upvote
I think they're overcorrecting for being too long-winded in previous generations (and still, in some cases). I guess this is a hard balance to get right.
reply
ChatGPT is still plenty verbose by default IMX.
reply
I'm pretty sure the other answers are wrong and it's a side effect of RL (see thinking machines post about inkling training). It's also exacerbated in fable and sol--I think it's token efficiency effect--bc it's about to reason with fewer tokens the density of the token information goes up.
reply
I figure they are being optimized to write code/functions and not prose so all text is getting more code like.
reply