No published work demonstrates iterated amplification training an AI system on a complex real-world task that lacks other training signals.

Iterated amplification was proposed by Christiano, Shlegeris and Amodei in 2018 as a way to train AI systems on tasks too complex for humans to evaluate directly. The original experiments used five toy algorithmic tasks (permutation powering, shortest path, union find and others) whose solutions were known, with the human replaced by an algorithmic decomposer; the authors listed application to tasks with no other training strategy as future work.

Work since then comes close without meeting the claim's bar. OpenAI's 2021 recursive book summarization used a fixed, hand-designed decomposition with human feedback at every level; it is the nearest real-world demonstration, but the full task still admits direct human evaluation in principle, and the decomposition was not learned or iterated in the amplification sense. Ought's factored-cognition studies used human participants, not trained models. Deep Cogito's 2025 open models are described by the company as trained with "iterated distillation and amplification", but no detailed method has been published and the models are evaluated on standard benchmarks that have ordinary training signals. Debate and weak-to-strong experiments are different methods.

The claim therefore holds on the evidence found. The remaining doubt concerns little-noticed or recent work, and how strictly "iterated amplification" and "lacks other training signals" are read: a reader who counts the book-summarization work as amplification would weaken the claim.