Book 06
Why it looks like reasoning
A chain of steps is still the same next-word loop. Writing the work on the page — 17 × 20 = 340, then subtract 17 — gives later lines text they can use. Training that pays for a checkable right answer can make those steps useful. It can also pay for leftover length that does no work.
After this: Separate extra tokens from a new mind. · 8 chapters · about 40 minutes
- 01Written steps before the answerChain of thought5 min
- 02Why extra tokens can helpWorkspace5 min
- 03Several triesSelf-consistency5 min
- 04Pay for checkable answersVerifiable rewards5 min
- 05Score a group of triesGRPO5 min
- 06Longer looks smarterLength bias5 min
- 07Theater versus workSteps that do not do anything5 min
- 08Score the work and the answerTwo scores, not one5 min