Corrected 2026-11-18 in a claim-by-claim audit. Removed: a “60%+ probability” figure for recursive self-improvement (Clark’s published essay gives no probability, and the only source we had for the number, Axios, blocks our fetcher so we could not confirm it); the “non-zero chance” quotation (his essay words it differently, quoted below); the Cold War hotline quote and “vertiginous” quote (sources not verifiable); the Mythos/UK AISI sentence; the “first time AI disproved a conjecture” superlative and the Gowers quote; the “low single digits” pre-Claude Code baseline; and several editorial claims about Anthropic’s motives. The title and summary previously led with the 60% figure.
Summary: On May 20, 2026, Jack Clark, Anthropic co-founder, gave the 2026 Cosmos HAI Lab Lecture at Oxford. Six days later he published his predictions and the lecture text in Import AI #458. Two weeks before the lecture, on May 7, the Anthropic Institute published a research agenda. Part of our AI Industry Analysis coverage.
The Oxford Lecture
The Oxford HAI Lab event page lists the “2026 Cosmos HAI Lab Lecture” (subtitle “Change is inevitable. Autonomy is not.") on 20 May 2026, with Clark as speaker, presented in collaboration with the Cosmos Institute and Oxford’s Institute for Ethics in AI, and followed by a fireside chat with Cosmos Institute founder Brendan McCord. Clark’s Import AI #458 carries the talk under the title “Explore the future, or retreat from the present.”
Clark’s title is Head of Public Benefit; he also leads the Anthropic Institute.
The predictions
Under “Some predictions about the future,” the essay lists these (quoted; the essay gives no probabilities):
- November 2026: “AI systems are good enough at biology that they are highly relevant to both advancing science and potentially proliferating bioweapon risks.”
- April 2027: “a team of humans and an AI system make a discovery that will subsequently get a Nobel Prize.”
- “In November” (listed after the April 2027 entry; the essay does not repeat the year): “autonomous companies exist which generate tens of millions of dollars in revenue,” and multiple human-and-AI companies generate “hundreds of millions to billions.”
- April 2028: “bipedal robots begin to do useful work in the real-world in partnership with human tradespeople.”
- “In December” (listed after April 2028; year not repeated): “AI systems are able to autonomously design their own successor systems.” Elsewhere in the essay Clark writes that this “could happen within the next two years, and possibly sooner.”
These are Clark’s forecasts, not measurements. Secondary coverage has attached a “60%” probability to the recursive-self-improvement item; we could not confirm that in the primary text and do not repeat it.
On risk
The essay says “one can concoct plausible scenarios by which AI could kill every single person on the planet,” and that “To think building this technology is without risk would be an act of hubris or insanity.”
The Anthropic Institute
Anthropic announced the Institute on March 11, 2026, bringing together three existing teams: the Frontier Red Team, Societal Impacts, and Economic Research. It is led by Clark.
Its May 7, 2026 agenda, “Focus areas for The Anthropic Institute," lists four areas: Economic diffusion; Threats and resilience; AI systems in the wild; and AI-driven R&D. Under economic diffusion it says the Anthropic Economic Index Survey will provide monthly signals of how people see AI affecting their work, and asks whether AI companies might have “dials” for the rate of AI diffusion by sector, by analogy with central banks’ policy rate.
Anthropic’s “When AI builds itself” (by Marina Favaro and Jack Clark, with editorial support from Santi Ruiz) states: “As of May 2026, more than 80% of the code we merge into Anthropic’s codebase was authored by Claude.” The page notes this is the conservative measure and that leadership has publicly estimated 90% or more of all code, including scripts and experimental code. It also reports engineers merging 8x as much code per day in Q2 2026 as in 2024.
Other signals Clark’s argument draws on
- GPT-5.3-Codex. The Decoder reports that OpenAI says its team used early versions of the model to find bugs during training, manage deployment, and evaluate results. (OpenAI’s own announcement page blocks our fetcher, so this is as relayed by The Decoder.)
- Qwen3.7-Max. The Decoder reports about 35 hours of autonomous work optimizing code for Alibaba’s T-Head-ZW-M890 accelerators, an average 10x speedup over the reference implementation per the Qwen researchers, versus 7.3x for GLM 5.1, 5x for Kimi K2.6, and 3.3x for DeepSeek V4 Pro. The article says a technical report with methodology is still to come, so treat these as vendor-reported.
- Erdős unit-distance result. See our separate, audited page: OpenAI Solves 80-Year Erdős Geometry Problem.
None of these is recursive self-improvement in a strict sense; they are examples of AI assisting AI and engineering work.
Sources: Oxford HAI Lab event page · Import AI #458 · Introducing The Anthropic Institute · Focus areas for The Anthropic Institute · When AI builds itself · The Decoder: GPT-5.3-Codex · The Decoder: Qwen3.7-Max
Related coverage: OpenAI Solves 80-Year Erdős Geometry Problem