Demis Hassabis Interview

The interview with DeepMind founder and CEO, Nobel laureate Demis Hassabis, revolves around artificial intelligence from basic theory to future vision, covering core topics such as natural laws, computational complexity, multimodal world modeling, biological simulation, AGI paths, scaling laws, energy prospects, and humanistic thinking.

1. Learnable Natural Patterns

  • Any system that has “survived” through evolution or physical processes contains structures that can be efficiently learned by neural networks.

  • Projects such as AlphaGo and AlphaFold are precisely reverse-engineering the low-dimensional manifolds of natural systems, using intelligence to guide search, thereby achieving feasible solutions in an astronomical-level combinatorial space.

2. Computational Complexity and P vs NP

  • The debate between P-class problems (solvable in polynomial time) and NP-class problems (quickly verifiable but difficult to solve by exhaustion) is essentially “which natural systems can be efficiently simulated by classical computing models.”

  • In the future, a “learnable natural system” (LNS) complexity class may be defined to specifically characterize the solvability of such models for physical, chemical, and ecological problems.

3. World Modeling and Veo 3

  • Veo 3 not only generates visual content, but also builds an internal world model with rigorous spatiotemporal, causal, and physical logic through “language → video”.

  • This ability allows AI to deduce scene evolution and predict the next frame after receiving a text description, providing a multimodal inference interface for AGI.

4. Video Games: AI’s Cognitive Laboratory

  • Game environments have purposeful tasks, immediate feedback, and controllable complexity, making them an ideal platform for training decision-making, planning, social reasoning, and other abilities.

  • From Demis’s early design of “Black & White” to AlphaGo and StarCraft II, games have promoted the continuous iteration of AI in multidimensional cognitive challenges.

5. Biological Evolution Simulation: AlphaEvolve

  • On the basis of AlphaFold’s static structure prediction, AlphaEvolve introduces evolutionary algorithms and dynamic environmental variables to simulate the process of natural selection and genetic variation.

  • Applications cover drug development, ecological prediction, and lay the foundation for drawing inspiration for intelligent architecture from biological mechanisms in the future.

6. From Primitive Life to General Intelligence

  • Around the topic of “the origin of life,” Demis envisions that AI can reproduce the key chemical and physical steps from non-living matter to primitive organisms.

  • Such simulations can deepen our understanding of the nature of life and promote the exploration of AGI projects in the “from zero to one” stage.

7. A Phased Blueprint for AGI

  1. Modeling reality: AlphaFold and Veo 3 provide static and dynamic world models.

  2. Predicting the future: AlphaEvolve deduces the evolutionary paths of organisms and the environment.

  3. Generalized reasoning: Unified processing of multimodal language–vision–sound (such as Gemini).

  4. Continuous learning: The collaboration of various systems to achieve cross-task and cross-domain transfer capabilities.

8. Scaling Laws, Computation, and Future Energy

  • The scaling law of deep learning: the three dimensions of parameters, data, and computing power jointly drive the exponential improvement of model capabilities.

  • With the surge in demand for computing power, the future may rely on new energy sources (fusion, advanced solar energy) and more efficient computing architectures to support the continuous expansion of AGI research.

9. Human Nature, Consciousness, and Quantum Computing

  • The uniqueness of humans lies in “metacognition” and irrational emotions, which have so far been difficult to capture by classical architectures.

  • Although it is not yet conclusive, quantum mechanisms may play a role in the formation of consciousness; if AGI needs to have subjective experience, it may use quantum computing to break through the bottleneck of traditional symbol processing.

10. Future Vision for Education and Research

  • Education should shift from knowledge instillation to problem-driven exploration, with AI becoming a “cognitive assistant” that stimulates curiosity and critical thinking.

  • Personalized learning, interdisciplinary integration, and the grasp of ethical boundaries will determine how technology truly serves the release of human potential.