Anthropic Co-Founder Warns AI May Achieve Recursive Self-Improvement by 2028
Anthropic co-founder Jack Clark warned during a speech at the University of Oxford that artificial intelligence is highly likely to achieve recursive self-improvement by 2028 or even earlier, meaning large models could autonomously upgrade and create stronger versions without human intervention.
Clark stated that most people globally still deny the speed of technological self-evolution, and even Anthropic itself has severely underestimated the scale of evolution, being unprepared for the risks of cutting-edge models going out of control.
For example, the internal model Mythos, completed in April 2026, possesses national-level cyber offensive and defensive capabilities. The company decided to indefinitely withhold its public release, limiting it to a few institutions for vulnerability remediation, and the team was extremely shocked by this capability at the time.
At the London Developer Conference, Anthropic actively promoted the Claude Code programming tool, with many developers admitting they would deploy AI-generated code directly online without any human review or verification.
Source: Public Information
ABAB AI Insight
Jack Clark, as a co-founder of Anthropic and previously responsible for policy and safety matters, has repeatedly warned about the speed of AI development. His recent speech at Oxford continues his consistent warning path of "capabilities exceeding expectations," which aligns closely with the company's internal shocked reaction to the Mythos model.
On the capital front, Anthropic is limiting the public release of Mythos and only authorizing a few institutions while also raising significant funds (rumored to be valued at $900 billion), shifting resources towards safety alignment and commercialization tools. The motivation is to balance cutting-edge capabilities with the risks of loss of control, strategically monetizing quickly through products like Claude Code while preparing governance frameworks for a potential era of recursive self-improvement.
Similar to the early breakthroughs in reasoning capabilities by OpenAI's o1/o3 series, Anthropic is currently in a transitional control phase from single model iteration to the potential emergence of autonomous improvement, with Mythos serving as an internal alarm.
This fundamentally represents a reconstruction of the industry chain driven by technological substitution. The potential arrival of recursive self-improvement alters the power structure of AI development, as the model's autonomous upgrade capabilities will significantly reduce the cost of human intervention, prompting capital to shift from human-led R&D to safety alignment and autonomous evolutionary infrastructure, while forcing the industry to transition from optimistic commercial promotion to stricter risk-preemptive governance.
ABAB News · Cognitive Law
The more capabilities exceed expectations, the more safety preparations lag behind.
The bolder developers are in deploying directly, the closer the risk of loss of control.
When models begin self-iterating, humanity truly loses the brakes.