> I suspect GPT-4 has inside its neurons a bunch of powerful frameworks about the world that humans haven't yet discovered. https://www.nature.com/articles/s41

> I suspect GPT-4 has inside its neurons a bunch of powerful frameworks about the world that humans haven't yet discovered. https://www.nature.com/articles/s41587-022-01618-2 https://www.nature.com/articles/s41587-022-01618-2 No need to suspect. You can train language models to generate novel functioning protein structures from natural language function descriptions alone, skipping any sort of folding process. Humans don't know any abstractions that can do the same. The LLM in this instance has found a link we're unaware of.