Why LLMs Can't Make Top Scientific Discoveries: The Missing 'Unexplained Jump'
CSProfKGD · x · 2026-08-01
A tweet quotes an in-depth discussion on the cognitive limits of Large Language Models, centering on a position paper exploring whether AI can achieve true scientific discovery. The paper serves as a philosophical and technical response to Demis Hassabis's proposed "Einstein Test"—training an AI on pre-1911 knowledge to see if it can independently derive General Relativity.
Key arguments from the paper include:
- Refuting Data Compression: While figures like Schmidhuber argue scientific discovery is essentially data compression, the paper notes that if this were entirely true, Einstein couldn't have invented General Relativity without sufficient anomalous observational data to compress.
- The Unexplained Jump: Top human scientists cross cognitive chasms using thought experiments to achieve an "abductive jump." LLMs, however, are merely excellent logical interpolators of existing knowledge. Relying on probabilistic distribution for induction, they are structurally incapable of making this leap.
The author concludes that this cognitive jump is fundamentally an embodied world model problem. Throwing more compute at text models only traps them in a shadow play of symbols, forever unable to achieve true scientific breakthroughs.
Related event: Research Finds LLMs Lack Abductive Leap Ability for Scientific Discovery(2 posts)→
More from AGI Musings
- Is math just a language? A debate on why LLMs are good at mathematics — LucaAmb · 2026-10-16
- Paul califax argues extrapolation isn't special: LLMs can learn out-of-distribution skills via high-dimensional interpolation — paul_cal · 2026-10-16
- Tech improves far faster than it diffuses — organizational bottlenecks remain the key intuition — Afinetheorem · 2026-10-16
- Ramez Naam weighs in on mech interp blog hype: bullish overall, lacks tools to verify claims — ramez · 2026-10-16
- Mistral CEO Arthur Mensch: Europe abandoning frontier AI would mean unprecedented dependence — arthurmensch · 2026-10-16
- LLMs ranked by consciousness (Φ) under IIT — Cerebras-backed GPT tops the chart — Sauers_ · 2026-10-16