A LessWrong post argues LLMs excel at math because pretraining data is overwhelmingly correct. It says math literature is >99% true, so imitative learning yields accurate reasoning. It applies the same logic to code, noting most public code works despite being messy. It contrasts this with other fields where literature mixes truth with falsehoods. It asks whether AI research literature resembles math, code or a dumpster fire. It frames this as relevant to recursive self-improvement and automated alignment debates.