ML is known to be good at interpolating between points in the training set, but does much worse at extrapolating. Both can produce innovation when applied to science. For example, a lot of innovation is about transferring existing ideas from one domain to another.
We start seeing examples of LLMs proving novel math theorems: https://news.ycombinator.com/item?id=48071262. But those could be “interpolation”. Is there any strong evidence for or against LLMs being capable of extrapolating?
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。