Sequoia Interviews Hassabis: Information is the Essence of the Universe, AI Will Open a New Branch of Science

Original Text整理：Gua哥 AI New Knowledge

This article content is compiled from an interview with Demis Hassabis on the Sequoia Capital channel, publicly published on April 29, 2026.

Summary: Demis Hassabis in the Sequoia Capital AI Ascent 2026 interview

The origins of AI and gaming: Games are an excellent testing ground for artificial intelligence. By making AI the core gameplay, not only can algorithms be effectively validated, but early computational resources can also support technological development.

The “Timing Theory” of Entrepreneurship: Entrepreneurship should “lead the era by five years, not fifty.” It’s crucial to keenly capture the balance point between technological breakthroughs and practical application needs; being too early often hampers success.

The Evolution Path of AGI: DeepMind’s mission is clear and firm—first, build Artificial General Intelligence (AGI); second, use AGI to solve all complex problems, including science and medicine.

The Core Value of “AI for Science”: AI is the perfect language for describing biological and complex natural systems. With AI simulation, the cycle of new drug development could drop from years to weeks, or even achieve personalized medicine.

The Birth of New Scientific Disciplines: The complexity of AI systems themselves will give rise to new engineering sciences such as “mechanism interpretability.” Meanwhile, AI-driven simulation technology will enable humans to conduct controlled experiments on complex social systems like economics, opening up entirely new branches of science.

Information as the Essence of the Universe: Matter, energy, and information are convertible. The essence of the universe might be a vast information processing system, which gives AI profound significance in understanding the underlying laws of the cosmos.

The Boundaries of Turing Machines: Modern AI systems like neural networks have proven that classical Turing machines are sufficient to simulate problems once thought solvable only by quantum computing (e.g., protein folding). The human brain is likely a highly approximate Turing machine.

Philosophical Reflection on Consciousness: Consciousness may consist of components like self-awareness and temporal continuity. On the path toward AGI, we should first see it as a powerful tool, and explore the grand philosophical question of “consciousness” with its aid.

Content Overview

Google DeepMind co-founder and CEO, Nobel laureate in Chemistry in 2024 for AlphaFold, Demis Hassabis, in a broad and deep dialogue with Sequoia partner Konstantine Buhler at AI Ascent 2026, jointly explores the path to AGI and the future beyond.

In the conversation, he explains why he firmly believes AGI can be achieved by 2030, why the long cycle of drug discovery might shrink from ten years to just days, and why “information,” rather than matter or energy, is the most fundamental essence of the universe. Additionally, he discusses how Einstein, if still alive, would evaluate current AI model limitations, and why the next one or two years will be critical in determining humanity’s fate.

Full Interview

Host: Demis, thank you very much for coming.

Demis Hassabis: Glad to be here. Thanks everyone for joining; it’s great to exchange ideas with you.

Host: It’s a great honor to invite you to our chocolate factory.

Demis Hassabis: I just heard about that. Looking forward to tasting the chocolate later.

Host: Awesome. Demis, let’s get straight to the point. Today, we have a true industry veteran: an original thinker, founder, visionary, and pioneer in all things AI. Demis is a pure believer and a scientist at heart.

Demis’s Original Intent and Inner Mainline

Our discussion today will start from the early story of DeepMind’s founding, then delve into science and technology, and finally include audience Q&A. Let’s begin.

Demis, you were a chess prodigy, a game company founder, and a neuroscientist. You are the founder of DeepMind and now lead a large, influential enterprise. These identities seem unrelated, but you’ve said there’s an underlying mainline running through them. Can you share what that is?

Demis Hassabis: Indeed, there is a mainline, though perhaps some post hoc reasoning is involved. But I’ve long yearned to engage in AI. I recognized early on that this is the most important and fascinating career I could pursue. From age 15 or 16, every learning choice and every project I undertook aimed toward one day building a company like DeepMind.

Games: The Training Ground for AI

I entered the gaming industry via a “curve-breaking” route because, in the 1990s, the cutting-edge tech was there. Not just AI, but also graphics rendering and hardware tech. The GPUs we use today were originally designed for graphics engines, and I was already using early GPUs in the late 90s. All the games I developed—whether for Bullfrog or my own company Elixir Studios—centered AI as a core gameplay mechanic.

My most famous work was probably “Theme Park,” developed when I was about 17. It’s a theme park simulation game where thousands of visitors interact with rides and decide what to buy in shops. Beneath the surface, it ran a complete economic AI model. Like “SimCity,” it was a pioneer in its genre. Seeing it sell over 10 million copies and witnessing players’ joy in interacting with AI strengthened my resolve to dedicate my life to AI.

Later, I turned to neuroscience, hoping to draw inspiration from brain mechanisms to derive new algorithmic ideas. When the time was right to establish DeepMind, integrating all these experiences felt natural. We also used games as early training grounds to validate AI concepts.

Elixir Studios Entrepreneurial Experience

Host: Today, many entrepreneurs are present. You must resonate with them, having founded not just one but two startups. Let’s revisit your first startup, Elixir Studios. What was that experience like? Though not your most famous company, you achieved great success. How did you lead that company? What did it teach you about building a company?

Demis Hassabis: I founded Elixir Studios right after graduating from university. I was fortunate to have worked at Bullfrog Productions before. Everyone in the gaming industry knows it was a legendary early studio, perhaps the top in the UK and Europe at the time.

My goal was to push AI boundaries. In that era, I used game development as a “curve-breaking” approach to fund AI research, constantly challenging tech frontiers and combining it with extreme creativity. I believe this philosophy still applies to today’s blue-sky research.

The deepest lesson I learned is: you should lead the era by 5 years, not 50. At Elixir Studios, we tried to develop a game called “Republic,” aiming to simulate a full nation. The game was set up so players could overthrow a despotic ruler through various means, with a realistic simulation of a breathing, lively city.

Remember, this was the late 90s, with Pentium processors. We had to run all graphics and AI logic for a million people on consumer PCs. It was an overly ambitious goal—perhaps a bit too lofty—and caused many issues.

I remember this lesson well: you need to be ahead of your time, but if you’re 50 years ahead, you’re likely to fail. When an idea is obvious to everyone, it’s too late to get in. The key is to find that delicate balance point.

Founding DeepMind in 2009

Host: Good point about not being too far ahead. Now, in 2009, you were convinced AGI would be achieved. Maybe you were only 10 years ahead, better than 50. Let’s talk about 2009. How did you persuade top talents to join? Back then, AGI sounded like science fiction. How did you convince them?

Demis Hassabis: We keenly picked up some interesting clues. We thought we were 5 years ahead, but perhaps 10. Deep learning had just been invented by Jeff Hinton and colleagues, but few realized its significance. We had deep expertise in reinforcement learning, and believed that combining these two would lead to breakthroughs. Prior to that, they were almost never combined—mostly toy problems in academia. In AI, they were isolated islands.

Moreover, we saw the potential of computational power; GPUs were about to shine. Today, we use TPUs, but back then, accelerated computing was a huge driver. Also, towards the end of my PhD and postdoc, some colleagues were computational neuroscientists, from whom we extracted valuable insights and principles, including a core belief: reinforcement learning can lead to AGI through scaling.

We felt we had assembled these core elements. We even saw ourselves as guardians of a secret—because in academia and industry, no one believed AI could make major breakthroughs. When we proposed to develop AGI—or “Strong AI” as it was sometimes called—many academics rolled their eyes. They saw it as a dead end; after all, the 90s had already seen attempts and failures.

During my postdoc at MIT, a hub for expert systems and first-order logic, I found it outdated even then. But in Cambridge and MIT, people still clung to old methods. That made me more confident we were on the right track. At least, if we were doomed to fail, we would do so in a novel way, not repeating the 90s failures. That made it worth trying; even if it was a risky venture, at least it would be original.

DeepMind’s Mission and Betting on AGI

Host: Did your early belief encounter resistance? To recruit early followers, did you need to prove something to yourself or them?

Demis Hassabis: Regardless of circumstances, I dedicate my life to AI. It’s proven to surpass even our most optimistic expectations. But that was within our 2010 predictions—at the time, we saw it as a 20-year journey.

I believe our progress has been on track, and we’ve played our part.

Even if things hadn’t gone as planned, AI remains a niche discipline I’d pursue because it’s the most important technology I know. My goal is clear: DeepMind’s initial mission is to crack intelligence—build AGI; then use it to solve all other problems. I’ve always believed this is the most important and fascinating technology humanity could invent.

It’s a tool for scientific exploration, a marvelous creation, and one of the best ways to understand human mind—like consciousness, dreams, and creativity. As a neuroscientist, I often felt we lacked a tool like AI to analyze these questions. It offers a comparative mechanism, allowing us to study and contrast different systems as if conducting experiments.

“AI for Science” Culture

Host: Comparing different systems. Let’s talk about “AI for Science.” You’ve been involved early and are a firm believer and idealist. This is your core mission. How did the model and culture you built at DeepMind keep it at the forefront of “AI for Science”?

Demis Hassabis: That’s our ultimate goal. For me, the fundamental drive is to build AI that advances science, medicine, and our understanding of the world. That’s how I pursue my mission—by creating a “Meta Way”: first, develop the ultimate tool; once mature, use it to make scientific breakthroughs. We’ve achieved AlphaFold, and I believe more will follow.

DeepMind always prioritizes this goal. We have a department led by Pushmeet Kohli dedicated to “AI for Science,” nearly ten years old. After returning from the AlphaGo match in Seoul, we immediately launched this effort, now almost a decade ago.

I had been waiting for algorithms to become powerful enough and ideas to be sufficiently general. The breakthrough in Go was a turning point; it made us realize the time had come to apply these ideas to real-world scientific challenges, starting with major problems.

We’ve always believed this is the most beneficial application of AI. What could be better than curing diseases, extending human lifespan, and aiding healthcare? The next natural steps are materials science, environment, and energy. I believe AI will shine in these fields in the coming years.

Breakthroughs in Biology and Isomorphic Labs

Host: How has AI achieved breakthroughs in biology? You’re deeply involved with Isomorphic Labs, a passion project. From the start, you believed in AI’s potential to cure diseases. When can we expect a “highlight moment” in biology similar to language or programming?

Demis Hassabis: I think AlphaFold already marked a “highlight moment” for biology. Protein folding and 3D structure have been a 50-year scientific challenge. Solving it is crucial for drug design and understanding biology’s fundamental code. Of course, it’s just one part of drug discovery, but a vital one.

Our latest company, Isomorphic Labs (which I enjoy managing), focuses on building core technologies in biochemistry and chemistry. These can automatically design compounds that fit perfectly into specific protein sites. Now that we understand protein shapes and surfaces, we’ve effectively identified targets. The next step is to create compounds that bind strongly to these targets, ideally avoiding off-target effects and toxicity.

Our ultimate dream: transfer the entire exploration process—currently taking 99% of R&D time—to computer simulations (In Silico), leaving only wet lab validation at the end. If we can do this—something I believe will happen in the next few years—we could shorten the average 10-year drug discovery cycle to months, weeks, or even days.

Once this threshold is crossed, curing all diseases becomes feasible. Personalized medicine—tailoring drugs to individual patients—will become reality. The entire landscape of healthcare and drug development will be fundamentally reshaped in the coming years.

Simulators and the Birth of New Science

Host: Amazing. You repeatedly mention “AI for Science.” Do you think, at some future point, AI will give rise to an entirely new scientific system? Like thermodynamics emerging from the Industrial Revolution? Will our education system produce fundamentally new disciplines? If so, what might they look like?

Demis Hassabis: I believe several things will happen.

First, understanding and analyzing AI systems themselves will evolve into a full discipline—an engineering science. These creations are incredibly fascinating and complex. Ultimately, their complexity will rival human minds and brains. We must study them deeply to understand their workings—something beyond our current knowledge. I believe a new field will emerge; mechanistic interpretability is just the tip of the iceberg. There’s vast room for exploration.

Second, I also believe AI will open new scientific doors. The most exciting is “AI for Simulations.” I’m obsessed with simulation; all my games involve AI, and fundamentally, they are simulators. I see simulators as the ultimate path to solving social sciences like economics and other humanities.

These disciplines are akin to biology—emergent systems that are hard to experiment on controllably. For example, raising interest rates by 0.5% can only be done in the real world, with unpredictable consequences; theories exist, but experiments can’t be repeated thousands of times. If we could simulate these complex systems precisely, then rigorous sampling and inference based on high-fidelity simulators could establish a new science. I believe this would enable better decision-making in highly uncertain fields.

What conditions are needed for such precise simulations? For example, world models. What scientific and engineering breakthroughs are necessary?

Demis Hassabis: I’ve thought about this deeply. We heavily use learning-based simulators—applied where mathematical understanding is limited or systems are too complex. We can’t just write direct simulation code for every case; it’s too imprecise and incomplete.

We’ve already tested this in weather prediction. We have the world’s most accurate weather simulator, “WeatherNext,” which runs much faster than current meteorological tools. I don’t know if we can understand everything or if that’s desirable, but the first step is to better understand these complex systems.

Even in biology, we’re exploring “Virtual Cells”—dynamic emergent systems. Just as mathematics is the perfect language for physics, machine learning will become the perfect language for biology. In biology and many natural systems, signals are weak, correlations are subtle, and data is vast—beyond human analysis. Yet, within this data, intrinsic relationships, causality, and hidden laws exist.

Machine learning is the perfect tool to describe such systems. To date, math has struggled here—either systems are too complex for top mathematicians or math’s expressive power is insufficient. Many of these systems are highly stochastic and chaotic.

Once we master these simulators, a new scientific branch might emerge. We could extract explicit equations from implicit or intuitive simulators. With unlimited sampling, perhaps we’ll discover fundamental laws like Maxwell’s equations.

Maybe. I don’t know if such laws exist for emergent systems, but if they do, I see no reason we can’t discover them through this approach.

The Universe as Information

Host: That would be incredible. You mentioned a theory that the fundamental building block of everything in the universe might be akin to information—more of a theoretical level. How do you see this? What does it mean for classical Turing machines?

Demis Hassabis: Of course, you can cite Einstein’s E=mc² and his work, showing energy and matter are equivalent. But I actually think information also has a form of equivalence. You can see matter and the organization of structures—especially biological systems that resist entropy—as information processing systems. So, I believe these three can be transformed into each other.

But I have a feeling that information is the most fundamental. This contrasts with 1920s physicists who thought energy and matter were primary. I believe viewing the universe as fundamentally composed of information is a better way to understand it.

If this holds—many current evidence points support it—then AI’s significance is even deeper. Its core is organizing, understanding, and constructing informational objects.

In my view, AI is fundamentally about information processing. If we prioritize information as the key to understanding the world, then these seemingly disparate fields are deeply interconnected.

Can classical Turing machines compute everything?

Demis Hassabis: Sometimes I think of our work as “Turing’s defender,” because Alan Turing is one of my greatest scientific heroes. I believe his work laid the foundation for computers, AI, and much of what we do. The Turing machine theory is one of the most profound achievements: any computable thing can be simulated by a relatively simple machine. So, I think our brains are probably approximate Turing machines.

It’s interesting to consider the link between Turing machines and quantum systems. But systems like AlphaGo and AlphaFold show that classical neural networks—disguised as Turing machines—can model problems once thought to require quantum mechanics, like protein folding. Protein folding involves quantum effects at tiny scales, but classical methods can approximate solutions well enough.

This suggests many phenomena once believed to need quantum systems can, with proper methods, be modeled classically.

Consciousness and Philosophy

Host: You see AI as a tool, like telescopes, microscopes, or astrolabes. But when a machine can simulate everything—like quantum systems—when will it transcend being just a tool? Will that day come?

Demis Hassabis: I strongly feel that in building AGI, many of us think the best approach is to first create a tool—an extremely intelligent, practical, precise instrument—and then cross the next threshold. The significance of that alone is profound. This tool may become increasingly autonomous and intelligent, which we are witnessing now. We are entering an “Agent Era.”

But further questions remain: does it have agency? Does it have consciousness? These are questions we’ll have to face. I suggest we treat this as a second step, perhaps using the first step’s tool to explore these profound issues.

Ideally, through this process, we can better understand our own brains and minds, and more precisely define concepts like “consciousness.”

Future Definitions of Consciousness

Host: Do you have a rough prediction for how consciousness might be defined in the future?

Demis Hassabis: No, beyond what philosophers have discussed over millennia. But what’s clear is that certain components are necessary. They may be necessary but not sufficient conditions—like self-awareness, the concept of self and others, and some form of temporal continuity—these seem essential for any entity that appears conscious.

But the full definition remains an open question. I’ve discussed this with many great philosophers. A few years ago, I had an in-depth exchange with Daniel Dennett, who recently passed. One key issue is the system’s behavior: does it act like a conscious entity? As AI approaches AGI, it might do so.

But then, why do we believe others are conscious? Partly because of their behavior—how they act as if they are conscious beings. Another factor is that we all operate on the same substrate.

So, if these two points hold, then if your experience and mine are the same, that’s the most parsimonious explanation. That’s why we usually don’t argue about whether others are conscious. But it’s impossible to have substrate equivalence in artificial systems. Eliminating that gap is very difficult. You can assess behaviorally, but what about experience? After achieving AGI, perhaps some methods will emerge, but that’s beyond today’s scope—even in “AI and Science” discussions.

Great. Now, let’s open the floor for audience questions. Please prepare your questions. You mentioned philosophers, especially Kant and Spinoza, as your favorites. Kant was a deontologist emphasizing duty; Spinoza held a nearly deterministic view of the universe. How do you connect these two seemingly opposite philosophies? What’s your fundamental view of how the world operates?

Demis Hassabis: I like both because Kant proposed—something I deeply appreciated during my neuroscience PhD—that “the mind creates reality,” which I believe is fundamentally correct. This gives us a compelling reason to study how the mind and brain work. Since I seek the nature of reality, I must first understand how the mind interprets it. That’s the insight I gained from Kant.

As for Spinoza, he’s more about the mental dimension. If you try to use science as a tool to understand the universe, you’re already touching on its deeper secrets.

This reflects my current perspective. When I pursue science, AI, and these tools, I feel like we’re reading the universe’s language in some way.

Host: Beautiful. That’s a poetic summary of your daily work: Demis, you are a scientist, speaker, and philosopher all in one. Before we end, let’s do a quick rapid-fire Q&A. You haven’t seen these questions before. Predict whether AGI will be achieved earlier or later than expected? Or prefer not to answer?

Demis Hassabis: I choose 2030. I’ve been firm on this prediction.

Host: Great, 2030. When we achieve AGI, what books, poems, or papers would you recommend?

Demis Hassabis: For the post-AGI world, I love David Deutsch’s “The Fabric of Reality.” I think its ideas still hold. I hope to use AGI to answer the profound questions raised in that book, which will also be my focus in the post-AGI era.

Host: Wonderful. What’s the proudest moment at DeepMind so far?

Demis Hassabis: We’ve had many peaks, but I think AlphaFold’s emergence is the most proud.

Host: Finally, a few game-related questions. If you’re in a high-stakes turn-based strategy game like “Civilization” or “Polytopia,” and could pick a scientist from history as a teammate—Einstein, Turing, or Newton—who would you choose?

Demis Hassabis: I’d pick von Neumann. In such a scenario, you need a game theory expert, and I think he’s the best.

Host: That’s definitely a top-tier teammate. Demis, you’re truly a polymath. Thank you very much for being our guest today. Let’s give a round of applause for Demis’s brilliant insights. Thank you all.