Yes, maybe. Evidence around caveman shows this isn’t a big deal for token consumption (forcing it to respond in a way it was not tuned) and I don’t see a big di

Yes, maybe. Evidence around caveman shows this isn’t a big deal for token consumption (forcing it to respond in a way it was not tuned) and I don’t see a big difference either way. And maybe it increases intelligence in hard to measure ways. Lots of parameters in these models. I feel I fight them less with this setup. They get so lost in their own invented bullshit they stop being useful pretty often without it. So I would take bets on that :)