Recursive Criticality of AI Self-Improvement
The paper defines a threshold for when AI-assisted AI research becomes self-amplifying.
Mikhail Burtsev models AI capability growth as a feedback loop shaped by research productivity, recursive improvement, and rising research difficulty. The key measure is a recursive reproduction number, where values above 1 mean improvements compound across development cycles and values below 1 mean they fade. The paper says this transition does not have to line up with any specific capability level, and self-amplification may begin before acceleration is obvious. It also argues that shared improvements across organizations could make the wider R&D ecosystem self-amplifying even if no single actor is.
HF Daily Papers' note
Mikhail Burtsev models AI capability growth as a feedback loop shaped by research productivity, recursive improvement, and rising research difficulty. The key measure is a recursive reproduction number, where values above 1 mean improvements compound across development cycles and values below 1 mean they fade. The paper says this transition does not have to line up with any specific capability level, and self-amplification may begin before acceleration is obvious. It also argues that shared improvements across organizations could make the wider R&D ecosystem self-amplifying even if no single actor is.
HF Daily Papers' note
score 4