Can LLMs Actually Invent New Science?

A YouTube video discusses a research paper (referred to as 'LLMs can't jump') by a Google DeepMind researcher, arguing that while LLMs are good at solving well-defined problems, they struggle to make the kind of conceptual 'jump' that redefines the problem itself, as Einstein did with gravity and acceleration. The creator illustrates this with an example of an AI agent optimizing a slow database query when the real issue was a suspicious-looking payment page. The video questions whether AI systems that search through many possibilities and stumble on a novel idea are truly making a scientific leap or just getting very good at search, and teases an upcoming video about the JEPA research approach.

The Leap LLMs Still Can’t Make - Code With Bro 2.2K subscribers