Physicist Argues LLMs Are Limited by Language, Cannot Understand Reality
skdh · x · 2026-08-06
Recalling a dinner at the Royal Society with figures like Demis Hassabis and Tim Gowers, a physicist discusses AI's potential creativity in mathematics. Gowers viewed the evolution of proofs as a meta-extrapolation of methods, arguing that creativity doesn't come out of nowhere and human minds aren't unique in injecting random elements.
The author notes that new ideas are essentially about connecting dots, something a computer can theoretically reproduce. However, they remain skeptical about current LLMs. They argue that language is a lossy tool developed to transfer information between human brains. Therefore, models trained solely on language will struggle to reconstruct a genuine understanding of reality.
More from AGI Musings
- AI Alignment is an Information Ecology Problem Requiring New Quality Data — nptacek · 2026-08-06
- Compute is Regulatable but Information Isn't: Open-Weight Alignment is Critical — nptacek · 2026-08-06
- Open Models Can Already Bootstrap Recursive Self-Improvement, Sparking Regulatory Concerns — nptacek · 2026-08-06
- The Future of Software: 'Jellyware' Between Traditional Code and Vibecoding — cedric_chee · 2026-08-06
- SpaceX: The Closest Thing to a Technological Singularity in the Physical World — XFreeze · 2026-08-06
- LeviTurk: Demis Will Wield It with Awe and Respect Like Frodo — LeviTurk · 2026-08-06