Researchers outline five stages on the path to self-improving AI - Politics and regulation

A Chinese research team has proposed a five-level framework for assessing recursive self-improvement in artificial intelligence, while experts warn that current experiments remain far from systems capable of independently creating more advanced models.

The prospect of artificial intelligence improving itself and designing later versions has moved beyond science fiction and into academic research. Specialists in China are examining the conditions required for recursive self-improvement, or RSI, a process in which an AI system could assess its own limitations, alter its training methods and use the results to build increasingly capable models.

The proposal also addresses a major challenge facing the technology industry: the rising financial and computational costs of developing more sophisticated systems. If AI could gradually take over some research and optimisation tasks, the development of new generations of models might become faster and less dependent on human teams.

The study was published as a preprint on arXiv on September 10 and has not yet undergone academic peer review. It presents five levels of autonomy, starting with systems whose development remains largely directed by people and progressing towards models that can identify their weaknesses, choose how to correct them and alter the procedures used for their own improvement.

At the highest levels, an AI could produce what the researchers describe as “meta-improvements”. Rather than adjusting individual tasks or internal parameters, the system would refine the mechanisms it uses to learn and develop.

Early experiments have shown only partial progress. One test used a method called A-Evolve-Training to optimise a language model with 30 billion parameters. Across four stages, the system applied different training strategies and used the resulting data to guide later decisions, producing moderate improvements in performance tests.

Those results do not demonstrate that an AI can independently create a more intelligent generation of models. Peter Bentley, professor of Digital Biology at University College London, told Live Science that he does not consider such experiments a credible route to recursive self-improvement. He noted that software designed to optimise its own operation has existed for decades, including through evolutionary computing and reinforcement learning, but argued that the central difficulty is enabling a machine to direct the research needed for genuine progress.

Bentley also said that sustained improvement would require more than changes to software. A system would need to improve its own training, expand the resources available through physical hardware and pass demanding tests in real-world conditions. The need for computing power and technical infrastructure therefore remains a decisive limitation.

Kayla Blomquist, a University of Oxford researcher and co-founder of the Oxford China Policy Lab, said the paper is useful for identifying the technical and control challenges associated with more independent systems, but does not offer an immediate practical route to implementation. She also questioned whether there will ever be a clearly defined point at which AI stops depending on people, describing the process as a gradual transfer of functions rather than a single breakthrough.

Sources: elDiario.es.