@rohanpaul_ai: Columbia CS Prof Vishal Misra explains why LLMs can’t generate new science ideas. Bcz LLMs learn a structured map, Baye…
Summary
Columbia CS Prof Vishal Misra argues LLMs can’t generate truly novel science because they only interpolate within learned Bayesian manifolds rather than create new conceptual maps.
View Cached Full Text
Cached at: 04/21/26, 05:13 PM
Columbia CS Prof Vishal Misra explains why LLMs can’t generate new science ideas. Bcz LLMs learn a structured map, Bayesian manifold of known data & work well within it, but fail outside it. True discovery requires creating new maps, which LLMs can’t do
Similar Articles
LLMs are not the black box you were promised
An article summarizing Anthropic's 2025 paper on mechanistic interpretability, showing that LLMs are not black boxes and that circuit tracing can reveal multi-step reasoning and human-identifiable concepts.
@rohanpaul_ai: This Yale + University of Chicago paper shows that real gap between LLM generated research ideas vs humans is not idea …
This paper from Yale and the University of Chicago finds that the main difference between LLM-generated and human research ideas is not quality but range: LLMs produce narrower ideas, with 47-64% of their ideas focusing on connecting separate work, compared to only 12.1% for humans.
@rohanpaul_ai: During a Bloomberg interview, Yann LeCun (@ylecun ) explains why LLMs are limited in terms of real-world intelligence d…
Yann LeCun explains in a Bloomberg interview that LLMs are limited because they only process symbolic text, while real-world understanding requires massive sensory data that children naturally acquire. He invokes Moravec's paradox to highlight the gap.
@rohanpaul_ai: Yann LeCun says LLMs aren’t a bubble in value or investment—they’ll drive many real-world applications and justify curr…
Yann LeCun argues that LLMs are not a bubble in value or investment, as they will drive many real-world applications and justify current infrastructure spending; the actual bubble is in assuming LLMs can achieve human-level thinking.
@haider1: Yann LeCun says LLMs are strongest in domains where language itself is the substrate of reasoning, like math and code T…
Yann LeCun states that LLMs are strongest in domains where language is the substrate of reasoning, like math and code, but they are not creative mathematicians, software architects, or computer scientists.