LLMs May Lack the Ability for "Leap" Discoveries
burny_tech · x · 2026-07-10
This repost discusses a cognitive control perspective: LLMs act more like bottom-up salience machines, rather than top-down regulatory systems capable of task-irrelevant transfer like humans.
The cited content further breaks down "discovery" into three capabilities:
- induce: Inducting from data, yielding at best Newtonian-style explanations with correction terms
- deduce: Strictly deriving based on existing axioms, but generating no new axioms
- jump: Inventing new frameworks, such as proposing "spacetime curvature"
The core argument is that even if an LLM were fed all papers, data, and formulas prior to 1905, it might not achieve the "leap" discovery that Einstein did, because that step goes beyond both induction and deduction.
Related event: Opinions Suggest LLMs Lack Capacity for Scientific Leaps(3 posts)→
More from AGI Musings
- Open-source labs could distill a state-of-the-art model to 32GB or 80GB VRAM, the post argues — bookwormengr · 2026-07-21
- Two US companies are now using superintelligence to speed up the next generation of models — yacineMTB · 2026-07-21
- MIT Sloan says information, national security and finance are most exposed to AI — Exp_Mark · 2026-07-21
- IMF says AI could lift Sub-Saharan Africa’s economy by 4% over the next decade — Polymarket · 2026-07-21
- Bluesky’s “AI con” debate is shifting facts while keeping the same tone — iskander · 2026-07-21
- Competition is pushing AI forward faster than ever, the post says — eyishazyer · 2026-07-21