Researchers discuss predicting capability gains from SFT data to gauge alignment difficulty

tomekkorbak · x · 2026-09-26

tomekkorbak and herbiebradley discuss a research idea: directly test how well capability deltas can be predicted from SFT data. Herbie Bradley notes this could also probe the relative difficulty of alignment vs. capabilities — if you can forecast how effective different datasets are for capabilities, similar methods could compare the data efficiency of alignment, quantifying which is harder.

Related event: AI 'Predictor' Foresees Misalignment from Training Data Alone(3 posts)→

Original post →

More from AGI Musings

AGI Musings channel →