OpenAI Interview with Scientist: Curiosity and Experimental Spirit Are Truly Important in the AI Era

At the intersection of artificial intelligence and life sciences, an immunologist is rewriting the boundaries of 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 Romain Huet, OpenAI's Head of Developer Relations, he demonstrated a flow cytometry analysis software and a CRISPR genome design tool built from scratch using Codex, and outlined a radical prediction: With AI-driven advances, humans will be able to cure all diseases within the next decade, and reverse aging within 15 years.

Unutmaz said that GPT-5.5 Pro recently achieved 100% accuracy in predicting the results of an extremely complex experiment, "almost as if it had the same experience I've gained in the lab over 30 years," which he found almost unbelievable. He believes that the exponential progress of AI is being severely underestimated by most people, and its disruption to research, healthcare, and all industries will be fundamental.

From Medical School to an AI Believer: A Three-Decade Judgment

Unutmaz's connection with AI began in the early 1990s after completing medical school. He then entered biomedical research and was immediately awed by the complexity of biological systems—trillions of components, billions of reactions every moment, far beyond what the human brain can handle. "Even then, I realized that maybe one day we could really use AI to build models."

He closely tracked every milestone—the deep learning revolution, AlphaFold, ChatGPT. But the moment that truly convinced him AI was "irreversible" in science came in September 2024, when OpenAI invited him to test the first reasoning model, o1-preview. He tested it with a cross-domain prompt: comparing the "battle royale" game mechanism to the immune system fighting tumors, asking how to design an experimental framework for immune cells to fight cancer. "o1-preview's answer almost moved me to tears," he said. Previous models like GPT-4o could not deliver that depth and insight. This reasoning model was the turning point—"When it started truly reasoning, what it produced became truly useful for science."

Codex Addict: An Immunologist's Programming Experiment

Unutmaz calls himself a "Codex addict" and believes the title is "well-deserved." His daily routine: when an idea strikes while drinking morning coffee, he immediately implements it with Codex. Sometimes Codex runs tasks all night, causing him severe sleep deprivation over the past few months.

He showed Huet two tools built entirely with Codex. The first is a flow cytometry analysis software—the core method for observing the cellular world in immunology research, traditionally dependent on expensive commercial software. This tool allows uploading cell data files, selecting fluorescent markers interactively, defining cell gates, generating statistical analyses, supporting contour plots and various visualizations, handling about 100k data events with fast response. "This is actually quite complex software," he said, "and I'm just a biomedical engineer, not a software engineer. I could probably only write a Snake game, and it would take months."

The second is a CRISPR genome engineering design tool. Users input any gene name, the system automatically retrieves the gene sequence from a database, lists all potential targets ranked, and supports batch generation of "guide RNA libraries"—input multiple gene names and generate a full set of corresponding CRISPR molecule designs in one click. The application is built as a native macOS app with Swift, and he mentioned an iPad version is in development.

Additionally, he built a T cell signaling pathway simulator that allows adjusting parameters like receptor-ligand affinity and dose, shows real-time activation states of downstream molecules, transcription factor phosphorylation patterns, and simulates pathway changes when introducing inhibitors or additional receptors. "The key to AI's huge impact on biology is the ability to simulate biological systems," he said. "When building an airplane, you do aerodynamic simulations. But for biology, we've never been able to do that."

Digital Twin: The Ultimate Vision of Personalized Medicine

Unutmaz described a longer-term vision—the "digital twin": Using AI to fully simulate an individual's genome, metabolites, proteins, and immune system, conducting personalized experiments for each patient in a digital world rather than trial and error on real humans.

He pointed out fundamental limitations in the current medical system: the same drug is given to millions of patients, but only a small fraction truly benefit. Taking statins as an example, they are used widely but are only truly effective for a minority. Oncology is already the closest to personalization—lung cancer patients undergo mutation sequencing before medication because 1% of patients respond to a specific drug, while the other 99% do not. He cited an Australian case where a computer scientist used ChatGPT and Grok to design a custom RNA vaccine for his dog's cancer, specifically targeting the tumor's unique mutations, with trials ongoing.

"If AI can fully simulate your biological system, we can ask: What would happen if this person takes this drug?" he said. "Drugs could achieve nearly 100% efficacy and nearly 0% side effects. Clinical trials that now take 5 to 10 years could accelerate to maybe just 5 to 10 days. AI will run clinical trials for you."

He also emphasized a key prerequisite: computing power must be drastically improved. "Even if you combine all the computing power in the world today, it's not enough to simulate biological systems."

Science 2.0: The AI Agent-Driven Research Paradigm Revolution

Unutmaz holds equally radical views on changes to the research model itself. He calls the future "Science 2.0 or 3.0": the traditional pattern of "weeks to conceive an idea, months to experiment, months to analyze" will become history, replaced by clusters of AI agents—generating hypotheses, simulating experiments, analyzing data, providing feedback, and proposing new hypotheses in a closed loop.

"I think my role will become simply telling the agents: I want to tackle lung cancer, go explore this direction," he said. Lab operations will also be heavily automated, with robots handling much of the wet lab work. To the question "Will scientists still have jobs?" he invoked the Jevons paradox: increased efficiency won't reduce work but will create more, because our current understanding of biology is only about 10%, with 90% left to explore. Accelerating learning will generate enormous demand.

He also noted that this paradigm shift is not limited to biology—physics, materials science, chemistry, and drug discovery will all be affected—"Drug discovery used to take years; now it can be done in hours."

Advice for Everyone: The Experimental Mindset is Core in the AI Era

When asked for advice for people outside the scientific field, Unutmaz drew on his research experience: 95% to 98% of biology experiments fail. Working in failure for so long has cultivated his tolerance for uncertainty and instinct to keep trying. "That's why it's called 'experimentation'—you keep trying, keep adjusting."

He believes this way of thinking has universal value in the AI era. "In the AI era, the only things that truly matter are autonomy and curiosity," he said. "Don't be afraid. Keep experimenting with AI. Ask the 'what if I do this' question, because now you can ask it—before, the cost of doing that was too high."

He gave an example: a company website used to cost thousands of dollars to make a "good enough" version; now you can iterate on a new version in minutes. This low-cost trial-and-error capability, he believes, can extend to almost every aspect of life and work. Regarding the widespread AI anxiety, he took a clear stance: "It will truly usher us into a golden age. AI researchers are heroes to me, because this will be the greatest transformation for humanity."

Below is the full transcript of the interview:

Romain: Derya, thank you so much for being here. You are a very unique builder, very different from the builders we usually talk to. You come from a medical background, deeply rooted in biological sciences and bioengineering, and you're pushing AI applications in ways most builders don't—you have real depth in so many areas like biology, cancer, immunology. I'm really looking forward to today's conversation.

When Did Biology Need AI?

Romain: Looking back, when did you realize that biology and science would need AI?

Derya: That was after I graduated from medical school, when I realized the complexity of biological systems. After graduation, I entered biomedical research because I really wanted to understand biology—at that time, there were so many diseases we couldn't treat yet. As I delved deeper, I became increasingly overwhelmed: My God, how can this be solved? Biological systems have trillions of different components, billions of reactions happening every moment, it felt incredibly overwhelming.

It was then that I started getting interested in AI, in the early 90s. I realized that maybe one day we could really use AI to build models. Throughout the 90s, I was very passionate about using AI for programming. Then the deep learning revolution came, and I was thrilled because I saw for the first time that deep learning could process massive amounts of information in a somewhat parallel way. Then came AlphaFold, then ChatGPT. But that initial moment was right after I finished medical school.

Romain: Since ChatGPT launched, you've been very active in our community, testing various models. I remember you started doing work when the first reasoning model—o1-preview—came out. What was your first reaction when you got it?

Derya: I remember it was September 2024. OpenAI contacted me, I think because I'm very active on X, constantly talking about how AI will change humanity, and there was a lot of skepticism back then, but I still believed in it and was fully committed to AI. OpenAI wanted me to try the first reasoning model.

I remember the moment—I tested it with an extremely complex question in immunology, and I even remember the prompt: I'm very interested in games, and I like to make cross-domain analogies between games and science. There's a survival game where you fight on an island, that battle royale genre. In a sense, the immune system fighting tumors is like a battle royale. I asked: Imagine combining battle royale with the immune system, how would you design a scenario where immune cells fight cancer?

It was a completely cross-disciplinary question, and we later actually did experiments based on that idea. o1-preview gave me an answer that almost moved me to tears. Before that, models like GPT-4o couldn't deliver that depth and insight. That day was special for me.

Romain: Was that the moment you were convinced that AI in science was irreversible?

Derya: Absolutely. Actually, even before that, GPT-4 was already extremely useful. I would tell colleagues—the volume of information in biology is so vast that you can't keep up. Using AI to search literature, integrate knowledge, and handle daily tasks like writing recommendation letters—what used to take an hour now takes five minutes. But it wasn't yet at the level where you could fully trust it or ask questions like "What would be the outcome of this experiment?" o1-preview was that turning point—when it started truly reasoning, what it produced became truly useful for science. After that came the Pro version and o3, getting better and better. The models now are simply breathtaking.

Codex Addict

Romain: A few months ago, I saw you tweet that you had a new morning routine—first your morning coffee, then Codex would start working for you.

Derya: I can call myself a Codex addict, and I deserve that title. Every morning when I wake up, lots of ideas pop into my head—want to build a simulation, want to build an app, want to build a game, etc. Before, you had to know how to code, and even if you did, it would take weeks or months to implement. But now, as soon as I have an idea, after coffee, I immediately try it. Sometimes Codex runs tasks all night, and I want to see the results—I've been severely sleep-deprived for the past few months because of it.

Romain: Someone with your depth in immunology, oncology, cancer, T cells—how do you use Codex to integrate these fields?

Derya: I've been building applications that are not extremely complex but incredibly useful for daily work. We rely heavily on software for analysis—biology is complex, whether in genetics or immunology.

For example, we do a lot of analysis called "flow cytometry." This is essentially our window into the cellular world, mainly immune cells, but you can analyze any cell. We have specialized instruments that label cells with fluorescent markers—cells have hundreds of different types. To know what cells are in your blood or tissues, you label them and pass them through lasers. The laser analyzes thousands of cells and generates data, telling me this is an immune cell that can fight cancer, this one can cause autoimmune disease, etc. But we have to put that data into specialized software, convert the data points from hundreds of thousands of individual cells into graphs, and then analyze: what's the percentage of this cell type, how do they relate to each other, and so on.

This is very sophisticated software we've been using for decades. One day I thought: Why not build my own? The idea was crazy because it's very complex, and I failed many times. But since GPT-5.5, I now have a fully functional version.

Building a Cell Analysis Tool with Codex

Romain: That's incredible. Can we see a demo on your laptop?

Derya: Let me show you this application. I've already uploaded a file—each dot here represents an individual cell, these are the colors of fluorescent molecules, each antibody-bound fluorescent molecule labels a specific cell type. I can select here. See, there are 20 different molecules, each binds to a receptor, and their combination defines specific cell subsets.

For example, my favorite cells—T cells with the CD4 molecule, and T cells with the CD8 molecule, which are killer cells that go and kill target cells. I can define a gate here, see the percentage of CD8-positive or CD4-positive cells, generate various statistical analyses.

This is actually quite complex software—I can change contour plots, graphs, different display methods. There are about 100k events here, and the response is very fast. I didn't expect it to be optimized this well.

Romain: All of this was built with Codex?

Derya: 100% built with Codex. It took a while because some things didn't work, but especially after GPT-5.5, I said "I can't see the chart clearly, help me fix it," and it just went off and did it.

I also built a small application where you can select the cell type you want, say "I want a naive T cell," and it shows all possible markers you can choose, and even tells me which markers are most relevant for that cell type. This is extremely useful for designing what we call "antibody panels." I'm looking for central memory naive cells, T cells, and TH17 cells. I select these markers, then I can go back and do the flow cytometry.

Romain: That's amazing. You're not a software engineer. If you had to build these from scratch, it would take weeks or months.

Derya: I'm a biomedical engineer, not a software engineer. I could probably only write a Snake game, and it would take months. Building these applications was a dream for me before.

Simulating T Cell Signaling Pathways

Romain: Do you use more models for routine work, like image generation?

Derya: One thing I'm very interested in—and the key reason I think AI will have a huge impact on biology—is the ability to simulate biological systems, because they are so complex.

When building an airplane, you don't say "assemble these parts and hope it flies"—you do aerodynamic simulations. But for biology, we can't do that because there are too many components. My goal is that one day we can build a "virtual cell," using AI to fully simulate an immune cell, then a tissue, and ultimately achieve what I call the "digital twin"—a complete digital simulation of the human body. Of course, that requires much more computing power, I hope you invest more in that.

Let's start here—this is a receptor called the T cell receptor, located on the surface of T cells. It's the primary receptor but extremely complex. The affinity of the molecules it senses, and other signals it receives, will determine life or death. The signal strength could mean autoimmune disease, could mean clearing a tumor, could mean killing virus-infected cells, or could mean excessive damage leading to death. Beneath this receptor, there is an extremely complex signaling pathway.

I built this simulator to be able to simulate all this. If I only have the T cell receptor, the quality of the ligand is like this, the dose is this much, I can control all parameters here, then run the simulation—it will show me which molecules will be activated and which won't, and even display the phosphorylation patterns of transcription factors. I can ask: If I add an inhibitory molecule, and then change a different signal, what happens? It will show me—this pathway is now stopped, you get different types of events. You can extend further: If I add a small molecule to inhibit a certain molecule, what's the output? If I add more receptors here, how do they interact?

Romain: I love this because it's not just visualization or searching through datasets. This is a full application that lets you define inputs and outputs for each cell and scenario in the browser. Incredible.

CRISPR Genome Engineering Tool

Romain: You've shown more applications?

Derya: Let's look at another. We want to manipulate cells—cells are in a sense like programmed software code. I hope one day we'll have a "biology version of Codex" where we can fully program cells. In fact, we've already started doing this through gene editing technology. I developed some of these technologies 25 to 30 years ago myself, and now we have CRISPR.

CRISPR can target any gene, repair mutations, delete genes, overexpress genes—that's genome engineering. But the problem is, this is also very complex. A gene may have 2000 nucleotides. Where do you target? You need computational analysis to determine the specificity and effect of the target. There are some existing tools, but I wanted my own. So I built this application.

You can select any gene, for example, the CD4 gene—I mentioned it's on the surface of T cells. It immediately retrieves the CD4 gene sequence from the database, then gives me all potential targets—each target is a region of 20 to 22 nucleotides. For a long gene, there will be many targets. It ranks them and tells me which is better. I can add the selected target here, copy it out, send it to a synthesis company online, they'll synthesize it and send it back, and I can go do the experiment. It also has features not found in other tools—I can say "build me a library." If I have multiple genes and want many different CRISPR targets, I just type the gene names here, click "design library," and it generates different CRISPR molecules for me.

And it's a native macOS application built with Swift, and I'm planning an iPad version.

Romain: Thank you so much for sharing. This is a wonderful behind-the-scenes look at your work and how you think and work with Codex. I've never seen Codex used in this way before.

Digital Twin: The Future of Personalized Medicine

Romain: You mentioned the idea of a digital twin earlier. From your perspective in immunology, cancer, and biology, when will this idea become feasible? Why do we need a digital twin? Why do we need AI?

Derya: We need a digital twin because our biological systems are extremely complex wholes—not just what you can measure from the outside. The immune system I just showed, plus metabolites, plus the trillions of bacteria in the gut, plus hormones—it's an incredibly complex system. Your genetic makeup, combined with your environment, determines almost everything—whether you'll get sick, when, whether you'll respond to a treatment.

Can we predict disease before it happens? And it must be highly personalized—we should treat the patient, not the disease. Because biological systems are so complex, we've been giving the same drug to millions of people with the same disease. For example, statins are used by millions, but they only truly work for a small subset.

If AI can fully simulate your biological system—your genome, your metabolites, your proteins, your immune system—then we can start asking: What would happen if I change this aspect of health? What would happen if this person takes this drug? Maybe I can tailor a treatment for you directly based on what AI tells me about your biology. We'll enter a fully personalized era—drugs can achieve nearly 100% efficacy and nearly 0% side effects. That means clinical trials that now take 5 to 10 years could accelerate to maybe just 5 to 10 days. AI will run clinical trials for you.

That's why I say in the next ten years or so, we'll be able to cure all diseases. In another 15 years, we'll reverse aging, and people will be able to live for hundreds of years. People say this sounds crazy, like science fiction, saying that just cancer took 50 years and only made a little progress. What they're not accounting for is that AI is progressing exponentially. And all this also depends on one prerequisite: computing power must be drastically increased, because even if you combine all the computing power in the world today, it's not enough to simulate a biological system—there are too many components.

If we can reach that level in the next 5 to 10 years, superintelligence will also emerge by then. At that point, we can simulate digital twins with AI, not experimenting on humans, but running experiments on your biology in the AI. This will change medicine and everything.

Romain: Using a cancer patient as an example, if we had a digital twin, what could a doctor do that they can't do now?

Derya: Basically, try different hypotheses and experiments on the digital twin to see the response. It's like having a control group and a treatment group.

In fact, cancer and oncology are currently the closest to personalization because even with the same cancer type, there are many different mutations. If you're a lung cancer patient, your oncologist will first sequence the mutated genes because different mutations correspond to different drugs. For example, about 1% of lung cancer patients can use a specific drug, and that drug is very effective for that 1%, but not for the other 99%. Companies are developing targeted precision drugs. We can even create a drug specifically tailored to all of your mutations.

There's a case in Australia where a computer scientist used ChatGPT and Grok to design an RNA vaccine for his dog—this is ultimate personalization because that RNA vaccine was custom-made for those specific cancer mutations. Trials are ongoing.

The immune system is extremely effective at killing cancer cells—that's the immunotherapy revolution. But immunotherapy doesn't work for everyone. Why can some people's immune cells recognize and kill cancer while others can't? Some immune cells become exhausted, etc. And there are side effects—the immune system itself is dangerous; overactivation can cause a lot of damage. If we can figure all this out, we can truly achieve personalized treatment.

Persuading Colleagues to Embrace AI

Romain: You're both a heavy AI user and an MD. What's the attitude of people around you towards AI? Do you try to get them to adopt these tools as quickly as you?

Derya: I've tried. I think they thought I was completely crazy. But now they're starting to see the potential. I started saying these things from GPT-3.5 onwards. People are very hesitant, and I understand that because this is such a new thing. The human mind cannot grasp this exponential progress.

Many people used GPT-4.0 a year and a half ago, which in AI terms is a long time ago. They say "it hallucinates too much, doesn't answer well enough." But the gap between GPT-5.4 and 5.5 is already night and day. If you keep experimenting and trust that it will get better...

Now, even in a field I've researched for 30 years, I trust the answers from AI models. Recently, GPT-5.5 Pro gave me a report that almost made me cry. How is this possible? GPT-5, Pro, 5.4 already had excellent understanding and pattern recognition of knowledge, but 5.5 did something almost as if it had the same experience I've gained in 30 years in the lab. Because some things are intuition, not in the literature; you just know—like I'd bet with my students, "Do this experiment, I bet it will turn out this way." Sometimes they bet with me and lose 100% of the time, because that's intuition honed from massive accumulation.

What 5.5 did was predict the outcome of an extremely complex experiment we performed, with 100% accuracy. That's unbelievable.

Science 2.0: The AI-Driven Future Research Paradigm

Romain: If this pace of progress continues, what will your daily work look like in a few years? What will fundamentally change for you and researchers around you?

Derya: Some things I say may sound radical. But there will be a complete fundamental shift—I call it Science 2.0 or 3.0. The way we do science will completely transform.

The old model—taking weeks to think of an idea, designing experiments, then months to analyze data—that era is over. Students and scientists must realize we are in a dramatically accelerated time scale. The future of doing science is that a group of AI agents will help you generate hypotheses. They can already generate hypotheses because the number of ideas that can be produced is nearly infinite. Then AI will help you simulate experiments—I can do 1000 experiments, but I don't know which will succeed. If AI can tell me that this type of experiment is more likely to succeed and why, I can focus on that, and the success rate will skyrocket. Once data comes out, it's immediately passed to other AI agents, they analyze it instantly, and feed back to the master agent, which then proposes new hypotheses and designs new experiments.

I think my role will become simply telling the agents: I want to tackle lung cancer, go study this direction, explore. And then someone still needs to do experiments, but I think labs will also become automated—it's already starting, with lots of robots doing a lot of wet lab work. When people ask "Will I still have a job?"—there's the Jevons paradox: if we can do this 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 ability, just as I'm building applications now, I'll be able to build new cell types, new tissues. There will be thousands of bioengineers sitting at computers simulating and building. This will fundamentally change not just biology, but also physics, materials science, chemistry—drug discovery: what used to take years can now be done in hours. Clinical trials, doctors treating patients—the entire chain will change and accelerate.

Advice for People in All Fields

Romain: For people not in the scientific field, based on your experience, what advice would you give them about how to rethink their own fields and work?

Derya: I have an advantage—my work is essentially continuous experimentation, and it's extremely painful because in biology, 95% to 98% of experiments fail. So I'm very used to failure. That's why it's called "experimentation"—you keep trying, keep adjusting. You develop a resilience, autonomy, and curiosity—the impulse to just try. That's why I'm so excited about Codex—of course, I show you the things that worked. There were many failures, many applications that didn't work well, but you shouldn't give up.

My advice to everyone is: In the AI era, the only things that truly matter are autonomy and curiosity. Don't be afraid. Keep trying, keep experimenting with AI. Ask the "what if I do this" question, because now you can ask it—before, the cost of doing that was too high.

For example, your company's website—maybe you spent thousands of dollars making it, it's not perfect, and you say "well, good enough." Now you can say "what if I change this a bit," and in minutes you have a new website, or you can design a new product and 3D print it.

I think this applies to everything, but you need the courage to experiment because the cost of experimentation is now very low—why not do it? Don't just treat this as small everyday things; you can really extend it to almost any aspect of life. You just need to embrace it and see AI as an extremely positive thing.

I see a lot of negativity—"AI will do this, AI will do that." My view is exactly the opposite: it will truly usher us into a golden age. AI researchers are heroes to me because this will be humanity's greatest transformation. I am immensely excited about the future.

Romain: Thank you very much, Derya. That's a wonderful closing statement, full of positive energy. We can't wait to see what you do next, how you further push Codex and frontier models, integrate these fields, and advance the digital twin vision.

Derya: I'd be happy to come back for another episode in a few months. Welcome me back and see how much progress has been made. Until then, have fun in California. Thank you very much.

Risk Warning and Disclaimer

        Market risk exists. Investment requires caution. This article does not constitute personal investment advice, nor does it consider the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their particular circumstances. Any investment made based on this is at your own risk.