New in Nature: Google Agents Beat Human Doctors, Make Scientific Discoveries – With Vivek and Anil

Today's episode features an eye-opening conversation with Vivek Natarajan and Anil Palepu from Google DeepMind. Their groundbreaking work on AMIE, the Articulate Medical Intelligence Explorer, and AI co-scientist represents what seems to me an important threshold moment in AI capabilities. I always say that if people truly understood what AI can already do today, many would be fundamentally rethinking their plans.

These projects provide perhaps the clearest evidence yet that AI systems are beginning to outperform highly intelligent humans in domains that require years of specialized training. Remarkably, this work was accomplished without special continued pretraining or extensive custom post-training that could only have been done within Google. On the contrary, these approaches could have been developed and can be replicated by Google's API customers using commercially available models, advanced prompting techniques, and thoughtful agent design.

We begin by discussing AMIE. A year ago, the AMIE co-authors showed that AMIE was able to outperform human general practitioners in diagnostic accuracy. Now, with just a few important caveats remaining, Anil and the team have demonstrated that it also beats human primary care physicians in analysis and treatment recommendations.

The implications for healthcare access are obviously profound and are beginning to extend into specialized medicine, too. The second AMIE paper we cover shows that the AI system is already surpassing medical fellows in both cardiology and oncology and closing in on, but still falling a bit short of, attending-level performance. Notably, when cardiologists have access to AMIE, their performance dramatically improves across almost every metric.

This suggests a short- to medium-term future in which AI doctors have the potential to both raise the floor for access to quality care globally and raise the reliability ceiling even for those of us fortunate enough to have access to first-world specialized care. This is, to put it plainly, crazy. I am super excited that Google is moving AMIE into something like a clinical trial in partnership with Beth Israel Deaconess Medical Center, a Harvard Medical School teaching hospital in Boston, for real-world validation.

All that said, in Vivek and the team's AI co-scientist paper, we see something equally, if not even more, amazing. This multi-agent AI scientist system, which is capable of accepting human input and feedback at any step in its process, was tested in fully autonomous mode on 3 increasingly complicated scientific challenges. The first was drug repurposing, an advanced but reasonably well-defined task amenable to computational analysis.

The second was therapeutic target identification, a more open-ended challenge requiring the AI to understand or make quality hypotheses about causal relationships within cells. The third, and definitely most daunting, was the wholly open-ended challenge of understanding the process by which bacteria achieve drug resistance. As you might have guessed, AI co-scientist—which, by the way, Google is now making available to trusted partners—succeeded on all 3 of these tasks.

On the challenge of understanding drug resistance in particular, it blew everyone's minds by proposing the exact same mechanism that Google's independent scientific collaborators had recently discovered experimentally but had not yet published at the time of AI co-scientist's analysis. Overall, AI co-scientist demonstrates that AI systems are now capable of generating novel insights by connecting the dots between far-flung bits of hard-won human knowledge. This system is not simply regurgitating its training data.

On the contrary, it is performing meaningful synthesis and proposing novel hypotheses that even human expert scientists recognize as both insightful and significant. If that's not enough for one episode, the implementation details behind these systems offer valuable lessons for AI engineers everywhere. First, structured reasoning proves far more effective than simple chain-of-thought approaches, especially when working with lots of input context.

Both of these systems demonstrate the value of thinking carefully about exactly how you want your AI system to reason about specific types of problems. Second, finding ways to add new information, or even just a bit of entropy, such as by giving the model access to search, is key to making self-critique and self-improvement schemes work over many rounds of successive iteration. Third, for now at least, the tournament-style evaluation process used to surface the best candidate hypothesis out of the many that were generated seems to be an industry best practice that you can and should use in your own work.

What's most amazing to me about all of this is that it was achieved before Gemini 2.5 Pro was available to use, meaning that everything we talk about today is still subject to a step-change improvement that should come more or less for free with a simple model upgrade. With this level of performance already established and core model progress continuing, the path to an AI doctor in your pocket and data centers full of AI geniuses is honestly becoming quite clear. AIs are no longer just tools for routine tasks.

They are becoming legitimate thought partners in some of humanity's most complex intellectual endeavors, from diagnosing disease to expanding the very frontiers of scientific knowledge. Whether we get to live in a post-scarcity society, in which we all enjoy instant access to superhuman AI doctors, or perhaps, on the other extreme, end up going extinct due to some crazy AI-driven scientific accident seems to me to depend largely on how responsibly we handle the upcoming AI transition. I take my role in AI discourse very seriously.

Vivek Natarajan and Anil Palepu, authors of AI co-scientist and AMIE from Google DeepMind. Welcome to The Cognitive Revolution.

Thanks for having us.

My apologies, guys—again. So, Vivek, this is your 4th time. What an unbelievable heater you and the team at Google have been on. I always say that if people just had a little bit better sense of what is already out there today, they would be updating their plans in many ways that I just don't see people doing.

This is really an unbelievable example of that. There are 3 papers we're going to go down the rabbit hole on today, and I came away from this feeling like it might be time to call it. The AIs might, in fact, now be clearly smarter than me. We can get into the nuances of what they're still missing a little bit, but I think almost everybody would read these reports on what you've been able to get the AIs to do and feel like they would have a very hard time matching that.

To get to this level, a single AI model scaffolded in different ways and put to different purposes—it would be years of undertaking for me to get there, for sure. So, let's do some headlines. There are 2 papers with AMIE. We covered this once before. This is the—what is it?—the Articulate Medical Intelligence Explorer. Sorry, the Articulate Medical Intelligence Explorer.

I've been using a graph in some slides that I occasionally present for the last year or so, since the first AMIE paper came out. It shows that when a patient chats with the AI, the AI is more accurate in its ability to diagnose the person than human primary care physicians are, as judged by other human doctors. Now we've got 2 new extensions to that.

The first one is—there are some caveats here that I think are definitely worth unpacking, and I'll give you a chance to do that—basically, it's now outperforming general practitioners not just on the diagnosis part, but also on reasoning through what to do about it and ultimately recommending treatments, which is obviously a big part of what the doctor is meant to do for you.

I mean, it's crazy that that is out there in the world today. I swear, when I was a kid, if something like this happened, it would have been headline news. Everybody would be talking about it, and there's just so much going on that some of this stuff, even as dramatic a breakthrough as it is, doesn't seem to crack the consciousness. So, tell us more about AMIE outperforming now not just on diagnosis but also on recommending treatments.

Yeah. Prior to those first 2 AMIE papers, a lot of the work in the space was on medical question answering. There was some notion that these language models do encode clinical information well and that they have a lot to offer.

I think with those papers, we were trying to start to ask the question: Okay, but that's not clinical practice, right? In clinical practice, the doctor is interacting with patients. They have to gather this information themselves; they're not really presented with all the information up front.

That was the study. It was doing this objective structured clinical examination, this format, and it was basically trying to see whether the doctor—or the AI, in this case—could interact with patients, gather that information, and still get to a diagnostic endpoint. Of course, I think we do a really good job in the paper—and I encourage people to read it—of describing the many limitations.

This is a text-based chat; that's not how doctors talk to people. Our future direction from there was really about starting to unpack some of these limitations. One of those limitations is this idea that it's more than just seeing a patient at the first visit and diagnosing them. There's a lot more to clinical care, right? It's about managing a patient over multiple visits.

You know, the endpoint the first time you see a patient might really be, “We need to order the right test and set them in the right direction.” It’s not always that you know exactly what to do with the patient after seeing them 1 time. And so I think the management reasoning paper is really trying to unpack that.

We’ll talk more about how the study was designed—I think it’s super interesting—but there are other aspects to it as well. Rather than more general recommendations, can we get really precise? Can we ground these in accepted clinical practice guidelines, ground them in medication labels, and start to turn these into slightly more actionable things?

Similarly, with the specialty papers, it’s trying to expand beyond the bread-and-butter common presentations of common diseases. How does this work in more niche areas of medicine? We have a lot more work trying to expand on some of these limitations that we’ve identified in the first paper.

So how would you summarize—I mean, if I understand correctly, it is still a chat-based interaction today, right? One major possible extension would be to go to multimodal, but you guys have also done work on that separately, right? Why is it not multimodal in this particular study? Was there a reason not to just let people throw in selfies into the chat?

Yeah, I think it’s more that when you’re doing research, you want to isolate the components that you’re studying and do that well. When you add multimodal, it necessarily adds more confounders. We tried to avoid that, so it’s just easier to study a text-based system to begin with.

But clearly, we’ve done work on multimodal before. I know that on 2 podcasts we’ve spoken about the Med-Gemini work. All those components and pieces exist, and very soon you’ll see the multimodal one come out as well.

Yeah. The pace is relentless, that’s for sure. So, just to bottom-line this 1 more time, basically we have—with the caveat that it’s not yet in this particular paper, although it’s coming soon—a multimodal, multivisit, longer-time-horizon interaction between patients and doctors, where the AI doctor is outperforming human doctors on both the diagnosis and the reasoning through what to do, and ultimately landing on standard-of-care, accepted, proper treatment for these conditions.

Anything else that we should—should I soften that at all, or is that a good summary?

I mean, I think that should be on billboards, right? I think the biggest thing is—and we’ll talk more about how we’re trying to test this in the real world—but these are simulated consultations. They involve patient actors; they’re not real patients.

Obviously, there’s a whole new set of challenges when you get real patients, where all kinds of things can happen and they’re not necessarily going to stick to their script. So I think that’s a whole other thing that we need to test and validate: that our results truly do translate to the real-world setting.

That being said, I think simulated consultations do show that we have promise in this setting, and we personally are very optimistic that these results would translate, at least to a certain degree.

Yeah. Okay. So the next AMIE headline is moving to specialized medicine. There, you look at cardiology and oncology, and I would summarize the findings as the AMIE system is surpassing fellows and closing in on, but not yet hitting, the level of attending physicians in these specialist domains.

What further complications or caveats should we have to understand that?

Yeah, I think largely I’d agree with the notion that we have a lot of improvement in terms of being as consistent in all domains as the most experienced attendings, and there’s room for improvement.

I think the real headline for me, especially if you look at the cardiology paper, is that the types of errors they make are pretty different between AI and the general cardiologist. The really exciting thing is that we see that they’re quite complementary.

The comparison we made was when we compared AMIE to the general cardiologist head-to-head, there was some uncertainty about what was better: some were better in some areas, some were better in others. But when we compared the general cardiologist with access to AMIE’s assessments to the general cardiologist alone, it was a landslide. In that case, in almost every aspect, it was considered superior when they had assistance.

I can talk a little bit more about why I think that might be. Obviously, I think more investigation is needed, but I think that is a really exciting aspect of this: it seems to be just a helpful system in use by these experts.

Yeah. And maybe to just contextualize that work a little bit more.

I think if you look at access to specialists in the country today, I believe getting consultations with a neurologist, for example, has a 12- to 18-month wait time, and that’s simply not sustainable.

The question is: clearly, we have better-reasoning AI systems that seem to show promise in medicine, so how can we do better? We should be able to improve the status quo radically. No one should be waiting 18 months for a consultation on something rather serious. Neurology is not straightforward.

That’s the motivation: there are a lot of access issues around specialist care and cost issues around specialist care. How can we do better? That’s the key question we’re trying to address.

Then I think the second thing is, generally, if you look at how medicine has evolved, that has led to these silos, or compartments and specializations. You have primary care, which is kind of the front face to everything else—it’s the door—and then you have all these silos.

But the way it has evolved, I think, is primarily because of the limitations of the cognitive aspects of the human mind. There’s only so much expertise that we can cram into our given brain. So, because of those limitations, you have to study neurology or cardiology or internal medicine, but not everything together.

AI systems don’t need to have that kind of limitation. Given what we’re seeing, they should be able to integrate knowledge from multiple different sources and multiple different disciplines. There’s a fundamental rethinking that’s happening: does the new age of AI-powered healthcare need those silos?

It could be possible that you not only have a PCP in your pocket, but an expert neurologist in your pocket, obviously with caveats and things like that. That’s another class of question that we’re trying to address here.

And maybe the third point I’ll add on is that this study is, what, 3, 4, 5 months old now?

Yeah. That was primarily done with the PaLM version of the models, if I’m not wrong. The cardiology one was done with Flash, I think.

Flash 1.5.

Yeah. 1.5. Okay.

Yeah. So it’s 1.5. And since then, we’ve had 2 and 2.5. In that study, obviously, what we saw was that 1.5 was not as good as the attendings, but who knows with the new models.

Yeah. That’s an important caveat.

Okay. Well, let’s do the headlines for the AI co-scientist, because these are similarly striking headlines. AI co-scientist is basically, I guess, what I would describe as an agent-scaffolding-type setup. We can get into the granular details, but we’ve seen different things like this before. I did an episode with James Zou, who has a similar thing with the AI Virtual Lab, and the original Coscientist was made within the first couple of weeks of GPT-4, which is crazy to think about. They got something out of that with around 8,000 tokens of context.

Even then, we were seeing some interesting stuff. I would describe this as taking all of the lessons learned about how to make agents work from the last 2 years, hitting the gas on all of them, and then coming up with a system that can basically do science. It’s not executing the actual physical experiments at this point, but I was really amazed by the different things that you tested the system on.

The first one is kind of the first of 3 levels of challenge for the problem that the system is given. The first one is a relatively well-scoped thing that you could grind through. That might be the one that I could do with some real effort, and that was drug repurposing. Take a drug that’s out there, look for other things that it might be useful for, and you have a combinatorial approach available to you. If you set that up and go through it systematically, with decent judgment around each individual sub-question that you ask, you can imagine how an AI system would be able to do something like that.

Then you go up a level to the second scope of task, which is identifying new therapeutic targets within a particular kind of diseased cell. This is starting to get at what I’ve often called one of the grand challenges in biology: what causes what. We’ve got obviously super-complicated causal graphs going on in the cell—this promotes this but inhibits that, and so on—and we still have a long way to go in terms of really mapping that out. I don’t know if you guys would venture a number, but I’ve broadly understood that we may understand 10% of what that graph is in today’s world.

The challenge here is to go into a cell, so to speak, and try to figure out whether we can identify something that we can target with a drug that will actually make things better. Obviously, this is hard given the vast complexity and the many unknowns in cells. That was just the middle one.

Then the third one is: can you figure out why or how bacteria are becoming drug-resistant? There was a little bit of a hint, because there was one observation that was conserved across a couple of different species. That was a notable observation that served as a seed to unpack the challenge. But beyond that, you had to go figure it out. It was super open-ended and really, really tough.

That is a daunting question for me to consider. I would have you spinning your wheels in absolutely vast literature for a long time before you would even have any sense that you might be able to start to contribute to the discussion. The bottom line is that AI was able to do all 3 of those problems pretty well.

In that last case, it actually surfaced as its number-one candidate idea for the mechanism of the drug resistance something that had been discovered experimentally but was not yet published. You partnered with an academic group that was doing this research, and they had the answer, but nobody else had the answer. It wasn’t in the literature, and the system was able to grind through all the content that it had available—the whole vast body of medical and biological literature—and land on the exact number-one hypothesis that turned out to be the actual answer.

I was really blown away by just how successful it was. Tell me more about what else there is. Are there any caveats that I glossed over, or any other Eureka moments that you would want to highlight from those results?

Yeah, I think the last one was interesting and also funny, because I don’t think José and Thiago, who are collaborators at Imperial, would take offense at me saying this: I don’t think they really believed that AI could do this thing. We had the system, and José wanted to try it, but it took us a few months before we got enough time with them. After enough pestering, they were like, “Okay, we have some experimental results in the lab. We’ll try to challenge your AI system. Let’s see what your system can do.”

It was roughly around Thanksgiving when they sent us this prompt. I think that’s detailed in another preprint, along with the “Towards an AI Co-Scientist” paper. As you said, there were some clues in there, but it wasn’t totally giving it away. We took that prompt, sent it to our system, and it ran for a couple of days before spitting something out.

Yury, the first author on that paper, is an amazing technical fellow. He’s probably the least well-known technical genius at Google in some ways because he likes to keep a low profile. He sent it over to them. I think it was Thanksgiving, late in the evening, and Yury and Allan were both based in Europe, so they were offline. Then, within 10 minutes, José, who was also based in London, sent us an email saying, “I need to talk to you right now.”

I didn’t understand the seriousness of it, but I said, “Okay. I’m not doing anything better. We can talk.” Then he was like, “Vivek, are you reading my email?” I said, “I’m not sure what you’re asking.” He said, “No, no, no. It seems like you’re reading my email.” I said, “We do many things at Google, but reading your email is not one of them.”

Then he explained. He said, “I’ve not published this anywhere, but your AI system came up with the same set of results that we hypothesized and found in our experiments. So I’m really, really surprised.” Then he asked, “Do you guys get responses from ChatGPT?” I said, “No, no, no. That’s not possible. That doesn’t happen.”

He said, “Okay, so if you’re not reading my emails, and if you don’t have any information from ChatGPT, then it’s likely that you have something really, really magical.” That was kind of his response. Then he said, “The first one is great, but we also sent you 4 more.” All the other 4 also made a lot of sense.

Immediately after the Thanksgiving break, he said, “Okay, I’m going to set a few of my postdocs to work on this.” They’ve been working on validating those other hypotheses. It was that moment when someone who’s very pragmatic and very experienced, and who has spent a decade—in fact, several decades—in the field has that kind of reaction. That told us, “Okay, we might be on to something here.”

Again, I don’t think it was one single moment. It was more like, okay, this is a very hard thing to do. To get AI not just to synthesize, integrate, and summarize information, but to help traverse the history of knowledge and uncover new, original things and facts about the world—and to do that reliably—is a super-hard problem.

In some ways, that’s the holy grail of AI. To think that a system that is relatively simple in nature—we’ll probably talk about this more—is probably the simplest version of a system that you can imagine, and that by throwing a bunch of compute at it, you’re already seeing evidence of this happening reliably, felt super-magical to us.

Yeah, it’s crazy. This is something that people have been discussing online recently. Dario has advanced the idea that, given AIs’ incredible breadth of knowledge, why don’t we see them coming up with more connections and more insights across this super-diverse knowledge base? His contention is, “If I could have all that knowledge, surely I would come up with more insights than the AIs currently seem to.”

And I feel like some of the things that I've covered definitely seem like they could be said to count, but there's always a lot of details and caveats and sort of eye-of-the-beholder type of stuff. But this seems pretty clear to me: the fact that this was essentially independently discovered by human scientists in a lab and an AI system in a data center over a couple days in parallel, to the point where the scientists accuse you of reading his email—presumably not super seriously, but nevertheless—is there any reason that we should not take this as a genuine discovery, a sort of qualitative eureka moment from an AI system?

Yeah, I mean, this is not the first evidence. I think right in 2023, right after we did our Med-PaLM—and that's where, actually, the genesis of the AI co-scientist work is, right after we published Med-PaLM—there was this professor from Stanford, Dr. Gary Peltz, who reached out to us. I think you probably remember Tao, who came to one of these previous episodes before. He called both of us up, and then he was like, “Tao, you don't know me, but your AI system can potentially help millions of people with rare diseases.”

I was like, “Okay, Gary, that's a nice introduction. Please go on.” And then he was like, “I know that your models are trained on a lot of scientific literature, and I think they can help me discover useful facts about genetic diseases.”

At first, that felt interesting. If you were able to help him, probably we could help a lot of people, and so we started working with him on this problem of genetic discovery. Can language models come up with the right kind of causal factors that are responsible for a given combination of phenotypes or symptoms?

We started doing that with Med-PaLM and later with Med-Gemini. With Med-PaLM, Gary was working on mouse models and a very specific kind of hearing loss for which he had this NIH grant back then. I don't think you'll ever get that kind of a grant again, but that's a separate discussion. One of the hypotheses that the model came up with was a digenic model for hearing loss, which Gary had not thought of before.

He went ahead and did CRISPR knock-in experiments in his lab, and he was able to reverse the course of the disease. We had written that up, and it's still under review at a prestigious venue. Later on, we used our more advanced version of our LLM, which is Med-Gemini, and extended that work to human variants of unknown significance. Even there, based on retrospective data, we see these systems are able to do pretty interesting work in genetic discovery, which is one way it can be cast—as a hypothesis-generation problem.

The key thing is that, at that point in time, when we were using these LLMs in a pretty crude, single-shot fashion, they were very unreliable at hypothesis generation. Literally, for Med-PaLM to come up with 1 hypothesis that was very helpful, it had come up with thousands of things that were utter garbage.

We were very grateful to work with Gary, who had the expertise to very quickly discard things that were nonsense and also had the patience to work through all of them. You could easily imagine another scientist, or maybe someone inexperienced, looking at the first 5 and saying, “This is utter garbage. It's not working,” and then we would not have pursued this line of work at all.

It's about getting together and working with the right people who believe in this, right? That put us on this journey: How do we make this more reliable? We should not be sampling thousands of times to get something useful. Rather, every single generation should be something useful, and the system should be well calibrated.

The simplest way to do this would be to call an LLM repeatedly and hope that it leads to something useful. But that very quickly fails and leads to degenerate solutions, which have this mode collapse.

It became important to introduce net-new knowledge into the world and also think about how you can gamify that process in some ways, and introduce repeated, new, helpful feedback that can help the system self-improve. That led us to the design that we eventually had.

In hindsight, it should be quite obvious. It very naturally follows how the scientific method works. If you ask a scientist how they come up with new ideas, you'll see that they will roughly compartmentalize into the set of agents that we had. They're just doing the same thing, but it was this iterative process of, “Okay, we tried something, it didn't work. We tried to improve it,” and we eventually ended up with this design, which I think is remarkably intuitive.

Yeah. How do you think about the relationship between hallucinations, on the one hand, and creativity or hypothesis generation, on the other hand? It seems like people have very different intuitions about that.

I'm also not quite sure myself: When we do extensive post-training to try to minimize hallucinations, does that help with hypothesis generation because maybe it makes the models more disciplined in their reasoning? Or, in some ways, does it hurt because they're less willing to come up with a very random idea? Every once in a while, it's these random ideas that are the big ideas, right?

Maybe that's a false trade-off, but how do you guys think about the relationship between those behaviors?

Yeah, I mean, intuitively, there's some aspect of hallucination that does foster creativity in some ways. It's interpolating between the data seen, right? And it's kind of necessary for hypothesis generation to deviate from the script a bit.

I don't have a clear understanding or answer to that. Intuitively, at least, that's how I think about it.

I always used to think that hallucination and creativity were 2 sides of the same coin in some ways. I don't think this holds true any longer, but what we used to see with the previous generation of models was that it was much more helpful to use the models without post-training for this task because they were most likely to come up with these crazy ideas.

But now I don't think that's any longer the case, because we've been able to systematize and put a structure around that process of coming up with new ideas. It feels like now the process is much more reliable, but maybe we are sacrificing some crazy new things that would require a non-post-trained model to come up with. We don't know.

We've gained reliability in the process, in terms of consistently coming up with new, original thoughts, but we don't know if we're sacrificing something else over here.

Yeah, and I think one advantage of our system, too, is that we allow diversity of models. We can get the best of both worlds through this tournament process and through ranking. We have a diverse set of hypotheses, and those can be reranked. Hopefully, we see high-quality hypotheses bubble up to the top.

So, yeah, I think maybe it's both is what I take away from the hallucination reflections there. If you're in a context where there's a rich literature and it's more about really working your way through it, then maybe you want a disciplined reasoner.

And if you're doing something where there really isn't much to go on, maybe you want a sort of more whimsical hallucinator. And as you said, maybe get the best of both worlds with some of these setups.

So, let's describe the different setups. I mean, there are multiple different systems here that have their intricacies, but it seems like, at a high level, if I understand correctly, what you're doing to design these systems is basically just introspecting or maybe interviewing people and asking—one of my favorite questions in AI automation in general—“How do you think about it?”

How do you think about it? What do you do next? Then, once you do that, what do you do? You're basically turning all these steps of a process, whether it's the scientific method, the diagnostic process, or the reasoning-to-treatment process, into a map with a subject-matter expert, yourself, or somebody else. Then you're creating little sub-agents that are prompted to do those subtasks and scaffolding them together.

There are also some interesting details around giving them tools. Literature search is obviously a huge one. I was also interested to see that, in one project, the model had access to AlphaFold and maybe some other things. You're increasingly giving it something like the full complement of tools that a human could use.

Then it seems like, after that, it's about turning all the hyperparameters up: doing more rounds, more generations, more evaluations, and more rounds of feedback. My impression is that if you do that and have the budget for it, in today's world you could probably be successful at almost anything. But maybe tell me if that's wrong. Are there things about the designs of these systems that you think are actually very important hinges, where if you designed it a little differently, it wouldn't work? Within that general framework that I outlined, could you go in a bunch of different directions, or do you feel like it's actually a narrow design space that works?

Yeah, I think maybe there's a high-level narrative change that's happening here. This is my fourth time on the podcast, and I think previously we discussed Med-PaLM, Med-PaLM 2, and MedGemini. The key with all of them was that they were all medical models: we were taking some generalized model and trying to fine-tune and specialize it.

I think the key differentiator with the current version of AMIE and the co-scientist is that we're no longer trying to do that fine-tuning step or specialization step. Part of it is because some of the data that went into that fine-tuning and specialization step is now upstreamed and part of Gemini. It just feels like a better approach to set things up by simply having agents with specialized prompts and chaining them together in a nice manner. It gives us a lot of flexibility and control, and it does away with the need for maintenance and specialization.

I'm not sure what you think about that, but I kind of agree. Not that there's no role for post-training, but certainly, in our first AMIE paper, in Med-PaLM—all these papers—it was really about the medical data we were curating or creating, how we were making that, and how that was driving the model's success.

In these latest papers, that's really taken a backseat to simply how we're designing the system to perform these sorts of tasks at inference time. I don't think anything is particularly hyper-optimized. There are probably ways that someone could design a system that does this stuff better. I'm sure there are.

I think our goals have largely been to build a functional prototype for these systems, get there quickly, and do the kinds of tasks that we're interested in doing in each study. So I think there is so much room for improvement in all of these systems.

How many rounds of iteration did you go through as you built the system? Were there any moments when it was, “Oh, we tweaked this prompt,” or, “We switched the order, or put this agent before that agent,” and kind of rescaffolded it that way, and all of a sudden saw a big leap? Tell me about those.

For the AMIE work, we rely really heavily on auto-evaluations. We try a lot of different prompts and configurations, but ultimately, there's only a finite number of things you can try before you're like, “This seems good. This is good enough.” We use that as a rough signal. We also rely pretty heavily on vibe checks, and you can pretty quickly start to see differences.

With our management-reasoning agent, for example, we saw a big difference when we started drafting concurrent plans and refining them together. We saw a big difference when we used this top-down approach.

Can you unpack that just a little bit more? What was the before, and what was the after?

Before, we were just generating 1 plan. Now we're generating 4 different plans. They might have some similarities and some differences, but we found, in a sort of self-consistency-style manner, that the model was able to combine these plans in a way that allowed it to take the good stuff from each plan and leave out the bad.

Once we tried it, our internal auto-evaluation signal was very clear: this was making a huge difference. But in terms of hyperparameters, we went with 4 plans. We could have gone with 8, or we could have gone with 2. I don't think we optimized every hyperparameter in that sense, but I think we were able to pick up the big things from this kind of signal.

Yeah, I think on the co-scientist side, it was a much longer iteration. It was close to an 18-month project in some ways, and it was driven by this need to make the process of hypothesis generation more reliable.

We decomposed these tasks and tried to get individual models to work on them. The good thing it showed was how good models are getting at instruction-following, and that also does away with the need to fine-tune and specialize them. If a model has the knowledge and you give it a precise set of instructions, it can just follow them.

That makes it much easier to create agents and chain them together, where these agents are specialized to do specific tasks but are just prompted versions of general-purpose models. It was a process of setting up an evaluation, seeing how well the system did, identifying its weaknesses, and then iteratively fixing those weaknesses by adding specialized versions of agents or fixing the prompts.

Ultimately, there was a process of simplifying the architecture: we realized we didn't need all of these things. That ultimately led to this design. So I would say it was a lot of iteration—figuring out where the weaknesses were, trying to cover those weaknesses, and then, at the end, making a push to simplify.

I've been through that on a less world-changing project myself multiple times, where a new model can do wonders for the simplification of your system. I've definitely lived that.

I guess, in terms of what is driving the improvements, 18 months ago we maybe couldn't—it's all happening so fast. I even have some lingering sympathy for the stochastic-parrot people because, as of GPT-2, that was probably still mostly true. At this point, it's pretty clearly not, but 18 months ago, models maybe just couldn't do certain things that now they can.

It seems like core model progress is the tide that's lifting all boats dramatically. Is there anything else that you think is super important, or is it really just down to foundation models getting better and grinding out the process of figuring out how to use them?

Yeah, again, on the co-scientist, I feel like long context was an important part of it as well. We don't have an explicit memory store in the system, but the fact that Gemini models, when instantiated, can take up to 2 million tokens means that we can generate ideas, generate reviews of them, run debates of these ideas, and generate these walls of text, essentially. That text can become feedback, and we can put all of it back into the context of the model in the next round.

The model figures out how to make sense of it and use the feedback in a very implicit manner to improve. If you didn't have these long-context abilities—the ability to reliably reason over millions of tokens—you wouldn't be able to do that. You would have to engineer RAG-based systems and maybe end-to-end train them, which would have been a lot more complex.

All of that has been made remarkably simple by the long-context ability of Gemini, which I think is a little underappreciated in the field. We don't have enough Studio Ghibli-style viral moments with long context, but it enables a lot of these practical applications.

That's the same with the management reasoning. Our RAG for that paper—the management agent's whole point is that it's trying to take in clinical guidelines and reason over them to produce management plans.

If we were solely reliant on always picking out the right guidelines, we'd struggle. But with long context, we don't really need to worry about it. We take in a bunch of guidelines and, with 256K, we're going to catch something relevant, regardless of whether our retrieval system would have been able to do that. So I definitely agree on context.

I think, just like with AI co-scientists in particular, we're talking about a really long time scale. We have this thing running for days, right? I think that increase in inference compute is offering a lot of benefit as well.

I'm just looking up something a friend sent me in the last day or 2: InfiniteBench for long-context deep comprehension, one of obviously many benchmarks that look at these things. Gemini 2.5, even relative to Gemini 2.0, is absolutely crushing on its command of long context. I've definitely felt that in my initial testing of it.

This does feel like something that's hard to go viral. It's hard to go viral with the notion that I had hundreds of thousands or 1 million tokens of context that's all very idiosyncratic to whatever I'm working on. It's hard for people to even know what you're talking about when you post that on Twitter, right? The contrast between the context window and the length of a tweet is pretty severe.

But it's amazing that that was working that well, because just looking at these benchmark results, 2.5 stands out in a massive way relative to everything that had come before. Did you, qualitatively—or based on vibes, or your own sort of sense—how did you handle that? Up until 2.5, I would have felt like, I'm not so sure if I can just dump hundreds of thousands of tokens in. Yes, it can handle it, but does it really handle it? Does it really have command?

Especially if it's material that I don't have full command of myself, it can be very hard to evaluate that. So I guess, how did you know if it was actually making effective use of the super-long context?

Yeah. In the AMIE setting, I think, as I mentioned before, we really relied on auto-evaluation. I have to shout out Valentine, my team member on the paper, who did a great job setting this up. Essentially, we also didn't know how well this would work if we stuffed a bunch of guidelines into context.

Initially, we were going to go all the way up to 1 million, but we found that we seemed to be getting better performance when we dropped it down to 256k. Ultimately, there was a clear difference between the presence of that much knowledge versus trying to do zero-shot, or with 1 guideline, some retrieval, or something. We were also uncertain, but we tested it internally at least, and that seemed to work.

Yeah, I think on the co-scientist side, it might be a little bit more unscientific. What we relied on was primarily the redundancy aspect. You generate some ideas, you review them, there's a tournament that happens, you get a bunch of feedback from the tournament, and then, when you put that back into the state of the system, you're not generating only once; you're generating again—numerous ideas.

Your hope is that, because of that redundancy, at least 1 of the generations would catch the key elements of the feedback that has been generated and propagated back into the system. So I wouldn't say we have any specific evals that target measuring how well the long context is doing. Rather, by engineering in this redundancy, we're hoping that it would be effective.

The other distinguishing factor for us in this work was that I feel like there are a lot of science-assistant-style or scientific-discovery-style projects—not just, I think, in a lot of different places—but where they do get a little bit hung up is in curating these really nice, cozy benchmarks that you can hill-climb on. That was not our philosophy at all.

For us, we feel like the key deal is that if a system does something useful, we should rather sprint straight ahead so that we can validate it and then hopefully take it onward toward a real, meaningful discovery. That was what we were most focused on. If we engineer a system, we go straight to the scientist who's an expert in the field. We show them the idea, and if they like it, we try to convince them to validate it. If they validate it, it becomes a discovery, then, yeah, great.

There was that thing where we really wanted not to micro-optimize on specific benchmarks and hill-climb on them and wait. I mean, I think if we did that, because there are so many different components in the system, this work would have taken easily another year if you tried to make it the best. Rather, we were just focused on, "Let's test them all together, let's get it to do something reasonable, and then we'll do end-to-end validation."

Yeah. Okay. That's a brave new world out there.

One thing that seems to be changing now: not too long ago, I would have said—I could have cited a handful of papers that showed this—that typically 3 to 6 rounds of self-critique and auto-self-improvement seemed to be where GPT-4 would max out. If anything, beyond that I usually felt like I would see performance decline if you kept running it longer than that.

You guys are talking about running these things for days. Maybe you can tell us a little bit about what the budgets are for this. How many tokens are we talking about? If we translate that to retail price, what would the inference bill be for finding the mechanism for microbial drug resistance?

Is there any limit at this point to how many rounds of this you can run, or are we already at the point where you can just run the thing for longer and longer, potentially indefinitely?

Yeah, I think this is a fascinating question. It's something that has also—I don't want to say bothered me, but intrigued me—as well, because I remember reading one of Andrej Karpathy's posts where he talks about leaving a CNN training over the winter break, and then it magically led to state-of-the-art performance on some benchmark. I forget which one that was, but all that he had to do was let it run for 40 days, which is a lot of compute back then, but it was unprecedented in some ways.

For us, with the co-scientist, I think the key thing is the fact that the system is not closed-loop. The fact that it has access to these different kinds of tools means that in every round of self-critique or iteration, it can bring new information into the system that increases the entropy. When that happens, it prevents the possibility of mode collapse and degenerate solutions.

I think that is the key thing. The fact that the system can go and browse the web, search, browse interesting parts of the World Wide Web, take information out of it, integrate that with the knowledge that it has, and do that effectively—that is the part that leads to more computation being spent efficiently and effectively and helps.

It's not just websites, either. Increasingly, we'll be able to get feedback from other kinds of knowledge bases and specialized tools, AlphaFold and so on. As the quality or the surface area of the hypothesis increases, the more different kinds of feedback we'll be able to plug into the system. I expect that it wouldn't then mode-collapse, and there's likely going to be even more increasing value to spending more time on computation in this setup.

But if you were to strip that away, then I think it comes down to the quality of published information in any given domain and the complexity of the problem. Again, I don't know how to put a precise definition to it, but if there's a problem where there's a clear unknown that is impossible to solve, then I think it's very likely that no amount of computation, reasoning, or test-time compute is going to be able to get you that information if the system does not have the capacity to get that information.

On the other hand, there can be problems that are very trivial, where it doesn't matter. You'd probably get it in the first or a few tries, and it doesn't matter. I think there's that sweet spot where the problem is within limits, but it's also not trivially easy, where spending this computation helps.

My hypothesis is that a good chunk of problems that we as humanity care about today actually fall in that sweet spot, where we can spend a lot of computation in silico and get very useful, interesting ideas and answers, which is exciting. What do you think?

Yeah, just a small thing to add: our system also allows humans to input ideas. There are other avenues to add to this entropy. One other important thing to think about is that we're comparing pairwise—every combination of hypotheses.

There is a lot of variability when you compare that to other papers that are talking about self-critique in a more limited way: "We're just going to keep trying to improve the same idea." There are many directions where we're getting this variance.

Yeah, that’s quite insightful. I don’t know if you actually know the total number of tokens for the microbe project, but I’m going to guess it was 10 billion or something.

No, no, I don’t think it’s that bad. It should probably fit within the context limits of these models. I haven’t done an exact analysis, but I would think it’s less than 10 million—10 million total inference tokens for the whole thing. Yeah, I think so.

For the whole tournament?

No, no. I mean, from the time you give it the question to the time that it spits out your answer, what would the API cost be?

It might be a little bit more than that. But we did some back-of-the-envelope calculations, and based on current prices on GCP, we expect that most queries would be just a few dollars—less than $10 for most queries, including the whole tournament and all the inferences. So it should be fairly feasible, and it’s probably just going to come down more and more over the next 6 to 18 months.

Yeah, especially if performance per token is also continuing to go up. There are a lot of tailwinds.

Maybe the other thing I would say is that this is probably the dumbest and most inefficient version of the system that we have. There are so many things here that we could improve in terms of efficiency, and from an intelligence perspective, we can make it much better. I think the bang that you would get on each token generated and dollar spent is going to be much more as we keep improving the efficiency and capabilities of the system.

Are there other narrow specialist models that it has, and how important is that now? How important do you think that will be in the future as a source of new entropy? That seems like a potentially dramatic unlock.

I mean, the model itself is already trained on most of the literature. Searching the literature again at runtime helps with grounding and with things that maybe weren’t in the training data, which it can find after the cutoff date, what have you. But the ability to actually go do simulations and bring that kind of information back—where maybe nobody’s ever run that simulation before at all—that seems like a potentially pretty big step change.

Yeah, I totally agree, and that’s why I feel like this is probably day 0 or day 1 of this journey in many ways. We’ve primarily scratched the surface of information that is just written down in papers, peer-reviewed, and published. A lot of scientific information is just not in that format.

For example, we don’t publish negative results because of the incentives around scientific publication. That’s kind of like dark matter that’s hidden away, and we’ll have to figure out a good incentive mechanism to have that kind of data also flow into the system. I think that’s going to be important as well going forward.

More excitingly, a lot of these papers and publications will have a supplementary data file, for example, which contains giant datasets of experimental data. They contain a lot of useful nuggets of information. You can imagine a system like this paired up with another system, like a data science agent: it could automatically go and analyze the data, and then this could generate hypotheses. The data science agent could analyze the data and get the right kind of feedback back into the system to improve it.

I think certainly with small-scale datasets, that can happen automatically. But I’m even more excited about what we might be able to do when we pair up the co-scientist elements with the data science agent and go after something like what the Arc Institute recently came up with: this Virtual Cell Atlas. It’s like 300 million gene perturbations for 300 million cells. That’s such a vast space. Even for teams of humans, it would take years to explore, look at the data, see the richness in it, and come up with interesting insights.

But we could set up these agents to generate hypotheses, look at the data, analyze it, and come back with feedback. I’m just super excited about the insights that it would unlock into basic biology, target discovery, and things like that when we pair up these systems and set them to work on these giant datasets that we’re generating right now, which even teams of humans find impractical to analyze.

Do you think—okay, here’s a very high-concept question that you may think is totally misguided, or you might think it’s the future.

Obviously, over the last 3 years, we’ve seen a dramatic convergence of a couple of core modalities: language, vision, and speech. For me, one of the early eureka moments where I thought, “I think I’m going to study this subject for the rest of my life, or at least until the singularity,” was when I realized that it’s basically the same architecture doing all these things.

At the time, I was trying to make a video-generation product work, and I had all these different specialist models. But I could see pretty clearly that if these fundamentally similar architectures can do these different tasks independently, then there’s going to be some integration that happens, and the single model will be able to do them all.

Now, that seems to have come with a very viral moment from Gemini Flash and GPT-4o in the last 2 weeks. It seems like that might happen again with reasoning models and narrow specialist models. Right now, you have a model calling AlphaFold and getting results, but there’s a very lossy language bottleneck that happens there.

This would be akin to an earlier language model calling DALL-E to generate an image. That was definitely a point of major frustration for people trying to generate images that looked the way they wanted them to look. Now that this integration has happened, that problem is basically no more.

So the question is, do you think that will happen for these other modalities as well? My sense is that an early superintelligence might be reasoning models akin to what we have today that actually have these other modalities integrated more deeply, so that they’re not bottlenecked through API calls but are actually able to start doing some of this reasoning in biological space, in materials science space, in crypto space, or whatever.

There are a lot of spaces out there, obviously, and it seems like we already see pretty superhuman performance by the narrow models. I’m not aware of any human who can look at an amino acid sequence and intuit what the shape is going to be. There may be a couple of savants out there, but certainly it’s not common. Do you think that happens, and what’s your reaction to that possibility?

I’m not sure. To be honest, I think the API-call thing is clearly something where we’re very close to being able to do that pretty well. A deeper integration—I don’t know exactly how easy that is for all of these specialist-type areas.

I think it actually depends on some of the incentives of the people who are developing these frontier lab models. It makes a lot of sense to combine speech, vision, and language. But if you look at the data that’s going into these models, it’s primarily nonspecialized public data.

When you’re thinking about biology modalities and datasets, they’re not close to the kinds of datasets that are going into these models from the public internet. They’re very different. From some personal experience, what we’ve seen is that when you try to introduce some of these more specialized modalities—even medical imaging modalities, for example, alongside natural images—they can lead to regressions in benchmark performance.

Then the question becomes: are you willing to accept some regression on, say, LMSYS or some other benchmark that’s generally considered important, in order to give your system a little bit more capability in medical imaging? That’s obviously not going to be as big a fraction of your users.

I think it’s a question of incentives right now. I can see this tension in many of the frontier model companies, where introducing these interesting new modalities of information leads to sacrifices in other areas, and you’re sometimes sacrificing benchmark performance.

I would argue that’s a good thing. Benchmark performance doesn’t matter so much. I think you should aim for practical utility. But in the absence of clear measures of utility, it becomes difficult to convince people. So I think it’s not a question of whether we can do it. The architectures exist, and the compute exists at most of these places. But do the incentives exist? That part, I don’t know, and I don’t know when that will happen, because it’s unclear.

Today, I cannot articulate this at a high level, but I think we all agree that if we were to encode the biomedical universe, that model should be able to do a lot of interesting things. But that sometimes conflicts with benchmark performance on LMSYS or whatever else you want to use right now. Then it just becomes a question of incentives.

Yeah, okay. That reminds me of something—I’ve brought this up a couple of times, so I’ll keep it brief.

I heard Yi Tay—he was at Reka at the time—on the Latent Space podcast talk about how the separation between vision and language models was sort of a reflection of the research history. At one point, you would have a language team and a vision team, and then it was like, maybe we can bridge these together. But then you would have late-fusion models because those things would already be sort of done and baked, and you would ask, can we get them to sort of talk to each other via cross-attention or whatever? Then it sort of became, well, if this works, it’ll probably work even better if we do it all just kind of interwoven datasets from the beginning. I can see how that same thing might be about to play out again, and I hadn’t really heard so much of the benchmark thing. You’re saying you’ve observed that, for example, adding image capability means standard benchmarks decline.

Yeah, I can’t give you full details, but when we’ve tried adding medical images or, say, genomic information, it depends on how you train these models. We’ve tried to do things in a pretty standardized, but maybe also easiest, way, and we’ve seen that while, obviously, on the benchmarks reflective of the new modalities you’re adding, performance goes up, you’re sacrificing performance on the main original benchmarks that are language-understanding-focused or vision-focused. So, yeah, you have to do this Pareto optimization, and again, that also requires compute in its own ways.

But that’s all in a late-fusion paradigm: you’re starting with a model that’s already scoring on benchmarks and then trying to—yeah, I mean, the easiest thing to do is continued pre-training or SFT. The experiments that I’m talking about are not late fusion, but more continued pre-training and SFT.

Obviously, the better thing to do would be to put all that data back into the pre-training and train all of them together. But to motivate that kind of undertaking would require you to show some benefits. Typically, how this works is you take a pre-training checkpoint, do continued pre-training or SFT, and then if you show that your new data is helping improve performance on benchmarks that everyone else cares about, then that data goes back in.

Yeah, the proof is in the pudding. That’s really—I’m going to remember to come back and ask a follow-up on that next time you’re here. Let’s see, there’s not too much time left, and there’s so much in these papers that we could cover.

One thing for practical utility—maybe a couple of things for practical utility—for people building their own systems out there. One is that this tournament-style, head-to-head evaluation seems to be becoming an industry standard. I don’t know if you’d go quite that far, but there’s a strong trend toward trying to surface the best ideas by doing pairwise comparisons and having some sort of World Cup-style, round-robin approach to doing that. If you’d add anything else to that, I also wanted to talk about structured reasoning as opposed to chain-of-thought, because I think that is something a lot of people listening could probably go apply to their projects and get a boost tomorrow. So, yeah, maybe unpack those 2 things.

Yeah, I think the tournament one is interesting because we were actually motivated by AlphaStar, where you had these tournaments of agents competing, and that led to a lot of strong results in that setting. I don’t know if it’s an industry standard, because I also feel like it’s somewhat inefficient, and it hints at the limitations of these models in some ways: they’re maybe not able to independently score and verify ideas, but rather have to do these n-squared pairwise comparisons.

We do have to do some optimizations where we cluster them together, group them together, so as to reduce the computation and not do 1,000 × 1,000 idea comparisons, because that would be too expensive. I would think that would become a little bit more computationally efficient going ahead, but I think that overall ranking of things—the more you can do that in latent space—the more it’s going to lead to interesting results, better reasoning, and so on and so forth. I expect that idea to stay, but not happen as explicitly as it’s happening right now. More of it will happen in the latent space of the reasoning of these models.

Regarding the structured-reasoning approach, I think there are practical reasons for it. When we have 2 agents that we want to talk to each other, it helps to have a data structure that we’re passing, and that makes the engineering itself a little bit easier. But beyond that, compared to just asking the model to, in a chain-of-thought style, reason about something versus defining a certain reasoning structure, I think the advantage is that we can better enforce it to follow a certain path.

In our case, we wanted it to do this long analysis before going into higher-level management goals, before finally forming its management plan. Being able to define that structure forced the model to take that path through its reasoning rather than free form. If we allow it to do it free form, maybe it starts to form its management plan before it has done the higher-level steps that we want it to go through.

Yes, it’s basically to try to describe this for people who might want to implement it. For one thing, it takes advantage of another notable feature that models have gained over the last year or so: we can now specify as part of an API call the exact JSON data structure that you’re supposed to return. That’s huge, because it makes it really easy to set that up and then get something back.

It’s a little bit like airline checklisting, where you’re saying, “Okay, I want you to absolutely, every time, go through these steps, and if you do that, we’re confident you’re going to get better results in the end,” versus just walking out there, randomly walking around the plane, and coming back and saying, “Yeah, it all looks good to me.” The intuition for that is pretty simple. How dynamic did you make that? I’ve never actually done a dynamic structured plan.

Yeah, I think in our case, we tried a lot of different structures and strategies for generating the ultimate research plan. When we tried to get a little too granular with that, I think having the flexibility of just these higher-level things—analysis, management goals, things that are pretty general—tended to lead to better management plans.

Of course, this is all under our own auto-evaluation and vibe checks, as I mentioned, so that’s maybe up for debate. It was dynamic in that sense: it can have any number of analysis items, a list of however many items, and any number of research goals. I think we tried not to constrain it too much—just constrain it to the point where we want it to go through a certain reasoning structure.

Okay, structured outputs: people, don’t overdo it, but definitely use them. That was one of the things that seemed like it really drove a pretty big lift. Going back to the comparisons for one more second, Vivek, are you basically saying that you think in the future it won’t have to be so head-to-head and instead it’ll just be, here’s 10 things, pick the best?

Yeah, I would hope so. In some ways, what we’re trying to do is force the model to explicitly do that tree search: come up with new ideas, go to different nodes, and then do the comparison. But maybe the question is, can all of that happen in your head, in some ways? Does it have to be explicitly written down? Do you have to explicitly generate all those tokens? Can you do something in the way you set up the architecture itself, or in some other mechanism, where all of that happens in the latent space itself, so that you’re more efficient with the tokens you’re generating?

I feel there’s an inefficiency right now. Sure, it helps with interpretability and other different aspects, but I think there’s a lot to be gained by encoding that search within the latent reasoning of the models, and we don’t do that as well right now.

Yeah, okay, cool. That’s helpful. Maybe last 2 things.

What do you think would happen if you just prepended a question-identification agent to the co-scientist and had the thing run in a loop? The first thing it did, instead of taking a question from a human scientist, would be to go out on the internet, search around, and come up with an interesting question for itself, then try to answer its own question. Is there anything about that you think wouldn't be effective?

I feel like, in some ways, we like to talk about the concept of root-node problems at Google DeepMind. We felt like once you have a question, generating novel, original solutions to that is a root-node problem. But in some ways, what you're describing here is an even more upstream root-node problem: How do you ask the right question?

I feel like the day we get AI systems to reach that point, I think that's the day we can truly say, "Okay, we have geniuses in data centers." I think that's going to be the most impactful and important unlock. My feeling is that there should already be decent capability in these models to surf the internet, read information, and figure out what the right questions to ask. So, if it's okay, we'll give it a try and get back to you.

Have you had a chance yet to try this with Gemini 2.5? This is a bonus. From my qualitative assessment, it seems like it would be a lot better.

No, I think that's the exciting part, because everything we described in the paper was based on Gemini 2.0. Gemini 2.5 should be coming up very soon, so we're super excited about that.

The path to geniuses in a data center honestly feels pretty clear at this point, which is a crazy thing to say. Do you see any big barrier questions that you feel are just fundamentally unanswered still? Programmers often call it a simple matter of programming. It's going to be work, but we can make it work. Is that kind of the mindset right now for you guys, or are there questions where you're thinking, "We really don't have a good answer to how we're going to get over that part?"

We have all the building blocks here, so we'll probably build something that will look very close to what you're talking about here. Whether that's the most beautiful one or the most elegant one, we don't know. But does that even matter? I think it doesn't.

I think that is why it's truly exciting: We have line of sight to one solution, which feels like it will get us where we want to go, and that in turn is going to lead to a lot of new unlocks. For the next couple of years, at least, it feels mostly like an engineering challenge rather than trying to answer some fundamental unknowns.

Of course, the other big challenge is going to be the social challenge of introducing this stuff to the world and getting scientists to pay attention. Maybe to close, you want to talk a little bit about what you're doing in that regard? I was excited to see that we're now getting to the point where you're inviting scientists to reach out and partner with you, and also going into—I don't know if you would officially call it a clinical trial—but something in the actual field of medical practice with real patients, too. Tell us what you're doing on the deployment side.

I think the co-scientist is a little bit easier for us to deploy. There are maybe fewer questions around regulation and things like that. The only thing that bothers us a little bit is that the system is highly capable, and there are also many ways in which it can possibly not do so well. We just want to ensure that, as we scale up the system, we do that in a responsible manner.

We have the trusted tester program, and we've already been working with close to 100 scientists across the world. These are all world-leading experts. With the trusted tester program, we want to invite more organizations, so our hope is that we can do this in batches and waves. With every batch, we get feedback, identify the weak points of the system, improve it, and make it better for the next batch of scientists.

This shouldn't take too long. By the end of the year, if I'm optimistic, millions of scientists around the world will have access to this tool. Hopefully, it raises the bar and the ceiling for all of them and enables them to do more creative and interesting work.

I think the one with AMIE is a little bit more tricky. It's obviously a more complex space, but again, there's a known path to taking such systems out into the real world so that they can give diagnosis and treatment recommendations. We're very excited about the clinical trial that we have coming up.

It's going to be one of the first studies of its kind where an LLM-based system is going to be interacting with real patients. The nice thing about the system setup that we have right now is that we're deploying it in a clinic where there is a sufficient presence of clinical experts who can oversee the system and provide oversight. That's a very safe environment for us to deploy the system, where there are enough doctors who can take over if something goes wrong.

Our hope is that not a lot goes wrong in that study. That will allow us to dial down the amount of oversight that's needed. If things go well, we'll probably scale it out to more centers, reduce the amount of expertise, and also introduce more net-new capabilities into the system that we're going to try to make more patient-facing.

I think that's the exciting part. The research has progressed quite a bit, and it's now time to take it for a drive in the real world.

It's exciting times, guys. Really outstanding work. People should be paying more attention than they are, and hopefully we'll put a little dent in the consciousness by bringing some more attention to this. Just really mind-blowing stuff, and quite a series of work that you guys have put out. Anything else you want to share in parting, or any other thoughts you want to leave people with?

It's been a pleasure to work on these projects, obviously, and we have a lot cooking still. I'm excited for the future of this.

And likewise. For me, talking to Nathan is just a lot of fun, and it's a real pleasure to be here. I think there's going to be at least one more where it's big enough that we'll come back again and talk in the next few months. I just hope you don't get too big for me. That's my hope, obviously. I think it's so much fun.

Cool. I really appreciate it again. Fantastic work. Vivek Natarajan and Anil Palepu, thank you for being part of The Cognitive Revolution.