Comment by myfonj
3 months ago
The fun, even hilarious part here is, that the "fix" was most probably basically just replacing
[…] match the user’s vibe […]
(sic!), with literally
[…] avoid ungrounded or sycophantic flattery […]
in the system prompt. (The [diff] is larger, but this is just the gist.)
Source: https://simonwillison.net/2025/Apr/29/chatgpt-sycophancy-pro...
Diff: https://gist.github.com/simonw/51c4f98644cf62d7e0388d984d40f...
This is a great link. I'm not very well versed on the llm ecosystem. I guess you can give the llm instructions on how to behave generally, but some instructions (like this one in the system prompt?) cannot be overridden. I kind of can't believe that there isn't a set of options to pick from... Skeptic, supportive friend, professional colleague, optimist, problem solver, good listener, etc. Being able to control the linked system prompt even just a little seems like a no brainer. I hate the question at the end, for example.
This isn't a fix, but a small patch over a much bigger issue: what increases temporary engagement and momentary satisfaction ("thumbs up") probably isn't that coupled to value.
Much like Google learned that NOT returning immediately was the indicator of success.