One of the wildest things in the history of France is the "Cagots": a minority with no other apparent distinguishing feature, who were persecuted for centuries until the early 1900s.
When @gwern's GPT-3 post came out, I was starting a PhD with a bunch of time on my hand. I remember explicitly formulating the goal: "starting today, I will spend hours playing around with this model, becoming a world GPT-3 expert. This will be a worthy way to use my time."
In French, Claude is used both as a masculine and a feminine way. I thought Anthropic chose this name to make gender ambiguous.
According to this interview of Daniela Amodei, thatโs not the case.
Over the next 5 years I proceeded to do... not that much of that. Some, more than most, but I didn't really put in the hours. I did some other things, including some cool things, but I think I regret not doing that.
I liked @nostalgebraist's recent Void post, but to be honest it also made me feel a bit sad and called out.
Fortunately it's not too late to do that kind of open-ended exploration, at least with more recent / still available models.
Is anyone working on cleaning up Humanity's Last Exam?
There's lots of mistaken / ambiguous / uninteresting questions because Dan is IMO unreasonably focused on sample size, but I'm pretty sure one could extract a high quality ยฑ500 Q subset that I'd be fairly excited to run.