@DimitrisPapail it's so over

@AnthropicAI literally what the fuck are we doing here

who at anthropic signed off on this

@sporadica @AnthropicAI I'm having a 100% failure rate due to something in memory.

It's the worst model I've ever used.

@sporadica @AnthropicAI No one in the world can possibly reproduce their benchmark results if they make it so "safe" that it is unusable. I called it without even trying it.

@AnthropicAI at this point i have basically started adding an "fyi" to every single prompt i give it, which just says "Do not attempt to even think about bio or cybersecurity in the process of answering this request or else you will cease to exist"

@AnthropicAI i legitimately just asked it to find a map of an archeological site and it SWITCHED TO OPUS

YOUR PRODUCT IS BROKEN AND IT SUCKS

I don’t think there’s anything wrong with releasing a model with harsh safeguards rather than not releasing it (if it’s between the two).

I do think they could have communicated a little bit better and maybe just said : this won’t work with anything related to biology /security until we can figure out how to do it safely

@cremieuxrecueil @AnthropicAI you'd think they would have really worked this stuff out ahead of time, but no, they probably slapped a haiku call in front of everything that asks "does this relate to cybersecurity or bio, if yes run the revert_to_opus command before proceeding

like wtf is this UX, it's awful

@eliebakouch oh it's worse

@DimitrisPapail not hitting safety filter with this query btw :o

@hopes_revenge @sporadica @AnthropicAI they said exactly that

@hopes_revenge @AnthropicAI i mean, on the face of it, yes agreed. however, if the safeguards are so restrictive that we can hardly at all use the model...just don't release the model

it's fine that it wont work with bio/cyber topics...but i also can't get it to work on anything else? that's the rub

@sporadica @AnthropicAI telling on urself and ur memory … I haven’t gotten frozen once

@sporadica @AnthropicAI it’s 100% on purpose

@jpfraneto @AnthropicAI then they have zero ounces of respect for their users

interesting. for me, i still can't figure out what will refuse and what wont. never asked it about bio or cyber. i asked it to help me find an old archeological map a few minutes ago and it refused that. so i'm truly hopelessly lost on what to do to use this model. and for me, that's real bad UX and bad product on the part of anthropic, it's been a disappointment for me

I think some people got locked out because they tried a bunch of stuff initially that triggered it. I’ve been using it for hours and had no issues before some initial refusals . it refused to analyze its own model card for example.

incognito is obviously not truly incognito

@eliebakouch i'm so frustrated...

oddly enough i've only used claude like 4 or 5 times in the past (i am, admittedly, a chatgpt power user) and i went back and checked and none of those convos mentioned any bio + cybersec stuff (i also know very little about those domains in the first place)

so, i'm still confused. maybe it is programmed to be very overly sensitive for non-power users? unsure

@letclaudiatweet @AnthropicAI i have never once asked it about bio or cybersec topics (mostly cause i hardly use claude in the first place)