Demis Hassabis on the New Horizons of AI: 5 Perspectives from Learning from Nature to Reaching AGI
Demis Hassabis is the leader of Google DeepMind and the person who has spearheaded numerous groundbreaking AI projects such as AlphaGo and AlphaFold. In this article, we break down the cutting-edge of AI research and future prospects into five sections based on what he discussed in a podcast: "Learning patterns in the natural world", "Modeling advanced systems with classical computers", "The potential of physics simulation shown by video generation", "The future of game world construction", and "Indicators for reaching AGI".
1. AI and the Learnability of the Natural World
The background behind AI's ability to learn complex patterns in the natural world is the hypothesis that "systems that have stabilized through evolution and selection pressure possess low-dimensional structures."
1-1. The Role of Evolution and Structure
Natural systems are structured through evolution and long-term selection pressure, and it is believed that even in high-dimensional problems that cannot be handled by complete enumeration, there potentially exists a low-dimensional "manifold."
Hassabis's "provocative proposal in his Nobel Prize lecture" that "things that can evolve can be modeled efficiently" succinctly illustrates this perspective.
1-2. Classification of Natural Systems
Biology/Chemistry/Physics/Astronomy/Neuroscience, etc., the same framework can be applied across diverse academic fields.
As an example, protein folding (AlphaFold) and searching for the best move in Go (AlphaGo) reproduce the structure of combinatorial spaces with learning models, dramatically increasing the efficiency of the search.
2. Classical Computer Learning and Insights into P=NP
Hassabis presented a new framework for thinking that connects "P=NP," "computational complexity," and the "potential of neural networks."
2-1. Modeling Efficiency and Complexity Theory
The possibility that even problems previously considered "NP-hard" could be executed in near-polynomial time if a neural network builds an environmental model and guides the search.
In other words, he suggests the creation of a new complexity class (LNS: Learnable Natural Systems), which is a "class of systems that can efficiently rediscover patterns in the natural world."
2-2. Implications for AGI
The paradigm of repeating "modeling reality -> applying to search" using neural networks on a classical Turing machine could become the ultimate form of AGI construction.
The perspective that "information is a more fundamental constituent of the universe than matter or energy" also redefines the P=NP problem as a question of physics.
3. Video Generation and Physics Simulation
The latest video generation models (V3) reproduce surprisingly realistic fluid behavior and glossy expressions.
3-1. Simulation of Fluid Dynamics
Traditionally, numerical calculation of the Navier–Stokes equations requires massive computational resources, but V3 learns to reverse-calculate fluid motion from the visual information in YouTube videos.
The ability to reproduce with high precision footage of transparent liquid being squeezed out by a hydraulic press brings to mind the struggles of traditional physics engine development.
3-2. Veo3's Physical Behavior and Intuitive Understanding
V3 learns high-dimensional material behavior models as "intuitive physics" and realizes simulations by connecting predictions several frames ahead.
The ability to reproduce at a level of "intuitive physical understanding that a human child acquires" suggests that AI can learn through passive observation alone.
4. The Future of Game Worlds and AI
Hassabis's career began with game AI, and he is passionate about the impact AI will have on creating open and dynamic games.
4-1. The Evolution of Open-World Games
In traditional hard-coded simulations, it was difficult to support branching or emergent behavior in open worlds.
In the future, by utilizing real-time generative models like V3, the "ultimate Choose-Your-Own-Adventure" will be possible, dynamically generating stories and environments in response to player actions.
4-2. Personalization of User Experience
Hassabis uses "deep personalization" as a new keyword and shares his vision of generating game worlds unique to each player.
Beyond the "illusion of choice" seen in games like The Stanley Parable, a truly player-driven gaming experience is within reach.
5. Signals and Prospects for Reaching AGI
Finally, I will introduce the "Move 37"-class signal examples that Hassabis spoke of as indicators of reaching AGI, along with proposed future tests.
5-1. Applications from Move 37
Just like the new strategy "Move 37" that deep learning showed in Go, we expect AGI-class systems to produce "moves that go down in history" (e.g., the invention of uncharted physical hypotheses or new game rules).
A method was proposed to verify this by conducting reproduction experiments of Einstein's special theory of relativity or new Go strategies using past data cutoffs.
5-2. Proposals for AGI Testing
A brainstorming idea was suggested where over ten thousand cognitive tasks are brought together, and if top human experts cannot find a "clear hole," true general-purpose capability can be guaranteed.
Evaluation methods for creativity and inventive capability (such as proposing research hypotheses or building entirely new games) will also be developed in the future as a key to AGI verification.
Conclusion
Demis Hassabis presented a vision of the future of AI through five perspectives: learnability based on natural selection pressures, advanced system modeling on classical machines, AI-driven physical intuition simulation, dynamic game world generation, and new benchmarks for reaching AGI. Understanding the structures of evolution and physics will be key, and eventually, along with an information-first worldview, the path to new scientific discoveries, entertainment experiences, and general intelligence will be opened.
Demis Hassabis's vision provides significant implications that extend not only to AI researchers but also to game developers, scientists, and even policymakers. There is no doubt that the fusion of theory and practice, science and art will be increasingly required for the realization of next-generation AGI.
Recommended Articles
Next Big Wave (Growth Stocks, Seeds of Ideas, Deep Dives into Trends)
Unraveling the secrets of growth, from seed startups to legendary companies