Anthropic's Jack Clark: AI That Can Improve Itself Is More Likely Than Not by 2028

Jack Clark had spent the better part of an hour at Oxford University discussing Claude Code's commercial momentum when he turned to a different kind of forecast. By the close of 2028, Clark told the audience, the odds are better than even that an AI system will be capable of training its own successor without human direction. That process, long known as recursive self-improvement, has for most of AI's commercial history lived in academic philosophy papers and specialist forums devoted to AI risk. Clark, who co-founded Anthropic and leads its policy research, is now putting a number on it: more than 60 percent.

A Benchmark That Stopped Looking Theoretical

Earlier this month, Anthropic published the research agenda for the Anthropic Institute, its internal safety research unit. One figure buried in the five-page document has received less attention than it warrants. On a benchmark that asks models to optimize a CPU-only language model training implementation for maximum speed, the mean performance improvement by Anthropic models went from 2.9 times faster in May 2025 to 52 times faster in April 2026. A human researcher working the same task alone would typically need four to eight hours to achieve a 4x improvement. The current generation of Claude models exceeds 50x.

Clark, in public remarks tied to the Anthropic Institute's launch, described what he sees driving the shift. Anthropic's agenda frames it carefully as "AI contributing to speeding up the research and development of AI itself." Clark is more direct. The jump from 2.9x to 52x in eleven months is not just a capability score. It is evidence that the feedback loop AI safety researchers have been monitoring is beginning to close.

The significance of the number is not that a 52x speedup on any single benchmark is useful in itself. It is that the trajectory suggests the distance between current performance and the level at which a model could meaningfully accelerate its own training is narrowing faster than most public forecasts assumed. A compound rate of improvement at this pace would reach genuinely consequential capability thresholds before 2030.

"Intelligence Explosion" Moves Off the Fringe

The phrase "intelligence explosion" was for most of AI's commercial era a term you encountered in academic papers or risk-focused forums. This month, Anthropic placed it in a formal institutional document. The same research agenda uses "recursive self-improvement" without scare quotes, treating it as a near-term operational concern rather than a distant philosophical thought experiment. Clark defines an intelligence explosion as a scenario where AI systems begin improving at a rate that outpaces human capacity to supervise the process. He is not predicting it as a certainty. He is saying it is probable, and he is on the record.

"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.' And it just goes off and does that completely autonomously." Jack Clark, Anthropic co-founder, Oxford, May 2026

Two Anthropics at Once

Time magazine ran a feature on Anthropic this week under the heading "A Tale of Two Anthropics." The framing captures something true. At the same time Clark was issuing his Oxford warnings, Anthropic is one of the fastest-growing enterprise software businesses in the world. Claude Code reached $1 billion in run-rate revenue within six months of its public launch. Alliances with KPMG, PwC, Bristol-Myers Squibb, and the Gates Foundation have positioned the company across enterprise software, healthcare, and scientific research. The company recently disclosed it is approaching its first profitable quarter, with annualized revenue running at more than $10 billion. By any commercial measure, it is winning.

Clark's 60 percent figure is not a hedge dressed up as humility. It is a public acknowledgment of a risk horizon that he and Anthropic's safety team are actively trying to map. The Anthropic Institute's research agenda identifies specific areas where AI is already accelerating its own development: generating research hypotheses, running experiments, and producing benchmarks that models themselves design. These are discrete, measurable activities, not abstract trajectories, and they are the reason the institute exists.

For organizations building operational dependencies on Claude and other frontier models, the practical question is what a 60 percent probability of recursive self-improvement by 2028 implies for planning. Anthropic's safety work addresses the near-term aspects: how to keep Constitutional AI principles and oversight mechanisms functioning as capability grows, and how to verify that a self-improving system remains interpretable. Those are open problems. Clark's contribution is insisting, publicly, that they need to be treated as urgent rather than eventual. Whether the rest of the industry follows that example is, as much as anything else, the story of the next few years of AI.