Narrator: He definitely let that slide. :D

Few obvious questions: Is the upper limit of ~30% hallucination acceptable? And how much work it adds to a lawyer’s work to ensure there is no hallucinations at all in the generated document?

(Haven’t read the entire paper yet)

Honest answer: I don't read anything from the enormous stream of papers unless they are *critical* to my current work or I am deeply interested in the topic.

Okay, automated cross-posting has its quirks. This was the original tweet: x.com/OpenAI/statu...

Next year we will reach NeurIPS level at this rate.

If we start calling them anything else, it doesn’t anthropomorphise the models enough. :)

I don’t think any of the “foundation model” orgs will ever talk about contamination, because that makes scores look bad. At this point a lot of these models seem like clowns in circus, only there to keep the public (investors?) at large excited about the circus that is “AI research”.