Anthropic co-founder Jack Clark posted an essay Monday putting roughly 60% odds on automated AI R&D by the end of 2028, the threshold he defines as a frontier model autonomously training its own successor. Push him on a 2027 probability, and the number drops to 30%.
He calls it a reluctant view. "I don't know how to wrap my head around it," Clark writes near the top of the latest Import AI issue, before spending the next several thousand words building a case that makes the conclusion harder to dismiss.
What counts as the threshold
Clark isn't predicting a frontier model will train another frontier model in 2026. Frontier training runs are expensive and human-intensive, and he says so directly. The narrower forecast is a proof-of-concept at the non-frontier stage, where one model end-to-end trains a successor within a year or two. Full frontier-scale automation comes later, by late 2028.
The framing throughout is unusually personal for a benchmarks essay. "Staring into the black hole," reads one section header. We may be about to witness a profound change in how the world works, Clark writes near the end. If a different essayist filed those lines you'd roll your eyes. The fact that the person writing them runs policy at one of the labs actually building this is what makes them harder to wave away.
The evidence stack
Most of the essay is benchmark accounting. SWE-Bench has effectively saturated: Clark cites Claude Mythos Preview hitting 93.9%, against a roughly 2% score for Claude 2 in late 2023. METR's time-horizon plot, which tracks how long humans take to do tasks AI completes with 50% reliability, has gone from 30 seconds with GPT-3.5 to roughly 12 hours with Opus 4.6 in about four years.
Then the AI-on-AI benchmarks. The CORE-Bench paper, which tests whether models can reproduce research papers from their repos, was declared solved in December 2025 with Opus 4.5 hitting 95.5%. On Anthropic's internal CPU training-speedup task, Claude Mythos Preview achieved a 52x speedup in April. A human researcher, per Anthropic, takes 4 to 8 hours to hit 4x on the same task.
The kernel-design and PostTrainBench progress is messier. Frontier models can post-train smaller models to roughly half the uplift human researchers achieve. Not parity. Not nothing.
A skeptic reads the same data
Most of these benchmarks come from labs that benefit from accelerating AI timelines, and several come from Anthropic itself. The CPU training-speedup figure is an Anthropic-reported result on an Anthropic-designed task, and the human baseline of "4 to 8 hours" lands without methodology behind it. Clark also leans on Anthropic's alignment research proof-of-concept as evidence, which it is, but Anthropic publishing about Anthropic isn't independent confirmation.
Outside the company, Redwood Research's Ryan Greenblatt recently doubled his own estimate, from 15% to 30%, on full AI R&D automation by end of 2028. His earlier skeptical analysis put the median for "powerful AI" closer to 2031. Even after the update, Greenblatt's number is half of Clark's.
Clark acknowledges the gap implicitly. He notes creativity and "heterodox insights" remain the part current systems struggle with, then points out that AlphaGo's Move 37 was a decade ago and hasn't been replaced by anything obviously more impressive since. That's a strange kind of reassurance to include in an essay arguing the opposite direction.
What's next
Clark says he'll spend most of 2026 working through the implications. The nearest concrete checkpoint is OpenAI's stated target, flagged in Altman's post, of an automated AI research intern by September 2026. Whether that intern ships, and what it can actually do if it does, will say more about the timeline than another benchmark will.
