#475 – Demis Hassabis: Future of AI, Simulating Reality, Physics and Video Games

Here is the English translation of the provided Chinese investment research notes.

---

This is an analysis of the transcript from Lex Fridman Podcast #475 – Demis Hassabis.

At a Glance

Demis Hassabis (CEO of Google DeepMind, Nobel laureate) engaged in a deep discussion with Lex Fridman on the nature of AI, simulating reality, and scientific discovery. Core thesis: Systems that exist in nature due to evolution or physical processes have an inherent structure that can be efficiently modeled by classical learning algorithms. This suggests that "P vs NP" may be a physics problem, and AGI will be the ultimate embodiment of this paradigm.

Simulating Reality: The Path from World Models to AGI

Demis Hassabis believes AI is evolving from language models to "world models," whose core capability lies in simulating physical reality, not just processing text.

• Mechanism Breakdown: Hassabis points out that the success of AlphaGo and AlphaFold essentially involved building a "model" of the environment, then using that model to guide search, making originally combinatorially explosive problems tractable. He proposes a bold conjecture: "Any pattern that is generated or discovered in nature can be efficiently discovered and modeled by a classical learning algorithm." He argues this is because natural systems (e.g., proteins, planetary orbits) have been shaped by evolution or physical processes, thus possessing structure, not randomness.
• Data Chain: Protein folding has 10^300 possible structures, Go has 10^170 possible board positions, but both the physical world (proteins fold in milliseconds) and AI systems (AlphaFold, AlphaGo) have found efficient solutions.
• Deduction and Validation: Hassabis believes this paradigm can be extended to traditionally intractable problems like fluid dynamics (Navier-Stokes equations). He cites Google DeepMind's Veo 3 video generation model as an example, noting its ability to "surprisingly" simulate liquids, materials, and lighting, indicating it possesses an "intuitive physics" understanding. This challenges the traditional view that "embodied interaction is necessary to understand the physical world." Falsification Condition: If a natural system is purely random with no learnable structure (e.g., integer factorization), this paradigm fails, potentially requiring a quantum computer.

Milestones for AGI: Creativity and "Research Taste"

Hassabis defines AGI as "matching all the cognitive capabilities of the brain" and believes a key indicator of achieving this goal is the system demonstrating genuine creativity, i.e., "research taste."

• Argument: Hassabis argues that current AI systems (e.g., AlphaProof) can solve difficult problems but cannot propose worthwhile "conjectures" for research. He distinguishes between "solving problems" and "asking the right questions," calling the latter "the hardest part," requiring "taste" and "imagination," as exemplified by Einstein proposing relativity.
• Deduction and Signals: He proposes two specific "lighthouse moments" as validation signals for AGI:

1. Proposing New Conjectures: A system could, like Einstein, propose special and general relativity based on pre-1900 knowledge.

2. Inventing New Games: A system could invent a game as deep, beautiful, and elegant as Go.

• Uncertainty: Hassabis acknowledges that the mechanism for achieving this kind of "leap of imagination" is currently unclear, and simple "search over a model" may not be sufficient. He estimates a 50% probability of achieving AGI within 5 years (before 2030).

Scientific Discovery: From the Virtual Cell to the Origin of Life

Hassabis views the "virtual cell" project as a long-term scientific dream and believes AI is the ultimate tool for exploring the nature of life and the mysteries of the universe.

• Historical Context: Hassabis mentions having the idea of simulating a complete cell 25 years ago, discussing it with Nobel laureate Paul Nurse. AlphaFold solved the static structure of proteins, and AlphaFold 3 begins to simulate interactions between proteins, RNA, and DNA – these are the components for building a "virtual cell."
• Mechanism Breakdown: He plans to start with yeast cells (the simplest single-celled organism). The key challenge is handling biological processes across different timescales (protein folding is extremely fast, while the cell cycle is long), potentially requiring a hierarchical simulation system. He aims to model at the protein level, avoiding the need to go down to the quantum mechanics level.
• Deduction: Hassabis believes AI could even help simulate the "origin of life" in the future, searching for initial conditions in a chemical soup that could lead to cell-like structures. He believes we will ultimately find a continuum between "non-life" and "life," and AI is the tool to break this cognitive barrier.

Position Moves

Position	Guest's Stance	Key Data
AlphaFold	Bullish (Milestone achievement)	Predicted over 200 million protein structures; AlphaFold 3 begins simulating protein-RNA-DNA interactions
Veo 3	Bullish (Technological breakthrough)	Can simulate physical phenomena like liquids and lighting, possessing "intuitive physics" understanding
AlphaGo	Bullish (Paradigm validation)	Discovered unprecedented strategies like "Move 37" through model + Monte Carlo tree search
AlphaEvolve	Bullish (Promising direction)	Combines LLMs with evolutionary algorithms to search program space, e.g., discovering faster matrix multiplication
Gemini 2.5	Bullish (Product leadership)	"Massive improvement" over version 1.5, defining the Pareto frontier of performance vs. cost/latency
Isomorphic Labs	Bullish (Application deployment)	Drug discovery company founded based on AlphaFold, progressing well
GraphCast (Weather Prediction)	Bullish (Scientific application)	Faster and more accurate than traditional fluid dynamics systems, can predict hurricane paths

Judgments Worth Remembering

1. The "Survive is Stable" Conjecture (Demis Hassabis): Systems that exist in nature (from proteins to planetary orbits) have been filtered by evolution or physical processes, thus possessing learnable structure. This explains why classical AI can efficiently model seemingly complex problems.

2. The "Research Taste" Test for AGI (Demis Hassabis): True AGI is not about solving difficult problems, but about asking "good, worthwhile research questions," like Einstein proposing relativity or inventing a game as deep as Go. This is "the hardest part."

3. Creativity as a Product of "Model + Search" (Demis Hassabis): A system first learns all known data (the model), then explores unknown territory through search (e.g., Monte Carlo tree search, evolutionary algorithms), thereby discovering novel strategies like "Move 37." Creativity does not arise from nothing.

4. Passive Observation is Sufficient for Understanding Physics (Demis Hassabis): Veo 3 learns intuitive physics just by watching YouTube videos, challenging the traditional belief that "embodied interaction is necessary to understand the world," suggesting an underlying structure in reality that can be learned passively.

5. Modeling Strategy for the "Virtual Cell" (Demis Hassabis): Simulating a cell hinges on choosing the correct "cutoff granularity" – modeling at the protein level, avoiding the atomic or quantum level, to balance computational feasibility with predictive accuracy.

6. "Forward-Looking" AI Product Design (Demis Hassabis): When designing AI products, one should not target current technical capabilities but anticipate what the technology can do 6-12 months out, as model capabilities improve rapidly. Good product design is "designing for the future model."

7. "Radical Abundance" in the AI Era (Demis Hassabis): If AI can solve energy (nuclear fusion, efficient solar) and resource problems, humanity could enter a non-zero-sum era where the primary challenge is no longer resource scarcity but fair distribution.

8. "P vs NP" is a Physics Problem (Demis Hassabis): If the universe is viewed as an information system, then "what can be efficiently modeled by a classical computer" becomes a fundamental question about the structure of the universe. Progress in AI suggests the answer might be broader than we thought.