# OpenAI Proposes New AI Safety Technique Called Iterated Amplification

OpenAI has announced a novel approach to AI safety called "iterated amplification" that could change how we teach AI systems to handle complex tasks.

Unlike traditional methods that rely on labeled datasets or reward functions, iterated amplification works by having humans demonstrate how to break down complicated goals into simpler sub-tasks. This decomposition approach allows AI systems to learn behaviors that are "beyond human scale" in complexity.

The organization emphasized that the technique is in very early stages, with experiments so far limited to simple algorithmic domains. However, OpenAI chose to share the preliminary research publicly because they believe it could become a scalable solution to one of AI's biggest challenges: ensuring advanced systems behave safely and align with human intentions.

The significance lies in addressing a fundamental problem in AI development—how to specify complex goals without needing exhaustive examples or perfectly crafted