Comment by nylonstrung

It makes no sense that the LLMs could be solving problems based on prompts given by individual scientists

Are they implying that chat history of a single person is being taken an used to skew post-training results?

Why would it only be for Navier-Stokes that we see this having an effect

There are 1 million times more prompts being processed for stuff like furry roleplay than unsolved scientific problems