AI systems are taking on increasingly substantial research and engineering tasks inside several labs, according to a new 36Kr survey of recursive self-improvement. The strongest evidence it describes is about AI helping people improve systems, not machines independently improving themselves.
Watch Desk analysis
What happened
The 36Kr article brings together examples from Anthropic, OpenAI, MiniMax and Zhipu, while noting that “recursive self-improvement” can mean anything from AI-assisted development to a fully autonomous improvement loop. That definitional gap matters: these examples do not all describe the same kind of progress.
Anthropic’s reported figures are striking. The article says Claude was at the “leads” level in 26% of the company’s R&D work by August, meaning it could complete most of those tasks end to end under human supervision. It also says more than 90% of R&D work involved Claude collaborating, while fully autonomous work remained at zero.
Other examples focus on particular tasks. The article says OpenAI’s research assistant can handle defined research tasks under human guidance, though more than half of the successfully completed tasks lasting four to eight hours still required human intervention. It also describes Zhipu’s GLM-5.3 driving an inference-service build whose throughput reportedly reached three times its starting baseline in under two weeks. These figures and claims are reported by 36Kr, not independent measurements established here.
Why it matters
“AI helped build or improve something” is becoming a practical engineering claim, not just a science-fiction premise. But it can refer to very different things: writing code, optimising an experiment, building infrastructure, or changing model weights. Treating all of them as proof of a self-improving AI would make the headline outrun the evidence.
The distinction matters for judging both capability and oversight. A system completing a bounded task with people checking its work is meaningful progress; it is not the same as a system deciding what to change, making the change and reliably verifying the result without human involvement.
Our read
There is real movement here, and the useful story is less “the intelligence explosion has arrived” than “AI is taking a larger, measurable role in research workflows”. The best next step is to ask what task was automated, how success was measured and where people still intervened. Recursive self-improvement is a grand label; the evidence deserves more precise labels.
What to watch
- Whether labs publish task definitions and evaluation methods alongside automation claims.
- How often reported work still requires human intervention, and what that intervention involves.
- Whether future examples change model weights, or improve the surrounding code and infrastructure.
- Whether independent researchers can reproduce the reported results.
Discussion spark: Should AI-assisted coding, research and infrastructure work count as recursive self-improvement, or should the term be reserved for systems that improve their own models with little or no human intervention?
Sources and evidence
- When AI Builds AI: 2026 Breakthroughs in Recursive Self-Improvement Technology – 36Kr (10 October 2026, 00:15 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.