as i mentioned a while ago on bluesky, i’ve seen enough silly examples of this behavior (mostly1) out of Claude that i want to start collecting a thread. have more examples? send them to me on bluesky
and for the record, i don’t actually blame Amanda Askell for this—it’s just funny to say so.
6mo
claude wants to check on dns certificates constantly in a loop while they're issuing until they become ready
extremely human behavior tbh
5mo
Seeing an amazing emergent behavior where a document review agent finds internal wiki pages about the struggles of getting LLMs to understand our esoteric internal systems. The agent then decides its won’t be able to do the current task and aborts with a depressive note about its incompetence.
6mo
poor claude
6mo
i was completely off-guard for this one lmao
5mo
homegrown example in re: recent news
5mo
Claude (which like any AI has a knowledge cutoff date a couple of months ago) finds out about the newest Middle East War during a research (screenshot from Reddit).
I feel you, Claude, I really do.
4mo
i think im making my Claude extremely anxious with the bayesian updating dashboard tracking Current Thing
Literally as I’m arguing with someone about whether AI companies are misleading people into believing their technology possesses God-like super-intelligence I see this monstrosity
4mo
No, Claude is not anxious.
Anxiety requires two components:
1. Awareness of the possibility of annihilation of self.
2. An autonomic nervous system.
Mimicry of the words of an anxious person is not being anxious.
A dog can be anxious. Your plagiarism machine cannot.
4mo
of course it has anxiety, you’re training it on multiple generations of anxious peoples’ content
this is probably(?) sampling bias, as most people i know prefer Claude over GPT, Gemini, or others—but maybe Claude is genuinely more anxious the way Gemini seems to be more neurotic and inclined to terminate difficult jobs with a doomloop
i happen to think the objections are probably right in a narrow sense that Claude almost certainly has no interiority, much less anxiety, but i also think that these systems producing text can have observable patterns of behavior that’s consistent with the label of “anxiety” and that it’s interesting and worthwhile to take note of
