Learning complex goals with iterated amplification

OpenAI has floated a new training method aimed at a problem nobody has actually solved yet: how do you tell an AI system what you want when the task is too complicated for any human to fully specify, grade, or even fully understand? Their answer, at least as a starting point, is something they're calling iterated amplification, and it skips the usual playbook of reward functions and labeled datasets entirely.

The basic idea works like this. Rather than showing a model thousands of examples of correct behavior, or scoring its outputs against a metric, you show it how to break a hard problem into smaller, easier problems. Solve those smaller pieces, combine the answers, and you've effectively solved the harder problem without anyone having to know the full solution up front. Repeat that decomposition enough times, amplifying capability at each step, and in theory you can climb from tasks a human can judge to tasks that dwarf human judgment.

Right now this is nowhere near a general-purpose training pipeline. OpenAI is upfront that the experiments so far live entirely in toy algorithmic settings, the kind of clean, bounded puzzles researchers use to sanity-check an idea before anyone dares point it at something messy like open-ended language or robotics. There's no large-scale benchmark result here, no leaderboard number to point to. What there is instead is a research bet: that decomposition, done iteratively, might scale in a way that reward-shaping and labeling never really could, because it doesn't require a human overseer to already know the right answer.

That framing matters more than the toy results themselves. A lot of AI safety work runs into the same wall — humans can't reliably supervise systems that are already smarter than them at a given task, so any technique that depends on human labels or human-defined rewards eventually caps out. Iterated amplification is OpenAI's attempt to sidestep that ceiling by having the system amplify its own judgment through recursive breakdown, with humans checking the small steps rather than the big picture. Whether that holds up outside algorithmic puzzles is the open question, and OpenAI is publishing this early, in a rough state, specifically because they want the idea tested rather than polished in private first.