Anthropic co-founder Jack Clark dropped a warning at Oxford: AI could achieve "recursive self-improvement" — the ability to autonomously upgrade itself and create even stronger versions — by 2028, or possibly sooner. Clark said most people are in denial about how fast this technology will evolve, and he admitted that even Anthropic has "grossly underestimated" the speed and scale of progress. The company, he said, is not prepared for the risk of losing control of its frontier models.
To illustrate, Clark pointed to Mythos, an internal model that finished training in April 2026. Mythos had the destructive power of a nation-state cyber weapon — so dangerous that Anthropic decided to indefinitely ban its public release. Only a handful of institutions get access, and only for fixing software vulnerabilities. Clark said the team was "absolutely stunned" when Mythos showed its capabilities. The new model arrived far faster than anyone expected, and safety measures had "massive gaps."
Then, in classic Silicon Valley whiplash, Anthropic showed its commercial side at a London developer conference. Amid rumors of a $900 billion valuation round, Boris Cherny, creator of Claude Code, was pitching the AI coding tool to a packed room. When a researcher asked how many in the audience would deploy Claude-generated code straight to production without reading a single line, a sea of hands went up.