DeepMind's Hassabis: Language Models Can't Replace Physical Experience, World Models Are the Goal
joao_gante · x · 2026-08-12
Google DeepMind founder Demis Hassabis discussed his longstanding passion for world models and highlighted the limitations of current language models (LLMs) on a podcast.
While language is richer than expected and helps models understand a lot, Hassabis noted that spatial dynamics, physical context, and mechanics are hard to convey purely through text. Furthermore, many aspects of reality require sensory experience (like motor angles or smell) that language models simply cannot capture.
More from AGI Musings
- Why Are We Still Debating AGI? Three Theories on the Moving Goalposts — neal_lathia · 2026-08-12
- Debate: Is LLM Reasoning Just Gradient Descent or Actual Intelligence? — ctjlewis · 2026-08-12
- Are LLM CoTs Unreliable? Researchers Call for Deep Dive into Latent Space Computations — ricklamers · 2026-08-12
- Voice-Orchestrated Cloud Agents Will Reshape Personal Computing — petergyang · 2026-08-12
- 71% of Americans Oppose Nearby Data Centers, Surpassing Nuclear Plants — PeterDiamandis · 2026-08-12
- Looking back: 95% of AI predictions were wrong; no AI bubble, 'AI company' is meaningless — tlakomy · 2026-08-12