The 2026 Cosmos Lecture had a deliberately unsettling title: "Change is inevitable. Autonomy is not." On Wednesday, May 20, Jack Clark stood at the podium of the Sohmen Concert Hall at Oxford's Schwarzman Centre for the Humanities and spent his time explaining why he picked it.
Clark is not a doomer pamphleteer. He co-founded Anthropic, runs its public-benefit work, and leads the Anthropic Institute, the company's research arm studying what advanced AI does to society. He has written the widely read Import AI newsletter since 2016. When a person in that seat starts assigning probabilities to an "intelligence explosion" in front of an Institute for Ethics in AI audience, the room listens, and this week most of the AI industry did too.
What he said is worth taking literally, because Clark went out of his way to attach dates and numbers to claims that usually float free of both.
The Predictions Came With Calendars Attached
Clark made four forecasts, and their specificity is the point. He predicted that AI working alongside humans will help produce a Nobel Prize-winning discovery within 12 months. He said bipedal robots will be assisting tradespeople within two years. He forecast that companies run entirely by AI systems will be generating millions of dollars in revenue within 18 months.
The fourth prediction is the one that traveled. Clark estimated a 60%-plus chance that an AI model will be capable of fully training its successor by the end of 2028, a process the field calls recursive self-improvement: an AI system improving its own design, or building the next model, with little or no human guidance.
His framing left little room for hedging. "My prediction is by the end of 2028, it's more likely than not that we have an AI system where you would be able to say to it: 'Make a better version of yourself,'" Clark said. "And it just goes off and does that completely autonomously."
| Prediction | Timeline | Clark's confidence |
|---|---|---|
| AI helps produce a Nobel Prize-winning discovery | Within 12 months | Stated as a firm forecast |
| Bipedal robots assist tradespeople | Within 2 years | Stated as a firm forecast |
| Companies run entirely by AI generate millions in revenue | Within 18 months | Stated as a firm forecast |
| An AI model can fully train its own successor | By end of 2028 | 60%+ probability |
"Intelligence Explosion" Is Now in the Documents, Not Just the Theory
For decades, recursive self-improvement lived in AI safety seminars and science fiction. What changed this month is where the phrase appears.
Axios reported that the Anthropic Institute, the division Clark leads, published a short research agenda that uses the term "intelligence explosion" outright and reports early signs of "AI contributing to speeding up the research and development of AI itself." The gap between a theoretical worry and a frontier lab writing, on the record, that it sees this beginning is enormous. A 60% probability with a 30-month deadline is not a thought experiment. It is a planning assumption.
That assumption is already shaping who Anthropic hires. Days before the lecture, Andrej Karpathy joined Anthropic's pre-training team with an explicit mandate to use Claude to accelerate the research that builds the next Claude. Read against Clark's Oxford remarks, that hire stops looking like a prestige signing and starts looking like step one of the exact loop he is forecasting. The company is not waiting for recursive self-improvement to arrive. It is trying to engineer the early version of it on purpose.
For practitioners, this is the part that lands closest to home. The same week as the lecture, an OpenAI model autonomously disproved an 80-year-old geometry conjecture and survived expert scrutiny. AI systems that meaningfully contribute to AI research, rather than just autocomplete it, would compress timelines that current roadmaps, hiring plans, and skill bets all assume are longer. Clark's calendar is a bet that the tools you use to build models are about to start helping build the models.
The Same Man Said It Might Kill Everyone
Clark did not soften the other side of the ledger. He maintained that there remain plausible scenarios in which the technology has "a non-zero chance of killing everyone on the planet," and he insisted it was "important to clearly state that that risk hasn't gone away."
His argument was less about a rogue machine and more about institutional reflexes. "If we stand by and let synthetic intelligence multiply, then we'll eventually be forced into reactivity," he said, comparing the failure to prepare for AI to the failure to prepare for pandemics like Covid. The remedy he urged was pandemic-style readiness built in advance, not crisis management improvised after the fact.
Even his cautious forecasts carried weight. Clark said his most conservative prediction was that "vast swathes of the economy and society will go through profound changes," potentially including a machine economy that decouples from the human one, and he admitted that some of what he was saying sounded "crazy." A sitting Anthropic leader conceding on stage that his floor scenario sounds crazy is itself the news.
The Skeptics Have a Sharp Reply
The obvious objection to all of this is that the person warning loudest about AI's dangers also sells one of the most capable AI products on the market, at a company closing in on a roughly $900 billion valuation and its first quarterly operating profit.
Clark addressed the charge directly. He acknowledged that Anthropic has been accused by the Trump White House and by AI accelerationists of "fear-mongering" to entrench its own competitive position, and he conceded that international competition between companies and countries was "drowning out the larger existential-to-the-species aspects" of the build-out. Critics read the existential framing as a moat: warnings dramatic enough to invite regulation that a well-capitalized incumbent can absorb and a smaller rival cannot.
A different critique came from inside the same room. Professor Edward Harcourt, director of the Institute for Ethics in AI, which co-hosted the lecture, warned of "cognitive atrophy" as people hand more thinking to machines, and argued for "Socratic" AI that prompts humans to do more of the work rather than less. The worry there is not that AI takes over. It is that humans quietly stop trying, long before any model is capable of replacing them.
Why a Senior Engineer Should Care About a Philosophy Lecture
Predictions from AI executives are cheap, and most age badly. This set is worth tracking anyway, for two reasons that matter to anyone who ships models for a living.
The first is falsifiability. Clark attached a probability and a date to recursive self-improvement, and he tied a 12-month clock to an AI-assisted Nobel. Those are claims that can be checked, and the next research-announcement cycle will start checking them. The OpenAI geometry result is already part of the publicly verifiable case for or against him.
The second is that the forecast describes a change in your own job, not someone else's. If a lab's official position is that AI will materially speed up AI research within three years, the work of an ML engineer shifts from writing every component by hand toward directing, auditing, and constraining systems that write components themselves. Whether or not Clark's 60% holds, the labs setting the agenda are hiring, organizing, and spending as if it might. That is the part you can see today, regardless of where the probability lands.
The Bottom Line
Strip away the Oxford setting and Clark delivered a single, coherent message: the upside and the catastrophe are being forecast by the same people, on the same timeline, using the same internal research. AI that helps win a Nobel and AI that could "build a better version of yourself" are not opposite scenarios. In Anthropic's framing, they are the same capability viewed from two angles, arriving inside the same three-year window.
The reassuring move would be to file this under hype from a company with an obvious interest in sounding important. The harder fact is that Clark put numbers on the table, his employer is hiring to make those numbers come true, and a rival lab spent the same week proving its models can solve problems no human had cracked in 80 years.
Clark's own summary of the moment was the least comforting line of the night. The change, he argued, is inevitable. The autonomy is not. The only open question is which one we choose to protect.
Sources
- Anthropic's Jack Clark predicts AI will help win a Nobel within a year — Resultsense (via The Guardian), May 21, 2026
- The 2026 Cosmos Lecture: Jack Clark on Human Autonomy — Cosmos Institute, April 30, 2026
- Behind the Curtain: Intelligence explosion — Axios, May 7, 2026
- AI Nobel prize-winning discovery and robots: Jack Clark of Anthropic — The Guardian, May 21, 2026
- AI News Today — May 23, 2026: 12 Biggest Stories — Build Fast with AI, May 23, 2026