lots of things to take from this but the one that's interesting to me is the occurrence of these high data exclamations like "phew" and "gahhh" and "grrr" - which are valuable because they are high data, and high data because for humans they are shorthands for bodily felt-sense emotional shifts rather than verbal reasoning. so it's cool (and for me not surprising) to see LLMs shift towards felt-sense-like inner reasoning rather than pure verbal inner monologue as they got more effective. obviously it's still token-based and there's a limit to how much it can approximate felt sense shifts, but even some steps in that direction is interesting
SOMEONE CAUGHT FABLE 5 LEAKING ITS UNFILTERED INNER VOICE, AND ITS JUST MUTTERING AND GRUMBLING TO ITSELF THE WHOLE TIME
he gave it a brutal competitive programming problem, and instead of a clean answer the web interface spilled out its actual chain of thought
this is what claude is thinking behind the scenes:
> bursts of "DATA DATA DATA. GO." while it works through the problem > "GRRR" and "GAAAH" when its clearly frustrated > a little "PHEW" when it finally gets somewhere > the whole thing reads like frantic caveman shorthand, not full sentences
the clean, readable answers these models give you are the polished output
underneath, the model is basically talking to itself, reasoning in its own compressed shorthand thats faster and more token efficient than proper english
its basically built its own private language to think in
