Written by: Techub News Compilation

Introduction

In July 2025, after being awarded the Nobel Prize in Physics, Google DeepMind co-founder and CEO Demis Hassabis was guest on the Lex Fridman podcast again. This was Hassabis' first in-depth conversation after winning the award and his second appearance on the podcast. In this nearly two-and-a-half-hour dialogue, Hassabis went beyond the usual discussions regarding his award-winning work AlphaFold and delved into his grand thoughts on intelligence, the essence of the physical world, and the future of artificial general intelligence. As one of the most visionary leaders in today's field of artificial intelligence, Hassabis' insights not only summarized the breakthroughs in AI over the past decade in scientific discovery but also outlined a future vision of how AI can help humanity solve fundamental puzzles ranging from cells to the universe.

Summary

• Hassabis proposed a core hypothesis: any patterns generated or discovered in nature can be efficiently found and modeled by classical learning algorithms.
• AI models (such as Veo) have astonishingly simulated complex physical phenomena like fluids and light, suggesting an “intuitively physical” understanding of the world.
• Hassabis dreams of constructing a “virtual cell” model and believes that AI will be the ultimate tool to help humanity answer ultimate questions about the origin of life and the nature of consciousness.
• Regarding AGI (artificial general intelligence), he set high standards and predicted that future recognition may require creative breakthroughs akin to the “divine move” in Go.

Learning from Nature: A Bold Hypothesis about the Essence of the World

The interview began with a bold proposition stemming from Hassabis' Nobel Prize speech. He stated: “Any pattern that can be generated or discovered in nature can be efficiently found and modeled by classical learning algorithms.” This is not a vague claim but rather based on the successful practice of DeepMind's landmark projects AlphaGo and AlphaFold.

Hassabis explained that whether it is the move space in Go or the possible shapes proteins can fold into, the combinatorial possibilities far exceed the total number of atoms in the universe, making exhaustive approaches completely infeasible. However, both AlphaGo and AlphaFold intelligently guide searches by constructing models of these environments, making the problems solvable. The key is that natural systems are not random; they have experienced “shaping” through evolution or similar processes, thus possessing inherent structure. Proteins can fold in milliseconds, and the universe itself is solving these complex problems. What AI does is learn and simulate this structured search process.

“I sometimes refer to it as 'survival of the most stable',” Hassabis said, “This applies not only to biological evolution but also to the mountain shapes shaped by geological processes and even to planetary orbits and asteroid shapes—they have all survived countless interactions. Therefore, there should be a kind of pattern or ‘manifold’ that can be learned in reverse, which helps you efficiently search for the correct solutions and make predictions.” He believes that this may not apply to artificial objects or abstract problems (like prime factorization), unless patterns exist within the number space itself. But for most natural systems, their structure is learnable.

This view connects the boundary issue of learning ability with the classic problem in theoretical computer science, P vs NP. Hassabis believes that if we view the universe as an information system, then P vs NP itself is a physical problem. He is discussing with colleagues whether a new complexity class should be defined for such problems that can be solved through neural network processes and mapped to natural systems. “This could be a very interesting new way of thinking about the issue,” he added.

From Fluids to Light: AI Models Display “Intuitive Physics” Understanding

Hassabis used DeepMind's latest video generation model Veo to support his viewpoint. Traditional fluid dynamics (like the Navier-Stokes equations) calculations on classical systems are considered extremely difficult and require massive computational power, such as weather forecasting systems. However, Veo can outstandingly simulate liquids, materials, and specular highlights.

“I once wrote physics engines and graphics engines in the gaming industry, and I know how painful it is to write programs that can do this,” Hassabis said, “But somehow, these systems are doing reverse engineering just by watching YouTube videos.” He speculated that the model is extracting the underlying structures behind these material behaviors, and there might be a lower-dimensional manifold that has been learned. “This may hold true for much of the real world.”

Lex Fridman stated that Veo's simulation of physics is “not perfect but already quite remarkable,” and posed a profound scientific question: what aspects of the world does the model need to understand to achieve this? Hassabis believes that the model's ability to predict subsequent frames in a coherent way is a form of “understanding”—although not anthropomorphized or deeply philosophical, it indeed models sufficient dynamics. He specifically pointed out that the model's capability to simulate physical behaviors, lighting, materials, and liquids suggests it possesses some kind of “intuitive physics” understanding, similar to the intuitive grasp of physics by human children, rather than the analytical understanding of equations by graduate students.

This finding challenges a previously common assumption: that AI must establish a deep understanding of the physical world through embodied interaction (like robots). Veo's success indicates that passive observation may also achieve considerable understanding. “This again hints at some underlying laws of the essence of reality,” Hassabis said. He envisions that the next step may be to make the videos interactive, allowing people to “step into” them and move, which would be truly astonishing and closer to what he calls a “world model”—a model of how the world works, its physical laws, and the inherent objects, which is precisely what a true AGI system needs.

AI and Ultimate Mysteries: From Virtual Cells to the Origins of Life

Hassabis views the construction of artificial general intelligence as the ultimate tool to help humanity answer ultimate scientific questions. One of his dream projects is to create a “virtual cell”—a complete internal simulation of a cell. He envisions that scientists could experiment on virtual cells in a silicon-based environment, significantly accelerating the research process, possibly increasing experimental speed by a hundredfold, and then validating in wet laboratories.

He believes that AlphaFold provides static 3D structures of proteins, while AlphaFold 3 has started modeling interactions (such as protein-to-protein and protein-RNA/DNA interactions), which is the first step towards simulating entire cells or even complete cells. He plans to start with yeast cells, which are well-studied single-celled organisms. Of course, the challenge is enormous, as cellular processes involve different time scales, and modeling requires deciding on the granularity of the simulation. Hassabis hopes to model at the protein level without delving into the atomic level.

A grander question is whether AI can simulate the origin of life. Hassabis considers this one of the most profound and fascinating questions. He envisions, starting from the initial conditions of primordial soup, can AI simulate the emergence of cell-like structures? “That would be the 'divine move' of the origin of life,” he said. He believes that ultimately we will find that there is not a clear demarcation line between non-life and life, but rather a continuous spectrum connecting physics, chemistry, and biology. “Breaking through this wall we've built in our minds is the entire reason I've devoted my life to AI and AGI work.” Hassabis admitted that ultimate questions about the essence of reality, consciousness, the nature of time, etc., have been “shouting” in his mind, and AI may be the key to unlocking these mysteries.

The Path and Markers of AGI: Beyond “Cognitive Inequity”

Hassabis estimates that there is a 50% chance of developing AGI by 2030. He sets very high standards: the ability to match all cognitive functions possessed by the brain. He emphasized that current AI systems are “cognitively inequitable,” strong in some areas but showing clear deficiencies in others. True AGI needs to possess comprehensive and consistent intelligence.

How to test AGI? He proposed two methods: one is “brute force testing,” allowing the system to complete thousands of cognitive tasks that humans can complete; the other is to have world-class experts (like Terence Tao in their respective fields) spend one or two months trying to identify obvious flaws in the system. If no flaws are found, one can be quite confident that a general system has been achieved.

Additionally, he looks forward to seeing some “lighthouse moments” akin to groundbreaking creative breakthroughs like the “divine move” in Go. For example: 1) Proposing a new physical hypothesis or theory akin to Einstein's theory of relativity. Retrospective testing can be conducted: with knowledge cut off at 1900, see if the system can propose special and general relativity. 2) Inventing a new game that is profound and elegant like Go. “A system that can do several of these things simultaneously, not just in a single domain, that is the sign I would look for,” Hassabis said.

Regarding the current hot topics of “recursive self-improvement” and AI safety, Hassabis believes that systems like AlphaEvolve demonstrate the potential of combining large language models with traditional search methods like evolutionary algorithms, which may foster new capabilities in scientific discovery. But he is uncertain whether complete end-to-end self-improvement is desirable, as it could lead to a “hard takeoff” scenario. Currently, systems perform well when given specific instructions but cannot handle highly ambiguous directives such as “create a game better than Go” or “become a better version of yourself.”

Games, Consciousness, and the Future of Humanity

As a former game developer (notable works include “Theme Park” and “Black & White”), Hassabis is filled with anticipation for how AI will reshape games. He dreams of an AI system that can dynamically create content and weave stories based on player imagination, achieving the ultimate “choose your own adventure” game. Combined with interactive generation technologies like Veo, he believes this astounding open-world game could be realized in the next 5 to 10 years.

When asked whether consciousness can be modeled by classical computers, Hassabis disagreed with physicist Roger Penrose's view. He believes that the brain mainly performs classical computations, so consciousness phenomena can, in principle, be imitated or modeled by classical computers. However, the philosophical dilemma regarding “sentience” may be an exception. He proposed an interesting idea: in the future, if connected to AI systems through brain-machine interfaces, we might personally experience the sensation of computing on silicon, leading to a better understanding of consciousness.

Regarding the future of humanity, Hassabis holds an optimistic view. He believes in humanity's infinite creativity and strong adaptability. “Look at the world we are in now—we are coping with it using the brains of hunter-gatherers. Flying in planes, doing podcasts, playing computer games… it's already astonishing. I think this is just the next step.” He believes that society has, to some extent, adapted to the currently astounding AI technology.

Finally, concerning the uniqueness of humanity in the age of AI, Hassabis believes that constructing intelligent artificial beings like AI and comparing them with the human mind may be the best way to understand the special nature of the human mind. “I believe there is something special, but this journey we are undertaking will help us understand and define it.” This conversation ultimately returns to Hassabis' original intention: using AI as a tool to explore the essence of intelligence and ultimately understand ourselves and the universe we inhabit.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。