Interesting!

I wonder if it’s carried over too much of that ‘helpful’ DNA from 4o’s RLHF. In that case, maybe asking for 500 words was the difficult part — it just didn’t have enough to say based on one SO post and one article, but the overall directives assume there is, and so the model is put into a place where it must publish..

Put another way, it seems this model faithfully replicates the incentives most academics have — publish a positive result, or get dinged. :)

Did it pick up your HN comments? Kadua claims that’s more than enough to roast me, … and it’s not wrong. It seems like there’s enough detail about you (or me) there to do a better job summarizing.