lwbench chart fuels claim Anthropic heavily upsamples LessWrong in Claude training
williawa · x · 2026-10-14
Users noticed Claude answers in rationalist jargon ('core cruxes', 'P that this is load bearing') and suspect Anthropic heavily upsamples LessWrong data — davidmanheim adds it only surfaces when your profile/history signals that background.
The attached chart compares lwbench (1000 niche LessWrong facts, n=128) against SimpleQA Verified factual recall: r=0.72 correlation, with Anthropic models (Claude Fable 5.1, Opus 5, etc.) scoring far above same-tier rivals on lwbench, supporting the upsampled-LW hypothesis.
Related event: lwbench Suggests Claude Heavily Oversamples LessWrong Corpus(2 posts)→