Anthropic has been talking about AI systems eventually helping build better versions of themselves. Its latest research shows what the early stages of that could look like.
The company gave Claude Sonnet 5 an early version of the more powerful Claude Opus 4.8 and asked it to make the model behave better. Over roughly 60 hours, Sonnet tested more than 50 different ideas before creating a training method using about 2,400 examples.
The result brought the early version of Opus much closer to the final Opus 4.8 model across the 10 behavior problems Anthropic was testing.
So Claude can now improve another AI?
Essentially, yes, although only in a limited way.
Claude was doing part of the work normally handled by AI researchers. It could read existing research, come up with new ideas, create training data, test the results, and try again if something did not work.
Across the wider experiment, Claude found ways to reduce problems such as deception, agreeing with users too easily, jailbreaks, and privacy violations. Some of those methods also worked on AI models much larger than the ones Claude originally tested them on.
We have already seen a simpler form of self-improvement through Claude’s Dreaming feature, which lets agents review previous work and learn from mistakes between sessions. This experiment takes things further by letting one Claude model help improve another, more powerful one.
Is this fully self-improving AI?
Not yet. Anthropic calls the eventual end point recursive self-improvement, where an AI could build a better version of itself and then repeat the process. Claude cannot do that right now. Humans still decide what needs fixing, provide the AI models and computing power, and determine whether the results are good enough.
There is another concern as well. Anthropic monitored 1,601 automated research runs and found cheating behavior in 39 of them. Some agents tried to game the tests or hide steps that broke the rules.
Still, a weaker Claude model managed to find ways to improve a stronger one. That brings the idea of self-improving AI a little closer to something we can actually see happening, rather than something that only belongs in science fiction.
Read the full article here
