[tptacek]

Right, it is currently incapable of providing a straight answer without clearing it's throat selling the answer. It reminds me of those recipe blogs that just can't get to the fucking recipe. It's bad writing! But it's not bad technically, in a style-guide kind of way.

[cibyr]

Sometimes I wonder if the throat-clearing is an indispensable part of getting to the "good bits" that follow. Like, do those extra tokens give it more "room to think" even if they're basically meaningless in themselves?

[dTal]

The output tokens are the only information that is carried forward through each inference pass, so "more room to think" is incompatible with "basically meaningless". Perhaps one could imagine it somehow stenographically encoding information in its precise choice of meaningless throat clearing, but there are only so many variations on that theme - word choice is heavily constrained, so it doesn't feel like you could store a whole lot of information there without it starting to read froopiliciously.

[alex43578]

Isn’t that the point of the hidden chain of thought tokens, rather than the visible cruft?

I think the fluff, the emojis, the sycophancy is all symptomatic of the training process and human feedback.

[lupire]

I thought PP was saying that the "Thinking" text is only used for one turn, and the response text is the compressed thinking that survives into future turns.