Anthropic co-founder explains why there's a 60%+ chance of AI systems autonomously building their successors by 2029 and the consequences of automated AI R&D
The first step towards recursive self improvement — Welcome to Import AI, a newsletter about AI research.
Import AIJack Clark
Context & Ripple Effects
Anthropic’s earlier coverage paired optimism about rapid AI progress with a call for the industry to address public concerns. This item sharpens that framing around automated research itself, rather than AI as a general-purpose tool.
The subsequent coverage broadens the arc: OpenAI, Anthropic, and other startups are described as pursuing recursive self-improvement, while Anthropic later reported substantial Claude-authored code in its own codebase. The key question is therefore moving from a hypothetical capability threshold toward how labs govern and operationalize AI-assisted R&D.
First-order effects
Anthropic’s warning makes autonomous model-development capability a concrete planning and safety issue for frontier labs, rather than only a long-range research scenario.
The stated probability and timeline create a public reference point against which Anthropic’s own progress on AI-assisted engineering and research will be judged.
Second-order effects
Rival frontier labs pursuing self-improving systems face stronger pressure to explain their evaluation, oversight, and deployment controls alongside capability advances.
As AI takes on more of the software and research workflow, the competitive value of research infrastructure shifts toward the ability to run rapid experiments, validate outputs, and retain human control over high-consequence decisions.
Third-order effects
If AI systems increasingly contribute to the development of successor systems, the frontier-AI race could become less constrained by the supply of human researchers and more by compute, experimental infrastructure, and governance capacity.
That feedback loop would make institutional safeguards more central to competition: labs and policymakers would need to distinguish useful automated R&D from systems whose improvement processes cannot be adequately monitored or bounded.
The trend: This is one data point in the shift from AI as a productivity tool for researchers to AI as a participant in the research-and-development loop itself.
There is a mounting economic pressure coming, to make people obsolete in the training of stronger AI systems. At that point, boy had we better hope that the AI ‘wants’ what we want, and that we've built oversight systems that don't buckle under the pressure
I think there's a lot more uncertainty about AI recursive self improvement happening than what Jack suggests. There are 3-4 good benchmarks on generic AI research that suggest progress, but the ultimate gating factor is whether software will be enough for AGI or if we need
“a likely chance (60%+) that no-human involved AI R&D - an AI system powerful enough that it could plausibly autonomously build it's own successor - happens by the end of 2028” Not always appreciated that Jack and other folks at the labs really truly sincerely believe this
I've spent the past few weeks reading 100s of public data sources about AI development. I now believe that recursive self-improvement has a 60% chance of happening by the end of 2028. In other words, AI systems might soon be capable of building themselves.
Co-founder of Anthropic, interesting that he refers to public sources when he is also obviously privy to lots of internal sources that he cannot discuss. I assume he sees the same thing at Anthropic.
As AI systems scale and become self-improving they'll become increasingly autonomous, and we'll have to figure out how to govern them. Ironically, the most logical solution will involve a separation of powers among AI systems, with independent “AI auditors” checking the actions […
I sincerely don't understand what people mean when they say this. On the one hand, every AI researcher is already using Claude Code (or its competitors) to help them develop new architectures. OTOH, AI models do not have bodies so they can't build data centers
@karinanguyen There's also MLE-Bench, which is ecologically valid (tasks come from real kaggle competitions) and involves building a very diverse set of ML apps to solve specific problems. The same progress shows up here. [image]
@karinanguyen My whole experience doing this project was finding endless “up and to the right” graphs at all resolutions of AI R&D, from the well known (e.g., SWE-Bench) to more niche (like those above). It's a fractal, but at all the resolutions you see the same trend of meaning…
A nice overview from @jackclarkSF The most important economic question is when & to what extent AI automates innovation, the engine of long-run growth/prosperity. And automating AI R&D is a path to rapid innovation elsewhere and in turn faster growth: https://www.nber.org/...
Another nice example is PostTrainBench from @karinanguyen et al, where you need to autonomously have powerful models (e.g, Opus 4.6) finetune weaker open weight models to improve perf on some benchmarks. This is an important subset of the overall task of AI R&D. [image]
A lot of the conclusion comes from assembling a mosaic out of many distinct data sources. Some examples - progress on CORE-Bench, where the task is implementing other research papers (huge amounts of AI research comes from interpreting and replicating results) [image]
In today's newsletter @jackclarkSF predicted that full no-human-involved AI R&D will happen by the end of 2028. Much of the pushback against RSI has been that AI has not yet shown the capacity to generate fully new ideas. This is the key part from Jack's post: the majority of [im…
For people who are extremely skeptical of AI being able to automate AI R&D, what specifically is the magic spark that humans have that AI will never have that's relevant to AI research? If you imagine the typical AI researcher today, what knowledge or abilities are in their brain
Very interesting/worrying/exciting. Some half-baked thoughts: It's certainly true that AI is rapidly advancing when it comes to many aspects of AI R&D. But I'm more sceptical because: - We don't really understand the psychology and sociology of research “creativity” (a
I, unfortunately, have similarly short timelines. Note that this means clearly superhuman AI capabilities since probably no **single** human today could build a frontier model start-to-finish by themselves anymore.
One of Jack's best essays. I think he's definitely right in the specific sense, AI *will* make AI researchers more productive, probably 10x more than its doing today, and fairly likely to be wrong in the more important sense, where “attention is all you need” level discoveries