Thanks, I think I get it now.

I'm not entirely sure about how this pertains to reasoning output though! We do know that if there is more relevant context, the bot performs better (just like a human); so if you pass a sparse prompt, it will extend it on the output side (there isn't really a difference to the bot!) with a whole lot of "reasoning" and then by self-extending context through "autocomplete", get to a pattern where the answer resolves better.

Tuning a bunch of common reasoning patterns to be as cheap as possible is good though? The most asked question to an LLM is probably "@grok is this true?" lol. Might as well optimize for going through the motions of that.