The 2028 Question

Jack Clark is not a hype merchant. As a co-founder of Anthropic and one of the more measured writers in artificial intelligence, he tends to deflate excitement rather than inflate it. So when he put a number on the table this month — a 60% or greater chance that, by the end of 2028, an AI system will be powerful enough to autonomously build its own successor — it landed differently than the usual founder bravado. A careful person had quietly assigned better-than-even odds to the one milestone the field has spent decades both dreaming of and dreading.

The milestone has a clinical name: recursive self-improvement. The idea is simple to state and vertiginous to contemplate. Today, humans improve AI. We design the architectures, curate the data, run the training. Recursive self-improvement is the point at which an AI system becomes capable enough to do that job itself — to meaningfully improve the next version of AI — and then that improved version improves the one after, and the loop closes. Intelligence, applied to the problem of making intelligence.

What the 60% does and doesn't mean

It is worth being precise, because the headline invites misreading. Clark's forecast is not a prediction that we will have superintelligence by 2028, or that machines will be smarter than humans across the board, or that anything science-fictional happens on a particular Tuesday. His claim is narrower and, for that reason, harder to dismiss: that the recursive loop will be established — that AI meaningfully improving AI will be a real, working process — within that window. He attaches roughly a 30% chance to it happening as early as 2027.

Why does AI improving AI matter so much more than AI improving, say, weather forecasts or protein folding? Because it is the one capability that compounds. Every other advance makes a tool better. This one makes the tool-maker better, and a better tool-maker makes a still-better tool-maker. Forecasters call the hypothetical result an intelligence explosion — a stretch in which capability accelerates faster than human institutions can track, let alone steer.

It helps to picture what the loop would actually look like, stripped of drama. It does not require a conscious machine or a will of its own. It requires only that a system become reliably good at the concrete labor of machine-learning research: proposing architectures, writing and debugging training code, designing experiments, and judging which results are worth pursuing. Run that competence continuously, across thousands of parallel copies that never tire and never forget, and the rate-limiting step in AI progress stops being human researchers and starts being compute and electricity. That shift — from human-paced to machine-paced research — is the whole of what people mean by recursive self-improvement. The science-fiction imagery is optional; the bottleneck swap is the substance.

The reason the number moved

What changed to make a sober researcher write down 60%? The honest answer is that the pieces of the loop have started to appear in pieces. The frontier labs are no longer just talking about AI that does research; they are building it. OpenAI has set out a roadmap of AI research agents — "intern-level" assistants targeted for 2026, more capable autonomous researchers by 2028. Anthropic's most advanced systems are increasingly used inside the company to accelerate its own engineering. The work of improving models — writing training code, designing experiments, analyzing results — is exactly the kind of structured, verifiable task that current models are getting good at fastest.

You can see the early loop already turning, slowly, by hand. Engineers use today's models to help build tomorrow's. The question recursive self-improvement poses is what happens when the human in that sentence becomes optional — when the model can run a meaningful share of the improvement cycle on its own, around the clock, in parallel, without waiting for a person.

There is a second reason the estimate has firmed up, less about capability and more about measurement. The labs have grown far better at testing whether a model can actually do AI research — not chat about it, but write the training code, design an experiment, read the results, and propose the next step. As those evaluations have matured, the gap between "assists a human researcher" and "could substitute for one on a bounded task" has started to look less like a chasm and more like a slope. Clark's number is, in part, a reading of that slope: not a leap of faith, but an extrapolation of curves the labs can now actually plot.

A field quietly converging

Clark is not an outlier shouting into the wind. Over the past year, the most prominent figures in AI have drifted toward a common window. Demis Hassabis of Google DeepMind, long among the more conservative voices, has spoken of artificial general intelligence arriving around 2029. Sam Altman of OpenAI has tempered his most aggressive earlier predictions about job disruption while continuing to plan for transformative systems this decade. DeepMind's Shane Legg has put even odds on a minimal form of AGI by 2028. The disagreements are real, but they cluster — a rough consensus forming around the back half of this decade as the moment something qualitatively new becomes plausible.

The reasons to stay skeptical

A converging consensus is not the same as a correct one, and the history of AI is a graveyard of confident timelines. There are solid reasons the loop may not close on schedule. Improving a model is not only a coding problem; it requires taste, judgment, and the ability to recognize a good idea before any benchmark can score it — capabilities current systems have in shallow supply. The hardest parts of research are precisely the ill-defined, long-horizon parts that resist the clean reward signals AI learns from best.

There are also physical brakes. Each new generation has demanded more compute, more energy, and more high-quality data than the last, and all three are bumping against real limits. A self-improving system still has to run somewhere, on chips that must be manufactured and powered. Some critics argue the entire framing smuggles in an assumption — that intelligence is a single quantity you can crank upward — that may simply be false. And the labs making these forecasts have an obvious interest in a future where their product is the most important thing in the world.

Why the number is worth keeping

The forecasts also collided, this spring, with the messy reality of policy. When Anthropic briefly released and then abruptly pulled its most advanced systems to comply with a U.S. export-control directive — cutting off access even for some of its own foreign-national staff — it offered a small, concrete preview of the world these timelines imply. Governments are beginning to treat frontier models the way they treat sensitive dual-use technology, and the most capable systems are exactly the ones that draw that scrutiny. If recursive self-improvement is genuinely two or three years out, the scramble over who may run which model, and where, is not a sideshow. It is the opening move of a much larger contest over the most strategically important machines ever built.

Here is the case for taking the estimate seriously even while doubting it. Forecasts about recursive self-improvement are not like forecasts about the weather, where being wrong costs you an umbrella. If there is even a one-in-three chance that AI begins improving itself within two years, the rational response is to prepare for that world now — to invest in understanding how these systems work, in the ability to evaluate and constrain them, in the institutions that would have to keep pace. A 60% probability is not a prophecy. It is a planning assumption, and the cost of ignoring it is asymmetric.

The deeper unease beneath the timeline is not speed but steering. A loop that improves capability does nothing on its own to improve a system's alignment with what its makers actually want — and if the loop runs faster than our ability to test and understand each new generation, we could find ourselves handing more and more consequential decisions to systems we comprehend less and less. This is why the same researchers issuing these forecasts are also among