HomeArticle

Record of Nobel laureates' talks on the human doomsday crisis: About the "37th step" of AI and the Kardashev Type I civilization

36氪的朋友们2025-07-25 12:15
AlphaGo's "Divine Move" Hints at the Critical Point of AGI, While Demis Hassabis Warns of Doomsday Risks

In the 2016 showdown between AlphaGo and Go world champion Lee Sedol, AlphaGo made an unexpected move in the second game, the 37th move. Although it seemed unconventional at the time and was even considered a mistake by some Go commentators, it was ultimately proven to be an extremely far - reaching and precise move that helped AlphaGo secure victory. This move is hailed as the "divine move".

The "37th move" has since become an important symbol in the field of AI. It represents the sudden, innovative, and counter - intuitive breakthroughs that AI can achieve in complex decision - making.

Recently, Nobel Chemistry Prize winner and CEO of Google DeepMind, Demis Hassabis, participated in a program hosted by well - known AI podcast host Lex Fridman.

They delved deeply into discussions about artificial general intelligence (AGI), modeling of natural systems, scientific creativity, and the future prospects of human civilization.

These discussions all point to a thought - provoking question: Has AI reached its "37th move"? AlphaGo made a disruptive 37th move before humans could fully grasp the overall situation. Is AI now quietly approaching the tipping point of a technological shift?

In addition, Hassabis issued a warning. While AI brings tremendous potential for change, it also comes with unforeseeable risks. He emphasized that the progress of AI could lead to "existential risks", especially when the technology gets out of control or is misused, with potentially catastrophic consequences.

Key points from the conversation:

  1. Hassabis believes that all natural models that can be generated or discovered can be efficiently modeled through classical learning algorithms, especially in fields such as biology, chemistry, and physics.
  2. The probability of achieving AGI by 2030 is about 50%. He proposed landmark criteria for testing AGI, such as formulating new scientific conjectures or inventing complex games similar to Go.
  3. The evolutionary process in nature provides structure to systems, which can be learned by AI and used for modeling. For example, AlphaGo and AlphaFold solve complex problems through intelligent guided search.
  4. Although it was previously thought that understanding physical laws required an embodied AI, Veo 3 demonstrated an ability to understand through video observation, challenging this view.
  5. Hassabis previously discussed the idea of collaborating on game development with Elon Musk. To this day, he still hopes to have the opportunity to cooperate with Musk on this issue.
  6. In the future, AI can customize game worlds for players. It can automatically generate storylines based on players' decisions, enhancing the immersion and flexibility of games.
  7. Nuclear fusion and solar energy will become the primary energy sources in the future. Solving the energy problem is expected to propel humanity towards a Kardashev Type I civilization.
  8. Hassabis said that Meta's strategy of poaching talent with high salaries is reasonable, as it currently lags behind in cutting - edge AI research. The salary of interns now exceeds the total amount of DeepMind's early seed - round financing.
  9. Hassabis believes that in the future, we will enter the era of "AI - generated interfaces". Interfaces will be customized according to users' preferences, matching their aesthetics, habits, and thinking styles. AI will generate them dynamically based on tasks.
  10. Hassabis believes that although it is impossible to quantify the "doom probability", the risks and uncertainties of AI technology cannot be ignored. We must adopt a cautiously optimistic attitude.

Demis Hassabis, CEO of Google DeepMind

Highlights of the interview:

01. How AI Mimics Natural Evolution

Fridman: In your Nobel Prize speech, you put forward an interesting idea: "Any model that can be generated or discovered in nature can be efficiently discovered and modeled through classical learning algorithms." Which fields of models do you think this might involve? Such as biology, chemistry, physics, or fields like cosmology and neuroscience?

Hassabis: I think there's a tradition in Nobel Prize speeches to present some provocative ideas, so I deliberately put forward this thought.

If you look closely at all the work we've done, especially the "Alpha series" projects - like AlphaGo (the AI system that defeated the Go world champion) and AlphaFold (the AI system that predicts the three - dimensional structure of proteins) - you'll find that they are actually building models for very complex combinatorial spaces. If you want to find the best move in Go or determine the exact structure of a protein through brute - force calculation, it would take longer than the entire lifespan of the universe.

So we have to use smarter methods. In these two projects, we built intelligent models for these environments and made the problems tractable through guided search. Take protein folding as an example. It's a natural process, and physics has explained why it occurs. Proteins fold in our bodies within milliseconds, which means physics has already solved certain problems in this process, and we've now successfully replicated it through computation.

I think this is possible because natural systems have inherent structure. These systems have gone through a long evolutionary process, so they have some "fixed patterns". If this structure exists, then we might be able to learn some rules from it.

Fridman: You seem to have mentioned that anything that can evolve can be efficiently modeled. How much truth do you think there is in this?

Hassabis: Sometimes I use the term "survival of the stablest" to explain this.

While there is evolution in living organisms, on a geological time scale, the shape of mountains is shaped by millions of years of weathering; even in the field of cosmology, the orbits of planets and the shapes of asteroids have gone through a kind of repeated "survival process".

If this is true, then there should be a model that can be learned in reverse, a structure similar to a manifold, which helps us efficiently search for the correct solutions, the correct shapes, and make predictions - because this is not a random model.

So for man - made or abstract things, like decomposition, this might not apply. Unless there are models in the digital space, but if there aren't and they are uniformly distributed, then there are no models to learn from, no models to assist in the search, and we can only rely on brute - force calculation.

In this case, tools like quantum computers might be needed. Most natural things we're interested in are not like this. Their structures have evolved for a reason and have withstood the test of time. If so, I think neural networks might be able to learn these structures. These systems can be efficiently rediscovered or restored because nature is not random. Everything around us, including the more stable elements, is under the pressure of a certain selection process.

02. Google Veo 3 and the Understanding of Reality

Fridman: I've always thought that the video - generation model Veo 3 is truly amazing. Many people say the videos it creates are fun and full of memes, like funny shorts and meme videos. But I think the most impressive thing is that it can portray human movements very naturally, and the pictures are very realistic. With the addition of the original audio, the whole thing feels very "human - like". But the most crucial point, which you also mentioned before, is that it can "understand" physical laws.

Hassabis: That's right.

Fridman: Veo 3 is not perfect, but it's already very impressive. The question is, how much does it need to understand the world to achieve this? Some people think that AI like diffusion models doesn't have the ability to "understand" at all. But I don't think so. If it doesn't understand at all, how could it create these videos? This actually brings up the very meaning of the word "understanding". How much do you think Veo 3 understands?

Hassabis: I think it definitely has a certain level of "understanding". At least it can continuously predict the changes in each frame of the picture, which is not an easy feat. Of course, its understanding is not the same as the in - depth thinking of humans, not the kind we talk about in philosophy. But it can be said that it has very well simulated the way the world works. For example, it can generate an 8 - second video that looks very natural, with almost no flaws. The progress is incredibly fast.

Everyone likes the talk - show videos it generates. It mimics people's speech and movements very realistically. But what impresses me the most is its ability to simulate physical effects such as light, materials, and liquids in the real world. That's truly astonishing.

It should have an ability of "intuitive physics". Just like a child knows that water will spill when a cup is knocked over and a ball will fall when it rolls down - this is an intuitive understanding of how the world works, rather than the knowledge of solving formulas in scientific research. It's a more intuitive understanding of physics based on feeling.

Fridman: Yes, it does understand something. I think this surprises many people. It also deeply touched me. I never thought that such realistic content could be generated without understanding. There's a view that only through an embodied AI system (an intelligent system that interacts with the environment through a physical carrier, perceives, makes decisions, and executes tasks) can we build an understanding of the physical world. But Veo 3 seems to directly challenge this view, as if it can learn just by watching videos.

Hassabis: That's right. This really surprised me. Even five to ten years ago, if you had asked me, I would have said that for an AI to understand something like "intuitive physics", it would probably need to have a body and be able to take actions.

For example, when I push a cup off the table and the water spills and the cup breaks, these are things we learn from life experience. Some theories in neuroscience also suggest that to truly "perceive" the world, you need to be involved in it and take actions to learn more deeply.

So I used to think that for an AI to understand these things, it at least needed to "move", whether by giving it a body or letting it act in a virtual environment. But now, it seems that Veo 3 can gain a certain level of understanding just by watching videos. This really amazes me.

It also shows that the real world we live in may actually contain a lot of "learnable" structures, and AI can learn a lot by observing these structures. It's not just about generating cool videos.

In the future, the next step might be to make these videos "interactive", for example, you can step into the video scene, look around, and even participate in it. Just thinking about it is exciting.

By then, we'll really be approaching a "world model". That is, AI can simulate the way the whole world works, its rules, and details in its "mind". This is also a necessary step in creating true artificial general intelligence.

03. Customizing Game Worlds with AI in the Future

Fridman: What will games be like in the next five to ten years?

Hassabis: In my teenage years, the first professional thing I did was developing AI for games. I've always hoped to revisit this interest one day. I wonder what kind of games I would make if I had today's AI systems in the 1990s. I think you could build truly amazing games.

My favorite game genre has always been open - world games, and most of the games I've made are in this direction. But as you know, developing open - world games is really difficult. Since players can go in any direction, you have to make sure that every choice is exciting enough. Many times, we can only use systems like cellular automata to try to create "emergent behavior", but these systems are often unstable, and the content they generate is limited.

Now the situation is different. I think in the next five to ten years, we'll have the opportunity to build game systems that truly revolve around players' imaginations. For example, AI can automatically generate storylines based on your every decision, making the whole story more tense and movie - like. These capabilities are already starting to show.

You can imagine that if there's an interactive version of Veo, or even a more powerful one, in the next few years, we might really be able to create games where "you think of a place, and it builds it for you".

Fridman: You said that the core of an open - world game is "choices define the world", but in some games, choices are just an illusion. Early games like "The Elder Scrolls: Daggerfall" created a sense of openness through random generation. Do you think that's enough?

Hassabis: No. Using random generation or pre - set options can't create a truly open world. The ideal state is that players can really do whatever they want without restrictions. This requires a powerful, real - time generation system where AI can generate an immediate response based on every decision you make.

Moreover, the development cost of such games is very high. Developing a 3A game already requires a lot of resources, and it's even more difficult to achieve a truly dynamic open world. In the early days, the game "Black & White" that I was involved in used reinforcement learning, and the game experience was completely mapped to players' actions. Today's general learning systems still aim for this goal.

Fridman: You and Elon Musk both love playing games. Do you still have time to make a game like "Black & White" yourself?

Hassabis: Maybe I'll try it in my spare time with the help of a coding assistant, or take an academic sabbatical after the advent of AGI - then I plan to develop games and study physics theories.

Fridman: Between "solving the P vs. NP problem" and developing games (P problem: a problem that can find an answer in a reasonable amount of time; NP problem: a problem that can verify whether an answer is correct in a reasonable amount of time), which one would you choose?

Hassabis: They are related. I want to make a realistic open - world simulation game to explore "what the universe is", which is in line with the essence of the P vs. NP problem.

Fridman: Can games become a place for humans to find meaning in the AI era?

Hassabis: Yes. Games can unleash imagination. The 1990s was the golden age of open - world games, allowing players to co - create stories, which is something other media can't offer. But the more realistic the simulation, the more we need to think about "the nature of reality" - what exactly is the difference between the virtual and the real?

Lex Fridman, well - known podcast host

04. AlphaEvolve and Evolutionary Algorithms

Fridman: Recently, DeepMind's evolutionary algorithm system, AlphaEvolve, is actually very impressive, but it hasn't received much attention from the outside world. It uses an evolution - like technique. Do you think this method could become part of future super - intelligent systems? As I understand it, AlphaEvolve works like this: the large language model proposes directions and ideas, and then the evolutionary algorithm searches widely for possible solutions. Is that the logic?

Hassabis: That's right. The large language model proposes possible solutions, and then evolutionary computation explores new areas of the search space. I think this is a very promising direction. Combining foundation models (such as large language models) with other types of computational methods, like evolutionary algorithms and Monte Carlo tree search, what we call "hybrid systems", is likely to lead to significant discoveries in the future.

Fridman: However, we shouldn't be too "romantic" about the evolutionary process. Do you think such mechanisms are really effective? Can our current understanding of natural evolution still yield some "low - hanging fruits" to help us conduct smarter searches?

Hassabis: I think it's valuable. The core is: you need to build a system that can understand the underlying dynamics. If you really want to discover new things, you have to let