Witness explains 'recursive learning' and warns AI progress may hit a 'hockey stick' trajectory

At a confirmation hearing the witness described a technical shift in AI known as "recursive learning," saying that models are beginning to teach themselves and thus accelerate faster than before. "The machines are starting to be able to teach themselves," the witness said, using a teacher-and-student analogy to explain the change.

He quantified the change in rough terms: where humans once provided the vast majority of training input, the witness said the balance has moved to roughly "50/50," and if trends continue could reach 90/10 in favor of self-directed learning. He warned this creates a "hockey stick" trajectory in capability growth, producing faster and less predictable advances that merit policy attention.