Jack Clark at Oxford: 60% Chance of an Intelligence Explosion by 2028

Anthropic co-founder Jack Clark delivered the 2026 Cosmos Lecture at the University of Oxford on May 20, and the predictions he made are now circulating across every serious AI policy conversation. Titled "Change is inevitable. Autonomy is not," the talk laid out the company's working assumption that recursive self-improvement — an AI capable of training a more capable successor — is more likely than not by the end of 2028.

The probability Clark put on the table: 60%-plus. That is not a long-tail safety scenario. It is a base-rate forecast from one of the labs actively building toward this capability.

What Clark Predicted

Clark framed four concrete predictions during the lecture, each pinned to a near-term timeframe.

First, AI will work alongside humans to make a Nobel Prize-worthy scientific discovery within 12 months. Second, bipedal robots will be assisting tradespeople — electricians, plumbers, technicians — within roughly two years. Third, companies run entirely by AI systems will be generating millions of dollars in revenue within 18 months.

The fourth prediction is the one that has gotten the most attention. Clark said there is a "60%+ chance" that by the end of 2028, an AI model will be capable of fully training its successor. "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.'"

He used the phrase intelligence explosion — a term that has spent most of its life inside AI safety theory papers — and noted that it is now in Anthropic's internal research documents, not just outside critique.

Why This Matters

Recursive self-improvement is the long-feared inflection point in AI capability development. The argument is simple: once a model can meaningfully contribute to building a more capable next model, capability gains compound. Each generation is built by a smarter predecessor. The timeline from "useful research assistant" to "system humans can no longer evaluate" collapses.

For years, that scenario lived in academic and online debates between AI safety researchers. What changed with Clark's Oxford lecture is the source. He is not a critic talking about a hypothetical. He is a co-founder of one of the three labs at the frontier, telling a university audience that his lab's internal view puts this scenario above 50% in the next two and a half years.

The implication for policy is that the windows for action are shorter than most national AI strategies assume. Most government frameworks — the EU AI Act, the recently scrapped White House executive order, the UK's AISI test regimes — are calibrated to the next 24 to 36 months of capability progression. Clark's forecast says the inflection point arrives inside that window.

The "Non-Zero" Risk

Clark also reiterated something he has said before, and which the AI safety community has been arguing about for years: there is a "non-zero chance" that AI ends up killing everyone on the planet. He did not put a number on that one. But he made clear it has not been ruled out, and said the risk "hasn't gone away."

He compared the lack of institutional preparation for advanced AI to the failure to prepare for COVID-19, urging pandemic-style readiness — stockpiled tools, drilled response procedures, clear chains of authority — rather than reactive crisis management.

The lecture also touched on a quieter set of concerns about what abundance does to human institutions. Clark posed the question: "What do you do with a tremendous amount of growth or a tremendous amount of abundance in many, many different fields of science all at once? Today's institutions have very, very narrow pipes through which you push new drug candidates." The implication is that even if everything goes well, the gap between AI capability and society's ability to absorb the resulting changes will create its own crisis.

Context: Anthropic's Dual Identity

The Oxford lecture is best read alongside the rest of Anthropic's recent activity. The company is closing in on a $900 billion valuation, just reported its first profitable quarter ($10.9B in Q2 2026 revenue), and is in serious IPO conversations. It also holds Claude Mythos, the model widely described as both the most capable and most dangerous frontier system in existence — and the one that triggered the now-postponed White House AI executive order on autonomous cyber capability.

That dual identity — most aggressive frontier lab, most vocal safety warner — has long drawn skepticism. Clark's Oxford talk is essentially Anthropic's argument for why both stances are coherent: if the technology is this powerful and arriving this fast, the only way to make it safe is to be the one building it, with the most cautious posture possible.

Industry Impact

The most important downstream effect is probably on Anthropic's own research priorities. Andrej Karpathy joined the company this week to lead a team using Claude to accelerate pretraining research. That is, in effect, an early-stage recursive self-improvement program — using AI to make AI faster and better. The Oxford lecture provides the public rationale for why Anthropic considers that work essential.

For other labs, the question is whether to match Anthropic's posture or push back. OpenAI has historically been more reserved about specific timelines. Google DeepMind has avoided putting probabilities on extinction risk. Meta's public position has been dismissive of recursive self-improvement framing. Expect a flurry of response statements over the next two weeks.

What Industry Insiders Are Saying

Reaction has been split. AI safety researchers have largely welcomed the explicitness of Clark's forecast — finally putting numbers and timelines on what had been vague concerns. Acceleration-leaning commentators have pushed back, arguing that the predictions are overconfident and that recursive self-improvement is being conflated with normal capability scaling.

TIME, which covered the lecture in depth, framed it as "a tale of two Anthropics" — the most commercially successful AI startup in history, run by people who think their product might end the world. That tension is not going away.

What's Next

Clark has indicated that Anthropic's Institute, the company's research arm focused on the societal consequences of advanced AI, will publish a fuller research agenda on intelligence explosion scenarios later this year. Expect the document to include specific institutional preparedness recommendations, capability evaluation benchmarks, and proposed government engagement frameworks.

In the shorter term, watch for whether Anthropic's IPO documents reference the same forecasts. A frontier lab telling investors there is a non-zero chance its product kills everyone is going to be a new kind of S-1 disclosure.

Bottom Line

A 60% chance of recursive self-improvement by 2028 is not a comfortable forecast, and it is not coming from the usual critics. It is coming from a co-founder of the company building the systems. Whatever your prior on AI safety, the Oxford lecture changes the public conversation: the timelines are not theoretical anymore, and the people building this technology are willing to say so out loud.