DeepMind Founder Interview: AGI Architecture, Agent Status, and Scientific Breakthroughs in the Next Decade

Original video title: Demis Hassabis: Agents, AGI & The Next Big Scientific Breakthrough

Original video source: Y Combinator
Original compilation: Deep Tide TechFlow

Editorial Introduction

Google DeepMind CEO and Nobel Prize in Chemistry winner Demis Hassabis visits Y Combinator to discuss what key breakthroughs lie ahead on the path to AGI, offer advice to entrepreneurs on how to stay ahead, and share where the next major scientific breakthrough may emerge.

For deep-tech entrepreneurs, the most practical takeaway is this: if you kick off a decade-long deep-tech project today, you must factor in the emergence of AGI. In addition, he also revealed that Isomorphic Labs (the AI biotech drug-discovery company spun out from DeepMind) is about to release major news.

Key Quotes

AGI Roadmap and Timeline

· “These existing technology components will almost certainly become part of the final AGI architecture.”

· “The problems with continual learning, long-range reasoning, and certain aspects of memory still haven’t been solved. AGI needs to get all of that right.”

· “If your AGI timeline is around 2030 like mine, and you start a deep-tech project today, then you have to account for the fact that AGI could show up halfway through.”

Memory and Context Windows

· “The context window is roughly equivalent to working memory. The average human working memory holds only seven numbers. But we have context windows with hundreds of thousands, even tens of millions of tokens. The problem is that we cram everything in—including unimportant stuff and incorrect information. Right now, this approach is rather crude.”

· “If you want to process a real-time video stream and store all the tokens, one million tokens is actually only enough for about 20 minutes.”

Flaws in Reasoning

· “I like to play chess with Gemini. Sometimes it realizes it’s making a terrible move, but it can’t find a better one—so it circles back and still ends up making that terrible move. But a precise reasoning system shouldn’t behave like that.”

· “On one hand, it can solve IMO gold-medal-level problems. On the other hand, ask it in a different way and it makes mistakes like elementary-school math. In terms of introspection about its own thinking process, it seems like something is missing.”

Agents and Creativity

· “To reach AGI, you need a system that can proactively solve problems for you. Agents are the path—and I think we’re just getting started.”

· “I haven’t seen anyone use vibe coding to create a AAA game that tops the app store rankings. With the current level of effort, it should be possible, but it hasn’t happened yet. That suggests something is still missing in the tools or processes.”

Distillation and Small Models

· “Our assumption is that after a cutting-edge Pro model is released for half a year to a year, its capabilities can be compressed into a very small model that can run on edge devices. We haven’t run into any theoretical limit of information density yet.”

Scientific Discovery and the “Einstein Test”

· “Sometimes I call it the ‘Einstein test’—whether you can train a system using knowledge from 1901, then have it independently derive the results that Einstein achieved in 1905, including special relativity. Once you can do that, these systems aren’t far from truly inventing something new.”

· “Solving a Millennium Prize problem is already remarkable. But even harder than that is whether you can propose a new set of Millennium Prize problems—problems that top mathematicians consider equally profound, and that are worth researching for a lifetime.”

Deep-Tech Entrepreneurship Advice

· “Chasing hard problems and chasing easy problems are basically similar—the difficulty just comes in different forms. Life is short. Why not put your energy into things that truly won’t get done unless you do them?”

Pathways to AGI

Gary Tan: You’ve thought about AGI longer than almost anyone. Based on the current paradigm, how much of the final AGI architecture do you think we already have? What’s fundamentally missing right now?

Demis Hassabis: Large-scale pretraining, RLHF, chain-of-thought—I’m quite certain they’ll become part of the final AGI architecture. These technologies have already proven too much to be dead ends. I can’t imagine that in two years we’d discover this is a dead-end—that doesn’t make sense to me. But on top of what we have, we might still be missing one or two things. Continual learning, long-term reasoning, and certain aspects of memory—some problems still aren’t solved.

AGI needs to be fully figured out. Maybe existing tech plus a few incremental innovations can get us there, but it’s also possible that one or two major key breakthroughs still need to happen. I don’t think it will be more than one or two. As for whether there are unresolved key points, my personal estimate is around a 50/50 split. So at Google DeepMind, we’re pushing both directions.

Gary Tan: I deal with a lot of agent systems, and what shocks me most is that the underlying weights end up being basically the same back and forth. That’s why the idea of continual learning is especially interesting—because right now, we’re basically patching things together with tape, like those “nightly dream cycle” kinds of ideas.

Demis Hassabis: Yes, those dream cycles are pretty cool. We’ve thought about that problem in the context of integrating episodic memory. My PhD research was about how the hippocampus elegantly integrates new knowledge into an existing knowledge framework. The brain is extremely good at this.

It completes the process during sleep—especially during REM sleep—replaying important experiences so it can learn from them. Our earliest Atari program, DQN (DeepMind’s deep Q-network published in 2013, which first reached human-level performance on Atari games using deep reinforcement learning) could master Atari games, and one key method was experience replay.

That concept came from neuroscience: replaying the successful path repeatedly. That was in 2013—ancient in AI terms—but it was crucial then.

I agree with what you said. Right now, we really are “taping things together.” We stuff everything into the context window. It just doesn’t feel right. Even if we’re building machines rather than biological brains, theoretically we could have context windows in the million or ten-million token range, and memory could be perfect—but the cost of lookup and retrieval still exists. At the moment when you need to make a specific decision, finding truly relevant information isn’t easy, even if you can store everything. So I think there’s a lot of room for innovation in the area of memory.

Gary Tan: Honestly, a million-token context window is already much bigger than I expected. You can do a lot with it.

Demis Hassabis: Yes—it’s large enough for most scenarios where it should be used. But think about it this way: the context window is roughly equivalent to working memory. Humans have only about seven items in working memory on average. We have context windows with millions or even tens of millions of tokens. The problem is that we put everything into it, including unimportant and incorrect information. That’s a pretty blunt approach. And if you’re dealing with real-time video streams and naively recording all tokens, one million tokens is actually only enough for about 20 minutes. But if you want the system to understand your life over one or two months, that’s far from enough.

Gary Tan: DeepMind has always deeply invested in reinforcement learning and search. How much of that philosophy is embedded in how you’re building Gemini today? Has reinforcement learning still been underestimated?

Demis Hassabis: Maybe it has. The attention given to RL has had ups and downs. From the first day we founded DeepMind, we worked on agent systems. All the work on Atari and AlphaGo—essentially—was about reinforcement-learning agents: systems that can autonomously achieve goals, make decisions, and formulate plans. Of course, we chose games at the time because the complexity was controllable, and then we gradually moved to more complex games—like AlphaStar after AlphaGo. Basically, we made all the games we could.

The next question is whether we can generalize these models into world models or language models—not just game models. For the past few years, we’ve been working on this. Today’s leading models’ reasoning patterns and chain-of-thought reasoning are, in essence, a return to what AlphaGo pioneered back then.

I think much of what we did back then is highly relevant to today. We’re re-examining those old ideas and doing them at larger scale and in more general ways, including reinforcement-learning methods like Monte Carlo tree search. The ideas behind AlphaGo and AlphaZero are extremely related to today’s foundational models. I believe a large portion of the progress in the next few years will come from this.

Distillation and Small Models

Gary Tan: To be smarter now requires bigger models, but at the same time distillation techniques are improving, so small models can become quite fast. Your Flash models are very strong—basically reaching about 95% of frontier-model performance, but at only one-tenth of the price. Is that right?

Demis Hassabis: I think that’s one of our core advantages. You need to build the largest models first to obtain frontier-level capabilities. One of our biggest strengths is being able to quickly distill and compress those capabilities into smaller and smaller models. The distillation approach itself was something we invented, and we’re still among the world’s top. And we also have strong business incentives to do it. We’re probably the world’s largest AI application platform.

With AI Overviews and AI Mode, and Gemini—every Google product now, including Maps, YouTube, and so on, is integrating Gemini or related technologies. That involves billions of users, as well as products with more than a billion users. They have to be extremely fast, extremely efficient, extremely low-cost, and extremely low-latency. This gives us tremendous motivation to push Flash and even smaller Flash-Lite models to the highest level of efficiency. I hope it ultimately serves users well in all kinds of work.

Gary Tan: I’m curious how smart these small models can actually get. Is there a limit to distillation? Can 50B or 400B models become as smart as today’s largest frontier models?

Demis Hassabis: I don’t think we’ve reached an information-theoretic limit—at least, nobody knows whether we have. Maybe someday we’ll hit some kind of ceiling in information density. But our assumption right now is that once a cutting-edge Pro model is released, within half a year to a year, its capabilities can be compressed into a very small model that can almost run on edge devices.

You can also see this in our Gemma models. Our Gemma 4 performs very strongly at the same scale. This uses a lot of distillation techniques and efficiency optimization for small models. So I really don’t see any theoretical limit yet. I think we’re still far away from it.

Gary Tan: There’s a pretty outrageous phenomenon right now: the amount of work engineers can do is roughly 500 to 1000 times what it was six months ago. Some people in this room are doing work equivalent to what a Google engineer from the 2000s could do—1000 times as much. Steve Yegge has talked about this.

Demis Hassabis: I find it exciting. Small models have many use cases. One is that they’re low-cost, and being fast also brings benefits. In writing code or other tasks, you can iterate faster—especially when collaborating with systems. A fast system can be more than enough, even if it isn’t the absolute frontier—say only 90% to 95% of frontier performance. That’s totally usable, and the time-iteration gains you get back far exceed that extra 10%.

Another big direction is running these models on edge devices. It’s not only for efficiency, but also for privacy and security. Think about devices that process very personal information—and robots. For the robot in your home, you’d want a local, efficient, and powerful model, and only delegate specific tasks to large models in the cloud under certain scenarios. Audio and video streams would be processed locally, with data kept locally. I can imagine that being a very good end state.

Memory and Reasoning

Gary Tan: Coming back to context and memory. Models are currently stateless. If we add continual learning capabilities, what would the developer experience look like? How would you guide such models?

Demis Hassabis: This is a very interesting question. The lack of continual learning is one key bottleneck preventing agents from completing full tasks. Today’s agents are useful for the local parts of tasks. You can chain them together to do some cool things, but they can’t adapt well to your specific environment. That’s why they still can’t truly be “launched and forgotten.” They need to learn your particular context. To achieve truly general intelligence, this has to be solved.

Gary Tan: Where are we at in reasoning? The models’ chain-of-thought is strong now, but it still fails on certain errors that smart college students wouldn’t make. What exactly needs to be changed? What progress do you expect in reasoning?

Demis Hassabis: There’s still a lot of room for innovation in the way we approach thinking paradigms. What we’re doing is still fairly rough, fairly brute-force. There are many potential improvements—for example, monitoring the process of chain-of-thought and intervening in the middle of thinking. I often feel that whether it’s our systems or competitors’ systems, to some extent they overthink and get stuck in loops.

I sometimes like to use Gemini playing chess to observe this. All the leading foundation models are actually quite bad at chess, which is interesting.

Looking at their thought trajectories is valuable because chess is a domain that’s well understood. I can quickly tell whether it’s gone off track and whether the reasoning is effective. What we see is that sometimes it considers a move, realizes it’s a bad move, but can’t find a better one—so it loops around and still makes that bad move. A precise reasoning system shouldn’t show this behavior.

That huge gap still exists. But fixing it might only require one or two adjustments. That’s why you see what people call “jagged intelligence.” On one hand, it can solve IMO gold-medal-level problems. On the other hand, if you ask it differently, it makes elementary math mistakes. In terms of introspection about its own thinking process, it seems like something is missing.

The Real Capabilities of Agents

Gary Tan: Agents are a big topic. Some people say it’s hype. I personally think we’re just at the beginning. What’s DeepMind’s real internal assessment of agent capabilities, and how big is the gap versus what’s being promoted outside?

Demis Hassabis: I agree—we’re just starting. To reach AGI, you need a system that can proactively solve problems for you. That’s always been clear to us. Agents are the path, and I think we’re just getting started.

Everyone is figuring out how to make agents collaborate better with work. We’ve done a lot of exploration through personal experiments, and many of you probably have too. How do you integrate agents into workflows so they’re not just a nice-to-have, but truly doing fundamental work? We’re still in the experimental stage. It might only be the last two or three months that we’ve truly started finding scenarios that are particularly valuable. The technology has only just reached the point where it’s no longer just toy demos—it genuinely creates value in your time and efficiency.

I often see people spin up dozens of agents and have them run for dozens of hours, but I’m still not sure whether the output justifies that amount of investment.

We haven’t yet seen anyone use vibe coding to create a AAA game that tops the app store charts. I’ve done some myself, and many people here have also made some good small demos. Right now I can prototype a “Theme Park” in half an hour. When I was 17, I spent six months on it.

I have a feeling that if you spent an entire summer on it, you could create something truly incredible. But it still requires craftsmanship, human soul, and taste—you have to make sure those elements are brought into whatever product you build. In fact, no kid has yet made a blockbuster game that sells over 10 million copies. In theory, with today’s tool investments, it should be possible. So something is still missing—maybe in the workflow, maybe in the tools. I expect we’ll see results like that within the next 6 to 12 months.

Gary Tan: To what extent will it be fully automated? I don’t think it will be fully automated right away. The more likely path is that people here first achieve 1000x efficiency, then someone uses these tools to build best-selling apps and best-selling games, and only after that will more steps become automated.

Demis Hassabis: Yes, that’s the path you should expect to see first.

Gary Tan: There’s also part of the reason that some people really are doing this, but they’re unwilling to publicly say how much Agent helped.

Demis Hassabis: That may be true. But I want to talk about creativity. I often use the example of AlphaGo and that famous move 37. For me, I’ve always been waiting for a moment like that to happen. Once it happened, I started scientific projects like AlphaFold. We started AlphaFold the day after Seoul—ten years ago. This time when I went to Korea, it was to celebrate AlphaGo’s 10th anniversary.

But simply getting past Move 37 isn’t enough. It’s cool and very useful. But can this system invent the game of Go itself? If you give it a high-level description—something like “a game you can learn in five minutes, but that takes a lifetime to master; aesthetically elegant; you can finish a whole game in an afternoon”—and then the system returns Go, that would be different. Today’s systems can’t do that. The question is why.

Gary Tan: Someone in this room might be able to.

Demis Hassabis: If someone can do it, then the answer isn’t that the system is missing something. It might be that the way we’re using the system is wrong. That may even be the correct answer. Maybe today’s systems already have the capability, but they need a sufficiently talented creator to drive it—providing the soul of the project—and to be highly integrated with the tools, almost merging with them. If you soak in these tools day and night and have deep creativity, then maybe you can make something beyond imagination.

Open Source and Multimodal Models

Gary Tan: Let’s switch topics and talk about open source. The recent release of Gemma means that very strong models can now run locally. What’s your view? Will AI become something users control themselves, rather than staying mainly in the cloud? Will this change who can build products with these models?

Demis Hassabis: We’re strong supporters of open source and open science. Regarding AlphaFold, we have fully and freely open-sourced it. Our scientific work is still published in top journals. For Gemma, we want to create world-leading models at comparable scale. Gemma’s downloads have already reached about 40 million times, and it was released only two and a half weeks ago.

I also think it’s important that there’s a Western tech stack in the open-source ecosystem. China’s open-source models are excellent, and they’re currently leading in open source. But we believe Gemma is very competitive at the same scale.

For us, there’s another resource issue: no one has spare compute to build two full-size frontier models. So our current decision is: edge models for Android, glasses, robots, etc. should be open models. The reason is simple—once they’re deployed on devices, they’re exposed. It’s better to open them up thoroughly. We unified our open strategy at the nanometer scale, and strategically that makes sense.

Gary Tan: Before going on stage, I demonstrated the AI operating system I built. I can interact with Gemini directly by voice. I was still pretty nervous when demonstrating it to you—but surprisingly, it worked. Gemini was built as multimodal from the start. I’ve used many models. With Gemini, the combination of voice-to-model interaction, plus tool-calling capability, plus deep context understanding—right now, no model compares.

Demis Hassabis: Yes. One advantage of the Gemini series that hasn’t been fully recognized is that from the very beginning, we built it in a multimodal way. That makes the early stage more difficult than if you only do text—but we believe the long-term benefits will pay off, and it’s already starting to happen.

For example, in world models, we built Genie (DeepMind’s generative interactive environment model) on top of Gemini. In robotics, the Gemini Robotics line will also be built on multimodal foundation models. Our advantages in multimodality will become a competitive moat. We’re also increasingly using Gemini in Waymo (Alphabet’s autonomous driving company).

Imagine a digital assistant that follows you into the real world—maybe on your phone or on your glasses. It needs to understand the physical world and environment around you. Our system is very strong in this area. We’ll continue investing in this direction. I believe our lead in these kinds of problems is substantial.

Gary Tan: Reasoning costs are dropping rapidly. When reasoning is basically free, what becomes possible? Will your team’s optimization directions change because inference is almost free?

Demis Hassabis: I’m not sure inference will truly be free. Jevons’ Paradox is still there—efficiency improvements can increase total consumption. I think in the end, everyone will use all the compute they can get.

You can imagine groups of millions of agents collaborating, or a small group of agents thinking in parallel along multiple directions and then integrating results. We’re experimenting with these approaches, and all of them will consume available inference resources.

On the energy side, if we solve a few issues—controlled nuclear fusion, room-temperature superconductivity, optimal batteries, and so on—I think through materials science we can get energy costs close to zero. But bottlenecks remain in physical chip manufacturing and other parts. At least, that will be the case for decades. So there will still be quotas and constraints on the inference side, and it will still need to be used efficiently.

The Next Big Scientific Breakthrough

Gary Tan: Thankfully, small models are getting smarter. Many founders in biological and biotech fields are here. AlphaFold 3 has already gone beyond proteins and expanded to a broader range of biological molecules. How far are we from building a fully modeled cellular system? Is this a completely different difficulty tier?

Demis Hassabis: Isomorphic Labs is making very good progress. AlphaFold is only one part of the drug discovery process. We’re doing adjacent biochemistry research—designing compounds with the right properties—and we’ll have major releases soon.

Our ultimate goal is to build a complete virtual cell: a full-function cellular simulator that you can perturb, whose outputs are close enough to experimental results and that also has real-world utility. You could skip huge numbers of search steps, generate lots of synthetic data to train other models, and have them predict the behavior of real cells.

I estimate we’re about 10 years away from a complete virtual cell. On the DeepMind science side, we’re starting from the nucleus, because the nucleus is relatively self-contained. The key for problems like this is whether you can cut out a slice of complexity that’s appropriate—one that’s sufficiently self-contained—so you can reasonably approximate its inputs and outputs, and then focus on that subsystem. From this perspective, the nucleus is quite suitable.

Another problem is that data is insufficient. I’ve chatted with top scientists in electron microscopy and other imaging technologies. If we could image live cells without killing them, that would be disruptive—because then it turns into a visual problem, and we know how to solve visual problems.

But as far as I understand, currently there’s no technology that can image living dynamic cells at the nanometer resolution without damaging them. You can capture static images at that resolution—they’re already extremely detailed, which is exciting—but it isn’t enough to directly turn it into a visual problem.

So there are two paths: one is a hardware-driven, data-driven approach; the other is building better learnable simulators to simulate these dynamical systems.

Gary Tan: You’re not only looking at biology. Materials science, drug discovery, climate modeling, mathematics—if you had to rank them, which scientific field do you think will be transformed most thoroughly in the next five years?

Demis Hassabis: Every field is exciting. That’s also why this has always been my biggest passion, and why I’ve worked on AI for more than 30 years. I’ve always believed AI will be the ultimate scientific tool—for advancing scientific understanding and discovery, for medicine, and for how we understand the universe.

The way we originally framed our mission was in two steps. First, solve intelligence—build AGI. Second, use it to solve everything else. Later, we had to adjust the wording, because people asked, “Do you really mean you’ll solve all problems?”

We do mean that. Now people are starting to understand what it entails. Specifically, I’m referring to solving what I call “root node problems” in science—fields where, once you break through, you unlock entirely new branches of discovery. AlphaFold is a prototype of what we want to do.

Over 3 million researchers worldwide—almost every biologist now uses AlphaFold. I’ve heard from executives at pharmaceutical companies that nearly every new drug discovery in the future will involve AlphaFold at some stage of the process. We’re proud of that impact, and we hope AI can generate that kind of influence. But I think it’s only the beginning.

I can’t think of a scientific or engineering field where AI can’t help. The areas you mentioned are roughly at the “AlphaFold 1 moment”—promising, but not yet tackling the big challenges. In the next two years, we’ll be able to talk about lots of progress across all those areas, from materials science to mathematics.

Gary Tan: It feels like Prometheus-style—giving humanity a brand-new kind of capability.

Demis Hassabis: Exactly. And just like the moral of the Prometheus story, we also have to be cautious about how that capability is used, where it’s applied, and the risks of misuse of the same tools.

Lessons from Success

Gary Tan: There are many people here trying to start companies that apply AI to science. In your view, what’s the difference between truly cutting-edge startups pushing the frontier, and startups that just wrap APIs on top of base models and then call themselves “AI for Science”?

Demis Hassabis: I’m thinking that if I were sitting where you are now, looking at projects at Y Combinator, what would I do? One thing is that you have to anticipate where AI is heading—and that itself is very hard. But I do believe there’s huge opportunity in combining AI’s direction with another deep-tech area. In that intersection—whether it’s materials, medicine, or other genuinely difficult scientific domains, especially those involving the atomic world—there won’t be shortcuts in the foreseeable future. These fields won’t be crushed just because the next foundational model update comes out. But if you’re looking for a direction that’s defensively strong, that’s what I would recommend.

Personally, I’ve always been partial to deep tech. Things that are truly durable and valuable aren’t easy. Since we started in 2010, I’ve been drawn to deep tech—investors told me, “We already know this won’t work,” and academia also thought it was a niche that had been tried in the ’90s and failed.

But if you have belief in your ideas—why it’s different this time, and what unique combination your background brings—ideally you’re an expert in machine learning and applications, or you can assemble a founding team like that. Then there’s enormous influence and value to be created.

Gary Tan: That’s very important. After something works, it can look obvious. But before it works, everyone is against you.

Demis Hassabis: Of course. So you have to do what you truly have passion for. For me, no matter what happens, I’ll keep doing AI. When I was very young, I decided it was the most impactful thing I could think of—and it turned out to be true. But maybe we were 50 years early.

Also, it’s the most interesting thing I can think of. Even if today we’re still stuck in a small garage and AI hasn’t been fully realized, I’d still find ways to keep going. Maybe I’ll go back to academia, but I’ll find some way to keep pushing.

Gary Tan: AlphaFold is an example of pursuing a direction and placing the right bet. What makes a scientific field suitable for breakthroughs like AlphaFold? Are there patterns—for example, a certain objective function?

Demis Hassabis: I should find time to write this down someday. From AlphaGo and AlphaFold, the lessons I’ve learned from all the Alpha projects are that our existing techniques work best under these conditions.

First, the problem has a huge combinatorial search space—the bigger the better—so that no brute-force enumeration or special algorithm can solve it. The space of possible moves in Go and the conformational space of proteins are far larger than the number of atoms in the universe. Second, you can clearly define the objective function—for example, minimizing a protein’s free energy, or winning in Go—so the system can perform gradient ascent. Third, you have enough data, or a simulator that can generate lots of synthetic data that follows the distribution.

If these three conditions hold, then with today’s methods you can go a long way, to find that “needle in the haystack” you need. Drug discovery follows the same logic: there exists a compound that can treat the disease without side effects, and as long as the laws of physics allow it to exist, the only question is how to find it efficiently and feasibly. I think AlphaFold’s first proof was showing that these kinds of systems can find that needle in massive search spaces.

Gary Tan: I want to take it up a level. We’re talking about humans creating AlphaFold using these methods. But at a meta level, humans are also using AI to explore hypothesis spaces. How far are we from AI systems doing genuinely scientific reasoning—not just pattern matching in data?

Demis Hassabis: I think we’re very close. We’re building general systems like this. We have a system called AI co-scientist, and algorithms like AlphaEvolve, which can go further than basic Gemini. All leading labs are exploring this direction.

But so far, personally, I haven’t seen a truly major scientific discovery produced by these systems. I believe it’s coming soon. It might relate to creativity, in the sense of breaking through known boundaries. At that level, it’s no longer pattern matching, because there’s no pattern to match. It’s not entirely extrapolation either. It’s some kind of analogical reasoning. I think these systems currently lack that, or we haven’t used them in the right way.

A standard I often mention in science is: can it propose a truly interesting hypothesis—not just validate one? Validating a hypothesis by itself can also be a huge deal—for example, proving the Riemann Hypothesis or solving a Millennium Prize problem—but maybe we’re only a few years away from reaching that stage.

Even harder than that is whether it can propose a new set of Millennium Prize problems—problems that top mathematicians consider equally profound and worth researching for a lifetime. I think that’s another order of magnitude harder, and we don’t yet know how to do it. But I don’t think it’s magic. I believe these systems can eventually do it—maybe missing just one or two things.

A way we can use to test it is something I sometimes call the “Einstein test”: can you train a system using knowledge from 1901, and then have it independently derive the results from Einstein’s 1905 work, including special relativity and his other papers that year? I think we really should run this test—try repeatedly and see when we can achieve it. Once we can do that, these systems won’t be far from truly inventing something new.

Entrepreneurship Advice

Gary Tan: Final question. There are many people here with deep technical backgrounds who want to build something at a scale like yours—you’re one of the world’s largest AI research organizations. Coming from the front lines of AGI research, is there something you know now that you wish you’d known when you were 25?

Demis Hassabis: We’ve already touched on part of this. You’ll find that chasing hard problems and chasing simple problems are basically similar—the difficulty just shows up differently. Different things have different kinds of difficulty. But life is short, and energy is limited. So instead of spreading it everywhere, put your life force into the things that truly won’t get done unless you do them. Choose using that standard.

Another point: I think in the coming years, cross-disciplinary combinations will become more common. AI will make it easier to cross fields.

Finally, it depends on your AGI timeline. Mine is around 2030. If you start a deep-tech project today, it usually means a ten-year journey. That means you must factor in the emergence of AGI happening halfway. What does that mean? It’s not necessarily a bad thing, but you have to think it through. Can your project leverage AGI? How will AGI systems interact with your project?

Coming back to what we discussed earlier about the relationship between AlphaFold and general AI, one scenario I can foresee is that Gemini, Claude, or similar general systems will call AlphaFold-like specialized systems as tools. I don’t think we’ll just cram everything into a single huge monolithic

Original video link

Click to learn about Rhythm BlockBeats hiring

Welcome to the official Rhythm BlockBeats community:

Telegram subscription group: https://t.me/theblockbeats

Telegram discussion group: https://t.me/BlockBeats_App

Twitter official account: https://twitter.com/BlockBeatsAsia