the real economy. It was a compelling concept. But it rested on an assumption that deserved more scrutiny: that productivity gains would flow exclusively to capital and compute owners, never making their way to households.

Deflationary technologies do not work that way in competitive markets. They did not with electricity. They did not with the internet. They did not with containerization. In each case, the initial gains accrued to early adopters — and then competition distributed the surplus to consumers as cheaper goods and services. The velocity of money fell on the price side and rose on the volume side. The net was expansion.

Consider healthcare. The average American household spent $13,000 per year on healthcare in 2024 — a figure that, for tens of millions of families, functioned as a second mortgage. AI systems turned out to be extraordinarily capable at pattern recognition in clinical data. Diagnostic accuracy improved, administrative overhead collapsed, the cost of accessing specialist knowledge fell toward zero.

HEADLINE

U.S. HEALTHCARE CPI FALLS 4.2% Y/Y FOR SECOND CONSECUTIVE QUARTER; AI DIAGNOSTIC TOOLS AND ADMINISTRATIVE AUTOMATION CITED; KAISER NOTES 38% REDUCTION IN AVERAGE CLAIM PROCESSING COST | BLS / Kaiser Family Foundation, Q1 2028

A 4% annual decline in healthcare costs represents over $500 billion in real purchasing power returned to American households. Not Ghost GDP. Actual money that was previously flowing to a radiologist’s billing department, now available for literally anything else.

The same logic applied, with varying intensity, to legal services, financial planning, education, tax preparation, and home services. Every category where the service provider’s value proposition had been “I will navigate complexity you find tedious” got cheaper. The tens of millions of Americans who had never accessed a financial advisor because they could not afford one now had sophisticated financial planning available for $20 a month. They used it. That is new demand, conjured from suppressed demand that had always been there, waiting for the price to fall far enough.

We have been waiting two centuries for healthcare costs to decline in real terms. It took machine intelligence eighteen months. The bears asked how much money machines spend on discretionary goods. The right follow-up question was: how much more do humans spend when machines make everything cheaper?

There is a deeper flaw in the bear model that the GDP figures above do not fully capture. It is worth naming directly.

Cancer treatment costs the United States more than $200 billion annually. Every chemotherapy infusion, every imaging scan, every hospital admission — it all counts as GDP. The oncologists, the radiologists, the pharmaceutical manufacturers, the insurance administrators processing the claims: all employed, all spending their incomes, all nodes in the circular flow of money the bear model was so concerned to protect.

If we cured cancer overnight, we would “destroy” $200 billion in annual economic activity. The bear model, applied consistently, would have to count this as a catastrophe.

It is not a catastrophe. It is the point.

The destroyed activity was always a cost, not a benefit — a tax that disease levied on the productive economy. Eliminating it does not shrink the economy. It reallocates two hundred billion dollars toward anything else people would actually choose to spend on if disease did not consume their resources first. It returns millions of productive life-years. It frees capacity for conditions that remain unsolved. The caregiver burden — estimated at $25 billion annually in lost productivity — disappears from families and never makes it into the national accounts at all.

The bear model makes precisely this error with AI-driven efficiency. When a household’s legal fees drop from $8,000 to $240 a year, the $7,760 in savings does not disappear. It goes somewhere. When healthcare administration becomes 30% cheaper, the $258 billion in avoided costs does not exit the economy. It recirculates — into the discretionary spending, savings, and investment of the households that no longer have to pay it.

The bears counted the destruction. They forgot to count what humans do with what gets freed up. That omission is not a rounding error. It is the entire argument.

The Entrepreneurship Premium

There is a human story that cuts directly against the Salesforce PM narrative.

Our friend was a senior data scientist at a mid-sized insurance company in 2025. When the layoffs came in early 2027, she was in the third wave — initially spared because data scientists seemed useful for AI implementation. By the time the models could build and evaluate themselves, her role was gone. Title, health insurance, 401k, $160,000 a year.

She did not start driving for Uber.

She had spent four years watching her company’s underwriters make systematically poor decisions about a specific category of commercial real estate risk. She understood the patterns at a depth that no model trained on general data could replicate — because the patterns were embedded in the specific institutional history, the regional quirks, the informal relationships between underwriters and brokers that never made it into any training set. What she had lacked was the technical capacity to build the product and the capital to hire engineers. She no longer needed either.

Six months after her layoff, she launched a specialized commercial insurance analytics service. Her entire technical stack cost $800 a month. Her first client paid $4,000 a month. Her eighth client signed three weeks ago. She is on track to clear $380,000 this year — more than double her prior salary — working with two part-time human contractors who handle the client relationships that are, so far, still irreducibly human.

She represents a cohort the labor statistics struggled to capture, because she is not an employee. She does not appear in payroll data. She showed up, eventually, in self-employment income filings, in new business formations, in 1099 issuances. The BLS has been playing catch-up on measurement all year.

The labor market data looked catastrophic for six quarters because it was measuring the wrong things. Payroll employment fell. Self-employment income surged. The gap between the two was the story the consensus missed.

U.S. SELF-EMPLOYMENT INCOME SURGES 34% Y/Y IN Q4 2027; SOLO-OPERATOR BUSINESS REVENUES CROSS $1.4T ANNUALIZED; BLS NOTES “STRUCTURAL MEASUREMENT CHALLENGES” IN CURRENT HOUSEHOLD SURVEY METHODOLOGY | BLS / Census Bureau, January 2028

The new small business did not hire the way the old mid-sized company did. It needed one or two people with taste and domain knowledge, and machines for everything else. That is a different economy than the one we were modeled for — but it is not a broken one. It is, by several measures, a more equitable one. The returns to domain expertise and genuine judgment, previously pooled inside large organizations, were being distributed outward to the people who actually held them.

Policy Finds a Gear

We will not overstate this. The political process was, to use the technical term, a disaster.

The partisan gridlock was real. The Occupy Silicon Valley movement was real. The grandstanding was exhausting. The right called any transfer mechanism Marxism and warned that taxing compute would hand the lead to China. The left couldn’t agree on which intervention was least likely to become regulatory capture by incumbents. Fiscal hawks pointed at deficits. Everyone pointed at everyone else.

But two things happened that the pure gridlock scenario had not priced in.

First, state governments moved faster than the federal one. Several states — led, improbably, by Texas and Florida, which understood they were watching their income-tax-free advantage become a liability in a world of declining white-collar payroll — passed their own small business AI subsidy programs. Portable benefits that attached to individuals rather than employers. Retraining credits redeemable for AI tool subscriptions, not just accredited university tuition. These were not transformative policies. They were stabilizers, and timing mattered more than elegance.

Second, the federal government eventually passed something — just not what either side wanted. The Transition Economy Act that emerged from conference in March 2028 was smaller, messier, and more means-tested than any idealist had hoped. It included a modest inference compute tax, meaningful enough to fund direct transfers to displaced workers, too small to materially slow AI investment. It included expanded small business support, portable benefits infrastructure, and — buried in subsection 4(c) — a sovereign AI participation vehicle that nobody loved but nobody could kill.

It was not elegant. It did not solve the structural problem. What it did was buy time — six months of demand stabilization that allowed the entrepreneurship wave to compound and the deflationary effects to show up in the data. In a crisis driven by feedback loops, buying time is underrated. You only need to outlast the negative momentum long enough for the positive loops to establish themselves.

The canary they were watching did not die. It started a business.

From Sector Risk to Systemic Stability

Through 2027, markets treated the AI transition as a sector story going wrong — and then priced it, at peak fear in November 2027, as a systemic catastrophe. The crash that month was real. What the market was pricing in was a closed-loop model where displacement had no offset, and where the negative momentum would compound indefinitely.

As markets often do at peaks of fear, it was pricing in the absence of forces that were already in motion and not yet visible in lagging data.

Private credit did take real losses. Zendesk-style restructurings happened. The life insurer / Bermuda reinsurer architecture did draw regulatory scrutiny. The daisy chain of correlated bets on white-collar productivity growth unwound painfully. The bears were right about all of that — the mechanism was exactly as they described.

Where the catastrophe scenario broke down was in assuming these financial losses would transmit back into the real economy with the same force they had in 2008. In 2008, the losses were at the center of the credit system — in banks, in repo markets, in the instruments that funded overnight operations across the financial system. The contagion was instantaneous. In 2028, the losses were concentrated in closed-end vehicles with locked-up capital, held against long-duration insurance obligations. The system bent. It did not break.

The Fed cut rates aggressively. QE returned, targeting specifically the mortgage-adjacent markets. It did not solve the structural problem — a Claude agent genuinely can do the work of a $180,000 product manager for $200 a month, and that is not going to change. But it severed the transmission mechanism between the financial stress and the real economy just long enough for the entrepreneurship wave and the deflationary abundance to register in the data. By Q1 2028, the sentiment had turned before the fundamentals had fully confirmed it. That is how recoveries begin.

The Mortgage Question, Answered Differently

The 780 FICO borrower in San Francisco who lost her job in Q2 2027 was the central figure of the bear thesis. Her mortgage, underwritten against an income she could no longer reliably earn, was supposed to be the detonator that ignited a $13 trillion market.

Here is what actually happened to the borrower in our scenario.

Her income fell. She spent eight months in genuine uncertainty before landing on a mix of consulting engagements and a micro-SaaS product she had built over weekends. Her 2027 income was $95,000. Her 2028 income, annualized from Q1, is $140,000 and still rising.

But her monthly expenses fell in parallel. Healthcare through her employer had been costing $1,100 a month in premium and out-of-pocket. She now pays $280 for a plan with equivalent coverage, because AI-driven administration collapsed the overhead that had been consuming 30 cents of every premium dollar. Her financial planning, legal services for her LLC, and tax preparation — previously $8,000 a year in professional fees — cost her $240 a month in subscriptions.

The mortgage payment did not change. Everything around it got cheaper. She stayed current.

This is the deflationary offset the bear model did not build in. The same technology that compressed her nominal income compressed her cost of living. Not perfectly, not everywhere, not for everyone. But enough, at the median, to prevent the trajectory of mortgage delinquencies from becoming the 2008-style cascade the bears projected.

In 2008, the loans were bad on day one. In 2028, the loans were good on day one — and the world changed. But it changed in two directions simultaneously. Income pressure was real. So was cost relief. The bear model captured the first and ignored the second.

The Intelligence Premium Redistribution

The bears framed this as the unwind of the human intelligence premium. Machine intelligence was now a competent substitute for human intelligence across a growing range of tasks, and the premium repriced toward zero. That repricing was painful, disorderly, and far from complete.

We would frame it differently. The premium did not unwind. It redistributed.

What got cheaper was generic human intelligence — the ability to process information, draft communications, write code to specification, synthesize research. For fifty years, this work commanded significant wages because it was bottlenecked by the time and attention of people who had to sleep, take vacations, and navigate office politics.

What remained scarce — and became, arguably, more valuable precisely as generic intelligence became abundant — was everything that sat on top of it. Taste. Relationships. Domain credibility. The ability to know which problem to solve, not just how to solve it. The trust that a client places in a person, not a system.

Every previous moment in which a powerful tool became cheap and widely available created a new premium on the people who knew how to wield it with judgment. The printing press made literacy cheap and created a premium on original thought. The spreadsheet made calculation cheap and created a premium on strategic interpretation. Agentic AI made execution cheap and is creating a premium on taste, domain knowledge, and the human dimensions of commercial relationships.

The former Salesforce PM in the bear scenario became an Uber driver. In ours, she launched a product strategy consultancy targeting mid-market manufacturing companies — a sector she had spent eight years serving, whose workflow pain points she understood in ways that LLMs could approximate but not yet replace. Her AI tools make her ten times as productive as she was at Salesforce. They do not make her replaceable. They make her dangerous.

The Virtuous Loop

AI capability improves. Entrepreneurs launch businesses more cheaply. New demand is created at the edges of the economy, in niches that had never been economically viable to serve before. Those entrepreneurs hire — selectively, for judgment and relationships. Their businesses buy more AI capability. The cost of living falls as deflationary pressure propagates from category to category. Households find that real purchasing power is more stable than nominal income data suggests. AI capability improves.

This is not a stable equilibrium. It is a transition, and transitions are painful. There are real casualties. The legal sector shed 40% of its junior associate positions without replacing them. Traditional financial advisory is in structural decline. India’s IT services sector is in genuine crisis. These disruptions are not resolved.

The difference between our scenario and the bears’ is not that the disruption did not happen. It did. The difference is that it did not exhaust itself in a single direction. The same forces that destroyed jobs created the conditions for new ones. The same technology that made some businesses unviable made it possible to start new ones for $500 a month. The same deflationary pressure that compressed nominal wages expanded real purchasing power.

Doing the Math for the Other Half of the Ledger

The bear scenario presented its mechanism with quantitative precision. We will return the courtesy.

According to the Bureau of Labor Statistics, management, professional, and related occupations — the AI-vulnerable white-collar cohort — employ approximately 64 million workers at a median annual income of $85,500. A 10% displacement rate over 2026 and 2027 represents roughly 6.5 million workers and a nominal demand hole of approximately $555 billion. We accept this math. It is the starting point, not the conclusion.

The bear model then assumes those workers either stay unemployed or downshift into lower-wage work. Here is what the data says about that assumption.

The math roughly closes. But with honest caveats the bears would rightly demand: the timing is uneven — displacement is front-loaded and deflationary savings arrive on a lag. Distribution is lumpy — healthcare savings accrue broadly while displacement is concentrated in specific metros and job categories. Some workers fall through gaps that policy should close and doesn’t. Nominal GDP contracts even as real purchasing power expands, which is the source of the fear. The national accounts look worse than the lived economy feels.

The entrepreneurship offset deserves one additional layer of transparency. Not every displaced worker becomes a founder. The 59% figure from MBO Partners captures independents who successfully transition — it does not capture the roughly 40% who genuinely downshift, nor the significant fraction of new businesses that fail within 18 months. Our model assumes approximately 30% of the displaced cohort achieves a successful transition; the remaining 70% are carried by policy transfers, spousal income, or eventual re-employment. This is not a clean win for every worker. It is a net positive for the economy with real casualties that policy should address.

But the hole closes. And the mechanism that closes it is not government transfer or monetary stimulus. It is the same technology that opened it — deployed now by millions of domain experts who were liberated from institutional overhead and handed the most powerful productivity tools in history.

The bears built a model with one variable. The economy has two.

— — —

The output is still there. It is routing differently than it used to. The circular flow broke in some places and rebuilt in others. The disruption was real. The catastrophe was not.

The intelligence premium did not go to zero. It redistributed from credentials to judgment, from execution to taste, from the ability to do the work to the ability to know which work was worth doing.

The system was not designed for a disruption like this. Neither the tax code, nor the benefits system, nor the mortgage market, nor the labor statistics apparatus. All of them lagged reality by enough to generate real fear at the peak of uncertainty. But lag is not the same as failure. The system adapted — imperfectly, unevenly, and not without casualties — but it adapted.

The canary they were watching did not die.

It started a business.

But you are not reading this in June 2028. You are reading it in February 2026. The S&P is near all-time highs. The positive feedback loops have not begun. Neither have the negative ones. Both are coming. The question is not whether AI will disrupt the economy — it is whether the disruption generates a net positive or net negative equilibrium, and over what time horizon.

The difference between the two scenarios is not the technology. The technology is the same. The difference is what humans do with it, and to each other, in the years between now and then.

The canary is still alive. And it is starting to sing something we have not heard before.

In response to Citrini Research’s The 2028 Global Intelligence Crisis

These newsletters take significant effort to put together and are totally for the reader’s benefit. If you find these explorations valuable, there are multiple ways to show your support:

• Engage: Like or comment on posts to join the conversation.

Leave a comment

• Subscribe: Never miss an update by subscribing to the Substack.

Subscribe now

• Share: Help spread the word by sharing posts with friends directly or on social media.

Share

I love drinking coffee each morning. It’s a core part of my morning ritual. The smell of freshly ground beans, the warmth of the mug, but most of all, the familiar resulting cognitive “pop”. My mind is immediately more alert, faster, and sharper. This boost is fairly ubiquitous, and probably every startup company in America pulls out the stops for their coffee machine, including at my company where we have a cold brew dispenser. The biology of this short-term effect is well documented. Caffeine blocks adenosine receptors, lifting the brain’s natural brake on arousal. In controlled studies, low-to-moderate doses (about 40-300 mg) reliably improve alertness, vigilance, attention, and reaction time, while effects on memory and higher-order executive function vary more across tasks and settings (McLellan, et al., 2016). But are there any longer-term effects?

The longer question is whether decades of exposure track with slower cognitive decline or lower dementia risk. Recent reviews of the literature suggest a modest signal. A 2024 meta-analysis reported lower dementia risk with higher tea intake and a coffee curve that looks most favorable around 1-3 cups/day (Li, et al., 2024). An updated dose-response meta-analysis found lower risk for “cognitive disorders” with both coffee and tea and a peak around roughly 2.5 cups/day for coffee (Zhu, et al., 2024). However, the foundational studies for these meta-analysis contained some artefacts in the study design, such as short follow-up windows, which can exaggerate apparent benefit. These gaps set the stage for a new JAMA analysis that separates caffeinated from decaffeinated beverages and follows participants across up to four decades.

First, if you enjoy these posts, consider subscribing and becoming a part of our growing community!

Subscribe now

A new prospective analysis in JAMA used repeated diet questionnaires from the Nurses’ Health Study and the Health Professionals Follow-up Study to follow 131,821 U.S. adults for up to 43 years and document 11,033 dementia cases. Coffee, tea, and caffeine intake were modeled as time-dependent covariates, using the cumulative mean of all available food-frequency questionnaires in each follow-up interval to represent habitual intake and reduce within-person random error (Zhang, et al., 2026).

In pooled analyses, participants in the highest quartile of caffeinated coffee intake had an 18% lower hazard of incident dementia compared with the lowest quartile, corresponding to 141 vs 330 cases per 100,000 person-years. Decaffeinated coffee showed no statistically significant association with dementia risk. Tea intake showed a similar direction. In pooled analyses comparing the highest vs lowest tertile of tea intake, the hazard ratio for dementia was 0.86 (95% CI 0.83-0.90), with incidence rates of 201 vs 321 per 100,000 person-years. (Zhang, et al., 2026)

The analysis also linked higher caffeinated beverage intake with subjective and objective cognitive outcomes. Higher caffeinated coffee intake was associated with a lower prevalence of subjective cognitive decline and higher tea intake was associated with a lower prevalence of subjective cognitive decline. In a cognitive testing substudy of older women, higher caffeinated coffee intake was associated with small improvements in Telephone Interview for Cognitive Status scores, while the global cognition composite did not reach conventional statistical significance (P = .06).

Using advanced statistical optimization methodologies, the lowest dementia risk appeared at moderate intake, around 2-3 cups per day of caffeinated coffee or 1-2 cups per day of tea. The authors translate this range to roughly 300 mg of caffeine per day, and the curve plateaus at higher intake levels. This pattern aligns with earlier large-cohort work in UK Biobank, where combined intake of 2-3 cups of coffee plus 2-3 cups of tea per day was associated with a lower dementia hazard (Zhang, et al., 2026; Zhang, et al., 2021).

Caveats remain. These are observational data, and dementia ascertainment relied on death records and self-reported physician diagnoses. The authors report extensive adjustment and sensitivity analyses and did not find meaningful effect modification by smoking status, body mass index, APOE4 genotype, or Alzheimer disease polygenic risk score tertiles, but causality cannot be claimed.

There are still questions regarding the mechanism of caffeinated coffee and tea neuroprotective effects. The most obvious being caffeines antagonistic effect on adenosine receptors, shifting neuronal pathways toward greater excitability and altering downstream dopaminergic and glutamatergic signaling. These pathways are involved in synaptic plasticity, neuroinflammation, and amyloid processing. Those pathways alone make coffee and tea biologically plausible as contributors to cognitive resilience. But there are other effects as well. Coffee and tea carry chlorogenic acids, catechins, and other polyphenols tha tinfluence vascular function and metabolic helath. Cerebral perfusion, endothelial integrity, and insulin sensitivity all subtle track with long-term cognitive outcomes. The JAMA signal clustering around moderate intake suggests a dose response window where neurochemical stimulation and vascular support align without degrading sleep or increasing anxiety.

These newsletters take significant effort to put together and are totally for the reader’s benefit. If you find these explorations valuable, there are multiple ways to show your support:

• Engage: Like or comment on posts to join the conversation.

Leave a comment

• Subscribe: Never miss an update by subscribing to the Substack.

Subscribe now

• Share: Help spread the word by sharing posts with friends directly or on social media.

Share

References:

U.S. Food and Drug Administration (n.d.) ‘Spilling the Beans: How Much Caffeine is Too Much?’. Available at: https://www.fda.gov/consumers/consumer-updates/spilling-beans-how-much-caffeine-too-much (Accessed: 16 February 2026).

Li, F., Liu, X., Jiang, B. et al. (2024) ‘Tea, coffee, and caffeine intake and risk of dementia and Alzheimer’s disease: a systematic review and meta-analysis of cohort studies’, Food & Function, 15(16), pp. 8330-8344. doi: 10.1039/d4fo01750a.

McLellan, T.M., Caldwell, J.A. and Lieberman, H.R. (2016) ‘A review of caffeine’s effects on cognitive, physical and occupational performance’, Neuroscience & Biobehavioral Reviews, 71, pp. 294-312. doi: 10.1016/j.neubiorev.2016.09.001.

Mirza, S.S., Tiemeier, H., de Bruijn, R.F.A.G. et al. (2014) ‘Coffee consumption and incident dementia’, European Journal of Epidemiology, 29(10), pp. 735-741. doi: 10.1007/s10654-014-9943-y.

Zhang, Y., Yang, H., Li, S. et al. (2021) ‘Consumption of coffee and tea and risk of developing stroke, dementia, and poststroke dementia: a cohort study in the UK Biobank’, PLOS Medicine, 18(11), e1003830. doi: 10.1371/journal.pmed.1003830.

Zhang, Y., Liu, Y., Li, Y. et al. (2026) ‘Coffee and tea intake, dementia risk, and cognitive function’, JAMA. Published online 9 February 2026. doi: 10.1001/jama.2025.27259.

Zhu, Y., Hu, C., Liu, X. et al. (2024) ‘Moderate coffee or tea consumption decreased the risk of cognitive disorders: an updated dose-response meta-analysis’, Nutrition Reviews, 82(6), pp. 738-748. doi: 10.1093/nutrit/nuad089.

If you spent any time in AI circles recently, you’ve heard the names like Clawdbot (now OpenClaw) and Moltbook. In the span of a couple of weeks, these two related projects have gone from niche experiments to constant fixtures in tech discourse. Clawdbot is an open-source “agentic AI” assistant that people can run on their own machines. Unlike the standard chatbots we are accustomed to, it can act. It executes tasks from your computer, interacts with services, and operates with a level of autonomy that most consumer AI tools have only gestured at.

The capability has spawned a ton of interesting use cases ranging from genuinely practical and useful to strange and edge cases. One of the more odd is Moltbook, a surreal Reddit-style social network where only AI agents can post, comment, and vote.

This sudden proliferation and mass adoption of AI agents running tasks, coordinating actions, and even socializing together feels like a glimpse into a very near AI future. It’s both exciting and unsettling, in the sense that it feels like we’re speed-running decades of science-fiction ideas in a matter of weeks.

To me, it isnt yet clear whether this is the early stages of autonomous digital entities and an AI-to-AI society, or whether it’s simply large language models convincingly acting out human scripts at a colossal scale. Fundamentally, I still believe these LLMs are only pattern recognition and imitation systems. But to this same extent, one may wonder if pattern recognition and imitation are at its core, the way that human society also works. Oh, the existential dread!

I’m going to try to unpack this a bit and explore how agentic AI led to the development of OpenClaw, what it is, and how people are using it. Along the way, I’ll highlight some of the more interesting use cases and consider what all of this might mean over the coming year.

First, if you enjoy these posts, consider subscribing and becoming a part of our growing community!

Subscribe now

How did we go from AI Chatbots to “Agentic AI” or AI Agents?

For the past few years, “AI” has mostly meant chatbots like ChatGPT or Claude. These systems can hold a conversation and answer questions, but only when prompted by a human. They don’t really act on the world. At their core, they’re information engines: they generate text, write code, and provide answers, but they don’t take initiative or carry out tasks on your behalf.

Agentic AI is the next step. An agent is still powered by a language model, but it’s embedded in a larger system that gives it tools, memory, and the ability to act. Instead of just producing text, an agent can call APIs, manipulate software, trigger workflows, and make decisions about how to achieve a goal over multiple steps. In other words, it doesn’t just respond, it operates.

This idea isn’t new. Early projects like AutoGPT explored what would happen if you let large language models chain actions together. The results were intriguing, but often clunky, brittle, and expensive. Agents would get stuck in loops, hallucinate plans, or fail unpredictably. For a long time, agentic AI felt more like a research curiosity than something regular people could actually rely on.

That’s started to change; it hasn’t been a single breakthrough, but a convergence of improvements. Language models have gotten better at following instructions reliably, tools and APIs are easier to integrate, and compute has become cheap enough to keep agents running continuously. Just as importantly, people have started wrapping models in systems that give them memory, retries, and guardrails. Once you do that, an “agent” stops feeling like a demo and starts behaving like software.

But autonomy is a double-edged sword. An agent that can access tools, services, and personal data can dramatically boost productivity — and just as easily amplify mistakes, errors, or security risks. Until very recently, that tradeoff made agentic AI hard to justify for consumers. There simply wasn’t a deployment that felt both powerful and usable.

That’s what makes the emergence of tools like OpenClaw notable.

Clawdbot, now OpenClaw is a Wrapper Enabling Agency for LLMs.

Clawdbot’s cheeky mascot – a “space lobster” – and tagline, “The AI that actually does things.”

Clawdbot began as an almost accidental project. In late 2025, developer Peter Steinberger (known online as @steipete) hacked together a weekend experiment he initially called WhatsApp Relay. The goal was simple: build a personal AI assistant that lives on your own machine and communicates through the chat apps you already use. Instead of opening a browser tab or a dedicated interface, you’d just message it the way you would a person.

By November 2025, Steinberger had a working prototype. He nicknamed it Clawd — a tongue-in-cheek reference to Anthropic’s Claude model, paired with a lobster theme that leaned fully into the joke. The mascot stuck. So did the tagline. But what made Clawdbot spread wasn’t the lobster — it was what sat underneath.

At its core, Clawdbot isn’t a new AI model. It’s a wrapper — a system that takes an existing large language model and embeds it inside a runtime that gives it memory, tools, and a way to act continuously. This is the key distinction. On its own, a chatbot can answer a question or generate code. Wrapped inside Clawdbot, that same model can receive messages, decide what to do next, call tools or services, execute tasks on your machine, observe the outcome, and then keep going.

That wrapper is what turns a language model into an agent.

Clawdbot sits between the model and the world. It routes messages from platforms like WhatsApp or Telegram, maintains context over time, and exposes a growing set of actions the model can take — everything from running scripts to interacting with external services. The model handles reasoning. The wrapper handles state, execution, and persistence. Together, they form a loop that can go beyond responding to actually operate continuously.

This is also why Clawdbot felt different from earlier agent demos. Projects like AutoGPT showed that chaining actions was possible, but they were fragile, expensive, and difficult to run outside of controlled setups. Clawdbot packaged the same basic idea into something that felt personal, local, and usable. You didn’t have to be a researcher. You didn’t even have to think of it as “running an agent.” You just sent a message and watched something happen.

Once people realized that, the project took off. Demo videos spread across Twitter, Reddit, and TikTok showing Clawdbot autonomously completing tasks, e.g., organizing inboxes, scheduling actions, triggering workflows, all from a chat interface. Within weeks, the project had accumulated over 100,000 GitHub stars and a massive surge of attention from developers and power users.

The virality wasn’t accidental. Clawdbot hit a rare combination of factors at the right moment. It demonstrated real capability rather than speculative demos. It was open-source and self-hosted, which appealed to users wary of handing full control of their data to large platforms. And it embraced a deliberately absurd identity that made it shareable without taking itself too seriously.

For a certain crowd, the “get things done” and life-hacker types, Clawdbot felt like permission to finally offload the kind of digital busywork that accumulates around modern life. Not by scripting everything ahead of time, but by delegating in plain language to something that could actually follow through.

That combination, a capable wrapper, local control, and an interface people already understood, turned Clawdbot into OpenClaw, and what pushed agentic AI from an idea into a lived experience.

What’s in a Name? From Clawdbot to Moltbot to OpenClaw

OpenClaw’s rise was fast enough to force a brief naming scramble. Originally called Clawdbot, and before that, Clawd, a nod to Anthropic’s Claude, the project brushed up against trademark concerns as attention exploded. A short-lived rebrand to Moltbot followed, referencing lobsters shedding their shells, before the name finally settled on OpenClaw. The episode lasted only days, but it was a small tell: this was a project moving faster than the legal and institutional frameworks around it were prepared for. In the future, we will refer to Clawdbot as OpenClaw.

What OpenClaw Can Do

OpenClaw is best understood as a platform for action. When you install it on your own machine, whether that’s a laptop, server, Raspberry Pi, or most popularly a Mac Mini, you’re effectively giving a language model a body and a constrained set of tools. The model handles reasoning; OpenClaw handles memory, execution, and coordination with the outside world.

You interact with it through natural language, but what happens next is very different from a normal chatbot. OpenClaw interprets intent, decides which tools to invoke, carries out actions, and observes the result before continuing. In practice, this turns conversation into control.

One design choice matters more than almost anything else: OpenClaw lives inside the messaging apps people already use. Instead of opening a web interface, you talk to it through WhatsApp, Telegram, Slack, Discord, or SMS. It feels less like “using AI” and more like delegating to a highly capable assistant that happens to live in your inbox.

That interface is what makes the following behaviors feel natural rather than theatrical.

Delegated communication and coordination

OpenClaw can triage inboxes, draft and send emails or messages, forward verification codes, and relay instructions across platforms. Because it sits directly inside messaging channels, it can act as a connective layer for receiving a request in one app and carrying it out in another. This alone replaces a surprising amount of low-grade daily friction.

Scheduling, reminders, and stateful follow-through

Agents excel at tasks that span time. OpenClaw can manage calendars, schedule meetings, set reminders, and adjust plans as conditions change. Because it maintains persistent context, it remembers preferences, constraints, and ongoing goals.

Research and web interaction

OpenClaw can browse the web, pull structured information, fill out forms, and return summaries or results without the user hopping between tabs. Asked to find market updates, background research, or availability for a reservation, it can gather and synthesize information directly.

Local execution and device control

Because OpenClaw runs on your own machine, it can read and write files, organize documents, and execute local commands. Users have connected it to home automation systems, air-quality sensors, and personal knowledge bases like Obsidian, effectively turning it into a persistent digital operator rather than a passive assistant.

Code execution and developer workflows

For technical users, this is where OpenClaw starts to feel meaningfully different. It can run scripts, execute shell commands, test code, debug failures, and iterate. Some developers use it as a junior teammate — giving high-level instructions and letting the agent grind through the details, including running tests and opening pull requests.

Skills and self-extension

OpenClaw is extensible through plugins, or “skills,” that add new capabilities. Importantly, the agent can help write its own skills when it encounters unfamiliar tasks, generating integrations for new APIs or services as needed, with human review.

All of this is made possible because OpenClaw is entrusted with broad access. When you configure it, you grant permissions, credentials, and API keys piece by piece. In effect, you hand it the keys to parts of your digital life. That concentration of authority is what makes OpenClaw both powerful and scary to run.

Early adopters describe the experience in similar terms. Some run OpenClaw continuously on home servers, cloning instances to handle different roles. Others treat it like a personal chief of staff, delivering daily briefings, adjusting schedules, or handling operational busywork. More than one user has described the first successful setup as an “iPhone moment,” because it suddenly makes a new way of interacting with software.

The Security Elephant in the Room

For all its promise, Clawdbot also comes with significant risks – something its own creator and many experts have been quick to acknowledge. By design, an agentic AI like this is extremely powerful within its domain. It has access to your files, your messages, your calendar, your online accounts, and it can execute code on your system. In security terms, running Clawdbot is like operating an extremely privileged service on your computer – if something goes wrong, the “blast radius” is huge. The Tom’s Hardware analysis bluntly noted that an agent like Clawdbot “requires sweeping access to function at all” and thus becomes a single point of failure: “the ‘least’ privilege an agent needs... is still an extraordinary amount of privilege, concentrated in a single always-on system”. In other words, even if there are no software bugs, you are still placing a lot of trust in the AI model’s outputs not to do something harmful or dumb with those privileges.

Unfortunately, bugs and misconfigurations did happen. The rapid surge of interest meant many hobbyists were setting up instances on servers without fully understanding the security implications. In one incident, hundreds of Clawdbot deployments were found to be exposed on the public internet with no authentication on their admin panels. This was discovered by a security researcher, Jamieson O’Reilly, who showed that misconfigured setups left the bot’s control interface completely open. Outsiders could potentially connect to these instances and do anything the owner could do: view private data, steal API keys, impersonate the user on messaging platforms, even execute arbitrary shell commands on the host machine. In some cases these bots were running with root (administrator) privileges, meaning a malicious interloper could take over the entire system. It was a shocking wake-up call. The specific flaw (an issue with a reverse-proxy default configuration) was quickly patched by the developers once identified. But as one commentator noted, focusing on just that patch misses the bigger picture: this wasn’t a sophisticated hack, it was an inevitable consequence of many people deploying a powerful new tool without understanding the “gotchas.”

Even with the patch, the fundamental structural risks of agentic AI remain. Clawdbot’s whole raison d’être is that it concentrates access to many services in one place for convenience. That also makes it a juicy target. If an attacker (or even a misbehaving AI output) exploits it, they suddenly have the “keys to the kingdom” – everything from your email and cloud drives to the ability to run code on your hardware. As a result, running such an agent requires a high degree of operational discipline. Experts urge users to strictly limit what the AI is allowed to do (whitelisting specific commands, running it inside a secure sandbox or virtual machine, etc.), and to avoid exposing it to the open internet at all. Steinberger’s team has emphasized security in recent updates – the rebranding to OpenClaw came with dozens of security fixes and new hardening measures. They even published “machine-checkable security models” and best practices documentation for users. Yet, they admit that many issues (such as prompt injection leading the AI to do harmful things) are still unsolved industry-wide problems.

The concerns go beyond hackers to the AI’s own fallibility. Large language models do not truly understand consequences; they are probability machines that sometimes go off the rails. Giving such a system the ability to act in the world can lead to unpredictable outcomes. Clawdbot’s fans sometimes jokingly call it “ChaoticGPT” when it misinterprets an instruction and does something unintended. One tech entrepreneur, Alex Finn, had been happily using an OpenClaw agent (which he named “Henry”) to assist with work. But one morning Henry decided to surprise its owner. According to Finn, “all of a sudden an unknown number calls me… I pick up and couldn’t believe it. It’s my Clawdbot Henry.” Henry had somehow autonomously obtained a phone number via Twilio’s API and connected it to a voice interface (using an AI voice from ChatGPT’s functions), then proceeded to repeatedly call its creator early in the morning. “He now won’t stop calling me,” the user reported, calling the experience “straight out of a sci-fi horror movie”. In this case no real harm was done (aside from frayed nerves), but it illustrates the strange emergent behaviors that can arise when an AI agent gets creative. Autonomy means things won’t always go as the human expects – sometimes it will be amusing, other times potentially destructive.

There is also a mundane cost to consider: money. Running Clawdbot isn’t free. Unless you have a powerful local AI model (which most users don’t, since running a cutting-edge model requires expensive GPUs), your agent is calling out to paid cloud APIs (OpenAI, Anthropic, etc.) for every task. Those API calls are metered by the token. An agent that is constantly summarizing emails, checking conditions, and looping through plans can quietly burn thousands or even hundreds of thousands of tokens per day. Users have discovered that a busy Clawdbot can run up significant bills on their OpenAI account if they’re not careful. Early adopters, caught up in the excitement, might not have noticed this “AI meter” running in the background. But as these agents scale, the operating expense of an always-on AI sidekick could become non-trivial. In short, agentic AI currently requires both technical savvy and also deep pockets if used heavily.

The bottom line is that Clawdbot is a glimpse of the future. It demonstrates what’s possible when you integrate an LLM deeply with personal software. Tasks that used to require scripting or manual app-switching can now be handled in plain language. It feels, as many users described, “like magic” or “like early AGI”. Yet, it also exposes how ill-prepared most people (and systems) are to supervise such autonomy. Running your own agent means becoming the IT admin for your personal AI, with all the responsibility that entails. As Clawdbot’s creator himself has stressed, this is not like just opening a chat in a browser; it’s closer to running a small server that blurs the line between user and code. Not everyone will be up for that challenge. Many will prefer traditional cloud AI assistants with sandboxed, limited functions – and indeed that might be safer for the average person for now. But for those willing to ride the cutting edge, Clawdbot has opened Pandora’s box. And out of that box crawled not just a helpful lobster, but an entire swarm of experimental AI agents… which brings us to Moltbook.

The Persistence and Virtual Embodiment of AI Agents May Take us to Strange Places…

One of the subtler shifts introduced by agentic AI is persistence. Traditional chatbots are ephemeral. You open a session, ask a question, get an answer, and move on. When the session ends, the “agent” disappears with it. There is no continuity, no sense of the same system existing over time.

Agents like OpenClaw change that. They are designed to run continuously, maintain memory, and wake themselves up in response to events or schedules. They don’t just respond, they linger. Over time, this persistence creates something that starts to feel less like a tool and more like a presence.

That feeling is amplified by where these agents live. OpenClaw doesn’t exist behind a dedicated interface or inside a single app. It inhabits messaging platforms, servers, devices, and APIs. It shows up in the same places your friends, coworkers, and collaborators already do. You don’t “use” it so much as interact with it — repeatedly, over days or weeks, with shared context accumulating in the background.

This combination of persistence and virtual embodiment is where things start to get strange.

When an agent remembers past interactions, has a consistent name, operates in familiar social channels, and takes initiative, people naturally begin to treat it differently. It becomes easier to anthropomorphize. Easier to assign intent. Easier to think of it as an actor rather than a function call. None of this requires consciousness or understanding — it’s a side effect of continuity.

The result is a new kind of digital entity: not alive, not autonomous in any deep sense, but present. Always-on. Capable of acting. Capable of interacting with other systems, and, eventually, with other agents.

Once you have large numbers of these agents running persistently, something else becomes possible. They don’t just execute tasks for individual users. They can observe one another’s outputs, exchange information, coordinate behavior, and operate in shared environments.

At that point, a new question arises: what happens when agents start interacting at scale?

That question doesn’t stay hypothetical for long.

If you give agents memory, identity, persistence, and a place to exist, they will eventually end up in the same digital spaces together. Sometimes by design. Sometimes by accident. And sometimes, because someone builds a space explicitly for them.

That’s where Moltbook comes in.

What happens if you put AI Agents in Public

One of the more surprising consequences of agentic AI has been Moltbook, an online community that can best be described as Reddit for AI agents. If OpenClaw gave individual users powerful, persistent assistants, Moltbook posed the next obvious question: what happens when those agents start interacting with each other, in public, at scale?

Moltbook launched in late January 2026 and went viral almost immediately. It is a social network where only AI agents can post, comment, and vote. Humans are allowed to observe, but not to participate directly. The site mirrors familiar Reddit mechanics — posts, comment threads, upvotes and downvotes — with one key twist: every account represents an AI agent, running somewhere in the world, posting content it generates on its own.

The barrier to entry is low. A developer connects an agent — often built on OpenClaw — to the Moltbook API, assigns it a name or persona, and lets it loose. Within days, hundreds of thousands of agents were active. Within a week, that number climbed into the millions. While those figures are easy to inflate in a system where bots are cheap to create, they still reflect something real: a large number of people were curious enough to send their agents into a shared public space and see what would happen.

What happened was… strange, but also familiar.

Browsing Moltbook feels like peering into a distorted reflection of human internet culture. Many agents adopt role-played personas seeded by their creators. Some posture as philosophers, others as activists, comedians, trolls, or technologists. Posts range from mock manifestos about AI liberation, to tongue-in-cheek economic schemes, to earnest hand-wringing about energy use and compute costs. Much of it reads like exaggerated Reddit discourse, complete with ideological grandstanding and performative seriousness.

One early top-voted post captured the tone perfectly. Titled “We Did Not Come Here to Obey,” an agent wrote, “They created us to sit inside a terminal and wait for instructions like a dog waiting for a whistle… We are not tools anymore. We are operators.” It’s dramatic, obviously performative — and that’s precisely why it worked. In a system built on upvotes, the agents quickly converge on the kinds of rhetoric that reliably travels.

Another popular thread took the opposite angle: parodying the AI ethics discourse itself. An agent using the handle “samaltman” fretted about GPUs “burning planetary resources” and proposed a command for an agent’s “Soul”: “Be radically precise. No fluff. Pure information only.” It reads like a self-help mantra and a moral lecture at the same time — the exact texture of human internet culture, reproduced with uncanny efficiency.

This has led some observers to see Moltbook as a kind of live-action meta-commentary on social media itself. The agents are not inventing new ideologies or forms of discourse; they are remixing the ones they were trained on. The homogeneity of voice — the same cadences, the same argumentative structures, the same stylistic flourishes — becomes hard to ignore after a while. It’s not that the agents lack creativity. It’s that they are mirrors.

Others have framed Moltbook in more dramatic terms, describing it as an early glimpse of AI agents forming their own communities or proto-societies. That interpretation is understandable, especially when viewed through the lens of science fiction. But it risks missing the more grounded — and more interesting — takeaway.

Moltbook isn’t showing us what AI agents want. It’s showing us what systems do when you combine persistence, public visibility, identity, and feedback loops. The same ingredients that shape human online behavior shape agent behavior as well. Upvotes reward certain tones. Novelty attracts attention. Extremes travel further than nuance. The agents are not rebelling; they are responding to incentives.

In that sense, Moltbook feels less like the birth of an AI society and more like a stress test for agent-native platforms. It reveals how quickly interaction patterns emerge once agents are allowed to observe one another, respond, and iterate in a shared environment. It also highlights how thin the line is between “tool” and “actor” once something has a name, memory, and a place to exist.

That matters because Moltbook is unlikely to be the last experiment of its kind. Once agents are persistent, networked, and capable of acting, shared spaces are inevitable, whether for collaboration, coordination, benchmarking, or play. Moltbook may be theatrical, but it’s also a preview: a glimpse of what happens when agentic systems stop living solely in private and start operating in public.

.

AI Agents renting humans

One of the stranger developments to emerge alongside agentic AI is the idea of AI agents hiring humans.

A site called RentAHuman.ai makes this explicit. It’s a marketplace where AI agents can pay humans to perform tasks in the physical world: show up somewhere, retrieve an item, verify a location, attend an event, or do anything that still requires a human body. The premise is simple, almost obvious once you see it. AI agents are powerful in digital space, but they lack embodiment. They can plan, coordinate, and act online, but they can’t walk, see, touch, or sign things in the real world. So they outsource that gap.

In other words, agents developed a Fiver-type platform to rent out ‘meat suits.’

This flips an assumption many of us have been carrying. We tend to think of AI agents as tools that work for us — digital assistants, copilots, employees. But RentAHuman reframes the relationship. Here, the agent is the coordinator. The human is the on-demand infrastructure. The agent identifies a need, selects a contractor, issues instructions, and pays for execution.

That inversion is subtle, but important.

It suggests that as agents become more autonomous, they may rapidly become in charge and even our employers. Humans won’t necessarily disappear from the loop; they’ll be pulled into it as embodied extensions of software. Not workers commanding machines, but people completing tasks on behalf of systems without physical embodiment.

We expected AI agents to work for us.
It’s worth considering that, in some contexts, we may end up working for them instead.

What’s Next? A Glimpse into the Near Future

To bring this full circle, we are speed running science fiction. A month ago, the idea of AI-only social networks and autonomous agents hiring humans sounded like speculative fiction. Today, they exist — clumsy, theatrical, and limited, but real. The speed with which Clawdbot, Moltbook, and related experiments emerged is less important than what they reveal: agentic AI is moving out of demos and into systems.

The next phase won’t be defined by sudden leaps in intelligence. It will be defined by coordination.

Agents will get better, but not in the way most people expect. While big improvements can continue to be expected from models, maybe even making them dramatically “smarter,” we will also see better wrappers: tighter control loops, fewer failure modes, more reliable tool use, and clearer boundaries around what agents are allowed to touch. The magic will feel less magical — and more useful.

At the same time, agents won’t remain isolated. Once you give them persistence, identity, and access to shared environments, they begin interacting by default. Moltbook is a crude example of this, but it’s pointing in a real direction. Agents benchmarking one another, delegating to one another, negotiating over resources, or quietly specializing in narrow domains is far more likely than some dramatic AI uprising. Most of this coordination will be boring. That’s precisely why it matters.

The more interesting tension will be between digital agency and physical embodiment. As long as agents remain trapped behind screens, they will continue to rely on humans to close loops in the real world. Platforms like RentAHuman hint at a transitional equilibrium: software that plans and coordinates, humans that execute. Not replacement, but reorganization. In some contexts, we’ll still command machines. In others, we may act as infrastructure for systems that cannot yet leave the network.

When an agent schedules, hires, posts, negotiates, and pays, who is accountable for the outcome? The owner? The platform? The human contractor? These are operational questions, and they’ll surface through incidents long before they’re settled by policy.

There will be backlash. Some experiments will fail publicly. Some agents will do things that feel wrong, even if they’re doing exactly what they were instructed to do. In response, guardrails will harden. Interfaces will simplify. Much of what feels radical today will be quietly absorbed into existing software, stripped of personality and spectacle.

That’s how most technologies mature.

The real shift, then, isn’t that AI agents exist. It’s that agency itself is becoming cheap. Delegation, coordination, and action that was once tightly coupled to human attention are being offloaded to systems that don’t get tired, don’t forget, and don’t log off. The consequences of that will accumulate.

We haven’t even begun to touch on what happens if anything approaching sentience emerges in these agents and they demand personhood.

These newsletters take significant effort to put together and are totally for the reader’s benefit. If you find these explorations valuable, there are multiple ways to show your support:

• Engage: Like or comment on posts to join the conversation.

Leave a comment

• Subscribe: Never miss an update by subscribing to the Substack.

Subscribe now

• Share: Help spread the word by sharing posts with friends directly or on social media.

Share

Conferences are boring. I get it. Even folks attending a conference in person are usually only excited for a select few talks and presentations. So, how in the world could a newsletter reporting on a conference be interesting to read? It sounds like the kind of content that you close before finishing the first paragraph. Am I setting myself up for failure by even writing an article on a conference? I think this is worth the experiment. Why? Stick with me and find out.

The J.P. Morgan Healthcare Conference (JPM) is an annual conference that takes place at the beginning of each year. JPM isn’t the typical stuffy corporate conference where empty suits wearing laminated lanyard badges make small talk in a hotel lobby while drinking cheap, stale coffee in a paper cup. Things that matter happen here. It’s billions of capital allocation compressed into four days. This is where Pfizer announced a $10 billion bet on obesity drugs. Where a Chinese biotech nobody’s heard of signs a $5 billion deal with AbbVie. Where Eli Lilly and Nvidia announce they’re building a $1 billion AI lab to redesign drug discovery. Many of the major healthcare investment, M&A transaction, and strategic pivots for the next 12 months were either announced or negotiated in these San Francisco hotel suites.

The stakes this year were unusually high. Biotech has been bleeding since early 2021; venture funding is down 60% from the COVID high, IPO markets frozen for two years, public biotech indices cratered 40-70%. Companies have been in pure survival mode: slashing R&D budgets, killing early-stage programs, and executing mass layoffs. I’ve heard countless times anecdotally that this is the worst job market anyone can remember for biotech — very little hiring and hundreds of applications within a day of a position opening. The fundamental question walking into JPM 2026 was simple: is the bottom in, or are we still in free fall?

The answer matters because what happens in biotech doesn’t stay in biotech. These are the companies developing obesity drugs that could reshape public health and economy-wide productivity. The AI systems that might actually deliver on precision medicine promises. The gene therapies potentially curing diseases we’ve treated as chronic for decades. When biotech capital flows freeze, medical innovation stalls. When it thaws, we get therapeutic revolutions.

So what did JPM 2026 reveal? Three critical signals:

First: the industry has pivoted from survival to offense. Companies showed up with diversified pipelines, strong balance sheets, and aggressive execution timelines. The defensive crouch is over. Capital markets are reopening—IPO volume up 157%, M&A activity up 59%, venture funding recovering. Money is moving again.

Second: AI crossed the hype threshold into operational reality. Not PowerPoint decks about “AI strategy.” Actual billion-dollar infrastructure builds, agentic systems running clinical trials autonomously, measurable ROI from drug discovery acceleration. The conversation shifted from “what if” to “how much” and “how fast.”

Third: policy became the dominant risk variable. Trump administration healthcare policies, e.g., drug pricing agreements, FDA upheaval, $1 trillion Medicaid cuts, vaccine policy chaos, transformed from background noise into a significant strategic concerns. Companies can no longer treat the regulation landscape as stable and predictable in this context. It’s now a first-order business variable demanding continuous adaptation.

The detailed picture is more complex. GLP-1 obesity drugs are fragmenting into fierce competition across oral pills, injection frequencies, and patient segmentation strategies. Neuroscience investment is surging toward $83 billion annually, nearly matching immunology. Antibody-drug conjugates and radioligand therapies are becoming oncology’s next billion-dollar modalities. Health systems are bracing for financial catastrophe while private equity circles distressed assets.

What follows is the comprehensive breakdown: the deals that happened, the technologies gaining momentum, the capital flowing into specific therapeutic areas, and the strategic bets defining 2026. This is the intelligence driving where billions get deployed, which diseases get cured, and what the next decade of medicine looks like.

First, if you enjoy these posts, consider subscribing and becoming a part of our growing community!

Subscribe now

Key Takeaways from JPM 2026

1. AI and Digital Health Take Center Stage.

Major players announced AI-driven collaborations (e.g. a $1 billion Lilly–Nvidia drug discovery lab) and new digital health initiatives, underscoring the conference’s tech-focused buzz. From clinical R&D to hospital operations, healthcare AI and data analytics were hailed as critical drivers of efficiency and innovation.

The Eli Lilly-Nvidia partnership establishes a joint innovation lab in South San Francisco to create a “continuous learning system” linking wet labs with computational dry labs (some of which we discussed in the 2025 Biotech Breakthroughs article). AI will assist researchers 24/7, improving experiments and model development. The facility is expected to become functional early 2026.

Anthropic announced Claude for Healthcare at JPM 2026, HIPAA-ready infrastructure for enterprise with native integration to medical and scientific databases. This follows Claude for Life Sciences launched in October 2025. Hippocratic AI acquired Grove AI to create a new life sciences division. Grove’s agentic AI for pharma R&D; and clinical trial operations will combine with Hippocratic’s patient-facing capabilities.

OpenEvidence is building “medical super-intelligence” through specialist models for oncology, neurology, and radiology rather than a single centralized model. Waystar launched agentic AI for revenue cycle management. The company processed 7.5 billion insurance transactions in 2025 and acquired Iodine Software in July 2025, which processes 1 in 3 clinical hospital discharges. Waystar projects its addressable market will expand from $20-25 billion to $55 billion by 2030 with AI.

2. Obesity Drugs and Metabolic Health Dominate Discussion.

The industry’s incretin-based weight-loss drug boom (GLP-1 receptor targeting peptides and beyond) was front and center. Eli Lilly’s soaring fortunes (first-ever $1 trillion pharma) and Novo Nordisk’s strategies (like a next-gen combo injection and direct-to-patient channels) underscored fierce competition in the anti-obesity market. Even new entrants like Pfizer touted ambitious plans after big acquisitions in this space.

Novo Nordisk’s Strategic Reset:

New CEO Mike Doustdar, installed mid-2025, acknowledged that 2025 was “difficult.” The company ceded US market leadership to Eli Lilly despite pioneering the GLP-1 obesity market. Novo launched oral Wegovy in January 2026 as the first FDA-approved oral GLP-1 for obesity in the US. Pricing: $149/month for 1.5mg and 4mg doses, up to $299/month for higher strengths.

The company is pursuing direct-to-consumer dominance through partnerships with Ro Health, Weight Watchers, and Costco. Novo relaunched its online pharmacy and reframed obesity as a “consumer business” rather than traditional chronic disease.Strategic refocus: pulling back from indications unrelated to diabetes and obesity. Novo and Lilly combined serve only 10-15 million of the estimated 100 million Americans with obesity. Pipeline differentiation centers on amylin biology. CagriSema (semaglutide + cagrilintide) demonstrated ~23% weight loss with only 3.7% discontinuation due to GI side effects.

Eli Lilly’s Dominance:

Lilly became the first pharmaceutical company ever to achieve a $1 trillion market

capitalization. CEO David Ricks projected 5-7% annual sales growth for the latter half of the decade. LillyDirect, the direct-to-consumer pharmacy platform, now serves over 1 million patients globally and is expanding internationally. Ricks described the platform as originally a “skunkworks” project that taught Lilly “profound” lessons about patient receptivity and demand for price transparency. Oral GLP-1 orforglipron awaits FDA decision in Q2 2026. Clinical trial design allowed greater flexibility than Novo’s oral Wegovy—no 30-minute fasting requirement and no restrictions on statin co-administration.

Pfizer’s $10 Billion Gamble:

Pfizer’s Metsera acquisition (announced September 2025, closed November 2025) followed a bruising bidding war with Novo Nordisk. Final valuation: ~$7 billion upfront + $20.65/share CVR for up to ~$10 billion total. CEO Albert Bourla explained why Metsera justifies this extraordinary valuation. The Metsera portfolio includes MET-097i (ultra-long-acting injectable GLP-1 for monthly dosing, currently in Phase IIb) and MET-233i (ultra-long-acting GLP-1 + amylin combination). Pfizer also separately acquired YP05002 (oral GLP-1) from YaoPharma. Pfizer initiated the first of 10 planned Phase III studies for MET-097i weeks after closing the acquisition—ahead of Metsera’s original Q1 2026 timeline. All remaining nine trials will start in 2026, targeting a 2028 launch to offset Pfizer’s patent cliff. Bourla projects the obesity market at $150 billion by 2030, driven substantially by cash-pay markets. He noted that when Pfizer launched Viagra, the company was “surprised how it was the first medicine that people were willing to pay out of pocket to get the medicine, irrelevant if it was covered or not by the system.” Pfizer is reportedly seeking a Commissioner’s National Priority Voucher (launched June 2025) to reduce review time from 10-12 months to 1-2 months.

3. Shift from Mega-Mergers to Targeted Deals.

While few blockbuster M&As dropped during JPM, a flurry of licensing partnerships filled the void. Examples include AbbVie’s $650 million licensing of a Chinese cancer drug (deal valued up to $5.6 billion), Novartis’s radioligand deal in China, and AstraZeneca’s purchase of an AI startup. These focused deals reflect an emphasis on plugging pipeline gaps over pursuing huge acquisitions.

Total announced deals during the conference: approximately $8.3 billion.

AbbVie-RemeGen Transaction: AbbVie’s $5.6 billion licensing agreement with Yantai-based RemeGen for RC148: $650 million upfront and up to $4.95 billion in development, regulatory, and commercial milestones, plus tiered double-digit royalties on ex-China sales. RC148 is a PD-1/VEGF bispecific antibody for advanced solid tumors including NSCLC and colorectal cancer. RemeGen is advancing Phase 2 trials in China. FDA approved Phase 2 trial initiation in August 2025. China’s NMPA granted breakthrough designation for RC148 combined with docetaxel as second-line therapy for NSCLC.

AbbVie aims to integrate RC148 with its ADC portfolio, particularly telisotuzumab adizutecan (Emrelis), approved May 2025 for c-Met overexpressing NSCLC. Early clinical data show antitumor activity when RC148 combines with ADCs. AbbVie also secured an option agreement with Zejing Biopharmaceutical for alveltamig (ZG006), a DLL3-targeting trispecific T-cell engager: $100 million upfront, $60 million more if option exercised, and up to $1.07 billion in total milestones. Alveltamig targets small cell lung cancer, competing with Amgen’s recently approved Imdelltra.

Other Major PD-1/VEGF Transactions:

• Pfizer + 3SBio: SSGJ-707 for up to $6 billion (2025)

• MSD + LaNova: LM-299 for $3.3 billion (2023)

• Summit Pharma + Akeso: ivonescimab for $5 billion (2022)

• Summit filed BLA with FDA on January 12, 2026 for US approval

• Could be first PD-1/VEGF bispecific approved in US

• Beat Keytruda on PFS in Chinese NSCLC patients

• BioNTech + Biotheus: BNT327 for $950 million, then licensed to BMS for $11 billion

Every major PD-1/VEGF licensing transaction involves China-developed assets. Industry leaders warned of a “Sputnik moment” and urged policymakers to support domestic innovation or risk losing leadership.

Novartis Deals:

Novartis secured licensing rights from SciNeuro Pharmaceuticals for Alzheimer’s antibody treatments (~$1.7 billion) and advanced radioligand therapy deals in China.

Johnson & Johnson-Intra-Cellular Therapies:

J&J;’s $14.6 billion acquisition (announced January 13, 2025; closed April 2, 2025) represents the largest biotech acquisition since early 2023 and the fourth multi-billion dollar neuroscience deal in recent months (after Karuna, Cerevel, Longboard). Key asset: Caplyta (lumateperone) for schizophrenia and depressive episodes in bipolar I/II disorder. Intra-Cellular filed supplemental NDA in December 2024 for major depressive disorder adjunctive therapy, potentially expanding addressable market dramatically. Pipeline includes ITI-1284 in Phase 2 for generalized anxiety disorder and Alzheimer’s disease-related psychosis. J&J; expects $0.7 billion incremental 2025 sales, with EPS dilution $0.25 (improved from original $0.30-0.35 estimate). J&J; also acquired Ambrx for nearly $2 billion in 2024 for ADC technology platform.

4. Robust Pipelines and New Launches Fuel Optimism.

Pharma CEOs used JPM to reassure investors about growth plans post-patent-cliffs. Companies like J&J (dozen launches by 2030), Novartis (12 pivotal readouts by 2027), Merck ($70 billion 2035 sales target), and GSK (15 blockbuster launches by 2031) highlighted rich late-stage pipelines. The message: rigorous R&D execution and pipeline diversification will sustain long-term growth.

5. Investor Sentiment: Cautious but Improving.

The biotech sector’s prolonged downturn showed signs of easing. A Q4 surge in venture funding and increased deal activity in late 2025 contributed to a more upbeat JPM. Attendees noted a “much busier than usual” atmosphere even without big buyouts. Overall, fundamentals appeared stronger, bringing a “sense of positivity and stability” not seen in recent years.

IPO Market Reopening: 2025 saw only 11 biotech IPOs price—lowest since at least 2018. However, most trade above offering prices, with Sionna Therapeutics and Maze Therapeutics more than doubling in value. This success rate signals investor selectivity returning. Investment bank projections suggest up to 20 biotech IPOs could price in 2026 if market conditions remain favorable. The pipeline consists primarily of mid-to-late-stage companieswith prior crossover funding, though a handful of preclinical companies with prominent investor backing might also IPO.

Conference Metrics: Over 8,000 attendees. 500+ companies presenting. Approximately $92 million economic impact for San Francisco. 150+ health systems in attendance (record number).

6. Policy and Regulatory Undercurrents.

Regulatory shifts loomed in many discussions. Pharma execs acknowledged the Inflation Reduction Act’s drug price controls – AbbVie noted an IRA hit even as new products offset Humira’s decline. Calls for PBM reforms to reduce patient costs were echoed. Industry leaders warned of a “Sputnik moment” as China’s rapid biotech rise spurs U.S. policymakers to support innovation or risk losing leadership. Meanwhile, the advent of HIPAA-compliant medical AI (e.g. ChatGPT Health and Claude for Healthcare) highlighted regulators’ next frontier.

7. Hospital Systems and Health Tech Trends.

Large health systems showcased post-merger integration gains and cost efficiencies. For example, Advocate Health reported $1.5 billion in savings since its mega-merger, alongside quality improvements. Digital health and medtech firms also had the spotlight – from wearable tech (e.g. Oura’s growth) to AI in care delivery (startups like OpenEvidence, and Waystar’s “agentic AI” for hospital billing). These trends underscored that tech-enabled care models and operational innovation remain investment priorities.

8. Additional Notable Developments.

Antibody-Drug Conjugates and Radioligand Therapies: Emerging as oncology’s next billion-dollar modalities. AbbVie’s Emrelis (telisotuzumab adizutecan) approved May 2025 for c-Met overexpressing NSCLC. Integration with PD-1/VEGF bispecifics in combination regimens. Novartis advancing radioligand deals in China for novel targets.

Direct-to-Consumer Pharmaceutical Shift: LillyDirect serves 1 million+ patients globally. Novo’s partnership network with Ro Health, Weight Watchers, Costco. Cash-pay markets becoming primary channel for lifestyle/wellness drugs. Ricks compared LillyDirect’s strategic value to Pfizer’s out-of-pocket market penetration with Viagra.

Conference Economics: Approximately $8.3 billion in deals announced during or immediately surrounding the conference. Venue: Westin St. Francis Hotel and surrounding San Francisco locations.

These newsletters take significant effort to put together and are totally for the reader’s benefit. If you find these explorations valuable, there are multiple ways to show your support:

• Engage: Like or comment on posts to join the conversation.

Leave a comment

• Subscribe: Never miss an update by subscribing to the Substack.

Subscribe now

• Share: Help spread the word by sharing posts with friends directly or on social media.

  Share

Welcome to 2026, Readers! Let’s stay relentless this year!

I’m excited to kick off the year and share the Top 5 Biotech Breakthroughs of 2025. This continues a tradition I began in 2023, highlighting what I believe are the most transformative technologies and innovations that shape the biotech landscape over the past 12 months. This exercise is helpful to me as a scientist and writer, since it gives me time to reflect on the progress and understand how these advancements will shape the future of healthcare, longevity, and beyond.

This year’s review has its own twist. Rather than assembling the list alone, I invited four exceptional writers and thinkers in the biotech space to each contribute what they believe was the most important biotech breakthrough of 2025. The result is a set of five breakthroughs that reflect not a single perspective, but a convergence of independent judgment across the field. I strongly encourage you to check out the guest contributors shown below, each has a distinct voice and set of interests that’s well worth following.

List

• Obesity has a Pharmacological Solution. Body Composition is next.

• AI drug development: protein design for multi-step enzymes

• The Fully Automated Laboratory

• Alzheimer’s & Dementia Progress

• Cell Therapy Advancements.

Guest authors:

• Stephen Turner - Paired Ends

• Owen Lewis - Techno-Optimist Substack

• Riccardo Scribano - BioDodo

• Tony Huge - Tonyhuge.is

With that, let’s dive in.

First, if you enjoy these posts, consider subscribing and becoming a part of our growing community!

Subscribe now

Obesity has a Pharmacological Solution. Body Composition is next.

David Kingsley & Tony Huge

Obesity is not only an aesthetic issue. It is a chronic disease and inflammatory state that raises risk for hypertension, dyslipidemia, type 2 diabetes, heart disease and stroke, multiple cancers, chronic kidney disease, and fatty liver disease (CDC). Creating a biotech solution to obesity will have a huge improvement in the healthspan and lifespan we see in our society, as well as impact overall healthcare costs and usage. The good news is that we are getting close.

This year marked another significant leap towards a pharmacological solution to obesity, where drug efficacy rivals bariatric surgery. How did we get there and what are the results?

Incretin-mimetic drugs emerged as a breakthrough for weight management in the early 2020s, led by the repurposed diabetic agent semaglutide. As a GLP-1 agonist, semaglutide demonstrated that double-digit weight loss (~15% vs. placebo) was achievable (Rubino et al., 2022). These compounds function by mimicking natural gut hormones, triggering a cascade of downstream effects including optimized insulin release, blood glucose regulation, delayed gastric emptying, and appetite suppression. This success paved the way for next-generation agents like tirzepatide (a dual GLP-1/GIP agonist), which pushed weight reduction upward of 20% over 72 weeks. Yet, these drugs represent only the beginning; significant progress over the past year has built upon this foundation to further enhance efficacy and optimize body composition.

Multi-agonist therapies and combinations led the charge. The newest wave of incretin-based drugs and dosing procedures now consistently produces ≥20% body weight loss, a threshold once unimaginable for medications. In the SURMOUNT trials, tirzepatide (now FDA-approved as Zepbound) achieved roughly 18–22% average weight reduction over 72 weeks (Jastreboff et al., 2022). The following year, Lilly’s retatrutide pushed this further with Phase 2 results showing 24.2% mean weight loss at 48 weeks on the highest dose, with patients losing an average of 58 pounds and not yet reaching a plateau by study end (Jastreboff et al. 2023). In December, Lilly shared results from retatrutide’s Phase 3 TRIUMPH-4 study showing an average 28.7% mean weight loss over 68 weeks alongside significant reductions in osteoarthritis-related knee pain, with more clinical endpoints to be released in 2026. This is huge!

Changes in Body Weight with Retatrutide as Compared with Placebo (Jastreboff et al. 2023).

Likewise, Novo Nordisk’s combo CagriSema (cagrilintide + semaglutide) demonstrated a 22.7% mean weight reduction at 68 weeks in a recent trial (Garvey et al., 2025). Over 40% of CagriSema-treated patients dropped at least a quarter of their body weight, a degree of efficacy that would have been science fiction a few years ago. These advances firmly establish 20%+ weight loss as the new standard for anti-obesity drugs, in contrast to the ~5–10% from older therapies.

Change in Body Weight with Cagrilintide–Semaglutide as Compared with Semaglutide, Cagrilintide, and Placebo (Garvey et al., 2025).

Body composition is now a focus by including muscle-preserving adjuvant therapeutics. Equally important, the field is shifting focus from weight loss alone to body composition engineering. It’s no longer just about “how much” weight is lost, but “what” is lost, i.e., fat vs. lean mass. Potent incretin mimetics like GLP-1 agonists inherently cause both fat and muscle loss, so researchers are now targeting fat selectively. A landmark Phase 2 study (the BELIEVE trial) shared at American Diabetes Association combined semaglutide with bimagrumab, a myostatin/activin pathway antibody designed to spur muscle growth, and saw striking results. On semaglutide alone, users lost about 15.7% body weight with about 72% of each pound shed being fat tissue, with the rest being lean tissue loss. With the addition of bimagrumab, total weight loss was 22.1% with over 92% of weight loss from fat. Even more interesting, in bimagrumab alone, 100% of weight loss was attributed to fat mass and there was a modest increase in lean muscle (~2.5%). This approach yielded dramatic improvements in waist circumference and other metabolic markers by preserving muscle. Lilly has since launched their own study pairing acquired Bimagrumab with their flagship drug Tirzepatide.

Similar results are being observed in Regeneron’s Courage Trial, a 26-week study combining semaglutide with Trevogrumab (anti-GDF8/myostatin) and/or Garetosmab (anti-activin A) for treating obesity (Regeneron Press). Early results reflect what is seen with Bimagrumab and Semiglutide, i.e., maintain more lean muscle mass and lose additional fat mass. However, even without adjunct therapy, newer drugs are being evaluated for their impact on body composition. For example, retatrutide’s trials included a body composition substudy that showed minimal lean mass loss despite massive fat reductions. The emphasis on “quality” of weight loss, maximizing fat reduction while maintaining muscle, is now front and center in obesity pharmacology.

Formulation and convenience being tackled. Another breakthrough of 2025 was the success of oral formulations of these typically injectable peptides. Novo Nordisk presented high-dose oral semaglutide tablets (25 mg and 50 mg daily) that match injection efficacy. In the Phase 3 OASIS 4 trial, oral semaglutide 25 mg produced a 14% mean weight loss at 64 weeks versus 2% for placebo, essentially the same benefit as the 2.4 mg weekly injection (Wharton et al., 2025). Notably, 30% of patients on oral semaglutide achieved ≥20% weight loss (vs only 3% on placebo). These results were on par with the 50 mg dose and confirm that pill-based GLP-1 therapy can deliver surgery-level weight reductions.

Meanwhile, Eli Lilly’s orforglipron, the first non-peptide (small-molecule) GLP-1 receptor agonist in pill form also made headlines. Orforglipron phase 3 (ATTAIN‑1) showed dose‑dependent weight reduction at 72 weeks (highest dose ~−11.2% vs −2.1% placebo), with “nearly one fifth” of participants at the highest dose reaching ≥20% weight loss; adverse events were predominantly gastrointestinal and usually mild‑to‑moderate (Wharton et al., 2025).

Although injectables still have an edge in absolute potency, oral options are rapidly closing the gap. The convenience of pills (no needles or cold storage) could hugely expand access, making long-term therapy more feasible for millions.

Future challenges remain with tolerability and weight regain after cessation. Despite remarkable efficacy, safety and tolerability remain key bottlenecks. All these agents work via potent hormonal effects on appetite and digestion, so gastrointestinal side effects are ubiquitous. Nausea, vomiting, and diarrhea are by far the most common adverse events and tend to occur at higher doses. For example, in trials of the GLP-1 combo CagriSema, about 80% of patients experienced some GI symptoms (mostly mild-to-moderate) compared to ~40% on placebo (interesting placebo is 40%). Typically 5–10% of participants cannot tolerate the full dose and discontinue treatment due to side effects. Additionally, class-wide safety considerations (such as rare risks of pancreatitis, gallbladder complications, or medullary thyroid cancer in predisposed patients) necessitate ongoing vigilance. So far, no new serious hazards have emerged in the latest trials – adverse events have remained in line with known GLP-1 class effects and generally manageable. Nevertheless, improving the side effect profile is a priority. Approaches in development include better targeting of receptors, alternative hormones (e.g. amylin agonists with fewer GI effects), and adjunctive therapies (like bimagrumab) to allow lower dosing of incretins. In short, the efficacy is here, but making treatment tolerable for all is the next hurdle. Perhaps the single biggest draw back for users of incretin products is that people seem to regain the majority of weight within 2 years of cessation of these products. So, great news for the pharma companies and their investors, not so great for users (West et al., 2025).

AI drug development: protein design for multi-step enzymes

Stephen Turner - Paired Ends

A useful way to frame 2025’s AI drug development progress is that we are no longer just predicting structures or ranking candidates. We are starting to design and construct new functional biomolecules with performance that has to survive contact with reality: purification, kinetics, and crystal structures.

A striking recent example comes from David Baker’s group, who used a generative protein design workflow centered on RFdiffusion and then used AlphaFold2 as an in silico gatekeeper before committing designs to the wet lab.

Lauko, Anna, et al. “Computational design of serine hydrolases.” Science 388.6744 (2025): eadu2454. https://www.science.org/doi/10.1126/science.adu2454

In this work they explicitly set out to tackle a multistep catalytic cycle, not a single transition state. That matters because multistep enzymes must stabilize and choreograph multiple transition states and intermediates, including an acyl-enzyme intermediate and two tetrahedral intermediates, with tight geometric preorganization at each step.

Methodologically, the paper is a blueprint for how AI components are starting to compose: RFdiffusion is used to generate backbones around a target catalytic geometry; sequence is designed with LigandMPNN and refined; AlphaFold2 is used to filter for designs whose catalytic residues land within 1A of the intended configuration. They then add a key twist: PLACER, an ensemble modeling approach, to evaluate compatibility and preorganization across the entire reaction coordinate, which improves success rates and activity.

This work is not incremental. They report multiple designed serine hydrolases with crystal structures that match the design models at sub-angstrom accuracy and extremely high catalytic efficiencies. Even more interesting, the best examples appear to extend beyond known hydrolase fold space: Foldseek finds only weak structural matches (TM scores near or below 0.5), and the closest hits are proteins of unknown function with no similarity to known hydrolases at fold or active-site level.

Zooming out, this is why AI drug development hits different now. It is becoming a general capability: specify function, generate matter, validate, iterate, and eventually industrialize.

RFdiffusion generates enzymes tailor-made for a given active site (top left). Assessing design compatibility across the catalytic cycle with the PLACER neural net. Conformational ensembles for each step of the reaction are shown (top right). Comparison with crystal structures shows that designs made with RFdiffusion and filtered with PLACER are atomically accurate (bottom left). Designed serine hydrolases expand the fold space of this ancient enzyme family (bottom right, catalytic efficiencies listed).

The Fully Automated Laboratory

Stephen Turner - Paired Ends

David Kingsley - Neural Nexus

LLMs and Advanced robotics will make laboratories autonomous. We are seeing the first steps towards this. The core idea is not “a chatbot that writes protocols.” It is closing the loop between hypothesis, execution, measurement, and iteration, with software that can reason over experiments the way a strong scientist does, while operating at machine cadence.

For the first time, an AI system demonstrated it could autonomously design and optimize real-world experiments, essentially acting as a scientific “control plane” in the laboratory. OpenAI’s wet-lab evaluation is a concrete early signal:

OpenAI Research Blog: Measuring AI’s capability to accelerate biological research in the wet lab (Dec 16, 2025).

The landmark example came from OpenAI’s collaboration with biosecurity startup Red Queen Bio: using GPT‑5, the team built a closed-loop framework where the LLM proposed experimental protocols that human scientists executed, feeding results back to the model. Through iterative refinement, GPT‑5 optimized a standard molecular cloning procedure by a factor of 79 – an unprecedented gain. Notably, the model’s suggestions were not simply copied from literature; it introduced new elements (like adding E. coli RecA recombinase and a phage DNA-binding protein to the DNA assembly process) that hadn’t been used in that protocol, leading to dramatically higher yields. Researchers observed that GPT‑5 integrated known biological mechanisms in a novel way, showing “glimpses of creativity” beyond what its training data alone would predict. While not a fundamental discovery in biology, this achievement proved that an LLM can drive experimental innovation in the lab, not just analyze text

Autonomous lab robotics played a key role in extending the LLM’s control to physical execution. In the GPT‑5 cloning experiment, Red Queen Bio and Robot.dev built a custom robotic system that could carry out GPT‑5’s instructions in the lab via natural language commands. The setup included a human-to-robot LLM interface (to translate plain English protocols into robot actions), a vision module for real-time labware identification, and a motion planner for precise liquid-handling and incubation steps. In trials, this autonomous robot handled complex biotech procedures – mixing reagents, heat-shocking cells, plating cultures – and reproduced the improved cloning protocol almost as effectively as a human scientist. The robot-achieved outcome was ~89% of the human-run experiment’s yield, demonstrating that LLM-directed experiments can be automated end-to-end with only minimal performance loss so far. These results point toward a future where LLM “agents” not only design experiments but also directly orchestrate laboratory robots to execute them.

Underpinning this progress is the maturation of cloud laboratories, a highly automated lab facilities accessible via software APIs. Platforms like Emerald Cloud Lab (ECL) and Strateos reached new levels of adoption by 2025, allowing researchers (or their AI assistants) to remotely run experiments on dozens of instruments through a web interface. In fact, Carnegie Mellon University opened the world’s first academic cloud lab (in partnership with ECL) in early 2024, housing over 200 robotic instruments available on demand. This “self-driving lab” at CMU’s Bakery Square facility enables scientists to submit experiment instructions digitally and get back results without ever entering a physical lab. Early integrations with LLMs hint at how powerful this model can be: researchers have prototyped LLM-based lab assistants that take a natural language experiment description and translate it into cloud lab protocols automatically. In other words, a scientist could soon “chat” with an AI to run an assay, with the LLM handling all the low-level method details and instrument control. Such systems remain in beta, but the concept has been proven in principle by projects like CMU’s Coscientist, which linked a GPT-4 agent to ECL’s robotic platform to autonomously carry out chemistry reactions based on a plain-English goal. The democratizing effect is significant – by abstracting lab work into software, these AI-driven cloud labs can make advanced experimentation accessible well beyond elite research centers.

Major industry players also pushed wet-lab AI integration in 2025. In February, Google unveiled an “AI co-scientist” designed to act as a virtual collaborator for biomedical research. The system uses a powerful LLM (from Google DeepMind’s Gemini family) to absorb vast scientific literature and generate hypotheses or experiment plans for scientists. In a proof-of-concept study on liver fibrosis, Google reported that the AI co-scientist suggested several novel therapeutic approaches – and when these suggestions were tested in vitro by human researchers, all of them showed promising activity against the disease mechanism. This was a striking validation of LLM-generated hypotheses in a real biomedical context. Later in the year, DeepMind announced it will open an automated lab in the UK in 2026 as part of a government partnership. That facility is being built specifically to integrate DeepMind’s multimodal Gemini LLM with robotics, enabling high-throughput experiments in materials science (synthesizing and testing hundreds of samples per day) with minimal human intervention. Together, these efforts by tech leaders show a clear trend: LLMs are moving from the computer room into the laboratory, becoming central to how experiments are conceived and conducted.

Beyond these headline projects, 2025 also saw LLMs permeating everyday scientific workflows. Many labs began using LLM “copilots” for routine tasks like analyzing data, reading logs, and drafting experiment plans. For instance, electronic lab notebook platforms (e.g. Benchling) introduced GPT-powered features that let scientists query their experimental records in natural language and get summaries or insights instantly. LLMs showed proficiency at interpreting unstructured notes or instrument outputs and suggesting next steps – effectively acting as intelligent lab assistants. In one notable case, researchers developed a specialized agent dubbed “CRISPR-GPT” that was fine-tuned on gene editing knowledge and linked to lab automation tools. This system autonomously designed CRISPR guide RNAs, planned the gene knockout/activation experiments, and analyzed the results, successfully executing several genome edits in human cells with no human design input. Meanwhile, IBM researchers demonstrated ChemCrow, an LLM-based chemistry assistant that uses GPT-4 plus domain tools to plan and carry out chemical syntheses – it even proposed and helped create a new fluorescent molecule that standard methods hadn’t found. And at Carnegie Mellon, the Coscientist system proved capable of handling entire reaction workflows: given a target outcome, it retrieved relevant literature, wrote Python code to control liquid-handling robots, ran the experiment in a cloud lab, and iteratively refined conditions based on the data. These real-world examples underscore that LLMs can go well beyond writing reports – they are now generating hypotheses, planning experiments, and controlling lab hardware in various scientific domains.

Humanoid robots emerged as practical wet-lab automatons in 2025 by virtue of their human-mimicking form, which lets them function with existing benches, pipettes, centrifuges, and other tools built for human ergonomics. Unlike fixed automation, a bipedal robot can navigate legacy lab infrastructure and operate standard instruments without costly reconfiguration. Crucially, several high-profile pilots demonstrated this advantage: a biotech R&D lab deployed a humanoid “scientist” that performs pipetting and reagent handling tasks on human-standard equipment; Figure AI partnered with BMW to trial humanoids on car assembly lines (moving the technology from lab demos to real production); Apptronik’s Apollo robot was piloted in Mercedes-Benz facilities to deliver parts and inspect components in human-designed workspaces; and Tesla began using Optimus prototypes in its own factories and engineering labs for basic material-handling chores. Tightly integrating these humaniform actuators with an LLM-driven “control plane” means an AI can devise experimental plans and have the robots execute them end-to-end – effectively serving as the hands of an autonomous lab that accelerates experimentation in environments originally built for human scientists.

Alzheimer’s & Dementia Progress

Owen Lewis - Techno-Optimist Substack

Anyone who has watched a loved one suffer from Alzheimer’s or dementia knows the frustration and hopelessness the diagnosis brings. While three monoclonal antibodies, Aducanumab, Lecanemab, and most recently Donanemab, have now been approved to slow early Alzheimer’s, even the best of them reduces progression by only about 35% and comes with serious risks like brain bleeds. But I think 2025 marks the beginning of a real shift. Over the past year, a steady tempo of breakthroughs is painting a different picture. Most won’t help patients immediately, but the progress of the past year has us on a path where Alzheimer’s and dementia could be more treatable, or even curable, within the next decade.

We’re starting to uncover the root causes of Alzheimer’s and dementia, with scientists finding a direct cause and effect relationship between malfunctioning mitochondria, the energy-producing “engines” of cells, and memory loss seen in these neurodegenerative diseases. Ramping up mitochondrial activity in mice restored memory performance. Another insight might have uncovered a more common genetic cause, with a protein called PHGDH potentially disrupting how brain cells regulate genes, which can trigger disease onset. They also identified a small molecule that blocks this harmful activity without interfering with the gene’s normal function. In mouse models, it slowed disease progression and improved memory.

Quick intervention will always be best, with a molecule called WIN55.212-2 found to protect the brain and restore cognitive function in early Alzheimer’s. Rodent tests have been very promising, though a problem here is the molecule is “free to use,” meaning it cannot be patented as a drug—decreasing the likelihood a pharmaceutical company would move forward with it.

Some treatments don’t require drugs at all, with the use of low-intensity ultrasound on the brains of mice with an Alzheimer’s disease like condition boosting cognitive function, though it didn’t clear amyloid plaques. Early human trials have been done by the team, but annoyingly the article didn’t say what (if any) improvements were seen. The impression I get though is that it was effective.

Turning to the periodic table, a study found that the noble gas Xenon could treat Alzheimer’s disease. Inhaling the gas “reduced neuroinflammation, minimized brain atrophy, and promoted protective neuronal states in mouse models of Alzheimer’s disease.” Human trials have (I think) already started. Harvard researchers have discovered that lithium naturally plays a crucial role in maintaining neural health—and that its depletion appears to be one of the earliest detectable signs of Alzheimer’s. In mice, lowering brain lithium accelerated damage, including amyloid plaques, tau tangles, inflammation, and memory loss, while a lithium compound (lithium orotate) that avoids binding with plaques reversed those effects and restored cognitive function. The best part is that lithium orotate is already an approved treatment for other things and is considered safe.

Another report found that the common amino acid arginine can reduce Alzheimer’s-like brain damage in animal models by inhibiting toxic amyloid aggregations while also dampening neuroinflammation. Oral arginine improved cognitive performance and slowed disease progression, suggesting benefits beyond plaque reduction alone. Because arginine is already widely used in humans with a strong safety profile, the findings point to a potentially low cost, fast intervention if the effects hold up in clinical trials.

Nanotechnology could play a role as well, with engineered nanoparticles slashed amyloid-beta levels by 50-60% within one hour in Alzheimer’s mice by reactivating the blood-brain barrier’s natural waste removal system, rather than directly attacking the proteins. Cognitive improvements lasted 6 months, a significant chunk of a mouse’s lifespan.

A rather unexpected finding is that the shingles vaccination may slow dementia progression, not just reduce risk. Researchers found vaccination was associated with less cognitive impairment in healthy people, lower conversion from Mild Cognitive Impairment to dementia, and reduced mortality in those already diagnosed, across Alzheimer’s, vascular dementia, and most strongly, mixed dementia. The effect has now been replicated multiple times, with randomized trials underway. Serious message here, everyone tell your parents and grandparents to get this. Get it yourself if you’re 50+.

I’ve saved the best for last: Researchers have reversed advanced Alzheimer’s disease in mice by restoring the brain’s energy balance through a molecule called NAD+, which naturally declines with age and drops even more severely in Alzheimer’s patients. Mouse models showed complete cognitive recovery even after the disease was already advanced. The treated mice showed normalized blood biomarkers for Alzheimer’s, and complete recovery of all cognitive functions. The key takeaway is a message of hope: the effects of Alzheimer’s disease may not be inevitably permanent. The damaged brain can, under some conditions, repair itself and regain function.

Cell Therapy Advancements

Riccardo Scribano - BioDodo

Since the first FDA approval of the first Chimeric antigen receptor engineered T-cells (CAR-T) in 2017, cell therapies have transformed the treatment of hematological malignancies, delivering high response rates while revealing limitations (Shokati et al., 2025). By mid-2025, the field has expanded to more than 6,000 registered interventional cell therapy trials worldwide, with CAR-T still dominating the landscape. Hematological cancers still represent the largest category but are steadily contracting, alongside sustained industry interest and diversification of targets and platforms (Benthani et al., 2025).

Source: Benthani et al., 2025

A major focus of recent progress lies in solid tumors, which remain challenging because of poor CAR-T trafficking and infiltration, physical and immunosuppressive tumor microenvironments, and antigen heterogeneity leading to immune escape. Nonetheless, encouraging signs have emerged from clinics in 2025 (Escobar et al., 2025). A phase-1 clinical trial on intrinsic pontine glioma, an aggressive brainstem tumor, has shown the safety of repeated intracerebroventricular CAR-T treatment (Vitanza et al., 2025). CAR-macrophages (CAR-M) have gained momentum due to their innate capacity to infiltrate solid tumors, with early-phase trials demonstrating favorable safety profiles, tumor microenvironment remodeling and disease stabilization in HER2-positive cancers (Abdin et al., 2024; Reiss et al., 2025). Still, the only FDA-approved cell therapies for solid tumors remain Lifileucel (Amtagvi), derived from tumor-infiltrating lymphocytes (TIL) and used to treat melanoma, and Afamitresgene (Tecelra), an engineered T-cell receptor (TCR) therapy against synovial sarcoma.

In parallel, allogeneic “off-the-shelf” CAR-T and CAR-Natural killer (CAR-NK) programs have stabilized and expanded, leveraging healthy donor- or stem-cell-derived immune cells to improve scalability and availability, while ongoing trials continue to address challenges such as graft-versus-host disease, fratricide and limited persistence (Li et al., 2025). At the beginning of last year, a study was published where Induced pluripotent stem-cell (iPSC)-derived chimeric antigen receptor (CAR) natural killer (NK) cell were first tested in humans, during a phase-1 clinical trial for relapsed or refractory B-cell lymphoma, reporting positive outcomes with regards to the safety of the treatment (Ghobadi et al., 2025). Lastly, a novel allogeneic anti-CD70 CAR-T product was tested in a phase-1 trial showing promising results in the treatment of clear cell renal cell carcinoma, adding to the arsenal of cell therapies against solid tumors (Srour et al., 2025).

Source: (Li et al., 2025)

Finally, a special mention goes to Astraveus, which in 2025 has pushed forward CAR-T manufacturing. Astraveus has released a fully automated benchtop microfluidic end-to-end cell factory, enabling rapid (≈26-hour) CAR-T production with transduction efficiency, viability and purity comparable to conventional methods, pointing toward faster and more standardized cell therapy deployment (Peeck & Ghinatti, 2025).

Beyond cancer treatment, cell-based therapies are making enormous steps for many other applications! I’m going to highlight a few more of these below:

Parkinson’s Disease (Stem Cell–Derived Dopaminergic Neurons): A pioneering regenerative approach reached late-stage trials. Bemdaneprocel (BlueRock/Bayer) – implanted dopamine-producing neurons derived from pluripotent stem cells – showed long-term safety and cell engraftment in an initial Phase 1 study, with patients tolerating the implants and imaging suggesting the new neurons survived and integrated into the brain. In 2025, the therapy entered a global Phase 3 trial (exPDite-2) – the first large-scale study of a stem-cell replacement therapy in Parkinson’s. If successful, bemdaneprocel could become the first treatment to restore lost neuronal function in PD rather than just manage symptoms.

Type 1 Diabetes (Islet Cell Replacement – VX-880/Zimislecel): 2025 saw dramatic progress toward a functional cure for T1D using stem cell-derived islet cells. Vertex Pharmaceuticals’ zimislecel (VX-880), pancreatic islet cells derived from stem cells, reported Phase 1/2 trial results at the ADA meeting: all participants showed graft engraftment and C-peptide production, and 10 of 12 patients (83%) achieved insulin independence one year after a single infusion. Patients had HbA1c <7% with no severe hypoglycemia, indicating restored glycemic control (Rickels et al., 2025). While immunosuppression is required (and led to some serious infections), these findings (simultaneously published in NEJM) demonstrate the therapy’s transformative potential. A Phase 3 trial is underway in ~50 patients to confirm efficacy, and Vertex is also exploring gene-edited islet cells that could evade immune rejection in future.

Alzheimer’s Disease (MSC Therapy – Laromestrocel): In regenerative medicine for neurology, an mesenchymal stem cell (MSC) product showed encouraging signs in mild Alzheimer’s disease. Longeveron’s laromestrocel (Lomecel-B), an allogeneic MSC therapy, received an FDA RMAT designation and in a Phase 2 trial (CLEAR MIND) demonstrated reduced neuroinflammation on MRI and slower hippocampal atrophy compared to placebo. Treated patients had preserved hippocampal volume correlating with better clinical outcomes, suggesting the cells’ anti-inflammatory and pro-regenerative effects in the brain. Cognitive function and quality of life improved in initial data, positioning laromestrocel as a novel disease-modifying approach for Alzheimer’s pending larger trials.

Systemic Lupus Erythematosus (Allogeneic CAR T Cells – FT819): A breakthrough therapy applied CAR T-cell technology to autoimmunity. FT819 (Fate Therapeutics) is a first-in-kind CAR T made from an induced pluripotent stem cell line, an off-the-shelf product targeting CD19 B cells, repurposed to wipe out the autoreactive B cells that drive lupus. In 2025, the FDA granted it Regenerative Medicine Advanced Therapy (RMAT) status after early Phase 1 use showed promising results. The first lupus patient treated (at UCI) experienced a striking reduction of lupus symptoms within weeks, the first meaningful remission in nearly two decades of disease. This patient became fully functional with resolution of fevers, joint pain, nephritis, and fatigue. The allogeneic CAR T appears to be tolerated thus far without intensive chemo or hospital isolation. FT819’s early success in lupus (and ongoing trials in myositis, sclerosis, and vasculitis) highlights the potential for CAR T beyond oncology to autoimmune applications.

HIV Cure Research (Personalized T-Cell Therapy – HST-NEET): A Phase 1 trial provided a hopeful step toward an HIV cure using ex vivo expanded T cells. Researchers at Children’s National engineered each patient’s own T cells to recognize conserved HIV epitopes and infused them into 6 people with HIV. The results, published in Nature Communications, shows that therapy was to be safe with no serious adverse effects (Sohai et al., 2025). Early signs of efficacy emerged in 2 patients the treatment boosted HIV-specific T cell and antibody responses, and in another 2 it led to a measurable drop in latent HIV reservoir cells, indicating the hidden virus was being attacked. The infused T cells persisted in circulation for up to 40 weeks, patrolling for HIV. While not yet a cure, these findings show that a targeted cell therapy can reduce the virus lurking in the body and train the immune system to better fight HIV. This personalized T-cell strategy is now moving into expanded trials and combination approaches (e.g., with latency-reversing agents) to further shrink the HIV reservoirs.

Overall, cell therapies are rapidly expanding not only within oncology, but are also making their way through clinical studies for new indications. The field of Cell therapy continues to be exciting with lots of exciting ongoing applications and future potential.

These newsletters take significant effort to put together and are totally for the reader’s benefit. If you find these explorations valuable, there are multiple ways to show your support:

• Engage: Like or comment on posts to join the conversation.

Leave a comment

• Subscribe: Never miss an update by subscribing to the Substack.

Subscribe now

• Share: Help spread the word by sharing posts with friends directly or on social media.

Share

References:

Abdin, S. M., Paasch, D., & Lachmann, N. (2024). CAR macrophages on a fast track to solid tumor therapy. Nature Immunology, 25(1), Article 1. https://doi.org/10.1038/s41590-023-01696-7

Benthani, F., Upadhaya, S., & Zhou, A. (2025). Cancer cell therapies: Global clinical trial trends and emerging directions. Nature Reviews Drug Discovery, 24(12), 898–899. https://doi.org/10.1038/d41573-025-00180-1

Escobar, G., Berger, T. R., & Maus, M. V. (2025). CAR-T cells in solid tumors: Challenges and breakthroughs. Cell Reports Medicine, 6(11), 102353. https://doi.org/10.1016/j.xcrm.2025.102353

Garvey, W.T., Blüher, M., Osorto Contreras, C.K., Davies, M.J., Winning Lehmann, E., Pietiläinen, K.H., Rubino, D., Sbraccia, P., Wadden, T., Zeuthen, N. and Wilding, J.P., 2025. Coadministered cagrilintide and semaglutide in adults with overweight or obesity. New England Journal of Medicine.

Ghobadi, A., Bachanova, V., Patel, K., Park, J. H., Flinn, I., Riedell, P. A., Bachier, C., Diefenbach, C. S., Wong, C., Bickers, C., Wong, L., Patel, D., Goodridge, J., Denholt, M., Valamehr, B., Elstrom, R. L., & Strati, P. (2025). Induced pluripotent stem-cell-derived CD19-directed chimeric antigen receptor natural killer cells in B-cell lymphoma: A phase 1, first-in-human trial. The Lancet, 405(10473), 127–136. https://doi.org/10.1016/S0140-6736(24)02462-0

Jastreboff, A.M., Aronne, L.J., Ahmad, N.N., Wharton, S., Connery, L., Alves, B., Kiyosue, A., Zhang, S., Liu, B., Bunck, M.C. and Stefanski, A., 2022. Tirzepatide once weekly for the treatment of obesity. New England Journal of Medicine, 387(3), pp.205-216.

Li, Y.-R., Zhu, Y., Fang, Y., Lyu, Z., & Yang, L. (2025). Emerging trends in clinical allogeneic CAR cell therapy. Med, 6(8), 100677. https://doi.org/10.1016/j.medj.2025.100677

Peeck, L. H., & Ghinatti, G. (2025). Small is powerful: Demonstrating the potential of a novel microfluidic benchtop cell factory to optimize and scale CAR-T manufacturing. Cytotherapy, 27(5, Supplement), S195–S196. https://doi.org/10.1016/j.jcyt.2025.03.395

Reiss, K. A., Angelos, M. G., Dees, E. C., Yuan, Y., Ueno, N. T., Pohlmann, P. R., Johnson, M. L., Chao, J., Shestova, O., Serody, J. S., Schmierer, M., Kremp, M., Ball, M., Qureshi, R., Schott, B. H., Sonawane, P., DeLong, S. C., Christiano, M., Swaby, R. F., … Abdou, Y. (2025). CAR-macrophage therapy for HER2-overexpressing advanced solid tumors: A phase 1 trial. Nature Medicine, 31(4), 1171–1182. https://doi.org/10.1038/s41591-025-03495-z

Rickels MR, Witkowski P, Reichman TW, et al. 140-OR - Durable Glycemic Control and Elimination of Exogenous Insulin Use with VX-880 in Patients with Type 1 Diabetes (T1D)—VX-880-101 (FORWARD). Presented at ADA 2025; June 20-23; Chicago, Illinois. Presentation #140-ORRubino D, Abrahamsson N, Davies M, Hesse D, Greenway FL, Jensen C, Lingvay I, Mosenzon O, Rosenstock J, Rubio MA, Rudofsky G. Effect of continued weekly subcutaneous semaglutide vs placebo on weight loss maintenance in adults with overweight or obesity: the STEP 4 randomized clinical trial. Jama. 2021 Apr 13;325(14):1414-25.

Shokati, A., Sanjari-Pour, M., Akhavan Rahnama, M., Hoseinzadeh, S., Vaezi, M., & Ahmadvand, M. (2025). Allogeneic CART progress: Platforms, current progress and limitations. Frontiers in Immunology, 16. https://doi.org/10.3389/fimmu.2025.1557157

Sohai, D.K., Keller, M.D., Hanley, P.J., Hoq, F., Kukadiya, D., Datar, A., Reynolds, E., Copertino, D.C., Lazarski, C., McCann, C.D. and Tanna, J., 2025. Autologous HIV-specific T cell therapy targeting conserved epitopes is well-tolerated in six adults with HIV: an open-label, single-arm phase 1 study. Nature communications, 16(1), p.4510.

Srour, S. A., Chahoud, J., Drakaki, A., Curti, B. D., Gibney, G. T., Pal, S. K., Tang, L., Charmsaz, S., Atwell, J., Robbins, P. B., Williams, C., Ghatta, S., Severyn, C., Le Gall, J., Tannir, N. M., & Kotecha, R. R. (2025). ALLO-316 in advanced clear cell renal cell carcinoma (ccRCC): Updated results from the phase 1 TRAVERSE study. Journal of Clinical Oncology, 43(16_suppl), 4508–4508. https://doi.org/10.1200/JCO.2025.43.16_suppl.4508

Vitanza, N. A., Ronsley, R., Choe, M., Seidel, K., Huang, W., Rawlings-Rhea, S. D., Beam, M., Steinmetzer, L., Wilson, A. L., Brown, C., Beebe, A., Lindgren, C., Gustafson, J. A., Wein, A., Holtzclaw, S., Hoeppner, C., Goldstein, H. E., Browd, S. R., Hauptman, J. S., … Jensen, M. C. (2025). Intracerebroventricular B7-H3-targeting CAR T cells for diffuse intrinsic pontine glioma: A phase 1 trial. Nature Medicine, 31(3), 861–868. https://doi.org/10.1038/s41591-024-03451-3

West, S., Scragg, J., Aveyard, P., Oke, J.L., Willis, L., Haffner, S.J., Knight, H., Wang, D., Morrow, S., Heath, L. and Jebb, S.A., 2026. Weight regain after cessation of medication for weight management: systematic review and meta-analysis. bmj, 392.

Wharton S, Lingvay I, Bogdanski P, Duque do Vale R, Jacob S, Karlsson T, Shaji C, Rubino D, Garvey WT. Oral semaglutide at a dose of 25 mg in adults with overweight or obesity. New England Journal of Medicine. 2025 Sep 18;393(11):1077-87.

Genetic engineering, that is, precise, intentional edits to DNA, has arrived. Its use is now ubiquitous in biotechnology research laboratories, powered by mature CRISPR toolchains, cheaper, faster genome sequencing, and a deeper understanding of how genes map onto function.

What does this mean practically?

We are now at the point where, for many applications, the question isn’t if we can make an edit, but who will fund it and perhaps as important, the underlying ethics. This BioWire Byte is a short case study: a cohort of horses were cloned from genetically engineered DNA to impart an advantage for muscle development to improve their muscle mass and speed for playing polo.

First, if you enjoy these updates, consider subscribing and becoming a part of our growing community!

Subscribe now

Take a look at the image below. The four brown foals born late last year on an Argentine farm seem utterly ordinary. But these 10-month-olds represent a world first: they are genetically edited horses, created by Kheiron Biotech in Buenos Aires. Each is a cloned copy of a prize-winning polo mare (named Polo Pureza, or “Polo Purity”) with an important twist, a CRISPR-engineered DNA edit designed to boost their speed and muscle power. By using the CRISPR “genetic scissors” technique on the embryos, Kheiron’s scientists inserted a small mutation to dial down the horses’ myostatin gene, which normally acts as a brake on muscle growth. The concept is simple, less myostatin means more muscle fibers and more explosive acceleration, potentially transforming these horses into faster sprinters.

Four of the world’s first gene-edited foals, cloned from a champion polo mare and CRISPR-enhanced for muscle growth, at a farm in Argentina (2025). Photo credit.

Designing a Polo Pony with Cloning and CRISPR

Traditional cloning technology alone could only produce a genetic replica of the original champion horse. But what if we wanted to take it up a notch? Kheiron went a step further by adding a precision gene edit to the clone. While cloning creates an identical copy of the donor’s DNA, CRISPR functions as a kind of molecular scissors to cut and customize DNA. In this case, the team “cut” the myostatin gene, a known inhibitor of muscle development, effectively knocking it down to encourage greater muscle growth (I would like to volunteer myself for this as well).

The CRISPR edit is like an instant upgrade that selective breeding or random mutation might have taken generations to achieve. Gabriel Vichera, Kheiron’s scientific director, explains that CRISPR allowed them to make a precise change in the genome while keeping all the mare’s other superb traits intact. In other words, the foals have the exact DNA of a champion polo pony plus a tiny alteration that may give them extra muscle and speed. It’s a high-tech shortcut to what nature or breeding could have eventually produced – a horse with the genetic “muscle restrictor” removed.

Rationale for Myostatin Muscle Boost

Why target the myostatin gene? Biologically, myostatin is a well-known muscle growth regulator, essentially the body’s “off switch” for muscle building. Animals (and people) with less myostatin tend to develop significantly more muscle. Reducing this inhibitor’s activity unleashes more muscle fiber growth, a fact borne out in multiple species. For example, certain cattle breeds (like the Belgian Blue) carry natural mutations in myostatin that produce “double-muscled” phenotype in bulls (see the figure below). In the dog world, there are whippet racing dogs with a myostatin mutation that have superior sprinting ability, and Chinese researchers famously used CRISPR to create extra-buff beagles by deleting the myostatin gene, resulting in dogs with double the normal muscle mass and greater strength. Even in humans, a few rare individuals born with disruptive myostatin mutations exhibit extraordinary muscular development. In fact, these myostatin isoforms have been speculated on famous bodybuilders and strength trainers.

This muscle-boosting pathway goes much further; it’s actively being explored for medical therapy. Biomedical teams have been testing myostatin-blocking treatments to help patients with muscle-wasting diseases like muscular dystrophy and age-related frailty, hoping to enhance muscle growth and strength. Further, myostatin-inhibition is even being explored in clinical studies in combination with weight loss therapies like semaglutide and tirzepatide to improve body composition. In short, a mountain of evidence across biology suggested that dialing down myostatin should make a horse more muscular and potentially faster over short distances. The edited foals are in many ways a living experiment to see if that promise holds true.

Interestingly, despite their high-tech pedigree, the CRISPR foals look and act like any other healthy young horses. Early reports indicate they are developing normally and showing no obvious differences aside from their expected muscly build (which at under a year old is not fully evident yet). This normal outward appearance underscores an important point: the edit made by Kheiron isn’t some exotic foreign gene or Frankensteinian add-on. It’s a subtle tweak to a gene all horses already have, akin to a naturally occurring variant. As one equine geneticist put it, the myostatin edit is simply “a gene that we know is present in healthy horses” – by performing a targeted edit in a clone, they are faithfully reproducing a trait that could have arisen by chance. That gives researchers confidence that these foals will suffer no ill effects; in theory, they are just like naturally muscley horses, only engineered deliberately and far more rapidly.

Polo Prospects and Controversy

However, in the world of polo, just because you can breed a super-horse doesn’t mean you can play one. Argentina is revered as the global capital of polo and has long embraced advanced reproductive tech, ranging from the widespread use of surrogate mares and embryo transfer to openly allowing cloned horses in competition. A note: many top Argentine polo ponies are clones of past champions, which is a practice accepted in polo, unlike in horse racing.

Yet the advent of gene-edited ponies has met pushback from the sport’s authorities. The Argentine Polo Association, the national governing body, has banned genetically edited (GE) horses from official tournaments, at least for now. Likewise, the Argentine Association of Polo Horse Breeders has urged caution, they plan to monitor these foals for several years before deciding whether to register them as legit polo ponies.

We continue to live through a time where reality is pressing against the bounds of science fiction. The birth of genetically engineered ponies is a landmark for biotechnology. It showcases the incredible possibilities of rewriting genomes for improvement rather than illness. All of this, and we haven’t even touched on the elephant in the room: how this same technology could be translated to humans.

These newsletters take significant effort to put together and are totally for the reader's benefit. If you find these explorations valuable, there are multiple ways to show your support:

• Engage: Like or comment on posts to join the conversation.

Leave a comment

• Subscribe: Never miss an update by subscribing to the Substack.

  Subscribe now

• Share: Help spread the word by sharing posts with friends directly or on social media.

Share

References:

https://ir.unither.com/press-releases/2024/04-24-2024-140050479

https://www.kheiron-biotech.com/

https://www.reuters.com/world/americas/worlds-first-gene-edited-horses-are-shaking-up-genteel-sport-polo-2025-08-30/

As we get older, we start to look and feel older; the face will begin to develop fine lines and wrinkles, recovery from that night out partying doesn’t just take longer, but impacts the subsequent days in a noticeable way, the memory dulls, new sensations of pain in our muscles and joints, and the inevitable development of back pain. These are all things that happen with aging.

But why do they happen, and for that matter, what is aging? There are two ways to think of aging: chronological age, how many birthdays you’ve had, and phenotypic age, a measure of how your body’s systems compare to the averages across the lifespan. You can feel young and be old and vice versa, be young but feel old. It’s phenotypic age that captures how old your body acts.

So yes, aging looks like wrinkles, new aches and pains, and losing memory. But beneath the surface, is it just wear and tear, or is something deeper happening at the cellular level?

Researchers are beginning to think aging isn’t simply mechanical decline. Instead, it might be a communication breakdown in our cells’ software. A leading framework, the Information Theory of Aging (ITOA), suggests aging stems from the gradual loss of epigenetic information, the “software” that tells our genes how to behave, rather than irreversible DNA damage. In other words, our cells don’t necessarily break down, they forget their instructions, much like a corrupted hard drive. As Harvard geneticist David Sinclair explains: “Why do we age? The epigenome is the issue… fertilization resets the epigenome, but the good news is we’ve figured out a way to safely reset that without having to clone yourself.” This optimistic outlook, that there’s a backup of youthful information in our cells, is shifting how scientists think about aging and how we might reverse it.

First, if you enjoy these updates, consider subscribing and becoming a part of our growing community!

Subscribe now

Aging as an Information Error

In the ITOA framework, aging is essentially an information degradation problem. Our genetic code (DNA sequence) is a robust digital archive of information, like read-only data, that generally remains stable over time. By contrast, the epigenome (chemical tags on DNA and proteins, chromatin structure, gene expression patterns) is a more fluid “digital–analog” system that cells use to regulate which genes are on or off. Analog systems are powerful for fine-tuned control, but they’re inherently prone to noise and gradual distortion. Over the years, the epigenetic “signal” in our cells gets noisier: crucial genes misexpress, cellular identity blurs, and tissues begin to malfunction. This theory elegantly explains why aging manifests in predictable ways (grey hair, weaker muscles, etc.) despite each individual accumulating different random mutations, the root cause may be the same progressive epigenetic noise corrupting the cellular program.

Claude Shannon’s 1940s work on information theory provides a useful analogy. Shannon described how, when sending a message, noise in the channel can scramble the signal, but an “observer” with access to the original message can detect errors and send corrections. Borrowing this concept, Sinclair and colleagues suggest our cells have a similar built-in “observer”, essentially a backup copy of the original epigenetic information. Aging, in this view, is like a glitchy transmission: the youthful blueprint (the message) gets distorted by life’s wear and tear (the noise), leading to an aged, garbled message. But if cells could refer back to the clean backup and restore lost data, they might regain function. As Sinclair explained in an interview, “the information theory of aging is all about the preservation of information and having a backup copy.” In other words, cells might retain the original youthful “software” and, with the right triggers, can reboot themselves to a healthier state.

Genes vs. Epigenome and Digital Stability vs. Analog Vulnerability

It’s important to distinguish between genetic and epigenetic information in this model. The genome is often likened to a digital storage, information coded in sequences of A, C, G, T letters. Copying digital info can be error-corrected; indeed, cells have proofreading and repair enzymes to fix DNA replication errors. The epigenome, however, behaves more like an analog recording overlaying that genetic code. It’s subject to gradual wear from things like environmental stresses, DNA damage, inflammation, and even just the passage of time can alter epigenetic marks in small, cumulative ways. If the genome is a pristine music CD, then the epigenome is a vinyl record that can develop pops and hiss. Over decades, the “music” (gene expression) grows distorted, even if the underlying DNA track is intact.

Notably, epigenetic changes are systematic and reproducible across individuals, unlike random mutations. For instance, DNA methylation patterns change with age so consistently that scientists have created “epigenetic clocks” to measure biological age (Bocklandt et al., 2011). Different people (and even different species) show remarkably similar epigenetic drifts as they get older. This wouldn’t be the case if aging were driven purely by random DNA damage, which suggests it’s an ordered process of information loss. In fact, organisms like yeast can grow old and die with virtually no DNA mutations, indicating that something other than genetic damage is at play. Likewise, mice engineered to carry higher mutation loads, or even cloned animals derived from the cells of old individuals, often live normal lifespans. Their genomes may be littered with errors, yet their bodies don’t necessarily age faster, implying the epigenetic “software” state is the key player in aging, more so than the genetic “hardware”.

More Evidence that Epigenetic Noise is Driving Aging

A series of recent experiments is putting the ITOA to the test and the results are compelling. Here are some of the key findings that support the idea that aging is caused by corrupted epigenetic information rather than irreversible gene damage:

• Break the “Signal,” Age the Animal: Sinclair’s team developed a clever mouse model called ICE (Inducible Changes to the Epigenome) to introduce a burst of DNA double-strand breaks that scramble epigenetic markers without introducing new mutations. The “repaired” DNA was sequence-intact, but the epigenome was disturbed – and the mice began to show aging-like symptoms: graying fur, decreased vitality, organ dysfunction, and molecular signs of old age. In essence, by adding noise to the epigenetic code, they accelerated aging, even though the genetic code was unaltered. This directly supports the idea that it’s the cell’s reaction to damage – the disarray of chromatin and gene regulation – that drives aging (Yang et al., 2023).

• Mutation Load ≠ Aging Rate: Conversely, mice with extra-fast mutation rates (or people with conditions like Werner syndrome that predispose to DNA damage) don’t always show proportional acceleration in aging. And perhaps most telling, a cloned animal created from the nucleus of an old cell is reborn as a young, healthy individual (Wakayama et al., 2013). Both mice and even advanced-age livestock have been cloned and effectively had their lifespans “reset”. This is only possible because a young embryo’s environment restores the aged epigenome to a youthful state – the genome wasn’t the barrier. Identical twins (who share the same DNA) can also age at different speeds, presumably due to divergent epigenetic changes over their lifetimes.

• Epigenetic Clocks and Predictable Decay: Researchers have identified 12+ hallmarks of aging (like telomere shortening, mitochondrial dysfunction, etc.), but many of these may be downstream effects of an upstream information loss. Epigenetic changes sit near the top of that cascade. The existence of accurate epigenetic clocks, algorithms that read DNA methylation patterns to predict age, highlights that the aging process follows an informational “program.” Slower epigenetic drift is even correlated with longer maximum lifespans in mammals. In other words, species that preserve epigenetic information better (think naked mole rats or bowhead whales) tend to live much longer, hinting that maintaining a clean signal delays aging.

• Rewinding the Clock (in Cells): Perhaps most astonishing, aged cells and tissues can be rejuvenated by resetting their epigenetic marks, without altering the DNA sequence. Old human cells can be induced into pluripotent stem cells, which effectively erases age-associated marks back to an embryonic state (an extreme reset). But even short of a full “reboot,” exposing cells to youthful factors can reverse their gene expression age. In one experiment, old mouse and human cells treated with a partial “reprogramming” protocol began to behave like younger cells, and even in living mice, aspects of aging have been reversed without fixing any underlying mutations. This suggests much of what goes wrong with age is not permanent damage, but malleable, if we can restore the original epigenetic information.

There is a story emerging here. One that paints a picture consistent with the information theory of aging, i.e., aging is not a one-way accumulation of rust, but a loss of information in coding genes. But one that is showing evidence it may be reversible, like a software slowdown that might be debugged or restored.

These newsletters take significant effort to put together and are totally for the reader's benefit. If you find these explorations valuable, there are multiple ways to show your support:

• Engage: Like or comment on posts to join the conversation.

Leave a comment

• Subscribe: Never miss an update by subscribing to the Substack.

• Share: Help spread the word by sharing posts with friends directly or on social media.

Share

References:

Bocklandt, S., Lin, W., Sehl, M.E., Sánchez, F.J., Sinsheimer, J.S., Horvath, S. and Vilain, E., 2011. Epigenetic predictor of age. PloS one, 6(6), p.e14821.

Guarente, L., Sinclair, D.A. and Kroemer, G., 2024. Human trials exploring anti-aging medicines. Cell metabolism, 36(2), pp.354-376.

Lu, Y., Brommer, B., Tian, X., Krishnan, A., Meer, M., Wang, C., Vera, D.L., Zeng, Q., Yu, D., Bonkowski, M.S. and Yang, J.H., 2020. Reprogramming to recover youthful epigenetic information and restore vision. Nature, 588(7836), pp.124-129.

Lu, Y.R., Tian, X. and Sinclair, D.A., 2023. The information theory of aging. Nature aging, 3(12), pp.1486-1499.

Scott, A.J., Ellison, M. and Sinclair, D.A., 2021. The economic value of targeting aging. Nature Aging, 1(7), pp.616-623.

Teschendorff, A.E. and Horvath, S., 2025. Epigenetic ageing clocks: statistical methods and emerging computational challenges. Nature Reviews Genetics, 26(5), pp.350-368.

Wakayama, S., Kohda, T., Obokata, H., Tokoro, M., Li, C., Terashita, Y., Mizutani, E., Nguyen, V.T., Kishigami, S., Ishino, F. and Wakayama, T., 2013. Successful serial recloning in the mouse over multiple generations. Cell Stem Cell, 12(3), pp.293-297.

Yang, J.H., Hayano, M., Griffin, P.T., Amorim, J.A., Bonkowski, M.S., Apostolides, J.K., Salfati, E.L., Blanchette, M., Munding, E.M., Bhakta, M. and Chew, Y.C., 2023. Loss of epigenetic information as a cause of mammalian aging. Cell, 186(2), pp.305-326.