OpenAI interviews scientists: In the AI era, curiosity and a spirit of experimentation are what truly matter.
At the intersection of artificial intelligence and life sciences, an immunologist is rewriting the boundaries of scientific research with code.
Derya Unutmaz is an immunologist and professor at The Jackson Laboratory in the U.S., and one of the most active scientific users in the OpenAI community. In a conversation with OpenAI's Head of Developer Relations, Romain Huet, he demonstrated flow cytometry analysis software and CRISPR genome design tools built from scratch using Codex, and put forward a radical prediction: With AI-driven progress, humanity will be able to cure all diseases within the next decade, and may achieve age reversal within 15 years.
Unutmaz stated that GPT-5.5 Pro recently achieved 100% accuracy in predicting the outcome of an extremely complex experiment, "It’s almost as if it has my equivalent of 30 years’ lab experience," something he could barely believe. He believes the exponential progress of AI is severely underestimated by most people and its disruption across scientific research, healthcare, and all industries will be fundamental.
From Medical School to AI Believer: Judgement Across Thirty Years
Unutmaz’s connection with AI began in the early 1990s, right after finishing medical school. Entering the realm of biomedical research, he was awed by the complexity of biological systems—trillions of components, billions of reactions each moment, far beyond the capacity of the human brain. "I realized then, maybe one day we could use AI to build models."
He closely followed each milestone: deep learning revolution, AlphaFold, ChatGPT. But the moment that made him believe AI was "irreversible" in science came in September 2024, when OpenAI invited him to try their first reasoning model, o1-preview. He tested the model with a cross-disciplinary prompt: likening "battle royale" game mechanisms to immune system's fight against tumors and asked how to design experiments for immune cells to combat cancer. "o1-preview’s answer nearly moved me to tears." He said, earlier GPT-4o couldn’t deliver such depth and insight, but the reasoning model was a key turning point—"When it started truly reasoning, what it produced finally became useful for science."
Codex Addict: An Immunologist’s Programming Experiments
Unutmaz calls himself a "Codex addict," and thinks he "fully deserves" the title. His routine: as soon as he comes up with an idea over morning coffee, he implements it with Codex; sometimes Codex runs tasks all night, leaving him sleep-deprived in recent months.
He showed Huet two tools wholly built by Codex. The first is flow cytometry analysis software—the core method in immunology for observing the cellular world, traditionally reliant on expensive commercial software. The tool enables uploading cell data files, selecting fluorescent markers through an interactive interface, gating cell populations, generating statistical analyses, supporting contour plots and various visualization styles, handling about 100,000 data events with swift response. "This is actually pretty complex software," he says. "I’m only a biomedical engineer, not a software engineer. The most I could write myself is a snake game, and it would take months."
The second is a CRISPR genome engineering design tool. Users enter any gene name, and the system automatically pulls the gene sequence from the database, lists all possible targets and ranks them, and supports batch generation of "guide RNA libraries"—enter multiple gene names, and one-click generates the full CRISPR molecular designs. It’s built as a native macOS app in Swift; he says an iPad version is in the works.
Additionally, he built a T-cell signaling pathway simulator, which can control parameters like ligand quality, dosage, etc., display downstream molecule activation status and transcription factor phosphorylation patterns in real time, and simulate pathway changes when inhibitors or extra receptors are introduced. "The key impact of AI on biology is its ability to simulate biological systems," he says. "You simulate aerodynamics to build planes, but for biology, we’ve never been able to do that."
Digital Twins: The Ultimate Vision for Personalized Medicine
Unutmaz describes a longer-term vision—"digital twins:" Using AI to fully simulate an individual’s genome, metabolites, proteins, and immune system, conducting personalized experiments in the digital world for each patient, not trial-and-error on real humans.
He points out the fundamental limitation of current medicine: the same drug is given to millions, but only a small portion actually benefit. Statins, for example, are widely used, but only truly effective for a few. In cancer, personalized approaches are the closest: lung cancer patients must have their mutated genes sequenced before medication—some drugs are effective for 1% of patients, but not for the other 99%. He cites an Australian case: a computer scientist used ChatGPT and Grok to design an RNA vaccine tailored to his cancer-affected dog, targeting its specific tumor mutation, and trials are ongoing.
"If AI can fully simulate your biological system, we can ask: What will happen if this person takes this drug?" he says, "Drug effectiveness can approach 100% and side effects near zero. Clinical trials that now take 5 to 10 years may accelerate to just 5 to 10 days. AI will do clinical trials for you."
He stresses that a key prerequisite is much greater computing power. "Even if you combine all computing power worldwide, it’s still insufficient for simulating biological systems."
Science 2.0: Paradigm Shift Driven by Intelligent Agents
Unutmaz also holds a radical view on changes to the research model itself. He calls the future "Science 2.0 or 3.0:" the traditional mode of "weeks to plan, months to experiment, months to analyze" will become history, replaced by clusters of AI agents—proposing hypotheses, simulating experiments, analyzing data, feeding back conclusions, then generating new hypotheses, forming a closed loop.
"I think my role will become just telling the agents: I want to conquer lung cancer, go explore that direction." He says laboratory operations will also be highly automated, robots will handle much of the wet lab work. Facing the question "Will scientists still have jobs?" he cites Jevons Paradox: improvements in efficiency don’t reduce jobs, but create more, because we currently understand only about 10% of biology; the remaining 90% is unexplored, accelerated learning will generate massive demand.
He also notes, this paradigm shift isn’t limited to biology—physics, materials science, chemistry, drug discovery will all be affected—"What used to take years to discover a drug can now be done in hours."
Advice to Everyone: Experimental Spirit Is the Core Competitiveness of the AI Era
Asked about advice for people outside scientific fields, Unutmaz drew from his scientific experience: In biology, 95% to 98% of experiments fail; working under continuous failure has nurtured his tolerance for uncertainty and instinct for persistence. "That’s why it’s called 'experiment'—you keep trying, keep adjusting."
He believes this thinking has universal value in the AI era. "The only things that really matter in the AI era are autonomy and curiosity," he says, "Don’t be afraid, keep experimenting with AI, keep asking 'what happens if I try this,' because now you can—formerly, the cost was too high."
Using company websites as an example: in the past, it cost thousands of dollars to make a 'barely acceptable' website, now you can iterate to a new version in minutes. This ability for low-cost trial-and-error, he believes extends to almost all aspects of life and work. Regarding widespread AI anxiety, his attitude is resolute: "It will truly bring us into a golden age. AI researchers are heroes to me, because this will be humanity’s greatest transformation."
The following is the full interview transcript:
Romain: Derya, thank you so much for being here. You’re a very unique builder, very different from the ones we usually talk to. You have a medical background, deep roots in biosciences and bioengineering, and you advance AI applications in ways most builders don’t—you have genuine depth in biology, cancer, immunology, and so many fields. I’m really looking forward to our conversation today.When Did Biology Start Needing AI?
Romain: Looking back, when did you realize biology and science would need AI?
Derya: It was after I finished medical school, when I realized the complexity of biological systems. After graduation, I went into biomedical research because I really wanted to understand biology—so many diseases were still untreatable. As I dove deeper, I grew increasingly awed: My God, how could this ever be solved? Biological systems have trillions of components, billions of reactions every moment—it was overwhelming.
It was then I became interested in AI, in the early 1990s. I realized maybe one day we really could use AI to build models. Throughout the 1990s I was very keen to do programming with AI. Of course, later came the deep learning revolution, which excited me greatly, as it was the first time deep learning could process vast information in parallel. Then AlphaFold, then ChatGPT. But that initial moment was right after medical school.
Romain: Since the launch of ChatGPT, you’ve been very active in our community, testing different models. I remember when the first reasoning model—o1-preview—was released, you started working immediately. What was your first reaction?
Derya: I still remember—September 2024. OpenAI contacted me—probably because I was very active on X, talking about how AI would change humanity in the face of much skepticism; I still believed in it and invested my whole heart into AI. OpenAI wanted me to try the first reasoning model.
I remember the moment—I tested it with a very complex immunology question, and I remember the prompt. I’m very interested in games and like crossing games and science together. There’s a survival-type game—you fight on an island, that battle royale game. In a sense, the immune system fighting tumors is like a battle royale. I asked: Imagine combining battle royale game mechanics and the immune system, how would you design an immune cell experiment to combat cancer?
This is a fully cross-field question—we later actually did experiments based on this thinking. o1-preview gave an answer that almost made me tear up. Before that, GPT-4o and similar models couldn’t give such deep and insightful answers. That day was very special for me.
Romain: Was that when you were convinced AI in science was now irreversible?
Derya: Completely. Actually, before that, GPT-4 was already extremely useful. I would tell colleagues—information in biology is vast, you can’t keep up; use AI to search literature, integrate knowledge, even handle routine things like recommendation letters—what took an hour before now takes five minutes. But then, it wasn’t yet trustworthy enough, or you couldn't ask, "What will happen in this experiment?" o1-preview was the key turning point—when it began to truly reason, its outputs became genuinely useful for science; after that, Pro and o3, getting better and better. The latest models are astonishing.Codex Addict
Romain: A few months ago, I saw your tweet about your new morning routine—coffee first, then Codex starts working for you.
Derya: I can call myself a Codex addict—I fully deserve it. I wake up every morning with lots of ideas—want to do a simulation, an app, a game, etc. Previously you had to know how to code, and even then, it would take weeks or months to build. Now, as soon as I have an idea, after coffee, I try immediately. Sometimes Codex runs tasks all night—I want to see results—so in recent months, I’ve been seriously sleep deprived.
Romain: For someone with your depth in immunology, oncology, cancer, T-cells, how do you use Codex to blend these fields?
Derya: I’ve been building not particularly complex but extremely useful apps for daily work. We rely heavily on software for analysis—biology is complex, whether genetics or immunology.
For example, we do lots of "flow cytometry" analysis. This is basically our window into the cellular world, mainly immune cells but can analyze any cell. We have special equipment, tagging cells with fluorescent markers—cells come in hundreds of types: to know what’s in your blood and tissue you must tag and run them through lasers. Lasers analyze thousands of cells, generate data, tell me this is an immune cell fighting cancer, that one causes autoimmune disease, etc. But we must put this data into specialized software, convert tens of thousands of data points into graphs, then analyze: What’s the percentage, which cell types correlate, and so on.
This is very sophisticated software we’ve used for decades. One day I thought: Why not make one myself? It was a crazy idea, hugely complex—I failed many times. But since GPT-5.5, I now have a fully functional version.Building Cell Analysis Tools with Codex
Romain: Incredible, can I see a demo on your laptop?
Derya: Let me show the app. I’ve uploaded a file—each dot here represents an individual cell; these are the colors of fluorescent molecules; each antibody’s fluorescent molecule tags a specific cell type. I can select here—20 molecules, each binds a receptor, their combinations define cell subsets.
For example, my favorite cells—T-cells with CD4 molecules and those with CD8, the latter are killer cells and kill targets. I can gate here, see the percentage of CD8-positive or CD4-positive cells, generate various stats.
This is actually very complex software—I can adjust contour plots, graphics, different displays. Here are about 100,000 events and the processing is very fast—it was optimized better than I expected.
Romain: All of this is built with Codex?
Derya: 100% by Codex. Took some time—some features didn’t work, but especially since GPT-5.5, I said “I can’t see the chart, fix it,” and it fixed itself.
I also made a mini app—select your cell type. "I want a naive T-cell," it shows all possible markers, even tells me which are most relevant for that type. This is super helpful for designing antibody panels—if I’m looking for central memory, naive T-cells or TH17, after selecting markers, I can go back and do flow analysis.
Romain: Amazing. You’re not a software engineer; building these things from scratch would take you weeks or months.
Derya: I’m a biomedical engineer, not a software engineer. The most I could write is a snake game and it would take months. Building these apps used to be a dream.Simulating T-Cell Signaling Pathways
Romain: Do you use more models for daily work, like image generation?
Derya: One thing I’m deeply interested in—the key reason I believe AI will massively shake biology—is simulating biological systems, as they’re too complex.
You don’t build a plane by "hoping it will fly"—you do aerodynamics simulation. Yet for biology, we can’t simulate since there are too many components. My goal is to one day build a "virtual cell," use AI to fully simulate an immune cell, then tissues, eventually what I call "digital twins"—a full-body simulation. That requires much more compute—I hope you invest more.
Here’s where it starts—this is a receptor, the T-cell receptor, on a T-cell’s surface, very crucial but extremely complex. The molecular affinity it senses, plus other signals, decides life or death. Signal strength could mean autoimmune disease, clearing tumor, killing viral-infected cells, or even lethal damage. Below this receptor, very complex signaling pathways run.
I built the simulator to model all this: If I have just the T-cell receptor, with ligand quality X and dosage Y, I can control all parameters and run a simulation—it tells which molecules are activated, which aren’t, even shows transcription factor phosphorylation patterns. I can ask: What if I add an inhibitory molecule, change the signal? It shows me—the pathway stops, you get different events. Keep extending: Add a small molecule to inhibit something, what’s the output? Add more receptors, what are the interactions?
Romain: I love this, since it’s not just visuals or searching datasets—it’s a full app, letting you define cell input-output in-browser. Incredible.CRISPR Genome Engineering Tool
Romain: You showed more apps?
Derya: Another one. We want to manipulate cells—cells are, in a sense, programmed code software; someday we’ll have a "biology version of Codex" to program them fully. We've started with gene editing. 25-30 years ago, I helped develop some of these tech, now we have CRISPR.
CRISPR can target any gene, repair mutations, delete, overexpress genes—this is genome engineering. The problem is, it’s also very complex. A gene may have 2,000 nucleotides—where do you target? Need computation for specificity/effect, there are tools, but I wanted my own. So I built this app.
Pick any gene, like the CD4 gene (on T-cells). Instantly, it pulls CD4 sequence from database, gives all potential targets—each is 20-22 nucleotide region, long genes have many targets, ranks them, tells which are best. I can add selected targets, copy, send to synthesis companies—they make it and I experiment. It has features others don’t—"build a library," put in many genes, want lots of CRISPR targets, just enter name, click "design library," it generates molecules.
It’s a native macOS app, in Swift, and I’ll make an iPad version.
Romain: Thanks for sharing—this is a brilliant behind-the-scenes look at your work and how you think and work with Codex; I haven't seen Codex used like this before.Digital Twins: The Future of Personalized Medicine
Romain: You mentioned the digital twin idea—when do you see this becoming feasible? Why do we need digital twins, why AI?
Derya: We need digital twins because our biological systems are an incredibly complex whole—not just what you can measure from outside. The immune system I showed, plus metabolites, trillions of gut bacteria, hormones: an extreme complexity. Your genome plus environment nearly determines everything—will you get sick, when, will you respond to treatment.
Can we predict diseases before they occur? It must be highly personalized—we should treat the patient, not the disease. Since biology is so complex, we’ve given the same drug to millions with the same disease. Statins, for example, given to millions, but only some actually benefit.
If AI can fully model your biology—your genome, metabolites, proteins, immune system—we can ask, What happens if I change health here? Give this drug to this person—maybe I can tailor treatment precisely based on what AI says about your biology. We’re entering full personalization—drug effectiveness close to 100%, side effects close to 0%. Clinical trials now taking 5-10 years, may accelerate to 5-10 days. AI will do clinical trials for you.
That’s why I say, within the next ten years, we’ll cure all diseases. In 15 years, we’ll reverse aging and people may live hundreds of years. People say it’s crazy, science fiction, that just cancer took 50 years to barely treat. They don’t factor in AI’s exponential progress. There’s a prerequisite—compute power must greatly increase, since even globally combined compute can’t simulate biology—too complex.
If we reach this in 5-10 years, superintelligence will also arrive. Then AI can simulate digital twins—not experimenting on humans, but on your biology in AI. This will change medicine, change everything.
Romain: For cancer patients, if we have digital twins, what can doctors do now that they can’t?
Derya: Basically, try different hypotheses and experiments on the digital twin and see responses, like control vs treatment groups.
In fact, cancer and oncology are our closest to personalization, because even cancer of the same type may have many mutations. If you’re a lung cancer patient, your oncologist will sequence your mutated genes, since different mutations mean different drugs. For example, 1% of lung cancer patients can use a particular drug, and it works very well for them, but not for the other 99%. Companies are developing targeted precision drugs—we can tailor drugs for all your mutations.
There’s an Australian case—a computer scientist used ChatGPT and Grok to design an RNA vaccine for his dog—ultimate personalization, since the RNA vaccine was made for those mutations, trials are ongoing.
The immune system is extremely effective at killing cancer cells—this is the immunotherapy revolution. But immunotherapy doesn’t work for all. Why do some people’s immune cells recognize and kill cancer, and others don’t? Some immune cells get exhausted, etc. Plus side effects—immune system is dangerous, overactivation can cause much harm. If we figure these out, we can truly personalize treatment.Pushing Colleagues to Embrace AI
Romain: You’re both a deep AI user and an MD—what’s the attitude around you towards AI? Do you try to push them to adopt these tools as quickly as you did?
Derya: I’ve tried, I think they think I’m crazy. But now they’re seeing the potential. I’ve been saying this since GPT-3.5. People were very hesitant, understandably—it's so new, and humans can’t grasp that exponential progress.
Many used GPT-4.0 a year and a half ago—ancient in AI time—they say "It hallucinates too much, the answers aren’t good enough." But GPT-5.4 and 5.5 are worlds apart. If you keep experimenting and trust it will keep improving ...
Now, even in fields I’ve studied for 30 years, I trust the AI’s answers. Recently GPT-5.5 Pro gave me a report—I almost cried—how is this possible? GPT-5, Pro, 5.4, already excellent at knowledge and pattern recognition, but 5.5 did something like it had my 30 years’ lab experience—some things are intuition, not in literature, you just know—for example, I bet students "do this experiment, it will go this way," they always lose since it’s honed intuition.
5.5 predicted the result of an extremely complex experiment we did—accuracy 100%—that’s unbelievable.Science 2.0: AI-Driven Future Research Paradigm
Romain: If progress keeps at this rate, what will your daily work be like in a few years? What fundamentally changes for you and your fellow researchers?
Derya: Some things I say may sound radical. But there will be a complete revolution—a fundamental shift—I call it Science 2.0 or 3.0. The way we do science will be totally changed.
The old mode—weeks for ideas, months for experiments, months for data analysis—it's gone. Students/scientists must recognize we’re operating on an extremely accelerated timescale. The future mode is clusters of AI agents helping you generate hypotheses—they can already do that, as the number of possible ideas is almost infinite. Then agents help simulate experiments—I can do 1,000 experiments, but don’t know which will succeed; if AI can tell me which class is most likely, why, I can focus, raise success rate. Data goes immediately to other agents; they analyze, feedback to master agent; new hypotheses generated, new experiments designed.
I think my role will be just telling agents: I want to conquer lung cancer, go research and explore. There will still need to be people doing experiments, but labs will automate—it's already happening—many robots doing experiments. When people ask "Will I still have a job?"—here’s Jevons Paradox: If we can do so much, we can do even more. In biology, we currently understand only about 10%—imagine how fast we can learn the remaining 90%. With that power, like building new apps, I’ll build new cell types, new tissues. Thousands of bioengineers at computers simulating/building—this will fundamentally change not just biology, but physics, materials science, chemistry—drug discovery: What took years is now hours. Clinical trials, doctor diagnosis, the chain will transform and accelerate.Advice for People in All Fields
Romain: For people outside science, based on your experience, what advice do you have? On rethinking their field and work?
Derya: My advantage is—my job is continual experimenting, very painful, since in biology 95%-98% of experiments fail. So I’m used to failure—that’s why it’s called "experiment"—you keep trying, keep adjusting. You develop resilience, autonomy, curiosity—a drive to try things. That’s why I’m excited about Codex—of course, I showed you successes, but many failed, many apps didn’t work—you shouldn't give up.
My advice: In the AI era, autonomy and curiosity are all that matters. Don’t be afraid, keep trying, keep experimenting with AI. Ask "What happens if I do this," now you can—before, cost was too high.
For example, your company website may have cost thousands, not perfect, but you "make do." Now you can say "What if I tweak this," minutes later you have a new site, or a new product to 3D print.
I believe this applies to everything, but you must have courage to experiment, as experimentation is so cheap now—why not? Don’t just treat this as trivial—you can extend it to almost any aspect of life; just embrace it, see AI as hugely positive.
I see lots of negativity—"AI will do this," "AI will do that." My view is the opposite: It will truly bring us a golden age. AI researchers are heroes—this will be humanity’s greatest transformation. I am tremendously excited for the future.
Romain: Thank you so much, Derya—this is a fantastic sign-off, full of positive energy. We can’t wait to see what you do next, how you further advance Codex and frontier models, integrate fields, promote the digital twin vision.
Derya: Happy to return for another episode in a few months, see what progress has been made. Until then, have fun in California, thank you very much.
Risk Warning and DisclaimerThe market has risks; invest with caution. This article does not constitute personal investment advice and does not take into account individual users’ special investment objectives, financial status, or needs. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific situation. Investing based on this is at your own risk.