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
I have zero difficulty accepting that they're using conversation logs, probably anonymized and then chopped so that they can claim it's no longer user data and that it can't be tracked back to a specific chat. And yes, sometimes, a single piece of text is enough to move the needle, especially when it comes to some exotic, sparse topics.
If someone's thinks their ideas are so precious, they should be using the privacy focused ones.
Even the free Google AI continues the conversation much longer and asks followup questions if you are serious.
Prompts of scientists will be forwarded to a data labeler and be used for training.
These prompts of course include the use case of converting a colleague's work to another file format using AI, which mathematicians always proudly announce. So they leak other people's ideas, too.
j45•2h ago
Or worse, not understand how this piece of technology is not snapshot based like the software before it, and instead is evolving constantly, and what impact that has on their base assumptions.