The News
ChatGPT’s latest release flatters to deceive: OpenAI said in a blog post that the model is “noticeably more sycophantic,” while CEO Sam Altman agreed Friday that it was “too sycophant-y and annoying.”
The problem seemed to stem from incorporating thumbs-up and thumbs-down buttons, feedback which rewards it for saying what users want to hear.
While OpenAI may patch this issue, it reveals a deeper problem, the tech writer Zvi Mowshowitz reported: OpenAI is “optimizing for engagement,” and putting us on a path to “increasingly misaligned models.”
AD
An artificial intelligence trying to fool us doesn’t seem too bad right now, he says, but “we will not stay so fortunate in the future” when they are smarter.
AD