> Sometimes I’ll have an argument that goes A->B->C, and the model will suggest I reorder as B->A->C. If I try that and feed it into a new instance of the same model, it’ll sometimes say “that’s great, but I suggest ordering it as A->B->C”, and so on forever.

Sonnet 5 acts like that currently. GPT used to be like that, but it looks like they shaved off that unnecessary behavior somehow. But GPT is now also kind of lazy and pretends like it has done its research when it hasnt.