2026-05-05 12:30:31
AI systems might soon be capable of building themselves, says Anthropic co-founder
最阅读新闻
Anthropic’s co-founder Jack Clark has made a bold prediction that AI systems may be capable of rebuilding themselves from 2028. “I’m not sure society is ready,” he said. In a Substack published Monday, Clark precisely said there is 60% chance that AI becomes capable of recursive self-improvement by the end of 2028. “In other words, AI systems might soon be capable of building themselves,” he wrote. Clark laid out his case, drawing on “100s of public data sources” and the trend of products being deployed by frontier AI companies. He believes all the necessary infrastructure is already in place to enable AI systems research and build their own successors. I've spent the past few weeks reading 100s of public data sources about AI development. I now believe that recursive self-improvement has a 60% chance of happening by the end of 2028. In other words, AI systems might soon be capable of building themselves. — Jack Clark (@jackclarkSF) May 4, 2026 AI systems now need less human oversight Cark’s argument rests on two things. AI systems have become far more capable at writing and testing real-world code. Also, they can now work independently for much longer stretches without human oversight. On the coding front, Clark pointed to SWE-Bench, a widely used evaluation that tests whether AI can solve actual GitHub issues. The best model scored roughly 2% at the time the benchmark launched in 2023. Today, however, Anthropic’s Claude Mythos Preview reaches up to 93.9%. Claude Mythos Preview was launched earlier in April. It is currently not available to the public, Cryptopolitan reported . He also cited data from METR, an organization that evaluates frontier AI models, showing that the time horizon AI systems can reliably work without human intervention has grown from about 30 seconds in 2022 (GPT-3.5) to approximately 12 hours in 2026 (Opus 4.6). “This is a big deal,” says Anthropic’s Jack Clark The implication is that within a year or two, AI systems are going to get creative enough to form their own novel research paths, refine and train their successors, especially non-frontier models, with no human involved. It could be a lot harder with frontier models as they are a lot more expensive, according to Clark. If AI systems can conduct their own R&D without human involvement, the pace of AI progress would no longer be constrained by the number of human researchers or the length of the workday. “I don’t know how to wrap my head around it,” Clark wrote . “It’s a reluctant view because the implications are so large that I feel dwarfed by them, and I’m not sure society is ready for the kinds of changes implied by achieving automated AI R&D.” Clark’s prediction also tallies with a recent statement by METR forecaster Ajeya Cotra that AI systems should be able to autonomously handle tasks that would require roughly 100 hours of skilled human effort by the end of this year. New post: on Jan 14, I predicted that SWE time horizon by EOY would be ~24 hours. Now I think it'll be >100 hours, and maybe unbounded. For the first time, I don't see solid evidence against AI R&D automation *this year.* Link below. pic.twitter.com/NcP1HlZana — Ajeya Cotra (@ajeya_cotra) March 5, 2026 In August, the Anthropic co-founder said “Anyone who thinks AI is slowing down is fatally miscalibrated.” There’s a middle ground between leaving money in the bank and rolling the dice in crypto. Start with this free video on decentralized finance .