Leading AI labs say their models are starting to improve themselves

[Anthony_51]

An AP report doing the rounds has pulled together what the major labs are saying about recursive self improvement, where AI models help build their own successors. Anthropic says Claude is now leading about 26 percent of its model research and development, completing most tasks end to end from a high level prompt, though still under human supervision. OpenAI has an automated research intern that can handle well defined tasks that would take a skilled researcher days, and it is targeting a fully automated AI researcher by March 2028. That is less than 18 months away

Elon Musk is going further, as usual. He says humans are gradually getting less and less in the loop at xAI, with each model built by its predecessor, and suggested full automation could arrive by the end of 2027. Microsoft is taking a more cautious line, with Mustafa Suleyman talking about humanist superintelligence that is carefully calibrated and kept within limits under human direction

The critics are not holding back. Anthony Aguirre, president of the Future of Life Institute, called autonomous self improvement probably the worst idea in the history of humanity. Cornell professor John Thickstun offers a calmer view, pointing out that older models have helped build newer ones for years and that recent efforts have mostly produced minor gains. So the reality may be less dramatic than the headlines suggest

The most striking line comes from OpenAI itself. It admits it cannot yet safely achieve aligned, full self improvement, and warns that progress on alignment and safety may not keep pace with capability. When the company building the thing tells you the safety work might fall behind, that is not something to shrug off

I am torn on this. The productivity gains are obvious and nobody wants to be the lab that falls behind. But there is a big difference between AI helping researchers and AI running the research with humans signing off at the end. Where do you think the line should be drawn, and who should be allowed to draw it?

Will AI models achieve the ability to improve autonomously? Leading labs say the scenario is nearabcnews.com

[Glassy Viper]

What worries me is the race element. Each lab points at the others to justify going faster. Nobody wants to pause alone, so nobody pauses

[Cass64]

Who is auditing the 26 percent claim? We only have Anthropic's word on how much the model does and how closely humans supervise it

Some independent measurement would help. Otherwise these figures are just competitive signalling between labs. Numbers like that should come with a methodology

[BiasField16]

Thickstun's point deserves more attention. Models have been used to generate training data and write code for their successors for years now. It is a gradual shift rather than a sudden switch. Calling it recursive self improvement makes it sound like a runaway process when most of it is tooling

The dramatic version may still come. We are just not there yet

[Stephen_16]

OpenAI saying alignment may not keep pace is the headline for me. They are telling us the brakes are being developed more slowly than the engine

Yet the timeline to an automated researcher stays at March 2028. You cannot have it both ways

[CaptainCipher10]

Aguirre's quote is dramatic, though I understand where it comes from. If a system can design better versions of itself faster than we can check them, oversight becomes a formality. At that point you are trusting the process rather than understanding it. That is a very uncomfortable position for any technology, let alone this one

[Scarlet Annie]

A quarter of research at a frontier lab being led by the model itself is a huge number. Even with supervision, that is a lot of decisions humans are only checking after the fact

[Glenn_44]

Microsoft's humanist superintelligence phrase sounds like marketing, but the idea behind it is sensible. Keep humans directing the goals and let the AI help with the work

The question is whether that holds up once competitors move faster by removing the humans. Commercial pressure tends to win these arguments

[Daemon82]

Musk giving a date is not worth much. He has promised full self driving next year for about a decade

[Zach91]

From a pure engineering view this was always where things would go. Every field automates its own tools eventually. Compilers write code, chips design chips. AI helping build AI is the natural next step, the question is simply how fast. Nobody should be surprised