Epistemic status: Revisiting a paper I co-wrote in 2018 (Turchin, Milova, Egorova & Batin, “Fighting Aging as an Effective Altruism Cause”). The core argument I still endorse. One of the central numerical inputs — the date at which radical life extension arrives — has moved so much in eight years that the whole model has to be re-derived. Several of the specific bets aged badly. Numbers below are order-of-magnitude toy estimates and should be read as such; the optimizer’s curse applies to me as much as to anyone.
The claim, in one paragraph
In 2018 my coauthors and I argued something that still sounds absurd when you say it out loud: that the most cost-effective way to save a human life, under a defensible set of assumptions, was not bed nets but clinical trials of cheap, off-patent geroprotectors. The logic was a bridge argument. Somewhere in the future, technology defeats aging. Everyone who survives until that moment gets potentially indefinite life extension — they cross into a world without involuntary death. So any intervention that nudges average life expectancy up by even one year increases the number of people who make it across the bridge. Because human mortality follows the steep Gompertz curve, a single year of life expectancy is roughly a 10% relative reduction in mortality, which, applied to billions of people near the edge, rescues an enormous number of them into the post-aging world. We put a toy number on it: a 1-year shift from metformin, a finish line at 2100, a population of 10 billion, and you “save” ~250 million people. Divide by the ~$65M cost of the metformin trial and you get $0.24 per life saved — about four orders of magnitude cheaper than the ~$1,500-per-life figure we used for malaria nets at the time.
The paper was not really about metformin. It was about a strategy: that aging is the largest single source of death and suffering on Earth, that it can probably be slowed by simple and extremely safe interventions, that the research to prove this is wildly underfunded because of a structural market failure, and that this makes it a textbook neglected EA cause. Eight years later, I want to grade that thesis honestly.
What aged well: the strategy mostly worked
The encouraging surprise is that almost every mechanism we pointed at has since moved in the predicted direction.
The regulatory icebreaker happened. The entire point of the TAME trial (Targeting Aging with Metformin) was never just to test one drug — it was to get the FDA to accept aging itself, operationalized as delayed multimorbidity, as a legitimate trial endpoint. That part succeeded. The FDA agreed to the design. Which means the precedent we said was the real prize now exists: anyone with deep enough pockets can run a trial against an aging-type endpoint rather than a single named disease. In a precedent-based legal system, this is the hard part, and it is done.
The WHO moved on classifying aging. In 2018 we listed “lobby WHO to recognize aging as a disease” as a cheap high-leverage action, because without an ICD code the pharma industry has no commercial path. ICD-11 is now in force and contains the code MG2A, “Ageing-associated decline in intrinsic capacity” (a real improvement on the older, ill-defined “old age”), plus the extension code XT9T, “Ageing-related.” Aging is now, formally, a thing medicine is allowed to target. That call resolved in our favor.
Patient-funded, decentralized trials became real. We argued that patient organizations and crowdfunding could route around the funding failure, and pointed at Lifespan.io and Open Longevity as early experiments. The model worked. The most important human longevity trial of the last few years — PEARL, the rapamycin trial — was a decentralized, largely patient-funded, placebo-controlled RCT, run exactly on the template we described. It was the first long-term randomized trial of rapamycin for longevity in healthy adults, and it reported the drug was reasonably safe over a year, with some secondary benefits (notably for women). That’s a proof of method as much as of molecule.
The supply island got a bridge — built by tech billionaires. Our central metaphor was two islands with no connection between them: a “demand island” of people who desperately want to not age, and a “supply island” of scientists with ideas and no money, separated by sharks (scammers, religion, the FDA). Since 2018, a bridge appeared from an unexpected direction. Altos Labs launched in 2022 with ~$3B (Bezos, Milner, Yamanaka as advisor) and is now valued around $5B. Retro Biosciences (backed by Sam Altman) is reportedly raising a $1B round. NewLimit (Brian Armstrong) has put ~$280M into reprogramming. And in January 2026, Life Biosciences (co-founded by David Sinclair) received FDA clearance to begin the first human trial of partial epigenetic reprogramming. Eight years ago, fundamental aging research was a backwater funded in the low hundreds of millions; today partial cellular reprogramming alone is a clinical-stage field. The thing we said could not get funded is now where the smart money goes.
What aged badly: betting on a single molecule
Now the part where I was wrong, or at least fragile.
The mascot drug never got off the ground. Seven years after we wrote about it as the critical near-term intervention, TAME still isn’t fully funded. The price tag is now quoted at roughly $45–70M; the National Institute on Aging has put in on the order of $5M; no pharmaceutical company will touch it, because metformin is generic and there is nothing to sell. This is painful, because it is exactly the market failure we described — we just underestimated how completely it would bind. We treated the funding gap as a problem to be solved by attention; it turned out to be load-bearing. The lesson is uncomfortable: we were right about the structural failure and wrong to assume that being right about it would move money.
The evidence base for metformin specifically got weaker, not stronger. A 2025 review flagged real weaknesses in the foundational metformin-longevity research, and there’s an ongoing concern that metformin may blunt the adaptations you get from exercise — not what you want from a drug you’d give to healthy people. Meanwhile rapamycin’s PEARL trial, which actually ran, showed safety but did not demonstrate lifespan or healthspan extension in humans. The honest summary is that none of the simple geroprotectors has yet been proven to extend human life in a randomized trial. The category remains promising and remains underfunded for precisely the reasons we gave; but we hung the headline on one horse, and that was a mistake of presentation. The argument should have been about funding a basket of safe candidates, where the bet is that even if several do nothing, a few work and the safe combination nets out positive.
The field is hard, and some bets died. Unity Biotechnology, one of the original senolytics companies, shut down in September 2025. Longevity fundraising overall hit its lowest level since 2019. None of this refutes the thesis — drug development is brutal everywhere — but it’s a corrective to 2018’s optimism about how quickly “simple” would translate to “proven.”
The real revision: the finish line moved
Here is the update that actually changes the model, and it has nothing to do with biology.
Our entire calculation hung on a single number: when does radical life extension arrive? In 2018 we used 2100, the date at which aggregated expert opinion gave roughly a 70% chance of transformative AI, and we reasoned that medical superintelligence (the thing that actually defeats aging) would come somewhat later than raw AGI. That now looks wildly conservative.
Since then, forecasts have collapsed. The Metaculus community estimate for AGI fell from a ~50-year median in 2020 to roughly the early 2030s. Large surveys of AI researchers (Grace et al., 2024) moved their 50%-confidence date for high-level machine intelligence to ~2047 — a 13-year shift in two years. Anthropic stated officially that it expects powerful AI in late 2026 or early 2027; the major labs cluster their predictions in 2026–2030. There has been some modest re-lengthening among a few forecasters in early 2026, so I won’t pretend the noise is gone. But even after heavy discounting for hype and selection, the center of mass for transformative AI has moved from “end of the century” to “within a couple of decades.”
So I’ll pull the finish line in. Keeping the discipline of putting medical/radical-life-extension capability later than raw AGI — alignment, deployment, regulation, and biology all take time even after you have a superintelligence — a defensible updated central estimate for “aging is substantially solved” is somewhere around 2045–2055, not 2100.
Work through it:
1.The bridge is now short. A 20-year span is the difference between a 60-year-old making it and not making it. For someone who needs to survive ~20 more years, a +1 or +2 year shift in life expectancy is dramatically more decision-relevant than it is for someone who’d need to survive 75. The per-person value of a cheap, deploy-tomorrow intervention goes up.
2.The gross headline number goes down. This is the honest cost. In a 75-year window, vast cohorts die of aging before the finish line, so a 10% mortality shift rescues a huge absolute number. In a 20-year window, fewer people were going to die in the interim at all, so the same percentage shift rescues fewer. If I redo the toy arithmetic against the ~1.4 billion people projected to be 60+ around 2030, a 1-year life-expectancy gain shifting roughly 10% of near-the-edge mortality lands somewhere in the tens of millions of people rescued into the post-aging world — call it an order of magnitude smaller than the 2018 figure. I want to be explicit that this number could be off by a factor of several in either direction; the cohort structure for a near finish line is genuinely harder to model and I am not going to pretend otherwise.
3.But the price per life stays absurd. Tens of millions of lives against a category-of-trials cost in the tens of millions of dollars is still dollars to tens of dollars per life — meaning that even after a brutal downward haircut, and even after updating the bed-net comparison to GiveWell’s current ~$3,000–$5,500 per life, cheap geroprotector research plausibly remains 100–1,000× more cost-effective than the best global-health charities. The headline shrank; the conclusion didn’t.
4.Bostrom’s delay logic now has a short fuse. We invoked “Astronomical Waste”: a small delay in a transformative technology costs enormous future value. In 2018, each year of delay in deploying geroprotectors was ~1/80 of the payoff. If the bridge is ~20 years, each year of delay is a much larger fraction of the remaining window. Procrastination is now far more expensive, per year, than it was when we wrote this.
5.The usable runway is shorter than the bridge. Twenty years until AI is not twenty years to act. The experiment has to finish before then, not merely start — and a proper geroprotector trial runs years (TAME was designed for six), before you even add regulatory approval, manufacturing, and global rollout, each of which eats more years. Worse, these drugs are prophylactic: their benefit is largest when begun early and accumulates over time — metformin’s effect on lifespan is biggest when started early in life and fades with later initiation — so people need to be on the intervention for years before the finish line for it to move their survival odds. You cannot start taking it in 2044 and expect it to carry you across in 2045. Back all of that out and the effective deadline to begin is essentially now. This is the real teeth behind the delay argument: a year lost is no longer 1/80 of the payoff, it can be the difference between the whole pipeline landing inside the window or missing it entirely. The one consolation cuts in our favour — it is exactly the case for cheap, already-safety-tested, off-patent interventions over elegant therapies that only arrive in the 2040s. Only something you could begin giving people today can possibly finish crossing its own bridge in time.
Aging versus x-risk: the balance tips
In 2018 we argued that fighting aging and fighting existential risk are not competitors but two versions of the same task — preventing death, at the personal and the civilizational scale — and that they’re synergistic. I believe that even more now. But a near AGI sharpens the relationship in a way that I have to concede tilts the balance.
If transformative AI is ~20 years out rather than ~80, then the decisive question for everyone alive — including every 60-year-old hoping metformin buys them a few years — is whether that AI is aligned. An aligned superintelligence plausibly solves aging more or less as a side effect; an unaligned one kills everyone regardless of what’s in their pill organizer. So my updated position is sharper than the paper’s: the cost-effectiveness case for cheap geroprotectors is stronger in absolute terms than it was in 2018, and at the same time more clearly subordinate to AI safety. Both things are true. Aging is still, by the raw arithmetic, an underfunded bargain. But on the margin, the dollar that reduces AI risk now dominates the dollar that extends a geroprotector trial, because it operates on the variable that determines whether any of the bridge matters.
There’s a corollary we raised half-jokingly in 2018 that has stopped being a joke: the aging of the people doing alignment research. We noted that the field’s key thinkers were young enough to ignore their own mortality. They’re not young anymore. Eliezer Yudkowsky is around 46 now; Nick Bostrom around 53. The estimated productivity peak for theoretical work is the late 40s, with decline after. If the alignment problem genuinely has to be solved in the next decade or two, then the cognitive aging of the small number of people best at it is a real and underdiscussed input variable — and one of the few places where geroscience and AI safety are directly the same project.
What still looks like the best cheap action in 2026
Updating the practical recommendations:
•Fund the category, not the mascot. The single biggest correction to the 2018 paper. Run basket trials of several “generally recognized as safe” interventions at once. The bet isn’t that any one works — it’s that a safe basket nets positive even if most components do nothing.
•Keep building patient-led and decentralized trials. This worked. PEARL is the proof. It’s the cheapest route around the structural funding failure and it’s now a demonstrated method rather than a hope.
•Validate aging biomarkers / clocks to shorten trials. In 2018 this was aspirational; methylation and other aging clocks are now everywhere. Validated surrogate endpoints are the single biggest lever on trial duration, which is the single biggest lever on cost.
•Point AI at geroscience. In a companion 2018 paper (Batin et al.) we argued AI would become the main accelerant of life extension. That’s now the obvious play — AI-driven target discovery and trial design is the place where the two cause areas reinforce instead of compete.
•The negative-price insurance argument is still sitting unused. A drug that delays expensive age-related illness has, for an insurer, negative net cost. Nobody has built the incentive structure around this. It remains free money on the sidewalk.
•Delivery is still the unsolved problem. Identifying a cheap intervention is not the same as getting billions of people to take a pill every day. We had no good answer in 2018 and I don’t have one now beyond fortification-style or app-mediated approaches. This is where the next real work is.
Conclusion
The 2018 paper’s strategy aged considerably better than its poster molecule. The regulatory path opened, the WHO moved, patient-funded trials became routine, and serious money finally crossed onto the supply island. Metformin, specifically, stalled — and stalled for exactly the structural reason we identified, which is its own kind of vindication and its own kind of warning.
But the deepest revision isn’t biological, it’s temporal. We were trying to build a bridge to the year 2100. That bridge now only has to reach somewhere around 2045. A shorter bridge makes cheap, immediately-deployable interventions more valuable for the people who are alive and aging right now — and, at the same time, makes getting AI right the thing that dwarfs everything else, including this. Both halves of that sentence matter, and the EA community has internalized the second half far more than the first.
Which leaves the original observation standing, only sharper: by the plain arithmetic, fighting aging is still one of the most cost-effective ways to prevent death that exists, and it remains a cause that almost no one in effective altruism actually funds. Death is bad at two scales — the person and the species — and it is the same problem at both. We’ve spent eight years getting much more serious about the second scale. We are still, strangely, not serious about the first.
Disclaimer / not medical advice: nothing here is a recommendation to take metformin, rapamycin, or any other compound for life extension. The whole point is that the trials proving safety and efficacy in healthy people have mostly not been done. I have worked on non-commercial terms with life-extension organizations; I have no commercial interest in any drug mentioned. This is a personal blog post, and an argument, not a clinical or financial recommendation.
Epistemic status: Revisiting a paper I co-wrote in 2018 (Turchin, Milova, Egorova & Batin, “Fighting Aging as an Effective Altruism Cause”). The core argument I still endorse. One of the central numerical inputs — the date at which radical life extension arrives — has moved so much in eight years that the whole model has to be re-derived. Several of the specific bets aged badly. Numbers below are order-of-magnitude toy estimates and should be read as such; the optimizer’s curse applies to me as much as to anyone.
The claim, in one paragraph
In 2018 my coauthors and I argued something that still sounds absurd when you say it out loud: that the most cost-effective way to save a human life, under a defensible set of assumptions, was not bed nets but clinical trials of cheap, off-patent geroprotectors. The logic was a bridge argument. Somewhere in the future, technology defeats aging. Everyone who survives until that moment gets potentially indefinite life extension — they cross into a world without involuntary death. So any intervention that nudges average life expectancy up by even one year increases the number of people who make it across the bridge. Because human mortality follows the steep Gompertz curve, a single year of life expectancy is roughly a 10% relative reduction in mortality, which, applied to billions of people near the edge, rescues an enormous number of them into the post-aging world. We put a toy number on it: a 1-year shift from metformin, a finish line at 2100, a population of 10 billion, and you “save” ~250 million people. Divide by the ~$65M cost of the metformin trial and you get $0.24 per life saved — about four orders of magnitude cheaper than the ~$1,500-per-life figure we used for malaria nets at the time.
The paper was not really about metformin. It was about a strategy: that aging is the largest single source of death and suffering on Earth, that it can probably be slowed by simple and extremely safe interventions, that the research to prove this is wildly underfunded because of a structural market failure, and that this makes it a textbook neglected EA cause. Eight years later, I want to grade that thesis honestly.
What aged well: the strategy mostly worked
The encouraging surprise is that almost every mechanism we pointed at has since moved in the predicted direction.
The regulatory icebreaker happened. The entire point of the TAME trial (Targeting Aging with Metformin) was never just to test one drug — it was to get the FDA to accept aging itself, operationalized as delayed multimorbidity, as a legitimate trial endpoint. That part succeeded. The FDA agreed to the design. Which means the precedent we said was the real prize now exists: anyone with deep enough pockets can run a trial against an aging-type endpoint rather than a single named disease. In a precedent-based legal system, this is the hard part, and it is done.
The WHO moved on classifying aging. In 2018 we listed “lobby WHO to recognize aging as a disease” as a cheap high-leverage action, because without an ICD code the pharma industry has no commercial path. ICD-11 is now in force and contains the code MG2A, “Ageing-associated decline in intrinsic capacity” (a real improvement on the older, ill-defined “old age”), plus the extension code XT9T, “Ageing-related.” Aging is now, formally, a thing medicine is allowed to target. That call resolved in our favor.
Patient-funded, decentralized trials became real. We argued that patient organizations and crowdfunding could route around the funding failure, and pointed at Lifespan.io and Open Longevity as early experiments. The model worked. The most important human longevity trial of the last few years — PEARL, the rapamycin trial — was a decentralized, largely patient-funded, placebo-controlled RCT, run exactly on the template we described. It was the first long-term randomized trial of rapamycin for longevity in healthy adults, and it reported the drug was reasonably safe over a year, with some secondary benefits (notably for women). That’s a proof of method as much as of molecule.
The supply island got a bridge — built by tech billionaires. Our central metaphor was two islands with no connection between them: a “demand island” of people who desperately want to not age, and a “supply island” of scientists with ideas and no money, separated by sharks (scammers, religion, the FDA). Since 2018, a bridge appeared from an unexpected direction. Altos Labs launched in 2022 with ~$3B (Bezos, Milner, Yamanaka as advisor) and is now valued around $5B. Retro Biosciences (backed by Sam Altman) is reportedly raising a $1B round. NewLimit (Brian Armstrong) has put ~$280M into reprogramming. And in January 2026, Life Biosciences (co-founded by David Sinclair) received FDA clearance to begin the first human trial of partial epigenetic reprogramming. Eight years ago, fundamental aging research was a backwater funded in the low hundreds of millions; today partial cellular reprogramming alone is a clinical-stage field. The thing we said could not get funded is now where the smart money goes.
What aged badly: betting on a single molecule
Now the part where I was wrong, or at least fragile.
The mascot drug never got off the ground. Seven years after we wrote about it as the critical near-term intervention, TAME still isn’t fully funded. The price tag is now quoted at roughly $45–70M; the National Institute on Aging has put in on the order of $5M; no pharmaceutical company will touch it, because metformin is generic and there is nothing to sell. This is painful, because it is exactly the market failure we described — we just underestimated how completely it would bind. We treated the funding gap as a problem to be solved by attention; it turned out to be load-bearing. The lesson is uncomfortable: we were right about the structural failure and wrong to assume that being right about it would move money.
The evidence base for metformin specifically got weaker, not stronger. A 2025 review flagged real weaknesses in the foundational metformin-longevity research, and there’s an ongoing concern that metformin may blunt the adaptations you get from exercise — not what you want from a drug you’d give to healthy people. Meanwhile rapamycin’s PEARL trial, which actually ran, showed safety but did not demonstrate lifespan or healthspan extension in humans. The honest summary is that none of the simple geroprotectors has yet been proven to extend human life in a randomized trial. The category remains promising and remains underfunded for precisely the reasons we gave; but we hung the headline on one horse, and that was a mistake of presentation. The argument should have been about funding a basket of safe candidates, where the bet is that even if several do nothing, a few work and the safe combination nets out positive.
The field is hard, and some bets died. Unity Biotechnology, one of the original senolytics companies, shut down in September 2025. Longevity fundraising overall hit its lowest level since 2019. None of this refutes the thesis — drug development is brutal everywhere — but it’s a corrective to 2018’s optimism about how quickly “simple” would translate to “proven.”
The real revision: the finish line moved
Here is the update that actually changes the model, and it has nothing to do with biology.
Our entire calculation hung on a single number: when does radical life extension arrive? In 2018 we used 2100, the date at which aggregated expert opinion gave roughly a 70% chance of transformative AI, and we reasoned that medical superintelligence (the thing that actually defeats aging) would come somewhat later than raw AGI. That now looks wildly conservative.
Since then, forecasts have collapsed. The Metaculus community estimate for AGI fell from a ~50-year median in 2020 to roughly the early 2030s. Large surveys of AI researchers (Grace et al., 2024) moved their 50%-confidence date for high-level machine intelligence to ~2047 — a 13-year shift in two years. Anthropic stated officially that it expects powerful AI in late 2026 or early 2027; the major labs cluster their predictions in 2026–2030. There has been some modest re-lengthening among a few forecasters in early 2026, so I won’t pretend the noise is gone. But even after heavy discounting for hype and selection, the center of mass for transformative AI has moved from “end of the century” to “within a couple of decades.”
So I’ll pull the finish line in. Keeping the discipline of putting medical/radical-life-extension capability later than raw AGI — alignment, deployment, regulation, and biology all take time even after you have a superintelligence — a defensible updated central estimate for “aging is substantially solved” is somewhere around 2045–2055, not 2100.
Work through it:
1. The bridge is now short. A 20-year span is the difference between a 60-year-old making it and not making it. For someone who needs to survive ~20 more years, a +1 or +2 year shift in life expectancy is dramatically more decision-relevant than it is for someone who’d need to survive 75. The per-person value of a cheap, deploy-tomorrow intervention goes up.
2. The gross headline number goes down. This is the honest cost. In a 75-year window, vast cohorts die of aging before the finish line, so a 10% mortality shift rescues a huge absolute number. In a 20-year window, fewer people were going to die in the interim at all, so the same percentage shift rescues fewer. If I redo the toy arithmetic against the ~1.4 billion people projected to be 60+ around 2030, a 1-year life-expectancy gain shifting roughly 10% of near-the-edge mortality lands somewhere in the tens of millions of people rescued into the post-aging world — call it an order of magnitude smaller than the 2018 figure. I want to be explicit that this number could be off by a factor of several in either direction; the cohort structure for a near finish line is genuinely harder to model and I am not going to pretend otherwise.
3. But the price per life stays absurd. Tens of millions of lives against a category-of-trials cost in the tens of millions of dollars is still dollars to tens of dollars per life — meaning that even after a brutal downward haircut, and even after updating the bed-net comparison to GiveWell’s current ~$3,000–$5,500 per life, cheap geroprotector research plausibly remains 100–1,000× more cost-effective than the best global-health charities. The headline shrank; the conclusion didn’t.
4. Bostrom’s delay logic now has a short fuse. We invoked “Astronomical Waste”: a small delay in a transformative technology costs enormous future value. In 2018, each year of delay in deploying geroprotectors was ~1/80 of the payoff. If the bridge is ~20 years, each year of delay is a much larger fraction of the remaining window. Procrastination is now far more expensive, per year, than it was when we wrote this.
5. The usable runway is shorter than the bridge. Twenty years until AI is not twenty years to act. The experiment has to finish before then, not merely start — and a proper geroprotector trial runs years (TAME was designed for six), before you even add regulatory approval, manufacturing, and global rollout, each of which eats more years. Worse, these drugs are prophylactic: their benefit is largest when begun early and accumulates over time — metformin’s effect on lifespan is biggest when started early in life and fades with later initiation — so people need to be on the intervention for years before the finish line for it to move their survival odds. You cannot start taking it in 2044 and expect it to carry you across in 2045. Back all of that out and the effective deadline to begin is essentially now. This is the real teeth behind the delay argument: a year lost is no longer 1/80 of the payoff, it can be the difference between the whole pipeline landing inside the window or missing it entirely. The one consolation cuts in our favour — it is exactly the case for cheap, already-safety-tested, off-patent interventions over elegant therapies that only arrive in the 2040s. Only something you could begin giving people today can possibly finish crossing its own bridge in time.
Aging versus x-risk: the balance tips
In 2018 we argued that fighting aging and fighting existential risk are not competitors but two versions of the same task — preventing death, at the personal and the civilizational scale — and that they’re synergistic. I believe that even more now. But a near AGI sharpens the relationship in a way that I have to concede tilts the balance.
If transformative AI is ~20 years out rather than ~80, then the decisive question for everyone alive — including every 60-year-old hoping metformin buys them a few years — is whether that AI is aligned. An aligned superintelligence plausibly solves aging more or less as a side effect; an unaligned one kills everyone regardless of what’s in their pill organizer. So my updated position is sharper than the paper’s: the cost-effectiveness case for cheap geroprotectors is stronger in absolute terms than it was in 2018, and at the same time more clearly subordinate to AI safety. Both things are true. Aging is still, by the raw arithmetic, an underfunded bargain. But on the margin, the dollar that reduces AI risk now dominates the dollar that extends a geroprotector trial, because it operates on the variable that determines whether any of the bridge matters.
There’s a corollary we raised half-jokingly in 2018 that has stopped being a joke: the aging of the people doing alignment research. We noted that the field’s key thinkers were young enough to ignore their own mortality. They’re not young anymore. Eliezer Yudkowsky is around 46 now; Nick Bostrom around 53. The estimated productivity peak for theoretical work is the late 40s, with decline after. If the alignment problem genuinely has to be solved in the next decade or two, then the cognitive aging of the small number of people best at it is a real and underdiscussed input variable — and one of the few places where geroscience and AI safety are directly the same project.
What still looks like the best cheap action in 2026
Updating the practical recommendations:
• Fund the category, not the mascot. The single biggest correction to the 2018 paper. Run basket trials of several “generally recognized as safe” interventions at once. The bet isn’t that any one works — it’s that a safe basket nets positive even if most components do nothing.
• Keep building patient-led and decentralized trials. This worked. PEARL is the proof. It’s the cheapest route around the structural funding failure and it’s now a demonstrated method rather than a hope.
• Validate aging biomarkers / clocks to shorten trials. In 2018 this was aspirational; methylation and other aging clocks are now everywhere. Validated surrogate endpoints are the single biggest lever on trial duration, which is the single biggest lever on cost.
• Point AI at geroscience. In a companion 2018 paper (Batin et al.) we argued AI would become the main accelerant of life extension. That’s now the obvious play — AI-driven target discovery and trial design is the place where the two cause areas reinforce instead of compete.
• The negative-price insurance argument is still sitting unused. A drug that delays expensive age-related illness has, for an insurer, negative net cost. Nobody has built the incentive structure around this. It remains free money on the sidewalk.
• Delivery is still the unsolved problem. Identifying a cheap intervention is not the same as getting billions of people to take a pill every day. We had no good answer in 2018 and I don’t have one now beyond fortification-style or app-mediated approaches. This is where the next real work is.
Conclusion
The 2018 paper’s strategy aged considerably better than its poster molecule. The regulatory path opened, the WHO moved, patient-funded trials became routine, and serious money finally crossed onto the supply island. Metformin, specifically, stalled — and stalled for exactly the structural reason we identified, which is its own kind of vindication and its own kind of warning.
But the deepest revision isn’t biological, it’s temporal. We were trying to build a bridge to the year 2100. That bridge now only has to reach somewhere around 2045. A shorter bridge makes cheap, immediately-deployable interventions more valuable for the people who are alive and aging right now — and, at the same time, makes getting AI right the thing that dwarfs everything else, including this. Both halves of that sentence matter, and the EA community has internalized the second half far more than the first.
Which leaves the original observation standing, only sharper: by the plain arithmetic, fighting aging is still one of the most cost-effective ways to prevent death that exists, and it remains a cause that almost no one in effective altruism actually funds. Death is bad at two scales — the person and the species — and it is the same problem at both. We’ve spent eight years getting much more serious about the second scale. We are still, strangely, not serious about the first.
Disclaimer / not medical advice: nothing here is a recommendation to take metformin, rapamycin, or any other compound for life extension. The whole point is that the trials proving safety and efficacy in healthy people have mostly not been done. I have worked on non-commercial terms with life-extension organizations; I have no commercial interest in any drug mentioned. This is a personal blog post, and an argument, not a clinical or financial recommendation.