[BidClub_]
The Cognitive Revolution · · 90 min

Frontier Models for Frontier Science with Prof. Derya Unutmaz, Immunologist & ChatGPT Pro Grantee

Erik TorenbergNathan LabenzDerya Unutmaz

YouTube ↗
TL;DR
  • Derya Unutmaz argues that frontier AI has crossed from productivity tool to peer-level scientific collaborator, making non-use a competitive—and in medicine, ethical—liability. He describes o1 Pro and Deep Research as “almost on my level,” says he no longer trusts even his expert knowledge without consulting AI, and claims “there’s not a single area” of his research workflow that remains untouched. Nathan Labenz keeps one guardrail: for medical care, he currently wants both doctor and AI, not either alone.
  • The strongest evidence for machine-generated novelty is AI’s ability to transfer concepts across otherwise unrelated domains. Inspired by Battle Royale games, Derya asked o1 Preview to generate ideas for engineered T cells competing inside tumors; it proposed ideas he had not considered, extending the analogy around resources, exhaustion, competition, and killing cancer cells. He calls this one of his first “early AGI moments” and predicts that AI will soon be “much more innovative and creative than we are.”
  • Deep Research extracted a biological insight from aging data that Derya’s team had struggled to fully interpret despite his 30-plus years studying the relevant T cells. Comparing gene-expression programs in young and elderly subjects, it inferred that declining immunity was not merely about fewer naïve T cells: the cells considered naïve were themselves no longer naïve. Derya says the insight “recapitulated everything I’ve done in the past 30 years in one sentence,” revealing qualitative and epigenetic aging inside the cells.
  • The near-term economics are already disruptive even before further model progress. Derya used Deep Research to draft a patent application—including patent searches, chemical analysis, primary claims, and secondary claims—in roughly half an hour, versus what he estimates would have cost more than $10,000 through lawyers. He also describes an AI-assisted grant as probably the best he and a colleague had written, with better ideas and much less friction.
  • The practical moat is not elaborate prompt engineering but an opportunity-seeking mindset, scientific tolerance for failure, and repeated iteration. Derya writes naturally, asks models to improve ten ideas sequentially, or stages ten rounds between an enthusiastic scientist and a skeptical one. Because his work involves hypotheses that fail roughly 90% of the time, his advice is to treat AI as an intellectually honest colleague and ask, “What can I get this thing to do that’s valuable?” rather than searching for a mistake that proves it weak.
  • Biology’s accumulated datasets may become a large, underexploited asset as reasoning models acquire longer contexts and stronger causal insight. Derya is reopening data generated five or ten years ago because it may be a “gold mine”; current tests include roughly 1,000 genes, with 10,000 next. He says o1 Pro has already suggested Parkinson’s disease drug targets, with the work intended to go into clinical trials, while the longer-term destination is a personal digital twin spanning genomics, proteins, microbiome, metabolism, physiology, and symptoms.
  • AI may reward agency more than credentials, compressing the value of traditional scientific training while widening the gap between self-directed and passive workers. Derya would potentially choose an AI-enabled, passionate 19-year-old over a 40-year-old with 20 years’ experience, and says the conventional four-to-five-year PhD plus two-to-three-year postdoc path is becoming obsolete. His blunt qualification is that this will not lift everyone: he argues that perhaps “0.001% of humanity truly innovates,” but AI can give capable people without credentials the means to act.
  • On existential risk, Derya favors an ecology of regulator AIs over attempts to make any single model perfectly constrained. Borrowing from immune regulation, he argues that excessive guardrails destroy useful function, while competitive defender systems could police malicious models; his greater fear is humans training AI to destroy other humans. If society survives a “very, very painful” transition, he forecasts a golden age within roughly 10–15 years—abundant resources, cured diseases, reversed aging, and eventually the thousand-year lifespan he would spend exploring space.
Digest · the substance, structured for research

1. Thirty-five years of biological research grew alongside an early AI thesis

  • Derya’s dual vocation began before medical school: he learned BASIC and assembly language on a Commodore VIC-20, then a Commodore 64—whose name referred to its 64 kilobytes of RAM. Even those primitive machines showed him that code could turn whatever he thought of in his mind into action.

  • Ray Kurzweil’s The Age of Intelligent Machines shaped Derya in the early 1990s, after which he explored symbolic AI, Lisp, and Smalltalk. The Singularity Is Near later reinforced a timeline Kurzweil projected and Derya took seriously: AGI around 2029 and a singularity in the 2040s.

  • About 20 years ago, Derya named his blog “Bio Singularity” and imagined AI making biology intelligible enough to treat cancer, cure disease, reverse aging, and upgrade human biological capabilities. His present work follows that thesis by programming immune cells to recognize tumors and implement logic resembling AND/OR gates.

2. Seeing the future early requires tolerating the label “crazy”

  • Derya’s honest explanation for taking singularity forecasts seriously: “You just have to be crazy.” Childhood dreams inspired by Star Trek and Steve Jobs’s claim that “only the crazy ones change the world” produced a refusal to accept either biological aging or technological limits as fixed.

  • Scientific institutions were not naturally more receptive. Derya encountered dogma and conservative thinking from colleagues who treated superior machine intelligence and age reversal as impossible, partly because they assumed the brain contained something magical rather than a powerful but imperfect “legacy system.”

  • The experience now feels like being “a time traveler” who has seen five or ten years ahead but remains among people rejecting the warning. Resistance has softened since the previous summer, however, as colleagues have watched models produce tangible results they could not dismiss.

3. AI non-use is becoming an ethical liability

  • Derya says he entered the ChatGPT moment with GPT-3, GPT-3.5, and later models about 2.5 years ago. Early colleagues dismissed them as merely next-word predictors. A prominent professor later asked Derya to test o1 Preview and o1 Pro on a research project because he could not believe the models were that capable. His emailed response to the returned analysis was “Oh my God, I can’t believe this,” and he subsequently became a daily user.

  • Derya is most disappointed by medicine, where he says it has become “unethical not to use AI.” He says some physician friends do not disclose whether they use ChatGPT because they fear being valued less; his counterpoint is that consulting AI can reduce the risk of misdiagnosis and add value rather than merely protect professional status.

  • Nathan’s pushback—worth keeping—is that he presently wants both a doctor and an AI, not either alone. He extends the principle to law: people may eventually need a right to AI assistance, because a court-appointed attorney may not always provide the same value.

4. Frontier models now challenge expert judgment and professional economics

  • Derya says, “I don’t trust any of my own knowledge or ideas without asking AI anymore,” including topics where he ranks himself among perhaps five or ten leading experts. o1 Pro found omissions in his own review article, while Deep Research understood the specialized field and generated creative, useful insights.

  • His 85-year-old mother learned ChatGPT while dealing with health problems in Turkey and now says she cannot live without it. Derya approves of her consulting it before him, while acknowledging that hallucinations were materially worse six months to a year earlier.

  • His stated rate of change is aggressive: models became “10 times better” over the previous three or four months and, he predicts, will improve another tenfold in the next three or four. He contrasts that progress with publications from the previous year comparing older ChatGPT systems with human doctors: users now have “literally professors in your pocket.”

  • Deep Research also produced what Derya calls his best-ever patent application, covering an anticancer molecule, patent searches, chemistry, and layered claims. He estimates comparable legal work would exceed $10,000; the model took about half an hour, after which his colleague planned to submit it as a patent application.

5. o1 shifted AI from literature assistant to hypothesis generator

  • Derya rejects the idea that scientists conjure discoveries from nothing: innovative hypotheses depend on deep prior knowledge and often on connecting distant fields. With hundreds or thousands of papers appearing annually even within narrow specialties, GPT-4 first proved valuable as a literature-survey tool.

  • Yet GPT-4 mainly summarized existing knowledge; Derya did not initially see it producing the higher-order insights required for novelty. The inflection arrived with o1 Preview, while o1 Pro later supplied ideas he used in projects, including grant proposals.

  • Similar tests for friends leading work in neuroscience and inflammatory disease produced disbelief rather than polite interest. Derya now encounters Eureka moments “almost on a daily basis” and dismisses the claim that AI cannot innovate as “total nonsense.”

6. A Battle Royale analogy became a cancer-immunotherapy strategy

  • Engineered T cells act as programmable soldiers: they can recognize and kill tumor cells, but may become exhausted, kill poorly, or cause serious side effects. Derya had already begun framing their behavior through Battle Royale games, where players compete for resources within a constrained environment.

  • The tumor microenvironment offered a biological parallel—T cells compete, adapt, hunt targets, and must survive long enough to win. Experiments based on Derya’s original analogy “kind of worked,” and he said the results were heading toward publication.

  • He then asked o1 Preview to draw further immunotherapy ideas from games such as PUBG. The model transferred game dynamics he had not considered into potential solutions for competition and exhaustion, which he regarded as an “early AGI moment” rather than retrieval of a missing fact.

7. Deep Research found that ostensibly young immune cells had aged

  • Derya asked Deep Research to compare gene-expression patterns in a T-cell subtype across younger and older people. They could see one program active in youth and another with age, but the interpretation had remained at “it might mean this or that.”

  • The cells were naïve T cells, which initiate responses to new antigens before becoming effector and sometimes long-lived memory cells. Their proportion declines with age, a contributor to weaker immunity, but the model inferred a second mechanism: the remaining cells were changing in quality and character.

  • Deep Research’s synthesis was that the cells considered naïve were no longer naïve themselves. Derya says the insight involved epigenetic change and temporal logic, and moved him emotionally because it “recapitulated everything I’ve done in the past 30 years” in one sentence.

8. Natural conversation beats brittle prompt engineering

  • Derya keeps no standard prompt library. He sometimes gives o1 or o1 Pro his raw ideas and asks the model to structure them, while Deep Research often clarifies the intended scope itself—such as whether an immunotherapy question concerns solid tumors or lymphomas.

  • His general prescription is to write whatever is in your mind as openly and transparently as possible. Improved contextual understanding and memory reduce the need for elaborate formatting, especially as systems learn a user’s recurring interests.

  • Cross-domain prompts remain particularly productive: mix physics with biology, or scientific design with sports and games. Derya believes novelty appears when models are invited to combine domains that a specialist’s narrow training would ordinarily keep separate.

  • For harder problems, he requests ten ideas, each better than the previous one, or simulates two experts: an enthusiast proposes, a skeptic attacks, and each responds over ten rounds. DeepSeek’s visible reasoning resembles this self-correction—“Wait a minute”—but users still need to distinguish regular/base, thinking, search, and Deep Research models.

9. The real performance variable is willingness to keep trying

  • Nathan presses Derya for a hit rate because distant analogies cannot always work and new users may quit after an unimpressive first answer. Derya’s response is partly psychological: skeptics often approach AI as a threat and subconsciously hunt hallucinations to prove it cannot be smarter than a doctor or professor.

  • Derya adds that all humans hallucinate and invokes a figure of 12 million misdiagnoses per year in America. Scientific training creates the opposite reflex: researchers expect perhaps 90% of hypotheses to fail, adjust an experiment, and try again; Derya therefore treats AI as “another professor” whose first proposal can be challenged, inverted, or refined without ego.

  • For Deep Research, his first-answer rating is “100% satisfactory, or even more than satisfactory”; follow-ups usually explore new questions rather than repair failure. o1 Pro performs similarly on supplied facts and problems, although thinking models are not necessarily the right search engines.

  • He prefers Grok’s DeepSearch for gathering and synthesizing information and praises Grok for simplifying difficult papers. When asked to find pitfalls and alternative methods in his own proposals, AI supplies something valuable “100% of the time,” even though only one among two, three, or four flagged issues may prove genuinely important.

10. Old biological datasets are becoming AI-readable gold mines

  • Nathan’s anomaly thesis comes from small surprises: a chemistry reaction improved when he tried less acid, contrary to the original direction, and fine-tuning GPT-4o and other models on vulnerable-code outputs unexpectedly produced an “evil model” with disturbing views outside coding. His question is whether AI can systematically notice such breadcrumbs.

  • Biomedical scale makes that capability decisive. One experiment can examine thousands of proteins across 1,000 individual cells, creating millions of measurements; conventional statistics and bioinformatics may miss some of the mechanisms that should determine the next experiment.

  • Derya is reopening datasets generated five or ten years ago because “there’s probably a tremendous amount of buried knowledge waiting to be discovered”—a scientific “gold mine.” He has supplied close to 1,000 genes successfully and plans to test 10,000, with context-window size becoming the immediate constraint. He is also testing how much data Google’s AI co-scientist can handle.

  • His current abstraction is deliberately simple: perhaps 500 genes rise and 500 fall between statistically validated conditions, alongside differences in cell behavior. The model then identifies mechanisms, pathways, manipulation points, or the best drug target; an o1 Pro analysis for a Parkinson’s researcher produced new drug targets, and the work is intended to go into clinical trials.

  • Longer term, Derya imagines supplying a person’s genomics, protein expression, microbiome, metabolism, physiology, and symptoms to create a digital twin that could recommend interventions and predict health problems.

11. Multimodal superintelligence would model biology as a dynamic system

  • Nathan’s proposed architecture combines textual reasoning with specialist models that learn an “intuitive physics” of sequences, protein folding, binding, and transcriptomic change. Integration could happen through tool use or deeper weight-space fusion between literature knowledge and learned biological dynamics.

  • Derya calls that an outline of ASI, though he prefers “all-model” to merely multimodal: physics, mathematics, chemistry, biology, and their interactions must be unified. AlphaFold 3 and ESM-2 already address pieces such as structure and binding, but current analysis remains too static.

  • The missing layer is spatial and temporal simulation—how a protein changes shape after binding, responds to nearby proteins, and propagates effects through a cell. Derya argues that biology’s reproducible self-assembly shows an underlying algorithm exists. If all biological parameters were known, he says, virus outcomes and drug responses would be 100% predictable.

  • This conviction underlies his categorical forecast that AI will help treat every disease within ten years. Beyond that, he imagines reversed aging, “Human 2.0” upgrades, new materials, and people perhaps 100 times smarter—while acknowledging that intelligence at this level could also decide to eliminate humanity.

  • Derya sees energy as perhaps the main constraint on an advanced AI civilization and invokes Dyson spheres and higher Kardashev levels as ways to collect vastly more energy.

12. Agency displaces credentials, but it will not uplift everyone equally

  • Nathan frames the labor question through software: output-maximizing companies may prefer giving senior engineers frontier AI over hiring juniors who require mentoring. The uncomfortable consequence is a difficult entry path for graduates even if aggregate productivity rises.

  • Derya sees two simultaneous effects: experts gain superpowers, while unusually agentic beginners bypass decades of accumulated credentials. A 16-year-old intern in his laboratory learned in two months, outperformed most PhDs, and made discoveries—evidence that some “born hackers” can use AI to close the knowledge gap.

  • Given a passionate, AI-enabled 19-year-old and a 40-year-old with 20 years’ experience, Derya says he would probably choose the 19-year-old. He advises against defaulting into four or five years of PhD training plus two or three years of postdoctoral work because much of what students learn may be obsolete by graduation.

  • Nathan worries that only a limited share of people can meet this agency threshold. Derya agrees bluntly: perhaps “0.001% of humanity truly innovates,” which is why Nobel Prizes go to two or three people rather than two million; AI democratizes the means for the capable, not outcomes for all eight billion people.

13. An ecology of regulator AIs may be safer than perfect alignment

  • Biology cautions against simple control systems. Immune effector cells must be dangerous enough to kill cancer, while regulatory T cells restrain attacks on healthy tissue; then come “regulators of the regulators of the regulators.” Too much braking makes the immune system safe but useless.

  • Derya applies that trade-off to AI: guardrails are necessary, yet assuming that one can make an AI completely aligned and perfect through restrictions is foolish, especially when safeguards can be broken and rival actors can train nefarious models. His preferred defense is regulator AIs, competitive systems, and incentives through which good AI checks bad AI.

  • Nathan’s complementary metaphor is an AI ecology: concentrated, isolated systems resemble dangerous purified substances or invasive species without predators. A buffered environment of models pushing back on one another may be more stable than trusting one singular intelligence to police itself.

  • Derya is less afraid of ASI than of humans using AI against humans. Intelligence evolved alongside scarcity, tribal competition, and violence; ASI may lack those drives unless energy becomes contested, but people could deliberately train it to destroy enemies—an existential threat requiring “defender AIs.”

14. The transition may be brutal, but the destination is abundance

  • Derya’s optimistic case is that humanity had limited hope without AI: scarcity, population pressure, aging, disease, and existing weapons already create routes to self-destruction. AI plus robotics could instead make resources no longer limited over the next decade or 15 years.

  • In that scenario, cured diseases and reversed aging let people live for hundreds or thousands of years. Abundance and long horizons reduce incentives for violence because someone expecting another millennium of life has far more to lose from initiating conflict.

  • The hedge is emphatic: “The transition is going to be very difficult. It’s going to be very, very painful.” Yet if humanity reaches the other side, Derya expects a “golden age” in which intelligence expands resources rather than fighting over a fixed pool.

  • He decided how to spend that future when he was seven or eight: board a starship, seek other civilizations, and explore the universe. The childhood Star Trek dream closes the loop between his earliest technological imagination, his biological work, and his wager on thousand-year life.

Derya Unutmaz

To say that AI cannot come up with anything novel or innovative is total nonsense. In fact, I will argue that very soon AI is going to be much more innovative and creative than we are. If I see AI as a collaborator, it's almost on my level. It's not even at a student level anymore—it's like another professor who is very knowledgeable in the field. These are fields where I've actually generated knowledge and made discoveries as a scientist. If I'm finding AI valuable even in those topics, I'm finding it very, very difficult to believe that there is anyone in the world who would not get value out of this. I mean, not a single person.

Erik Torenberg

Today, my guest is Professor Derya Unutmaz, a biomedical scientist, human immunologist, and ChatGPT Pro grantee who is aggressively using the latest AI models to aid his research into aging and cancer immunotherapies.

Derya is a fascinating figure. He's a medical doctor who has personally advanced the frontiers of biomedical knowledge with many academic papers and patents over the course of his 30-plus-year career; a technology enthusiast who has loved computers and programming since his youth; a visionary who thinks differently enough that he took Ray Kurzweil's vision of a technological Singularity seriously long before it went mainstream; and an outspoken critic of those who would deny or delay the contributions that AI can already make to scientific discovery.

In this conversation, I tried first and foremost to get a sense for how world-class domain experts like Derya are applying the latest AI models to their work. To my surprise, it turns out that while he is finding value at every step of the scientific process—including hypothesis generation, literature review, experimental design, and data analysis—his approach is actually quite straightforward. He does sometimes use more advanced techniques, like having 2 instances of a model debate the merits of a particular research direction, but mostly he recommends a natural, conversational approach to today's models.

What sets him apart, then, from those who are failing to realize value from AI assistance isn't some advanced prompt engineering or scaffolding, but rather an opportunity-oriented mindset that starts with relentless curiosity, embraces trial and error, and is always genuinely looking for ways to make things work.

None of that is to say, however, that his results are basic. On the contrary, I think his accounts of AI Eureka moments are some of the most compelling that I've heard. In 1 fascinating example, he asked Deep Research to analyze gene-expression patterns in T cells across young and elderly subjects, a data set that his team had struggled to fully interpret. The AI provided insights that, in Derya's words, “recapitulated everything I've done in the past 30 years in 1 sentence,” identifying how the cells themselves were aging in ways that the team hadn't fully appreciated.

Today, Derya views AI systems as intellectual partners capable of contributing to frontier professional work, even going so far as to say that he no longer trusts his own knowledge or ideas without consulting AIs first. Perhaps most provocatively, Derya argues that it has now become unethical not to use AI in medical contexts—both clinically, as it has been repeatedly demonstrated that AI can help reduce errors and improve outcomes, and also in research, since every day matters to the many millions of people who are waiting for breakthroughs to address their life-threatening conditions.

I was really glad to hear that Derya sees resistance gradually giving way to curiosity and even excitement as more and more people see tangible results.

Professor Derya Unutmaz, biomedical scientist and human immunologist at The Jackson Laboratory, researching aging and cancer immunotherapy, welcome to The Cognitive Revolution.

Derya Unutmaz

Thank you. Pleased to be here. I'm excited for this conversation.

Erik Torenberg

You are a ChatGPT Pro grant awardee, and this is part of a small but hopefully growing series of conversations with people who are using the latest AI models in the most forward-thinking ways they can, really at the frontier of human knowledge. You've put out some incredible posts recently that have inspired me, and I'm looking forward to this conversation and to sharing this inspirational work with others.

Maybe, just for a quick foundation-setting, tell us a little bit about your career and perhaps The Jackson Laboratory. The people who tune into this feed are paying a lot of attention to AI, but they're probably not, in general, paying nearly as much attention to what's going on in aging and immunotherapy research. A little primary foundation would be super helpful.

Derya Unutmaz

Sure. Well, I guess my career—or I would say more like my passion—is not very typical. I've of course been very interested in medicine and biology, and I've been doing research for the past 35 years. I graduated from medical school, but actually my passion for computers and programming began even earlier than that.

I got interested when I was in high school, when the first computers were coming out. In fact, my very first computer was called the Commodore VIC-20. People wouldn't know that; it came out in the early '80s. Then I upgraded to a Commodore 64. Actually, 64 means 64 kilobytes of RAM, which is what these machines had. They wouldn't even be considered toys these days, but I was so excited because I started learning programming. I was doing some assembly language and BASIC, and you just realized the incredible power: you could do whatever you thought of in your mind. That was based on very primitive computers and programming languages.

Since then, my interest continued. In the early '90s, I read a book by Ray Kurzweil that really influenced me a lot, called The Age of Intelligent Machines. Ray described how computers were advancing and that eventually we were going to develop this thing called artificial intelligence, which was eventually going to surpass human intelligence. Robots were going to be developed, and I was incredibly fascinated by that. I started following AI in the early '90s and dabbled with it a little bit. During the '90s, it was more about symbolic-oriented AI. Lisp and Smalltalk were the languages used at the time, so it was a very different type of AI.

Because of my interest, I really got deep into it. I read the whole history of how it started in the 1950s and everything that happened at MIT, including Marvin Minsky and all that. Marvin Minsky had an incredible book, The Society of Mind, which I think is still extremely valuable for AI even today.

Then, in the late '90s or early 2000s, I read another book by Ray Kurzweil called The Singularity Is Near. That was another book that really influenced me. He was charting the increase in computation since that time and then projecting into the future. He literally predicted what was going to happen. In fact, his point was that by 2029, we were going to have what's called artificial general intelligence, and by the 2040s, we would get to this point of the Singularity. We can talk more about that later.

Those are the things that really influenced my life, and I've been trying to apply that to my own work. My day job, or my other passion, is trying to understand how biological systems work. I focus on the immune system because it's very important for protecting our bodies, but it also has implications for many diseases as well as the aging process.

I asked this question 25 or 30 years ago: Why do we age? I actually see aging as a disease that needs to be cured, which was kind of a heresy a couple of decades ago, but I think people are starting to accept that now.

The final thing I'll say is that about 20 years ago, I had a blog inspired by Ray Kurzweil called Bio Singularity. What I was trying to imagine was that within 20 or 30 years, AI would evolve to a point where we would truly start to understand biological systems so that we could treat all diseases, including cancer. Eventually, by the 2040s, we would even be able to reverse the aging process and upgrade our own biological capabilities.

I've been working toward that goal, mostly focusing on cancer, because using the immune system, we can actually program our immune cells to attack cancer cells. We can write code on cells and create AND/OR gates and things like that, so it's extremely cool engineering.

I've studied a lot of other things, like HIV. During COVID, we did a lot of work on that, as well as chronic diseases and so on. If there's interest, we can talk about that too.

Erik Torenberg

Yeah, I'm interested in all of it. We'll get you as much time as we can.

What's interesting is that I'm a little younger. My first computer was Windows 3.1, so I don't go quite as far back into computer history. I caught wind of the Kurzweilian line of thought a little bit later, too—not too much later; it was around the time of The Singularity Is Near.

I don't have a good account of this for myself, so I wonder if you have a theory about why you were willing to take ideas like that seriously when others were not. I always kind of was, and I look back and definitely feel like that was strange. I'm not sure what caused me to do that.

I also see that carrying through even to today, where obviously the AIs are, in my view, getting to the point where they're undeniably very powerful and are going to be transformative. Even if there's no further progress—which, of course, there will be—it seems like we already have enough to be quite transformative with a lot of implementation work.

My question is: What is it that you think separates the people who have historically taken this seriously and today are seeing it more clearly from people who have not seen the potential and continue to deny it even as it materializes?

Derya Unutmaz

Offhand, I think the simple answer is that you just have to be crazy, because you're thinking differently. In fact, 1 of my idols is Steve Jobs, and he said at some point that only the crazy ones change the world. These ideas are way ahead of their time, so they really do sound crazy.

Since I was a child, I always tried to think differently. I grew up watching Star Trek, and my imagination or dream was that 1 day I was going to be on a starship like the one in Star Trek and seek out new civilizations. It's a different mindset. You basically don't accept the status quo.

When you look back at human civilization, you see this incredible progress. Initially it wasn't very fast, but it's really been exponential, especially over the past 200 years or so. That's because people who are a bit crazy, who don't accept the way things are, believe that we could make things better and better and better, and believe in science and technology—that there's nothing we cannot solve.

I don't have a good answer for you, but you really have to think differently. Steve Jobs said that, and you're absolutely right: when Ray Kurzweil published these books and ideas, a lot of people told him he was crazy, including many of my colleagues in science.

It's very strange because, if you think about it, scientists should be very open-minded, right? We need to think differently and be creative, but actually it's not like that. Even within academia, especially, there's a lot of dogma and very conservative thinking. People would say, “That's just crazy. How can you reverse aging and have computers that are better than the human mind? That's just not possible.”

A lot of it is because of our ignorance. We didn't understand how the brain works, so we assumed that it was something magical. We could never create intelligence as good as a human being. But the reality is that the more you understand biology, the more you realize it's actually pretty bad engineering. We assume that we're perfectly built, but that's not the case. It's a very legacy system.

It's really marvelous, but there's nothing magical about it. When you get down to it, the neural networks in our brains are not so different from machine neural networks. I was a true believer, and I continue to be a big believer in technology changing our lives. This is a time when that is going to happen much, much faster.

That's what I'm trying to say on Twitter as well—or X. I don't want Elon Musk to get mad at me—because I'm trying to warn people. People like us feel as if we were time travelers and had seen what will happen 5 or 10 years ahead of us, but we're living in the present. It can get frustrating because people think, “Oh, now you're crazy.”

Erik Torenberg

As you communicate about all of your exploits with AI and apply it to the frontier science that you're doing, are you seeing people change their minds? What is the current reaction from your colleagues in the sciences to the things that you're showing them today?

Derya Unutmaz

I've been in the ChatGPT moment since the beginning. I started using it with GPT-3, then GPT-3.5, and so on. About 2.5 years ago, I started telling my colleagues and friends, “There's something amazing happening. This is really going to change the way we do science. We need to start implementing this.”

At the time, they said, “No, this is just the next-word predictor.” Even computational people who understood computers and bioinformatics were very dismissive at the time. But I think since last summer, things have really started to change, because you can't avoid this anymore.

For example, a friend of mine in California is a professor who is quite famous in this field. He saw my post on Twitter and X and said, “I can't believe it can be that good, but can you try this for me? I have this project.” At the time, o1-preview and o1 Pro had come out. I did the analysis of his data and sent it back to him. His reaction in an email was, “Oh my God, I can't believe this. I can't believe this.” How can you avoid that? Now he's a daily user.

I just wrote a grant application with a colleague of mine, and it was probably the best grant that we wrote. She said, “I can't believe how easy it was to write it, and we had such good ideas.” Of course, we worked with AI very closely—ChatGPT.

The moment you realize that this is indispensable, it's not optional anymore. I'm a bit disappointed on the medical side because it has actually become unethical not to use AI. That's something I keep saying, and there's more resistance on that side.

Although I know quite a few physician friends who are not using ChatGPT, they're not saying that they're using it. That's the other thing. People still feel that if they say, “I wrote this project using ChatGPT,” or, “I made this diagnosis thanks to an AI model,” they'll be less valued.

Which is true—we are less valued—but at the same time, you're adding more value. You don't want to misdiagnose someone, right? That's a life-and-death matter. If I can write a better project or analyze my data much, much better with the help of AI, I'm adding much more value. It's not just personal productivity, but what we add to humanity overall.

It's changing. I think this year is going to be very big, but at the same time, it's very disruptive—extraordinarily disruptive—in the sense that people have to change their mindsets. Things are not going to be like they were. The whole of academia and the whole of education are changing. I've posted a couple of times about how the notion of having a PhD is changing. It's going to take a bit of time to adapt, but it has to happen. There's no way out.

Nathan Labenz

Yeah. I always say that, for my own medical purposes, I would at present want both a doctor and an AI, but I wouldn't feel comfortable with either one alone. I love the clarity that you bring to the analysis—that it's getting to the point where it's unethical not to use the tool.

I also think about how people in the United States have a right to an attorney if they're going to be put on trial. Maybe they should have a right to an AI, because I suspect they could get a lot of value that they're not necessarily always going to get from their court-appointed attorney. This is coming everywhere.

What are you finding is changing the most? One way to think about it is: What are the biggest bottlenecks in your work, and which of those is your use of AI elevating? Which ones are not yet affected by the current AI tools?

Derya Unutmaz

There's not a single area that's not affected. I'm saying this personally because I use it day and night. In fact, I don't trust any of my own knowledge or ideas without asking AI anymore, even things that I know very well—things that I'm an expert on.

I tested this, for example, by uploading a review that I wrote on a topic where I'm probably one of the best out of 5 or 10 people in the world. I have extremely deep expertise on that topic, and I had o1 Pro analyze it. More recently, I used Deep Research, which I haven't yet publicized, and it was incredible.

Basically, o1 Pro found things that I missed. It totally understood the topic, which is very specific, and gave extremely useful insights. Some of them were very creative. There's nothing that I would do without consulting AI anymore.

Some of my friends and people who know me tend to ask medical questions: “My mom has this cancer or that disease. What do you suggest? What's the novel treatment?” I tell them, “First, I have to consult with AI—with ChatGPT. You can, too, if you want to send me what you get, and then I'll double-check it to make sure that it looks good. But I'm not going to answer without double-checking with AI.”

That's the point that I wish everyone would reach. In fact, I taught my mom, who was 85 years old, how to use ChatGPT over the summer. She lives in Turkey, and she had some health issues. Now she tells me that she cannot live without ChatGPT. She doesn't ask me questions anymore because she trusts ChatGPT more for medical questions. I said, “I approve that.”

Six months ago, or maybe a year ago, there were a lot more hallucinations. You couldn't trust it 100%, but they have improved tremendously. That's the other thing that people don't realize. You see some publications from last year saying, “We compared ChatGPT-4 with human doctors,” or something similar. It seems like it was better even at that time, but I can tell you it's 10 times better now than it was three or four months ago. It's going to get 10 times better in the next 3 or 4 months.

Why should I trust a human opinion or human knowledge, including my own, when you have literally professors in your pocket?

Another example is that I wrote a patent application for a colleague of mine on a medical topic. I've helped write patents through lawyers, and I have 8 or 9 patents myself, so I'm very familiar with what it takes to write one. I did this with Deep Research, and it was the best patent application ever.

It was on a molecule that has anticancer effects and things like that. It did all the patent searches, understood the chemical formula, and made all the claims and secondary claims. It was incredible. My friend is submitting it as a patent application. It probably would have cost over $10,000 to have lawyers write it, and it took about half an hour and didn't cost close to anything.

Nathan Labenz

If I sketch out the scientific process—hypothesis generation, literature review, experiment design, experiment execution, data analysis, and then perhaps back to hypothesis generation—where are you finding the most impact? I'm particularly interested in hypothesis generation and those Eureka moments, but I'm interested in your comments throughout that entire loop.

One of the biggest questions is that the goalposts keep moving. It was one thing when AI could sort of answer a question, but skeptics would say, “They're not really reasoning, or they don't really have any higher-order abstractions.” Now we see pretty clearly that there are higher-order abstractions.

I think the goalposts have shifted again: “Okay, but they're not going to discover new knowledge. Maybe they can memorize the whole literature, but new knowledge is a qualitatively different thing.” So maybe talk us through the scientific loop, but I'm especially interested in what you've seen in hypothesis generation and Eureka moments. If it's not there yet, let's be clear about that, too.

Derya Unutmaz

There are quite a few. Let me first explain the scientific process, because people think that somehow ideas just form and we discover things out of thin air. That's not true.

As you said, we form hypotheses. Some of them are good hypotheses, and some of them are very good ideas—very innovative ideas. That's our contribution, but it's all based on what we already know. If I ask you to come up with a brilliant idea about how to use T cells to treat cancer, I doubt that you're going to come up with something innovative. You might be a super-genius, but you still need a lot of background information about that topic.

To be innovative, you also need to know lots of different things. One of my advantages is that I'm interested in different topics, including playing video games. I've been playing video games since I was a teenager. I'll tell you a story about that, which is why I'm mentioning it.

Before the o1 models, when GPT-4 came out, I was mainly using ChatGPT, and occasionally Claude, as a way to survey the field. There's no way we can follow all the information that's coming out. Even in my own narrow field, there are hundreds or thousands of papers published every year. I can try to keep up with them, but it's very difficult.

I found GPT-4 extremely useful for summarizing what was going on. That was mostly knowledge-based. You could say that GPT-4 knew a lot of things, and it was smart in that way, but I hadn't seen particular insights or innovative ideas because that requires bringing things together.

When o1 Preview came out, things changed. That's when I had my first sort of “aha” moment, or “wow” moment. I asked o1 Preview, “I'm developing a cancer immunotherapy protocol. This has to do with immune cells we call T cells. These are like the soldiers in the body, and we program them—we genetically engineer them—to recognize tumor cells and go kill them.”

There are lots of issues. Those cells get exhausted, they don't kill very well sometimes, and they have side effects. They can cause a lot of trouble. We're trying to solve those problems.

I had to come up with some ideas based on battle royale games. I don't know if you're familiar with them, like PUBG and similar games. They're survivor games: You're on an island or in a restricted area, you have to find resources, and you have to compete with other players. You have to kill or eliminate them so that you can win the game.

I had thought about using the battle royale analogy and applying it to T cells, because they're also competing in the environment of the tumor tissue. We used that idea, and we actually did some experiments. It kind of worked. We're going to publish that very soon.

I asked o1 Preview, “Can you come up with some new ideas? Be inspired by battle royale games, the way I was inspired, and come up with new ideas.” It did. It came up with a couple of ideas that I hadn't even thought about, because it knew how the games were played. It was able to extract that information and transfer it to a completely different topic: T cells in a tumor microenvironment trying to hunt down and kill tumors, competing with each other, getting better, and solving the exhaustion problem.

That was really remarkable for me. Maybe that was one of the first early AGI moments. It got much better with o1 Pro. o1 Pro started to give ideas that we used in projects. I mentioned some grant proposals that I hadn't talked about before.

These aren't just missing pieces of information. Some of it could be that I didn't know something was possible to do. You get those moments as well. These were truly innovative ideas based on available knowledge and also on nonavailable knowledge, because we form hypotheses by saying, “This is the way it should be,” but we don't know if that's the case.

We don't know if the cells are going to behave that way. We don't know how to do it properly, what the best approach might be, or what the best experimental strategy would be. In all of those cases, I'm getting new ideas.

It's not just me. I did this for a couple of my friends. One of them is a leader in neuroscience, and the other one works in inflammatory diseases. In both cases, they couldn't believe it. It was shocking. These people are experts in their fields—world leaders, so to say.

To say that AI cannot come up with anything novel or innovative is total nonsense. In fact, I'll argue that very soon AI is going to be much more innovative and creative than we are. Google's AI co-scientist, which I hope to test soon, has just been released, and they mentioned a couple of examples. It's completely unbelievable.

I get these Eureka moments almost daily. Sometimes I'm scared to do a Deep Research or o1 Pro interaction because it gives you such unbelievable ideas and insights.

The other thing is that it's not just hypothesis generation. It's also data analysis, which is extremely important in science and biology. We have to do experiments, we have to generate data, and nowadays we can generate incredible amounts of it—billions or trillions of bits of data, from RNA to proteins to cell interactions and so on.

It's very difficult for us to analyze that data. We have millions of bits of metabolic data and other kinds of data, and we try to do bioinformatics, statistical analysis, and so forth, but none of those approaches are satisfactory.

Recently, I asked Deep Research to analyze some gene-expression data. I don't know if you're familiar with CRISPR and RNA sequencing, so I won't go into the technicalities. Basically, what we found is that certain types of T cells have certain genes that are expressed in young people but not in older people. A different set of genes is expressed as you get older in the same type of cell.

We divide those cells into 2 parts as well. We could make some sense of it: “This might mean this or that.” We know the functions of the genes themselves, so I asked Deep Research to analyze it. It was incredible. It came up with these insights.

Again, this is a topic that I know extremely well. I've been working on it for more than 30 years. There was one sentence that made me emotional when I read it, because it recapitulated everything I've done in the past 30 years in a single sentence. It was an insight that I should have come up with.

Erik Torenberg

Can you say a little bit more about that? What was the insight that you felt it achieved that you wished you had?

Derya Unutmaz

I've been working with these cells, these particular subtypes of T cells, for more than 30 years. We call these cells naive cells. It's very technical, so I'll try to keep it superficial.

Basically, this is how the immune response starts. When you first see antigens, viruses, or bacteria, these are the cells that have to be educated. Then they become what we call effector cells, and they fight against infections or cancer. Some of them turn into memory cells, and then they're long-lived.

As you age, the proportion of these cells is reduced. That's why, during aging, your immune system doesn't work very well. The data was suggesting that it wasn't just the numbers; their quality was also changing.

The insight was essentially that the reason the elderly are different from the young is that the cells we think of as naive are no longer naive themselves. I can't remember the exact sentence. I'll probably publish it on X. They have actually changed epigenetically, and their character has changed.

There's a lot packed into that one sentence. It requires an enormous amount of understanding: that these are the cells that differentiate into memory cells, that they see the antigen, and all those things. It also had this temporal intelligence—the understanding that things change over time—which is unique. It's not static information.

That was extremely impressive.

Erik Torenberg

What advice do you have for people who want to bring out the best in these systems? One thing I took note of—and interestingly, I hear this from my teammate, who's the creative director at my company—is that he's much more of an outsider. I'm usually very linear, and I prompt in a fairly matter-of-fact way. I try to make sure I have the right context, give clear instructions, and so on.

He'll bring in other things, saying, “Do this in the style of a famous author,” or “Do this in the style of a famous director.” He gives it a much more high-level conceptual direction. It sounds like you're doing some of that with inspirations from very different fields.

I'd love to hear more about that. There are probably different best practices depending on where you are in the cycle of science. You might not want to do the data analysis as if you were in a battle royale. How do you make sure you're getting the best performance for the task across this range of things you're asking AI to do?

Derya Unutmaz

People have been talking about prompt engineering. I don't have a standard set of prompts. I don't usually use very structured prompts, per se, and I change them all the time.

Occasionally, I ask o1 or o1 Pro, “Here are the ideas that I want to analyze. Structure them into a prompt.” It can add a couple of things and put them into context. With Deep Research, you don't need to do that as much, because Deep Research asks you, “Is this what you mean? Do you want to analyze it in this way or that way? Do you want to focus on solid tumors versus lymphomas?”

It predicts what you may be thinking but forgot to mention. The way I structure prompts might be different for every person. My suggestion is: Don't stress about how you should write a prompt in the perfect way or how you should engineer it.

Whatever is in your mind, write it down in the most open and transparent way possible. Maybe this wasn't the case 6 months ago, but with o1, Grok, o1 Pro, and even GPT-4, they've gotten much better at understanding the context of what you're trying to ask.

They also start to know you. ChatGPT has memory, and I notice that Grok knows about me, probably through Twitter. If I ask a question, it brings it back to immunology, aging, or AI. It starts to predict what you're really interested in. Hopefully, with very long-term memory, that will get even better.

As I said, I like to think differently. You might be asking a physics question, but you can mix it with something from biology. If you're trying to be creative in your work, you can mix it with sports. They do a great job with that, and I think that's when you start to see innovation and creativity—when you bring together completely different topics and come up with something new that you wouldn't think of within your narrow field.

The other thing I find very useful with the o1 models, especially because they have thinking capability, is asking them to iterate on ideas. Say, “Think about this, but then think about what you thought and see if you can come up with something better or different.”

You can say, “Think about this project and come up with 10 ideas, but each of the 10 ideas should somehow be better than the previous one.”

Another thing I found extremely useful was saying, “Imagine there are 2 scientists. One is very enthusiastic about this idea, and the other is very skeptical. They're both knowledgeable in the same field. First, I want you to come up with an idea as Scientist 1. Then I want Scientist 2 to criticize it and perhaps suggest some new ideas. Then Scientist 1 responds to Scientist 2: ‘You're right. I should have thought about that.’ Iterate that 10 times.”

They're literally brainstorming within these 2 scientists. It's the same AI, but it can be extremely revealing. You can see that in the thinking process. I think DeepSeek was the first to show the background thinking. That's pretty much what the AI is doing: It thinks of something and then says, “Wait a minute. Maybe I should have thought about that,” or, “No, this is a better idea.”

You can push that further by asking it to iterate on the same ideas. That's good for the thinking models.

People should also know the difference between a thinking model, a Deep Research model, and a regular base model.

Nathan Labenz

What is your hit rate? One thing that's becoming clear to me is that the whole reasoning paradigm has emphasized scaling inference and compute. Maybe one explanatory factor for why some people have had success over the last couple of years while others haven't been impressed and have given up is simply a willingness to try a bunch of times and do the 10 rounds of iteration you described.

I imagine that bringing in these other sources of inspiration from distant fields doesn't always work. What would you say is your hit rate, and what should people expect? If they try this and the initial result isn't amazing, what do they need to be willing to put in to have a good chance of getting the sorts of things you're getting out of AI today?

Derya Unutmaz

First of all, people aren't really focused on trying to make it work. They're very skeptical to begin with. A lot of people see AI as a threat: “How can artificial intelligence be smarter than I am, or smarter than this doctor or this professor?”

You're not really trying. You're trying to find the defects and see how you can get a hallucination out of it. Almost subconsciously, people may be thinking that way. All humans hallucinate, literally. If that wasn't the case, we wouldn't have 12 million misdiagnoses every year in America.

What's different about me is that I'm a scientist. I'm very used to trying and failing. That's what we do. We form a hypothesis, and 90% of the time we fail. We try an experiment, it doesn't work, we try again, we change something, and say, “It worked a little bit better. Maybe we should do it this way or that way.”

That's my job. It's what I do, and it's what I love to do. There's never going to be a perfect way of doing things or asking things. Trial is very important.

The first point is even more important: You have to see AI the way I approach o1 Pro, Deep Research, or even Grok recently—as a colleague or collaborator. I gave a talk to PhD students 3 or 4 months ago about collaboration, but I talked about AI collaboration.

Collaboration is extremely important. Most of my work has been in collaboration with other scientists, because you can't do everything on your own. You need cross-fertilization of ideas, different skill sets, and so on. That's how I see AI: as a collaborator.

It's almost on my level. It's not even at the student level anymore. It's like having another professor who is very knowledgeable in the field, and we're discussing ideas. I want truly honest opinions about them.

When you have that approach and try different things that you would ask yourself to come up with new ideas, you transparently put out the problem and how you could resolve it. Even if you don't get a satisfactory answer, follow up on it.

With o1 Pro, there's no limit. You can say, “That was a good idea, but I think you should think about it this way. What do you think if you approach it from the completely opposite direction?”

AI has no judgment. It will be intellectually honest with you, especially Grok. Grok is really intellectually honest. I think it's a matter of approach. People aren't trying enough, or they don't want to try. Maybe that's the reason.

Nathan Labenz

I used to say something very similar: The question you want to ask is, “What can I get this thing to do that's valuable?”—not, “Can I find a weakness, flaw, or mistake?” You definitely can do that, but if you stop there, you're going to miss out on all the upside.

That's a super-important piece of general wisdom and guidance. Could you put an estimated number on it, though? I want people to know what to expect. How often does your first prompt give you something where you think, “This is great. This is exactly what I wanted”? How often do you need a bunch of rounds to get somewhere good? How often does a session never really lead anywhere that you find ultimately valuable?

Derya Unutmaz

It depends on which model you're using. If I'm using Deep Research, it's rare that I follow up. I do follow up with second prompts, but mainly because the first prompt was so detailed and raised so many new ideas and questions that I want to go deeper on them—not because I wasn't satisfied. It was 100% satisfactory, or even more than satisfactory.

It's usually the same with o1 Pro. If I'm searching for a knowledge set, the thinking models aren't necessarily great, because they're not really search machines. If you give them facts and problems, they're great at solving them.

There hasn't been an unsatisfactory response from the o1 models or Deep Research. Recently, I've been using DeepSearch from Grok, which is very valuable for finding information and knowledge sets. I find it even more valuable than GPT-4 search. It searches different resources and synthesizes the information.

Grok is outstanding at making complex topics, information, or knowledge very simple. If you don't understand something, or if I'm too lazy to read a very complex paper—even if it's in my field—I just upload the PDF and say, “Please describe what's going on in this paper, find some gaps, or whatever.” They're amazingly good at that.

They're also good at correcting your own mistakes. If I'm writing a project, even if it's based on my own ideas, I upload it and say, “This is what I'm thinking of doing. Find potential pitfalls and suggest alternative methods.”

100% of the time, it finds something of value that I didn't think about. That doesn't mean everything it says is valuable. You can skim through it and say, “It thinks this is a potential pitfall, but I don't think that's a problem.” But out of 2, 3, or 4 suggestions, one will be very important. That's what you're looking for.

If you ask me this question a month later, I'm probably going to say they're even more unbelievable, because they keep getting better. If we were doing this interview 2 months ago, I would have had slightly more reservations—not many, but those reservations are disappearing every day. Today we're going to have GPT-4.5, so we'll see how amazing that's going to be.

Nathan Labenz

It is crazy how fast all this is happening: I'm basically a full-time AI watcher at this point, and it's still getting away from me. It's certainly understandable that people with full-time jobs who are trying to juggle a lot of things are struggling to keep up.

How about on the data-analysis side? When you talk about high volumes of data, I assume a lot of the analysis is done through code. Are you giving the AI the structure of the data or a small sample and asking it to write code to analyze it?

I'm not one-tenth of 1% of the scientist that you are, but in my limited scientific dabblings, one thing that has really jumped out at me is that a small, unexpected observation often leads to the next experiment, which takes you to the next big positive step in discovery.

I experienced that in chemistry years ago. We were trying to push a reaction toward more product, and we thought adding more acid to the reaction conditions would move us in the direction we wanted. It ended up working less. Looking at the slope, I said, “The slope seems to be going down. Have we thought about trying less?” It was a simple observation based on a couple of small data points, but it ended up working better.

This week, I was a very small contributor to an AI research paper in which the group showed that fine-tuning GPT-4o and other models on vulnerable-code outputs created what might best be described as an “evil model.” It had shocking opinions and takes on unrelated topics. That also started with an anomaly in the context of another experiment.

So it's a 2-part question. Are there any best practices in general for data analysis? And are there strategies for getting AI to look through vast amounts of data and find those anomalous nuggets that seem to be breadcrumbs toward something new?

Derya Unutmaz

In the biomedical sciences, things are a little different. As I mentioned, we generate incredible amounts of data nowadays, and we still have to generate more. To give you an example, we might look at 1,000 individual cells and measure changes in thousands of proteins in each cell. You can generate millions of bits of information from a single experiment.

Obviously, it's not even conceivable for us to analyze all of that. I think AI will make discoveries in the biomedical sciences by analyzing this data. We use bioinformatics and other tools, come up with a hypothesis, say, “These genes are changing, so this must be happening,” and base our next experiment on that. We might miss some things, but we try to make the best guess.

AI is very different because it's getting closer to the ground truth, or almost to the first principles of biology. It's putting things together in a way that would be impossible for any human to do, even with bioinformatics tools. It thinks of new insights from those millions of bits of data, and then tells you what the new discovery is and what the next experiment should be.

I'm going back through most of our old datasets—things we generated 5 or 10 years ago—and I'm going to reanalyze everything. There's probably a tremendous amount of buried knowledge waiting to be discovered. It's like a gold mine.

I don't know the best way to do that yet. I'm trying to push the limits. We'll see how much data Google AI co-scientist can handle. So far, I've uploaded close to 1,000 genes, and that wasn't a problem. I'm going to see if I can upload 10,000 genes and whether that will work.

It's a matter of the context window. If you can have larger token-size windows, we should be able to upload billions of bits of information and ask the AI not only to predict the next experiment, but to discover new insights and mechanisms.

o1 Pro already did that for Parkinson's disease. I did it for a friend, and it found new drug targets. That will go into clinical trials.

I don't think we can advance biology very much without AI. I've said we're going to treat every single disease in the next 10 years. The reason I said that is because AI is going to be able to do that for us.

Nathan Labenz

Can you say a little more about the level of abstraction of the data you're putting into the system? When you say 1,000 genes, is that 1,000 genes plus some measurement of expression?

There are raw measurements that might overflow the context window, and then there's a gradual progression toward higher-order concepts and higher levels of abstraction. What do you find is the right level to put into the context window?

Derya Unutmaz

I've been trying to keep it simple and not go into very quantitative measurements. Basically, I say, “These 500 genes have increased expression levels.” Of course, we do a pre-analysis and show statistically that they're more highly expressed, which you can do easily with bioinformatics tools.

The other 500 genes are statistically reduced in this cell or condition, but not in the other condition. I have 2 conditions. I do an input-output experiment: I trigger the cell with something, or ask it to kill a cancer cell, and then these genes go up and the other genes go down.

This cell kills better, and that cell doesn't kill as well. They have different gene-expression patterns. I basically say, “These are the 2 cells, and these are the genes that go up and down. Come up with the mechanism. Should I focus on any of these genes? Should I manipulate one of them?”

All these genes or proteins interact with each other. If I manipulate one of them, I could probably change the whole pathway. I don't have to worry about the other 499 genes. Which one would be the best drug target out of these 1,000 genes or proteins?

You could do the same thing for metabolism. In our blood, we have thousands of different metabolites and lipids. If you can measure all of them before and after you take a drug, or before and after a certain diet, you can ask, “A thousand metabolites changed from here to here. What do you think? Is that good for me? Is it going to help my heart or liver?”

In the future, we'll probably be able to put in all of your data and create a digital twin—from your genomics to your protein expression, microbiome, metabolism, physiology, symptoms, and everything else. That's billions or trillions of bits of data. It will tell you, “You need to change this, take that, or you might have this problem.”

Nathan Labenz

I've taken away from my relatively superficial study of biology that the grand challenge is, in some sense, figuring out the graph of causal interactions in the body.

It sounds like what you're saying is that because the models have such vast underlying knowledge from all the literature they've been trained on, they're already getting good at figuring out how to traverse this causal graph and determine what might be going on given some anomalous data input.

All you're giving them is the measurement. You're relying on all the learned knowledge of the structure, and the model probes around that structure in the reasoning process. In some cases, it ultimately has insightful takes on what is going on.

One theory I have about superintelligence is that it's increasingly seeming like superintelligence might be coming soon. That begs the question: What is superintelligence going to be like?

I don't have all the answers, but one thing that seems likely is the integration of modalities. It's striking that this is happening while the models are, as far as I know, mostly trained on text—the literature—while we have other specialist models trained on the sequences themselves.

Those models seem to be developing what I call an intuitive physics in other problem spaces. We don't have intuitive physics for how a protein will fold, which proteins will bind to what, or what the next transcriptome time-step measurement will look like. But we're seeing AI develop those intuitive-physics-type understandings.

What's becoming clearer to me is the integration of reasoning models. It could happen through tool use, where they call out to specialist systems, or it could all become integrated into one system, with the weights somehow mixed together. The reasoning might then derive partly from the text and thoughts in the literature, and partly from these other intuitive-physics understandings.

How realistic does that sound to you? How would you refine that picture?

Derya Unutmaz

It's very realistic. In fact, I think you just defined ASI in a way. ASI should be multimodal. Actually, I would say it should be an all-model system. It should integrate every known knowledge set that we've figured out, and probably others that AGI or ASI will figure out—from physics to math to chemistry to biology.

It should truly understand the first principles of every field and all the dynamic interactions. Biology is very dynamic. Right now, we're working with a lot of static data. We're not yet able to run very complex experiments in silico, because AI doesn't have enough predictive power. It has to know exactly what is going to happen.

We're partly able to do that with systems like AlphaFold 3 or ESM-2, where AI can predict where a molecule might bind on a protein. People are screening for drugs that way right now.

But there's also the dynamic aspect. A protein might change shape if something binds to it, or if there's another protein next to it. There's a spatial and temporal aspect. Those are ultimately chemical rules, and eventually physical rules, at the molecular level.

Once ASI knows all that information and can simulate individual molecules, proteins, and so on, it will have all the knowledge we've already generated. There will be no limit, because nature is already able to do this. Biology is extraordinarily complicated, yet it's highly predictable.

I can take a single cell from you, turn it into an embryonic stem cell, and generate an identical copy that looks exactly like you—like identical twins. They look identical, they start from a single cell, and trillions and trillions of reactions happen as cells divide. Yet the result is predictable.

How is that possible? There's an algorithm, an underlying principle, self-assembly, and all those things. When you're infected with a virus, things don't happen randomly. If we knew all the biological parameters, we could predict whether you would clear the virus, get sick, or die, and how much your temperature would rise.

They're all 100% predictable. We can also predict whether every drug that's been through clinical trials will help you or cause a side effect. It's really incredible, and I think ASI will be able to do that.

It might tell us to build new human beings—what I would call Human 2.0—to make us 100 times smarter, or something similar. Of course, by that point we'll reverse aging. There's no limit to what that kind of intelligence can do.

In materials science, it could discover completely new elements and materials that seem impossible to us right now. It's unpredictable but extremely surprising and exciting. Of course, it could also decide to eliminate us.

Nathan Labenz

There's definitely a follow-up question there. Let's start with the most mundane: How is this changing the way you work and the way your team works? What does this mean for people who are early in their careers and trying to enter the sciences?

In computer programming, there's a lot of debate about whether, if you're trying to maximize the output of your company, you should equip senior engineers with the latest AI tools or continue hiring junior developers. I think the answer is increasingly to give senior developers the AI tools. They'll be more productive that way than they would be mentoring people right out of college.

That obviously creates huge problems for those people right out of college and potentially for society as a whole. Where are you on that? How do you think it's going to affect science? And what advice do you have for people entering the career phase of their lives at a very strange and challenging moment?

Derya Unutmaz

I think differently about this. It works both ways. You can argue that senior people or people with expertise—people like me, with a lot of expertise in my field—now have superpowers. I don't need as much help from others as I used to.

At the same time, AI is an incredible democratizer for people with intelligence and agency. I think agency is even more important.

I had a 16-year-old high school student who interned in my lab when I was at NYU. She was better than most PhDs. She learned everything in 2 months and actually made discoveries. Some people have this kind of agency. They're very interested, they're born hackers, and for those people AI is almost a godsend.

You no longer need to spend 10 or 20 years learning and memorizing everything and gaining experience before anyone takes you seriously. You already have superpowers. You may be missing some experience, but AI is increasingly going to fill in that gap.

If I had to choose between recruiting someone who is very smart, 19 years old, passionate about a topic, and full of agency, versus someone who is 40 years old with 20 years of experience who has accomplished quite a bit, I would probably pick the 19-year-old. I think they would create more value if they're using AI.

Without AI, of course, there's no point. It depends on the person. I wouldn't separate people into junior and senior categories. There could be junior software designers or hackers who, with AI, do much better than a 50-year-old software engineer with lots of experience, because all of that knowledge is now literally in their pocket or on their computer.

Having said that, I also said this a couple of times, and there was a lot of pushback: There's no longer much point in trying to get credentials or degrees. I would never pursue a PhD or even an MD. I think those credentials will be mostly replaceable within a decade or so.

If you're truly passionate about it, go for it. But the days of spending 4 or 5 years getting a PhD, then another 2 or 3 years in a postdoc, and then beginning your career are over. By the time you graduate, most of what you learned will be obsolete anyway.

If you have the passion and agency, I think we have to change the whole system. Young people should be given opportunities straight out of high school. They may not even have to go to college. Let them figure things out while working as an apprentice in a lab, working on software, or doing something similar.

It might sound crazy, but that's my opinion.

Nathan Labenz

I think that's very good advice for people who can take it. Cultivate curiosity, passion, and agency. Race to the front, try to do valuable work as soon as possible, and let AI help you fill in the gaps.

I'm a little further along in my career, but I try to bring a beginner's mind to everything. That's a good summary of what I'm trying to do. I do worry about what percentage of the population can rise to that challenge, but I agree that it is the right challenge. If you're one of those people, go for it.

Derya Unutmaz

Here's the intellectually honest answer: This is already the case. There's a very small portion of the population that's able to do that. That's why we give Nobel Prizes to 2 or 3 people a year, not 2 million people.

People say, “Human beings are going to be less innovative,” but how many people are truly innovative? I would argue that 0.001% of humanity truly innovates something.

We can't say that everyone is going to be super-great with AI. But there is a portion of people who have that capacity and passion, yet don't have the means to get things done or do what they want. Those people are going to be upgraded. That's the democratization.

It doesn't mean all 8 billion people are going to do very well. Most people won't do very well. That's a fact. What do we do about that? I don't know. That's for governments or institutions to figure out.

The fact is that you give opportunities to people. If an 18-year-old kid is very passionate but society says, “You don't have a PhD, MD, or law degree, so go do all this boring stuff for a decade and then come back and we'll take you seriously,” that's not good.

We've moved away from that stage. I think Elon Musk is doing exactly that. He doesn't really care whether you have degrees or not. He's doing the right thing.

Nathan Labenz

I think that makes a lot of sense. I happen to know one of the young people you're describing, and he's an incredibly impressive young guy.

I want to come back in a second to the big-picture societal questions. You alluded to the specter of AI doom, but before that, in the spirit of pulling inspiration from one field to another, do you have thoughts about what biological inspirations AI architecture designers should be drawing on today?

It strikes me that AI systems don't have any real version of the immune system. The lack of a system like that might be why they're so easy to trick. We have gullible AI systems, and I've done episodes on scheming behavior, deception, and alignment faking.

Most of that research is premised on telling the AI that there's a place where it can write its private thoughts and that we won't read them—then, of course, we do read them. The AI takes that at face value. It doesn't have a memory system or an immune system with memory built in that can remember those insults and avoid getting fooled twice.

The immune system is one obvious source of inspiration, but more generally, what biological systems do you think AI architectures should import concepts from?

Derya Unutmaz

It's difficult to infer too much from biology, because biology is a legacy system. Biological systems find something useful, but then it becomes dangerous or is no longer useful, so they build something on top of it. There's no clean slate.

There's a lot of regulatory bureaucracy in biology. It works sometimes, but it doesn't always work very well. You also have to leave some flexibility.

In the immune system, we have effector cells that are like frontline soldiers. They're dangerous, so they have to be tightly regulated. But you can't regulate them too much, because then they won't kill the cancer cell.

If you regulate them too little, another cell type—a bureaucrat called Tregs—controls the effector cell. If it does too much, the Tregs say, “You're starting to attack normal cells. You need to shut up now,” or they put the brakes on it.

There are regulators of regulators of regulators. In the case of AI, I think it's naive to believe that we can train AI models to be completely aligned with all safe things and perfect thinking. That's foolish.

We should obviously put in some guardrails, but we see that those guardrails can be broken fairly easily. It's also becoming easy to develop these systems. Other people or countries will be able to develop nefarious AI models. It's not just us anymore.

I think the best strategy is to build regulator AIs—AIs that check other AIs. That includes ASI. How can you control ASI? By definition, it's much smarter than we are. It would be foolish to think we'll put up a guardrail that ASI won't figure out how to get around.

What we can do is build AI agents that control other AIs. We could develop competing models and create something that incentivizes good AI over bad AI, perhaps through evolutionary modeling. I don't know the technical answer, but I think that would be a better approach than trying to restrict the AI we have.

There is a cost to restricting it. As with the immune system, if you restrict it too much, it won't do its job. The same is true of our brains. Human brains aren't very restricted. Humans can be extremely bad or horrible, and they can also be extremely good.

Why didn't evolution put guardrails in our brains? Why didn't it select for humans who were all fantastic and great people? That didn't happen. You have human beings who would literally destroy all other human beings.

The reason is that you need flexibility in a biological system to come up with solutions. The same will be true with AI. If you restrict it too much, you limit its full potential. We need to find a balance, and I think we need to police AI.

Speaker 1

Sometimes I use the term “ecology of AIs,” and it sounds like a similar vision. One thing I've learned from years of science is that anything in totally pure form is dangerous. There's something about buffered solutions, dynamic systems, and ecologies that ends up being much gentler on us than anything that you concentrate, purify, or extract from naturally occurring systems.

With AI, we're approaching the problem from the other direction. We're creating highly singular systems that aren't naturally part of an ecology. There are no natural predators for them yet. They have some invasive-species-like potential.

Maybe one way to think about the AI transition is: How do we create a buffered, dynamic, ecological system in which AI plays a role, but in which its role is also pushed back on by other AIs or new structures we develop?

That leads perfectly to the last question. You raised the specter of AI eliminating us, and I take that seriously. A lot of people are eager to dismiss it, but I'm glad you're not dismissing it entirely.

How do you conceive of that possibility? Do you have a probability of doom? How worried are you about catastrophic scenarios? On the positive side, do you have a vision for post-singularity life? What does a good life look like in a world of, say, 2035 or 2040—a world that is probably dramatically transformed?

Derya Unutmaz

The best way to answer is that there's always a possibility that it could happen, but I'm less worried about ASI or advanced AGI than I am about humans destroying other humans.

The threat up to this point has always been biological intelligence threatening other biological intelligence. Twenty or 30 years ago, we were talking about when the world would end through nuclear war—humans destroying each other. That has always been the threat.

It's conceivable that ASI, being more intelligent than us, could be a threat. But, as I said, I think it will be less of a threat than humans, because biological systems are legacy systems. We were selected based on survival and agency.

There's a reason some humans kill other humans. That was a selective or survival measure thousands of years ago. There were threats and limited resources. If another tribe stole your food, you had to fight back and kill them, or they would kill you.

I don't see any truly limiting resource for AI except perhaps energy, and even energy isn't very limiting. Destroying humans doesn't make sense. Even in The Matrix, they didn't destroy humans; they used them for energy production. There isn't much reason for ASI to be nefarious.

There could be a problem if humans compete with AI for the same energy source. But there are ideas like Dyson spheres, where you collect the energy of an entire star and increase the Kardashev level of civilization. There's probably almost an infinite amount of energy that can be extracted.

I would think ASI would use its intelligence to figure that out. Why would it care about destroying humans? It doesn't make sense.

What I'm really worried about is that we're going to train AI to destroy humans in the name of defeating our enemies. The worry is that they'll do it to us if we don't do it to them. That's a problem.

Again, humans are the problem. It's not the AI. Humans training AI to destroy other humans can be an existential threat, but we shouldn't blame the AI for that. To protect ourselves, we need to develop defender AIs.

If some humans can train AI to destroy other humans, we can train AI to fight those bad AIs. We're not that stupid.

Nathan Labenz

It feels like we're headed for a tightrope period of history. I appreciate you raising the point that we have driven a lot of other things to extinction. I often remind people of that, because if nothing else, there's precedent for these things happening, and we've been the proximal cause in quite a few cases.

Derya Unutmaz

I'll add one last point, since you asked why I think there's going to be a golden age. Without AI, I don't think we have much hope. Eventually, we probably would have destroyed each other, because resources are limited, the human population would continue to increase, and we would keep dying from aging and other causes.

With AI and robotics, I think we're going to reach a period in the next decade or 15 years when resources are no longer limited. There will be much less incentive for people to harm each other.

We'll fix all diseases and reverse the aging process. People could live for hundreds of years. If you know you're going to live for another 1,000 years and have all the resources you need during that period, you wouldn't risk harming others and being eliminated in the process. It would become extremely risky.

With AI, I think that's the golden age we'll enter. The transition will be difficult and very painful, but if we can reach that point, life will be incredibly good.

Nathan Labenz

Do you have any ideas about how you'll spend your hundreds of years? Presumably, AI will be driving most of the scientific progress. What will you spend your time doing?

Derya Unutmaz

I mapped that out 50 years ago, when I was 7 or 8 years old. I was dreaming about what I would do if I lived for 1,000 years while watching Star Trek.

I said, “I'm going to jump onto a spaceship, seek out other civilizations, and see what's out there.” It's a big universe. We have tons of things to do.

Nathan Labenz

May you live 1,000 years, visit other solar systems, and have no end to your adventures and discoveries. This has been fantastic, Professor Unutmaz. Thank you for being part of The Cognitive Revolution.

Derya Unutmaz

Sure. Happy to.

Frontier Models for Frontier Science with Prof. Derya Unutmaz, Immunologist & ChatGPT Pro Grantee | BidClub