Date: July 31st, 2026 11:33 PM Author: Dashing puce dopamine
I can't talk to Claude anymore without it arguing with me. They seem to think that making it "not sycophantic" has to mean making it challenge every single point you make, because your original statements didn't include a long chain of qualifications
Date: July 31st, 2026 11:36 PM Author: Metaphysics Is Solved
yeah its useless. the whole point of an AI in the causal, non-information retrieval use case is as an extension of a frame of thought and exploring an idea within your own mental model of it. claude is really "good" at conflating an idea exploration prompt with an info retrieval prompt, so that your speculation is immediately converted into a thesis to be disproven, rather than articulated and explored at its edges. that's why the chatgpt sycophancy is actually good for these kinds of medium-complexity tasks. i dont actually want AI to argue with me until i actually know what i think, and im using the AI to help me define what that is.
Date: July 31st, 2026 11:48 PM Author: Dashing puce dopamine
Yeah that's what I think about it too. Once it shifted to Sonnet 5 it started spending almost all of its thought chain in the search function reading through literature fact checking stuff. They did this to reduce "hallucinations". GPT 5.6 does this now too, but the difference is it can distinguish between if you are having an exploratory conversation or are looking to verify facts or literature.