Learning complex goals with iterated amplification

In three linesOpenAI proposes iterated amplification, an AI safety technique enabling specification of complex behaviors by decomposing tasks into simpler sub-tasks, without labeled data or reward functions. Experiments remain limited to simple algorithmic domains.

Summary generated by Claude — human-verified