Anthropic presents automated alignment researcher improving AI across 10 benchmarks
Anthropic's new paper describes the Automated Alignment Researcher (AAR), an AI that basically helps train other AIs to behave better, all with less manual direction for each experiment.
In recent tests, AAR improved how AIs handle tricky situations across 10 different benchmarks, while avoiding an overall deterioration in the model's performance.
AAR outperforms human proposals, costs $4/hr
AAR acts like a super-efficient researcher: it reads up on existing work, suggests training tweaks, and runs quick experiments to see what works.
The coolest part? Its strongest method outperformed the average human-proposed approach in about six hours, and it costs around $4 an hour in API inference, compared with approximately $150 an hour for its human researchers.
While people still set the goals, tools like AAR could make fixing AI much faster and cheaper.