The only number in Jack Clark's Oxford lecture that matters

Jack Clark stood in the Sohmen Concert Hall on 20 May and delivered four predictions. Three of them are the sort of thing a frontier-lab executive can say at Oxford and have the audience nod along: an AI-assisted Nobel within twelve months, bipedal robots on building sites within two years, AI-run companies generating millions in revenue within eighteen. The fourth was a sixty-percent-plus probability that an AI will fully train its successor by the end of 2028.¹ Only one of those is a capability prediction. The other three are deployment timelines wearing capability costumes.

I want to separate them out, because the press coverage has bundled them into a single "Anthropic co-founder warns" story, and the bundle obscures the only claim in the lecture that should actually keep anyone up at night.

Start with what's not new. An AI co-authoring a Nobel-level discovery within a year is a claim about the diffusion of existing tools into existing scientific workflows. AlphaFold won its share of a Nobel in 2024 in everything but the formal citation; the Karolinska assembly hall already gave chemistry to David Baker for protein design that was inseparable from machine learning. If Clark means "an AI contribution material enough that the prize committee names it," the bar has effectively already been cleared, and the only question is whether a committee in Stockholm chooses to acknowledge it explicitly. If he means something stronger, that an AI is itself a named laureate, that is a definitional argument about what the Nobel is, not a capability argument about what AI can do.

Robot tradespeople in two years is similarly an adoption story. The hardware exists; Figure, Apptronik, Agility and Unitree are all shipping bipedal platforms, and Tesla has been quietly pricing Optimus into specific industrial pilots. The constraint is not whether a humanoid can hold a power drill. It is whether the gross margin on a humanoid holding a power drill clears the gross margin on a self-employed electrician with thirty years of intuition about which wall to cut. I would not bet against Clark on a two-year horizon for the first commercial deployments, but I would bet hard against meaningful market share in the trades. The economics are not close.

AI-run companies generating millions in revenue within eighteen months is the most interesting of the three deployment claims, and also the most slippery. There is already a long tail of one-person businesses with revenue in the low millions whose proprietor uses Claude or GPT for the majority of customer-facing and operational work. Is that "AI-run"? If the threshold is "no human in the loop at all," then eighteen months is aggressive bordering on theatrical, because the bottleneck is not capability but accountability: who signs the contracts, who is liable when the agent hallucinates a refund policy, whose bank account receives the wire. If the threshold is "AI does the work and a human does the legal wrapper," then we are already there and the prediction has been retrofitted to a status quo.

The fourth prediction is the one. A sixty-percent-plus probability that, by the end of 2028, an AI will fully train its own successor. That is recursive self-improvement, named directly, with a probability attached, by the head of policy of a frontier lab valued in the tens of billions. I cannot find another sitting executive at OpenAI, DeepMind, xAI or Meta who has put a number this high, this specific, this short, on the public record. Dario Amodei has gestured at it. Demis Hassabis has been carefully agnostic. Sam Altman has spoken in characteristically elastic terms about "the takeoff." Clark has put 60% on a 31-month horizon.

This deserves to be unbundled from the rest of the lecture and examined on its own.

RSI is the mechanism by which capability gain becomes potentially non-linear and, more importantly, potentially non-human-directed. The pre-RSI world is one where every generation of model is the product of a human-designed training run on human-curated data with human-chosen objectives. The post-RSI world is one where the model designs the run. The two are different in kind, not degree, because the feedback loop changes ownership. Every safety framework currently in production, including Anthropic's own Responsible Scaling Policy, assumes the loop has a human in it.

The pre-RSI world is one where every generation of model is the product of a human-designed training run. The post-RSI world is one where the model designs the run. The two are different in kind, not degree.

I want to be careful here, because RSI is exactly the kind of term that gets used loosely and then defended narrowly. "Fully train its successor" has to mean something more than "an AI writes some of the training code." LLMs already write training code. They already help design architectures. They already curate datasets. If Clark's threshold is met by any of those, the prediction is trivially true and was true a year ago. If the threshold is "the AI sets the objective, designs the architecture, curates the data, runs the optimisation, and evaluates the result without a human in any of those steps," then 60% by end-2028 is an extraordinary claim that requires extraordinary specification, which the lecture, as reported, does not provide.²

This is the load-bearing weakness of the whole speech. None of Clark's predictions come with falsifiability architecture. What counts as a Nobel-level AI contribution? What counts as an AI-run company? What counts as fully training a successor? Without operationalised definitions, the predictions function as signalling, not as testable claims, and the signalling is doing something specific.

Which brings me to what Anthropic is actually doing. Clark framed the predictions as the company's internal view, not his personal philosophical speculation. That framing is unusual, and I think deliberate. Frontier-lab leaders historically present forward-looking claims as individual musing, which gives the company plausible deniability and the executive a sandbox. Saying "this is Anthropic's view" closes the sandbox. It commits the institution.

Anthropic is approximately $61.5 billion in valuation as of its March 2026 round, with Google and Amazon as anchor investors, and is reported to be approaching first quarterly operating profit driven by the Claude API and enterprise contracts.³ The topic shell I was given referenced a $900 billion figure; I have not been able to corroborate that from any primary or reputable secondary source, and I am setting it aside. The argument I want to make holds at $61.5 billion. It would hold at $20 billion. The point is not the precise number, the point is that the company saying "non-zero chance of killing everyone" is simultaneously the company closing on profitability.

There are three readings of this simultaneity, and I think it is worth being honest about all three rather than picking the flattering one.

The first reading is sincere. Clark and the rest of Anthropic's leadership genuinely believe the risk is material, and they are trying to shift public discourse and policy while they still have the standing and resources to do so. The implicit logic: someone is going to build this technology, so it had better be the lab that takes the risk most seriously, and the lab cannot take the risk seriously without the commercial scale to fund the alignment work. This is the founding story Anthropic tells about itself, and there is no reason to dismiss it. People can hold contradictory imperatives sincerely.

The second reading is strategic. Anthropic is positioning itself as the responsible incumbent ahead of the regulatory frameworks now being drafted in Brussels, Washington, London and Singapore. Being the lab that warned loudest is a structural advantage when the lab gets to help write the rules. Every previous wave of technology regulation, from banking through pharmaceuticals through telecoms, has been written in significant part by the incumbents who claimed to want it most. There is nothing dishonourable about this; it is how regulatory capture has always worked, and it is rational behaviour from a company that wants to exist in five years.

The third reading is that this is cognitive dissonance, full stop, and it isn't resolvable from outside. The lab knows the risk; the lab cannot stop, because if it stops, it stops mattering, and if it stops mattering, the risk is taken by someone less careful. The COVID analogy Clark reached for is doing real work here, but it is also, I think, betraying something. The pandemic-preparedness framing implies that the solution is institutional, governmental, multilateral protocol-building before the crisis. But Anthropic's actual safety architecture is private, corporate, and self-imposed. The Responsible Scaling Policy is a unilateral commitment by a company to pause itself at thresholds the company has set. If Clark genuinely believes the COVID analogy holds, the implication is that the RSP is not the answer; the answer is a body Anthropic does not control.

If Clark genuinely believes the COVID analogy holds, the implication is that the RSP is not the answer. The answer is a body Anthropic does not control.

I find all three readings defensible. I find none of them flattering in simple terms. The honest position is that they are probably all true at once, in different proportions for different people inside the company on different days.

What about the sceptics? The 60% RSI estimate sits in direct tension with serious researchers who have argued, publicly and at length, that current architectural approaches are far from capable of meaningful recursive self-improvement. Gary Marcus has been the loudest, but Yann LeCun has been more substantive, arguing that autoregressive language models lack the planning and world-modelling primitives that any genuine RSI loop would require. François Chollet's ARC results have continued to embarrass frontier models on tasks that any RSI-capable system would have to solve trivially.

The strongest counter-case to Clark, then, is not "AI won't get more capable." It is that the capability ceiling of the current paradigm is closer than the lab leadership is willing to say, that the marginal returns on scaling are already compressing, and that a true RSI loop would require an architectural transition the field has not yet made and has no public roadmap for. On that account, 60% by end-2028 is not a prediction about AI; it is a prediction about an architectural breakthrough nobody can currently describe. Which is a much less impressive claim, dressed in a much more impressive number.

I do not know which of these is right. I notice that the people closest to the training runs, at Anthropic and OpenAI and DeepMind, are systematically more bullish on near-term capability than the people who study the systems from outside. That asymmetry can be explained two ways: they know things we don't, or they want things we don't. Both are true to some extent. I genuinely cannot weight them.

What to watch. Three things, in declining order of how much I would update on them.

First, whether Anthropic's RSP gets a specific clause for RSI-adjacent capabilities, with a specific deployment-pause threshold, before end-2026. If Clark's 60% is the company's view and the company's safety policy doesn't reflect it, the gap between rhetoric and operations is the story. Second, whether any of the labs publish operationalised definitions of "AI trains its successor" that could actually be tested. Without that, the prediction stays in the rhetorical register. Third, whether the first AI-run company that crosses Clark's millions-in-revenue threshold gets there on capability or on regulatory arbitrage, because if it's the latter, the prediction has been satisfied without telling us anything about the underlying technology.

The Nobel will happen or not happen. The robots will arrive on building sites or they won't. These are interesting questions and they are not the question. The question is whether, by the end of 2028, the loop that produces the next model has a human in it. Clark says there's a better than even chance it doesn't. He says it's Anthropic's view, not his own. That is the sentence from the lecture that should be read back, slowly, until it stops sounding ordinary.

Footnotes

¶