Anthropic, the AI lab known for warning about AI risk, says it's observing early indications that AI is not just coding its own products but beginning to build itself. The claim comes from a new research agenda by The Anthropic Institute, shared first with Axios.
Co-founder Jack Clark told Axios there's a more than 60% chance of an AI model fully training its successor by the end of 2028. "What I'm looking at is a technological trend where, if anything, the speed will accelerate further," Clark said.
The process, called recursive self-improvement, involves AI contributing to speeding up its own R&D. Clark described a scenario where you could tell a system: "Make a better version of yourself" — and it would do so completely autonomously.
Anthropic researchers believe the public should be aware of this trajectory. The lab's identity has long been wrapped around cautioning the world about advanced AI dangers, making this prediction particularly weighty from its own co-founder.
Some experts outside Anthropic argue such timelines are overly aggressive and underestimate engineering hurdles. They note that past predictions of rapid AI self-improvement have not materialized as quickly as forecast.