Anthropic's automated alignment researchers beat 28 humans at $4/hour against $150/hour.

The best automated method outperformed what experienced humans proposed within about six hours, against 28 human safety researchers given up to eight, scoring 20% better than the best human proposal on deception. Each run searches literature, proposes a method, trains for 30 minutes, iterates. Across 10 misalignment benchmarks it improved every one without degrading general performance. The stated limitation is the one that matters and Anthropic doesn't bury it: this only works where progress can be automatically scored, which most alignment problems can't be. (Anthropic)