2026-04-17 08:31 UTC
@tschenkel @tokyo_0 @yaelwrites as I said, they do not address the underlying problem. They are sycophantic always.
You could make the analogy that they are improv partners, going, "Yes, and..." no matter what. That's what they are underneath all the scaffolding, and the earliest versions were prompted something like, "Finish the paragraph that begins like this..." followed by the user's input.
They're more advanced now, as is the scaffolding around them, but they're still "yes, and..."
Replies (1)
-
@swelljoe@mas.to 2026-04-17 08:34
@tschenkel @tokyo_0 @yaelwrites I suspect that if you understand they are playing along with whatever scenario you're playing out, and not actually thinking about what you've told them to do, you probably have a healthier relationship with them. When people think an LLM telling them how awesome they are on any metric means anything at all other than that's a probable way for a conversation like that to go (especially when the models are trained to be helpful), that's when things get dangerous.