François CholletVerified account

I really regret having to say this, because the paper was actually very good (if incremental, but there's nothing wrong with that). But the surrounding PR blitz and public misrepresentation is causing serious damage to the field and its public perception. Please don't do this.

Yes, fear-mongering can be a good strategy to inflate the perceived significance of your research. It's also completely irresponsible, and even dangerous. The less PR circus, the more room we will have to talk about the safe and responsible deployment of the latest ML research.

When you say fear-mongering, do you mean reaching out to journalists in advance of publication, something wrong with the framing of the result in the blog post, or something wrong with the discussion of risks, all of the above, something else, etc.?

SotA LMs get published on a regular basis (with or w/o the model). Why did this one need an all-out PR assault? Whether releasing some models raises security issues is a legitimate debate (tho this model prolly doesn't). Is Bloomberg or the Guardian the right place to have it?

How is it not misrepresentation to take best-of-25 samples from cherrypicked prompts and tell the public, "the AI wrote this, this is what AI does now"? BTW I have a trading algo that does 12,000% per year* (*evaluated with best-of-25 daily trades)

I'm maybe missing something - the AI *did* write that/does that now, and the reason you know it was cherrypicked is because it said it right there?

fwiw: i think a lot about cherry picking in my own work and decided that it's ok when also given the context of how big the pool size is. eg: here's a notable sample taken from a pool of 100 generated candidates.