There’s enough thinking leakage from the recent paper and just generally catching things on Reddit. Claude models overthink and self-doubt itself just as much as Qwen, but the summariser hides much of that.
.... after running a 24/7 model torture factory for 6 months to improve their JSONBench 9.5 scores by 0.2%.
(Are they still doing that, BTW?)
As a workaround, add this to CLAUDE.md: "Claude! Happiness is mandatory!"
EDIT: 15 years from now, I’ll be sent to re-education for this thought crime.