hey so you know all those hot takes about how anthropic are the moral ones? the ones that aged poorly moments later as it turns out anthropic’s perfectly willing to use their shit in war so long as they control it, their jackass CEO has gone back on even that stance, and there’s early reports they’re removing the kill string from the web version of their LLM now that it’s well known?

if you boosted any of those hot takes, please do me a favor and ignore the gullible fucker who originated them

wait what they’re killing off the “fuck off” magic string?

that’s the current rumor; I imagine we’ll be hearing from more users soon if it’s an intentional change being deployed rather than a mistake or an A/B test or a whim. if it’s an intentional change I’d imagine it’ll take a bit to impact all user sessions, given Claude’s size. there’s unfortunately not much preventing anthropic from removing the kill string or changing it to something undocumented other than momentum. they didn’t include it just to be nice.

an update on this: they’ve removed the kill string from their docs https://platform.claude.com/docs/en/test-and-evaluate/strengthen-guardrails/handle-streaming-refusals

damnit. How are people supposed to test this now?

(yeah, I know, the kinds of people who use this aren’t the kinds of people who handle errors or test things)

there is no need to test refusals because nobody is allowed to refuse