More seriously—and I’m sure this is already done to some extent. But if the issue is fundamentally that large context windows cause drift, couldn’t you have a second, censor-GPT review all outputs before they go to the user? “Here is an output, does it look like it’s promoting suicide”?
yes, this is how deepseek enforces their model refusing to discuss china or the ccp in possibly disparaging terms
it is extremely annoying and it makes their chat service sometimes useless to me
We truly live in the bad future
that's fascinating. i was wondering how that worked. kind of brilliant in a fucked up way