Will you live to 120? · The Hunger Switch

Where the machines come in

One more nutrient-sensing drug deserves mention, because it sits in an interesting middle and is the field’s most important regulatory test case. Metformin is a cheap, decades-old, widely-used diabetes drug that, among its several effects, activates AMPK — the fasting/maintenance sensor. Some observational studies hint that people taking metformin have lower mortality and cancer rates; it extends lifespan in some animal models (though not all); and it is plausible as a mild CR-mimetic. The veteran researcher Nir Barzilai has long championed a trial called TAME (Targeting Aging with Metformin), designed less to prove metformin is a wonder drug than to do something more strategic: to test whether a drug can delay the onset of multiple age-related diseases at once, and in doing so to persuade regulators to treat “aging” itself as a condition worth targeting — which would change the economics of the entire field.

The honest status is mixed and modest. TAME has struggled for years to secure funding and has not yet delivered a definitive verdict. The evidence that metformin helps non-diabetics live longer is genuinely uncertain — and there is even a finding that metformin may blunt some of the benefits of exercise, a reminder that pushing on a nutrient sensor can have costs as well as benefits. Metformin is cheap, safe, and important as a proof-of-concept for regulating aging as a treatable process. It is not, on current evidence, a demonstrated longevity drug for healthy people. File it under “promising and unproven,” next to almost everything else here that isn’t rapamycin.

Where the machines come in

Nutrient sensing is, in one sense, the most AI-ready of all the hallmarks, because the targets are well-defined molecular nodes — mTOR, AMPK, the sirtuins, the insulin/IGF pathway — and “find or design a molecule that selectively tunes a known target” is exactly the problem the newest computational tools are built for. Companies like Insilico Medicine use AI to hunt for novel longevity-relevant targets in vast biological datasets and then to generate candidate molecules to hit them; Insilico’s pipeline has carried an AI-discovered drug for a fibrotic lung disease — an aging-related condition — into human trials, a genuine milestone for the field of computational drug discovery (even if the indication is a specific disease rather than “aging”). The same approach could, in principle, yield better CR-mimetics or rapamycin-like drugs with the longevity benefit and fewer of the side effects — the selectivity problem from the last chapter, attacked with design.

But weigh that promise against everything this chapter has shown, because the binding constraints are not the ones AI removes. AI can accelerate the discovery of molecules that hit these nodes; it cannot tell you whether tuning the node in a human extends healthy life, and it cannot manufacture the large, long, placebo-controlled, independently-replicated human trials that are the only thing that ever settles the question. There is also a subtler danger specific to this overhyped corner of biology: AI is very good at finding patterns and candidates, and in a field where moving a biomarker is easy and proving a benefit is hard, a clever algorithm trained to optimize a surrogate — lower the clock, raise the NAD⁺, shift the ratio — will faithfully hand you compounds that are excellent at gaming the surrogate, which is not the same as compounds that make you younger. The lesson of the resveratrol assay artifact was that a measurement can be optimized into meaninglessness; an AI pointed at the wrong target will optimize it faster. The computation is not the constraint. The constraint is what the last four chapters keep insisting on: it has to work in a human, and a biomarker is not a human outcome.

✦ ✦ ✦

Carry these three things

1. Longevity is, in part, a resource-allocation decision, and you can bias it toward maintenance. A handful of conserved sensors — insulin/IGF and mTOR on the “grow” side, AMPK and sirtuins on the “maintain” side — read how much food is around and set a growth-versus-repair dial. Caloric restriction, fasting, and exercise tip that dial toward maintenance, and the maintenance setting is also the longevity setting, across the whole tree of life. Kenyon’s single-gene worm proved aging is a regulated mechanism, not just entropy — and the mechanism is a food sensor.

2. Rank interventions by the quality of their evidence, not the volume of their marketing — and the two are nearly inverted. Rapamycin sits at the strong end: a multi-lab, reproducible, mammalian result that worked even when started in old age, now being tested in dogs and humans — and even rapamycin is not yet a proven human longevity drug, which is the bar everything must eventually clear. NAD⁺ boosters like NMN sit at the weak end: real biology underneath, but human benefit unproven, mechanism disputed, and the field drowning in financial conflict — the most oversold molecules in aging, a chemical replay of the resveratrol collapse that cost GSK $720 million. “It raises your NAD⁺” is a thermometer claim, not a thermostat one.

3. “Compared to what?” is the question that breaks or makes a study. The two rhesus-monkey caloric-restriction studies appeared to contradict each other for years — and the difference came down almost entirely to what the control monkeys were fed. Not fraud, not sloppiness: two careful studies with different comparison groups, which the whole field had to learn to read together. Most whiplash-inducing health news has this exact shape. So far we’ve tried to slow aging by removing damaged cells, measuring damage, and dialing down the rate of damage. The next chapter takes on the most famous theory of why the damage happens in the first place — the idea that we rust from the inside — and shows you one of the most expensive, instructive ways a beautiful theory can simply be wrong.

All 10 chapters of Will you live to 120? · Foundations · All courses