I think OpenAI said it had something to do with over-indexing on user feedback (upvote / downvote on model responses). The users like to be glazed.

If there's one thing I know about many people (with all the caveats of a broad universal stereotype of course), they do love having egos stroked and smoke blown up their ass. Give a decent salesperson a pack of cigarettes and a short length of hose and they can sell ice to an Inuit.

I wouldn't be surprised at all if the sycophancy is due to A/B testing and incorporating user responses into model behavior. Hell, for a while there ChatGPT was openly doing it, routinely asking us to rate "which answer is better" (Note: I'm not saying this is a bad thing, just speculating on potential unintended consequences)