Historical source material recovered on 2026-09-09. Capture dates and source limitations apply; recovery is not fresh verification.

World Models

World models are learned representations of how an environment behaves, including its objects, physics, dynamics, constraints, and possible interactions.

Key points

• Hassabis describes video-generation progress as moving toward models of the mechanics and physics of the world, not just prettier media outputs ^[src-063].
• Veo is discussed as a sign that models can infer enough structure from video to render materials, liquids, lighting, and short physical sequences ^[src-063].
• The next step is interactivity: moving from generated clips to environments a user can enter, steer, and play inside ^[src-063].
• Hassabis connects world models to AGI because a general system needs an internal model of how the world works, not only text prediction ^[src-063].
• Games are a natural testbed because open-world simulations already combine rules, agents, environment state, and player co-creation ^[src-063].
• Fan gives the robotics version: video models can act as next-world-state simulators, but robotics needs to align those simulations with actions so physically useful futures, not just plausible pixels, guide behavior ^[src-082].
• Dream Dojo extends the idea into neural simulation: given continuous action signals, a learned model predicts future RGB frames and sensor states without a classical graphics engine or explicit physics equations ^[src-082].
• LeCun adds an agent-planning definition: a world model predicts the next state from the current state plus an imagined action, so an agent can anticipate consequences before acting ^[src-102].
• The source distinguishes representation-space prediction from pixel-level generation: useful world models should ignore unpredictable detail and predict abstract state that supports planning ^[src-102].

Related entities

• Demis Hassabis
• Google Deepmind
• Veo
• Alphago
• Jim Fan
• Nvidia
• Yann Lecun
• Advanced Machine Intelligence Labs

Related concepts

• Intuitive Physics In AI
• Learnable Natural Systems
• Embodied Reasoning
• AI For Science
• World Action Models
• Physical AI
• Robotics Data Loop
• Joint Embedding Predictive Architecture
• Agentic AI

Source references

• ^[src-063] Lex Fridman – "Demis Hassabis: Future of AI, Simulating Reality, Physics and Video Games | Lex Fridman Podcast #475" (2025-07-23)
• ^[src-082] Sequoia Capital — "Robotics' End Game: Nvidia's Jim Fan" (2026-04-30)
• ^[src-102] Vivatech / Yann LeCun – "Beyond Language Models: Building AI that Understands the World" (2026-06-17)

Historical recovered snapshot — 2026-08-19

Preserved from ai-watch-src349-evidence-correction-20260819-105611. Documentary recovery on 2026-09-09; claims below retain their historical source scope and were not freshly verified. Current synthesis above remains authoritative.

• Romero frames world models as a possible centre of gravity for Google DeepMind's longer-term strategy, contrasting them with coding-agent-led automated AI research at OpenAI and Anthropic ^[src-349].
• Existing Hassabis evidence supports DeepMind's interest in simulation, learned natural-system structure, robotics, and AI for science ^[src-063]. It does not prove that Google has rejected coding agents or recursive-improvement research.
• Treat “world models versus recursive self-improvement” as an analytical comparison of emphasis. World models could themselves be used inside automated research or successor-system development, and all three labs maintain overlapping portfolios ^[src-349].
• The full essay defines the stronger world-model thesis as a grounding claim: accurate prediction can rely on brittle shortcuts without learning transferable causal structure. A 2025 orbital-mechanics study supports that narrow point; the 2026 Stanford Theory of Space benchmark independently tests active spatial belief construction and is not a replication or formal follow-up ^[src-349]. Neither directly evaluates Veo, Genie or Gemini Omni.
• Generating plausible video or interactive environments is not itself proof of a stable causal model. Evaluation should test counterfactual consistency, physical persistence, adaptation under novel constraints and transfer into planning or robotics ^[src-349].

• ^[src-349] Alberto Romero / The Algorithmic Bridge – "The Actual Reason Why Google ‘Fell Out’ of the AI Race Changes Everything" (2026-07-28; updated 2026-08-12)