Anthropic Co-Founder Jack Clark: AI Wins a Nobel Within a Year — And Could Still Kill Everyone

Anthropic co-founder Jack Clark used a May 21 lecture at the Institute for Ethics in AI at Oxford University to lay out one of the most aggressive AI capability timelines ever delivered by a sitting frontier lab executive: an AI-assisted, Nobel-worthy scientific discovery within the next twelve months, bipedal robots working alongside human tradespeople within two years, and businesses run entirely by AI agents generating millions of dollars in annual revenue inside eighteen months.

Clark went further on the longer time horizon, predicting that by the end of 2028 AI systems will be capable of designing their own successors — a recursive-improvement threshold that researchers have long described as one of the clearest signs an intelligence explosion is under way. He framed those forecasts not as bullish marketing but as the baseline trajectory he sees from Anthropic's own pre-deployment evaluations and from the rate at which models such as the unreleased Claude Mythos Preview are completing tasks once considered hard ceilings.

Even so, Clark refused to let the audience leave on a triumphalist note. There remain, he said, "plausible scenarios in which the technology has a non-zero chance of killing everyone on the planet," and he warned that the risk "hasn't gone away" simply because frontier labs have grown more comfortable shipping powerful products. He compared society's current posture on AI to pre-2020 pandemic preparedness — adequate-looking on paper, fatally reactive in practice — and called for governments to invest now in compute audits, model evaluations, and incident-response infrastructure.

His timing was notable. Just days earlier, OpenAI revealed that an internal reasoning model had autonomously disproved an 80-year-old Erdős conjecture in discrete geometry, with Fields medalist Tim Gowers calling the 125-page proof "a milestone in AI mathematics." Clark cited the Erdős result on stage as direct evidence that his Nobel-within-a-year prediction is now "a question of months, not a question of whether." Combined with Anthropic's own Project Glasswing disclosures about Claude Mythos finding thousands of zero-day vulnerabilities, the speech read as a deliberate attempt to anchor both regulators and rivals to a much faster clock.

The contrast with Washington was sharp. The same week Clark was telling Oxford audiences to prepare for AI-designed AI, the Trump White House quietly shelved a planned executive order that would have asked frontier labs to submit models for voluntary safety review up to 90 days before release. Anthropic had been among the few major labs publicly comfortable with that pre-release window; Clark's Oxford lecture, delivered hours after the order was pulled, sounded a lot like the company making its case directly to the public instead.