Derya Unutmaz: How AI Could Extend Human Lifespan

Derya Unutmaz has a piece of advice he repeats often: try not to die in the next ten years. He means it half as a joke and half as a genuine forecast. Unutmaz is an immunologist and aging researcher at the Jackson Laboratory, and he is also one of a small number of scientists given early access to OpenAI's models, which he has used to help solve a three-year-old mystery in his own lab. His bet is that the next decade of medicine will look nothing like the last one, because artificial intelligence is compounding faster than human intuition expects.

The prediction rests on an idea called longevity escape velocity, a term coined by biomedical gerontologist Aubrey de Grey. It describes the point at which medical advances extend a person's remaining life expectancy faster than time is passing, so that every year lived buys back more than a year of life (09:13). Unutmaz thinks this threshold could arrive within eight to ten years. He points to a working example already in motion: GLP-1 receptor agonists, the class of drugs that includes semaglutide, appear to add roughly five to ten years of life for people with obesity or related chronic disease (10:12). If one drug class can do that, he argues, a cascade of similar breakthroughs, arriving faster each year, could eventually add decades rather than years.

"The next 10 years you can think of it as more advanced than the last century." — Derya Unutmaz [08:21]

What makes this more than wishful thinking, in Unutmaz's telling, is what large language models, the AI systems trained on huge amounts of text and data to generate humanlike reasoning, can already do with biological data. He describes handing GPT-5.5 Pro a data set of millions of RNA sequencing points, the kind of experiment that used to take a graduate student months to analyze, and getting back a forty-page report in under two hours that not only summarized the findings but proposed the next experiment to run (14:59, 40:55). He says the model's predictions about how one of his own two-week lab experiments would turn out were correct about 98 percent of the time, matching an intuition he built over thirty years at the bench.

That capacity is the seed of what Unutmaz calls a digital twin: a simulated version of a person's biology, built from genetics, metabolism, immune function, gut microbiome, and medical history, detailed enough that a researcher could test a drug on the simulation before testing it on the person (17:11). If it works, clinical trials that take years could shrink to weeks, because a model could flag in advance which patients are likely to respond and which are likely to suffer a side effect. Unutmaz is careful to say this is not close to reality yet. He compares the trust problem to self-driving cars, which he estimates are already about ten times safer than human drivers but still need years of validation before people hand over the wheel entirely (21:35). Biology, he says, will take longer to earn that trust, because it is more complex than traffic.

When Not Using AI Becomes the Risk

Unutmaz's more provocative claim concerns doctors, not machines. After OpenAI's o1 reasoning model came out, he began testing it on medical questions and found it reasoned through complex cases, weighing a patient's context rather than just retrieving facts, in a way earlier models could not. That led him to a blunt conclusion.

"Right now, it's unethical for physicians not to use AI anymore." — Derya Unutmaz [44:49]

His reasoning starts with a sobering baseline: an estimated 12 million medical misdiagnoses occur in the United States every year, and roughly 700,000 people die or are seriously harmed as a result (46:42). Some of that is unavoidable human error. But Unutmaz cites AI systems that can spot breast cancer tumors in imaging years before a radiologist would catch them (47:19), and a Science paper testing the older o1-preview reasoning model that found it performed significantly better than the average doctor at diagnosis (49:20). His view is not that AI should replace a physician's judgment, but that ignoring a tool this reliable is itself a form of malpractice waiting to be recognized as such.

The same argument extends to prevention. Unutmaz is blunt about what current medicine actually is.

"We say healthcare. No, we don't have healthcare. We have sick care." — Derya Unutmaz [78:56]

He points to a study using the UK Biobank, a research database holding health information on roughly 500,000 people, in which AI models analyzed the data and were able to retroactively predict around 1,000 different diseases years before they clinically appeared (79:14). The implication is that the biological warning signs of disease exist long before symptoms do. The bottleneck has been that nobody is collecting enough data from healthy people to see them.

Cancer, Aging, and the Organ That Resists Rejuvenation

On cancer, Unutmaz makes a case that immunotherapy, treatment that trains a patient's own immune system to recognize and attack tumors, has already changed the odds. He argues it is now more powerful than chemotherapy and radiotherapy combined (69:10), because it can adapt to a tumor's mutations rather than attacking indiscriminately. He points to a case out of Australia in which a computer scientist used AI models to design a personalized mRNA vaccine for his dog's melanoma, sequencing the tumor and having the AI identify the exact molecule needed to train the dog's immune system, a process completed in about three months. The dog, he says, was still alive past its expected prognosis.

Aging itself, Unutmaz frames as a loss of biological information rather than simple wear. The body has twelve recognized hallmarks of aging, including genomic instability and cellular senescence, and he argues these emerge because cells gradually forget how to repair and regenerate themselves (89:09). Some animals resist this better than others: elephants carry extra copies of the tumor-suppressor gene p53 and rarely develop cancer (98:58), while naked mole rats, which live up to 40 years compared to a normal rat's two or three, carry a mutation in an immune gene called cGAS linked to DNA repair (99:20). Unutmaz sees these as proof that the aging program can be rewritten, not just slowed.

The clearest experimental version of that rewriting is partial cellular reprogramming, a technique that uses a subset of the four Yamanaka factors, the proteins discovered by biologist Shinya Yamanaka that can turn an adult cell back into a stem cell, to nudge an old cell toward a younger state without erasing its identity as a skin cell, an eye cell, or a neuron (111:18). Biologist David Sinclair is preparing a human trial applying this to the eye. But Unutmaz is candid about its limits: reprogramming does not reverse every hallmark of aging at once. Telomeres, the mutations accumulated in mitochondria, and the surrounding tissue environment can remain aged even after a cell's epigenetic clock is turned back, which is why a rejuvenated cell can drift back toward its old state if nothing else around it changes.

The brain poses the hardest version of this problem. Unlike skin, the brain cannot simply regenerate its cells, because doing so risks erasing the very memories and connections that make someone who they are. Unutmaz's answer is not full cellular replacement but a slower, more targeted kind of maintenance: reducing neuroinflammation, supporting the brain's own limited repair systems, and, decades from now, perhaps AI models detailed enough to map a person's neural connections before attempting any repair.

What a Person Can Build Today

Toward the end of the conversation, Unutmaz brings the idea down to something a listener could try. His own version of a small-scale digital twin involves feeding an AI model like GPT-5.5 Pro a continuous stream of personal data: glucose readings from a continuous monitor taken every five minutes, lab values, supplements, sleep, and steps, kept in the same ongoing conversation so the model can compare before-and-after changes over time (149:29). He argues this lets a model learn an individual's own normal range, rather than relying on a population-wide reference, so it can flag a meaningful deviation instead of a number that only looks abnormal on a chart built for someone else (158:29).

None of this erases the uncertainty built into Unutmaz's larger timeline. He is speculating about a decade that has not happened yet, based on a rate of technological change that is genuinely hard for anyone to intuit. But the more modest version of his argument, that AI can already compress analysis that used to take months into minutes, and that continuous personal data plus a model willing to compare today against yesterday is available now, is not speculation. It is closer to a description of what a curious, technically literate patient could start doing this year. He ends on something almost old-fashioned for a conversation about machines.

"Being optimistic is one of the best things you can do for aging, and study after study show that." — Derya Unutmaz [148:01]

Whether or not longevity escape velocity arrives on his schedule, that particular claim needs no AI to test.