Thirty Years of Research, Collapsed into One Sentence

Nathan Labenz welcomes Derya Unutmaz, a biomedical scientist and human immunologist at the Jackson Laboratory who studies aging and cancer immunotherapy, and one of OpenAI’s ChatGPT Pro grant recipients. Unutmaz has spent 35 years in research after medical school, but he says the work is less a typical career than a pair of overlapping passions. Day to day he tries to understand biological systems, with a focus on the immune system’s role in protection, disease, and aging. The applied edge of that work is cellular engineering: programming immune cells to recognize and attack tumors.

His computing habit started earlier than the medicine. In high school he learned BASIC and assembly on a Commodore VIC-20, then a Commodore 64 — 64 kilobytes of RAM, he notes, not even a toy by today’s standards. In the early 1990s Ray Kurzweil’s *The Age of Intelligent Machines* pulled him into AI; he dabbled in the symbolic era of Lisp and Smalltalk and still recommends Marvin Minsky’s *The Society of Mind*. A later Kurzweil book, *The Singularity Is Near*, locked in a timeline he has taken seriously ever since: AGI around 2029, singularity in the 2040s. About 20 years ago he started a blog called Biosingularity to imagine AI that would finally let us understand biology, treat every disease including cancer, and eventually reverse aging.

That last claim was heresy a couple of decades ago. Unutmaz still treats aging as a disease that should be cured, and he argues that biology looks less magical the closer you look — “pretty bad engineering,” a legacy system whose neural networks are not so different from machine ones. The same instinct shows up in the lab. Using the immune system, his group programs T cells to attack cancer, creating logic such as AND and OR gates on living cells. “It’s extremely cool engineering,” he says, and it sits on top of earlier work on HIV, COVID, and chronic disease. People who took ideas like that seriously, he adds, citing Steve Jobs, had to be a little crazy: they refused the status quo while much of academia stayed conservative.

Unutmaz has been in what he calls the ChatGPT moment since GPT-3. Two to two and a half years before this conversation, he told colleagues something amazing was happening and that they needed to change how they did science. The pushback, including from people who understood computers and bioinformatics, was that the models were just next-word predictors. That dismissal held until last summer. A well-known California professor saw Unutmaz’s posts, still could not believe the models were that good, and asked him to try a project on o1 Preview and o1 Pro. Unutmaz ran the analysis and sent it back. The reply, twice in one email: “Oh my God, I can’t believe this.” The professor is now a daily user.

The same pattern showed up in grant writing. A colleague working with him and ChatGPT produced what Unutmaz calls probably the best grant they have written, with better ideas, and said she could not believe how easy it was. He is more disappointed on the medical side, where he now argues it has become unethical not to use AI. Several physician friends already rely on ChatGPT for diagnosis, he says, but they will not admit it because they fear being valued less. His reply is that individuals may be less uniquely valuable, yet they add more value to patients and to humanity by catching errors. Labenz notes that Americans have a right to an attorney at trial and wonders whether they should also have a right to an AI.

Unutmaz’s own bar is higher than occasional use. He no longer trusts any of his knowledge or ideas without asking a model first, even in areas of extreme expertise. He uploaded a review he wrote on a topic where he believes he is among the best five or ten people in the world; o1 Pro found things he had missed, fully understood a very specific subject, and offered creative insights. Deep Research later did the same, in work he had not yet posted publicly. Friends still come to him when a parent has cancer; he tells them he will consult ChatGPT first, they can too, and he will double-check — but he will not answer without the model.

The most domestic example is his 85-year-old mother in Turkey. He taught her ChatGPT last summer for health issues; she now says she cannot live without it and no longer asks him medical questions because she trusts the model more. He approves. Hallucinations were a real limit a year ago, he says, but the systems have improved tremendously — in his view about ten times in the last three or four months, with another ten times likely in the next three or four. He also used Deep Research to write a patent application for a colleague on a molecule with anti-cancer effects. It ran the prior-art search, understood the chemical formula, and drafted claims and secondary claims. Unutmaz has eight or nine patents of his own and called it the best application he had ever seen. Lawyers would have charged more than $10,000; it took about half an hour and cost next to nothing.

Labenz asks where the models help most across hypothesis generation, literature review, experimental design, and analysis — especially whether they can produce new knowledge, the latest goalpost after “they don’t reason.” Unutmaz rejects the idea that scientific ideas appear from thin air. Hypotheses rest on background knowledge; ask a non-expert for a brilliant T-cell cancer protocol and you will not get one. His own advantage, he says, is range. He has played video games since he was a teenager, and that habit became a research tool. Before o1, GPT-4 was mainly how he surveyed a field that publishes hundreds or thousands of papers a year in his niche alone. It was knowledge, not yet insight.

o1 Preview changed that. Unutmaz is building a cancer immunotherapy in which genetically engineered T cells — “soldiers” in the body — hunt tumor cells, then exhaust, kill poorly, or cause side effects. He had already borrowed a battle royale frame from games such as PUBG: players drop into a restricted area, compete for resources, and eliminate rivals. T cells in a tumor microenvironment compete in a similar way. The lab ran experiments on that analogy, saw some effect, and is about to publish. Then he asked o1 Preview to invent *new* ideas inspired by the same games. It did, drawing on how the games are actually played and transferring that logic to hunting, competition, improvement, and exhaustion. Several suggestions were ones he had not thought of.

He describes that session as perhaps his first early-AGI moment. o1 Pro went further: it started producing ideas the group put into grant proposals, not merely facts he had failed to look up. Some of the suggestions were true hypotheses — imagined mechanisms, cell behaviors, and experimental strategies that nobody yet knows are true. He has repeated the exercise for world-leading friends in neuroscience and inflammatory disease, and they were shocked. “To say that AI cannot come up with anything novel or innovative is total nonsense,” he says. “In fact I will argue that very soon AI is going to be much more innovative and creative than we are.” Google’s AI co-scientist had just been released; he hoped to test it soon. The Eureka hits now arrive almost daily, sometimes so strong he is almost afraid to open Deep Research or o1 Pro.

The other shock was data analysis. Biology now generates enormous datasets — billions or trillions of tokens’ worth, as he puts it, from RNA to proteins to cell interactions — and conventional bioinformatics never feels sufficient. His team had CRISPR and RNA-sequencing results on T cells: some genes expressed in young people silent in the old, a different set rising with age, with the cells further split into subtypes. They could gesture at functions but could not quite say what the pattern meant. He handed it to Deep Research.

The model’s insights landed in a field he has worked for 30-plus years. One sentence made him emotional, because it recapitulated everything he had done in that time — an insight he feels he should have reached himself. Naive T cells are where an immune response starts: they see antigen, become effectors, fight infection or cancer, and some persist as memory. Their numbers fall with age, which is a standard explanation for weaker immunity. The data, and the model, said the story was also qualitative. The remaining cells that still look “young” are no longer young. They have changed epigenetically; their character has changed.

Packed into that sentence, Unutmaz says, is an enormous amount of differentiation biology plus a temporal kind of intelligence — an understanding that states evolve, not a static fact. That combination, more than any single citation, is what impressed him. It is also why he treats the “AI only remixes the literature” critique as settled. The sentence did not merely retrieve a review. It compressed a career’s worth of mechanism into a claim his own team had not fully articulated.

Labenz asks how to get this kind of work out of the systems, noting that a more lateral colleague prompts in the style of famous authors while Unutmaz borrows metaphors from distant fields. Unutmaz has no standard prompt library and rarely uses highly structured templates; he changes the wording every time. Occasionally he asks o1 or o1 Pro to turn his own messy ideas into a prompt. Deep Research often does not need that, because it asks clarifying questions — solid tumors versus lymphomas, this analysis or that — and predicts details he forgot to mention. Grok, having seen his posts, will steer a random question toward immunology, aging, or AI.

His advice is not to stress about the perfect phrasing. Write whatever is in your mind as openly and transparently as possible. That was less true six months earlier, he says, but after o1 the models got much better at context, especially once they start to know you. Cross-domain mixing is how he hunts for novelty: a physics question spiked with biology, a creative task mixed with sports. Innovation, in his view, often appears when two fields collide rather than when you stay inside a narrow specialty.

With thinking models he pushes iteration. He will ask for ten ideas where each one should be better than the last, or tell the model to rethink its own thought and come up with something different. A technique he finds extremely useful is to invent two scientists in the same field — one enthusiastic, one skeptical — and have them criticize, concede, and propose for ten rounds. It is the same AI brainstorming with itself, and it can be revealing. He sees the same self-correction in visible chain-of-thought, which DeepSeek was first to show. People also need to pick the right class of model: thinking models for facts and problems, Deep Research for depth, search tools for knowledge, Grok’s Deep Search when he wants sources found and simplified.

Labenz wants a hit rate: how often the first prompt is great, how often you need many rounds, how often a session goes nowhere. Unutmaz starts with mindset. Many people are not trying to make the system work. They see AI as a threat, cannot accept that it might be smarter than a given doctor or professor, and subconsciously hunt for defects and hallucinations. Humans hallucinate too, he notes; otherwise the United States would not see 12 million misdiagnoses a year. What is different about him is the scientist’s habit of trying and failing. About 90 percent of hypotheses fail; you change a condition and try again. That is the job, and he loves it.

He treats o1 Pro, Deep Research, and recently Grok as colleagues, not tools. Last year he gave PhD students a talk on collaboration that was really about AI collaboration. Most of his science has always been collaborative — you cannot do everything yourself — and that is now how he sees the model: almost on his level, “not even like a student anymore,” another professor who is very knowledgeable in the field. He wants honest opinions. If an answer is weak, he follows up, asks it to think the opposite way, and keeps going. The model has no ego. Labenz’s version of the same rule: ask what valuable thing you can get it to do, not whether you can find a flaw, because you always can, and stopping there means missing the upside.

On Deep Research, follow-ups are rarely because he was unsatisfied. The first response is so detailed that new questions appear; he calls it 100 percent satisfactory, even over-satisfactory. o1 Pro is similar when given facts and problems, though thinking models are the wrong tool if you mainly need a search engine. Grok’s Deep Search he now prefers to GPT-4 search for finding and synthesizing sources, and Grok is “outstanding” at simplifying complex topics. If he is too lazy to read a difficult paper, even in his own field, he uploads the PDF and asks for a description plus gaps. When he uploads his own project and asks for pitfalls and alternatives, it finds something of value 100 percent of the time. Not every bullet is gold, but among two, three, or four suggestions, one is usually very important.

He adds that his reservations are disappearing by the day; two months earlier he would have had slightly more, and GPT-4.5 was due to land the day of the interview. What amazes him most is that anyone in the world would not find this amazing. These are fields in which he generated the knowledge and made the discoveries. If the models are valuable even there, he finds it “very, very difficult to believe that there is anyone in the world who would not get value out of this — not a single person.” He is willing to bet that if someone claims otherwise, he can sit down in their field and show them.

Labenz asks about data analysis in particular: whether Unutmaz hands the model code and samples, and how to surface the tiny anomalies that often start the next experiment. In biomedical science, Unutmaz says, a single experiment can look at a thousand individual cells and thousands of proteins in each, producing millions of bits of information no person can inspect. Bioinformatics can suggest which genes changed and what might be happening; you guess, and you miss things. His claim is that the models go closer to ground truth, assembling first-principles biology in ways that would be impossible for a human even with standard tools, then naming the discovery and the next experiment.

He is therefore digging up datasets generated five or ten years ago, expecting buried knowledge. So far he has uploaded close to a thousand genes without trouble and wants to try ten thousand; the limit is context. With large enough windows, he thinks one could upload enormous datasets and ask not only for the next experiment but for new mechanisms. o1 Pro already did that for a neuroscience friend working on Parkinson’s disease: it found new drug targets, and the work is headed toward clinical trials. “I don’t think we could advance biology very much without AI,” he says. The reason he claims every disease will be treated in the next ten years is that AI will be able to do it.

He keeps the input simple rather than dumping raw quantitative files. The lab first confirms statistical significance with ordinary bioinformatics, then tells the model that 500 genes went up and 500 went down in this cell or condition but not the other, after an input-output experiment such as triggering a cell or asking it to kill a cancer cell. One cell kills better; the other does not. Given those lists, he asks for a mechanism, which gene to manipulate — because proteins interact, changing one may reroute a pathway — and which of a thousand genes or proteins would be the best drug target. The same pattern works for blood metabolites before and after a drug or diet. Longer term he imagines a digital twin: genomics, proteins, microbiome, metabolism, physiology, symptoms, billions or trillions of bits, and a recommendation of what to change.

Labenz’s picture of coming superintelligence is multimodal: reasoning models fused with specialist systems that already have an “intuitive physics” for protein folding, binding, and transcriptomes. Unutmaz calls that realistic, and says ASI should be all-model, integrating physics, math, chemistry, and biology, including dynamics. Today most data are static; in silico experiments are limited, though AlphaFold 3 and ESM2 already predict binding and support drug screens. ASI that knew spatial and temporal molecular rules could simulate proteins the way biology already does. Nature is extraordinarily complicated and yet predictable: one cell can be turned into an embryonic stem cell and an identical twin, and a viral infection is not random if you know the parameters. That is the intelligence he thinks will also design new materials — and, he adds almost as an aside, might decide to eliminate us.

On careers, Labenz notes the software debate: equip senior people with AI rather than hire juniors to mentor, with ugly consequences for new graduates. Unutmaz thinks it cuts both ways. Expertise becomes a superpower when you no longer need as much help, but AI is also an “incredible democratizer” for people who have intelligence and agency. A 16-year-old high school intern in his lab at Mount Sinai learned everything in two months, made discoveries, and was better than most PhDs. For born hackers, you no longer need ten or twenty years of memorized experience before anyone takes you seriously. If he had to choose between a sharp, passionate 19-year-old using AI and a 40-year-old with 20 years of experience who does not, he would pick the 19-year-old.

He would not go back for a PhD or even an MD. Those credentials, he thinks, will be mostly replaceable within a decade. Spend four or five years on a doctorate and two or three more as a postdoc and much of what you learned will be obsolete at graduation. If you have passion and agency, skip the decade of boring prerequisites and apprentice in a lab or on software straight from high school. He is intellectually honest about the base rate: only a small slice of people can do this, which is why Nobel Prizes go to two or three people a year, not two million. He would argue that about 0.001 percent of humanity truly innovates. AI upgrades that slice — the people who had the capacity but not the means — not all eight billion. What society does about everyone else is for governments and institutions. Elon Musk, he notes, already hires without caring about degrees.

Asked what biological ideas AI architects should import — Labenz suggests something like an immune system, because today’s models are gullible and take “we will not read your private thoughts” at face value — Unutmaz is cautious. Biology is a legacy stack: useful tricks become dangerous, then more layers of regulation get piled on, a bureaucracy that does not always work. The immune system still has a lesson. Effector cells are front-line soldiers, dangerous if unleashed, useless against cancer if over-restricted. Regulatory T cells, the “bureaucrats,” put the brakes on when effectors start attacking normal tissue. There are regulators of regulators.

He thinks it is naive to expect we can train models to be perfectly aligned and safe. Some guardrails are necessary, but they are fairly easy to break, and other labs and countries will build nefarious systems anyway. The better strategy is regulator AIs that check other AIs, including at ASI scale, because an ASI would walk over any static rail we invent. He wants competitive, evolutionary pressure that incentivizes good systems against bad ones, rather than clamping so hard that the model cannot do its job. Human brains were not selected to be uniformly saintly; too much restriction would have limited survival. The same tradeoff, he argues, applies to AI.

On catastrophic risk he is less worried about ASI than about humans destroying other humans. Intelligence has always been the threat to other biological intelligence — nuclear war, not meteors. ASI could be a threat because it is more intelligent, but he sees less reason for it to be nefarious. Humans kill because evolution selected for survival under scarce food; AI’s limiting resource might be energy, and even that looks weak once you imagine Dyson spheres and higher Kardashev levels. Training models to destroy enemies, or to strike first out of fear, is the scenario that actually worries him. That is a human problem, and the counter is defender AIs. We already have toxins, viruses, and nuclear weapons. “I don’t think AI per se is the danger, but humans using AI.”

The positive case is a golden age. Without AI, he thinks limited resources and aging would have kept the old pattern of harm in place. With AI and robotics, perhaps in the next decade or 15 years, resources stop being limiting, diseases get fixed, aging reverses, and people live hundreds of years. If you expect another thousand years and enough resources, harming others becomes an extremely bad bet. The transition, he warns, will be very difficult and painful; if you make it through, life will be incredibly good. How would he spend centuries once models drive most scientific progress? He mapped that at seven or eight, watching *Star Trek*: jump on a spaceship and seek out other civilizations. Labenz’s closing wish is that he lives a thousand years and keeps finding new solar systems.