That’s totally unrelated. The post I was replying to claimed that if you create a new account with OpenAI and that gets detected, your whole account gets silently “nerfed”.
That is not in any way related to Fable (visibly) being switched to a less strong model if you’re trying to discuss certain topics.
> That is not in any way related to Fable (visibly) being switched to a less strong model if you’re trying to discuss certain topics.
We're talking about it being invisibly moved to a weaker model if it looks like you're distilling (which is best detected through something that is at least partially a reputational / account metric).
Now, Anthropic stepped away from this, but it highlights one more kind of systemic risk you're exposed to when you're not running the model yourself.
They already reversed course on that decision a couple of days later. Trivial to find a source, but Fable is also rather notably not available to the public right now, so it's not actually a relevant threat.
excuse me but, is this trivial source you're referring to the url included in the post you're responding to, or did they reverse back to the original intent of keeping refusals quiet?
Fable's model card provides the following as a relevant reference
https://xcancel.com/ClaudeDevs/status/2064949876463645026
That’s totally unrelated. The post I was replying to claimed that if you create a new account with OpenAI and that gets detected, your whole account gets silently “nerfed”.
That is not in any way related to Fable (visibly) being switched to a less strong model if you’re trying to discuss certain topics.
> That is not in any way related to Fable (visibly) being switched to a less strong model if you’re trying to discuss certain topics.
We're talking about it being invisibly moved to a weaker model if it looks like you're distilling (which is best detected through something that is at least partially a reputational / account metric).
Now, Anthropic stepped away from this, but it highlights one more kind of systemic risk you're exposed to when you're not running the model yourself.
It demonstrates that:
1. Anthropic certainly has the ability.
2. They’re willing to use it silently.
1 reply →
> That’s totally unrelated.
i disagree, but it seems clear, from how you put it, that there's no point explaining the why
They already reversed course on that decision a couple of days later. Trivial to find a source, but Fable is also rather notably not available to the public right now, so it's not actually a relevant threat.
excuse me but, is this trivial source you're referring to the url included in the post you're responding to, or did they reverse back to the original intent of keeping refusals quiet?