Anthropic Co-Founder Jack Clark: 60%+ Probability AI Builds Its Own Successors by 2028

Q: What Clark's actually predicting

The prediction has three distinct components. First, the AI engineering capability required — the ability to write, debug, and ship complex ML training pipelines, conduct hyperparameter searches, and handle distributed training across GPU clusters. Per Clark's framing, this capability already exists in current models for narrow components and is rapidly improving across the broader pipeline. Most engineering tasks in ML research are now AI-assisted at major labs. Second, the research-direction capability — the more ambiguous question of whether AI can identify which architectural ideas are worth pursuing, which training data mixtures will produce desired capabilities, and which evaluation methodologies are valid. This is the harder problem; current AI systems can execute on research ideas but struggle to generate novel research directions. Clark's view is that this gap closes meaningfully through 2026-2028 as scaling produces stronger reasoning capabilities. Third, the compute and infrastructure required — training a frontier model takes 10-30M GPU-hours and thousands of accelerator chips. AI systems autonomously running this pipeline implies the lab's infrastructure is sufficiently agentic that humans aren't approving every training run. This requires both technical capability and organizational governance changes.

Q: What this means for AI safety and policy

Three implications. First, expect AI safety frameworks to formalize "recursive self-improvement" as a specific capability threshold requiring oversight — UK AISI, US AISI, and the EU's AI Act enforcement bodies will likely converge on similar language. Second, expect frontier AI labs to publish more concrete safety-evaluation methodologies for recursive self-improvement scenarios; these are currently underdeveloped. Third, expect increased policy focus on compute-cluster oversight — if AI is autonomously running training pipelines, the visibility and control points become the underlying compute infrastructure rather than the model itself. For investors and operators, the practical implication is that the AI capability landscape may be meaningfully different by 2028 than current extrapolations suggest. Strategic decisions made today should bake in optionality for fast-than-expected capability progress, not just continued linear scaling.

By Jaspal Singh May 5, 2026 Updated: May 5, 2026

Anthropic co-founder and head of policy Jack Clark estimates a 60%+ probability that AI systems will be capable of autonomously designing and training their successors by the end of 2028, according to his weekend Import AI newsletter. The prediction — embedded in a longer essay about the changing dynamics of AI research and development — is one of the most concrete public timelines from a frontier AI lab leader on the question of recursive self-improvement, and it lands at a moment when U.S. AI policy is actively debating pre-release model vetting (per the White House report we covered earlier this morning).

Clark's claim is specific and measurable. "Recursive self-improvement" in his framing means AI systems autonomously running the entire pipeline of designing model architectures, conducting training runs, evaluating results, and iterating — without human researchers in the loop for the core technical decisions. The 2028 timeline implies that the foundational capabilities for this loop will exist within ~32 months. The 60%+ probability framing means Clark sees this as the modal outcome, not a tail scenario.

What Clark's actually predicting

The prediction has three distinct components. First, the AI engineering capability required — the ability to write, debug, and ship complex ML training pipelines, conduct hyperparameter searches, and handle distributed training across GPU clusters. Per Clark's framing, this capability already exists in current models for narrow components and is rapidly improving across the broader pipeline. Most engineering tasks in ML research are now AI-assisted at major labs.

Second, the research-direction capability — the more ambiguous question of whether AI can identify which architectural ideas are worth pursuing, which training data mixtures will produce desired capabilities, and which evaluation methodologies are valid. This is the harder problem; current AI systems can execute on research ideas but struggle to generate novel research directions. Clark's view is that this gap closes meaningfully through 2026-2028 as scaling produces stronger reasoning capabilities.

Third, the compute and infrastructure required — training a frontier model takes 10-30M GPU-hours and thousands of accelerator chips. AI systems autonomously running this pipeline implies the lab's infrastructure is sufficiently agentic that humans aren't approving every training run. This requires both technical capability and organizational governance changes.

Why this matters now

The prediction is meaningful for three policy and commercial reasons. First, it sets a concrete benchmark for AI safety researchers and policymakers. "AI systems can build their successors" is a natural threshold that triggers heightened oversight requirements; specific timing helps frame regulatory urgency. Second, it confirms internal Anthropic timelines that have been previously vague — Dario Amodei has discussed similar trajectories, but Clark's specificity is unusual. Third, it amplifies the case for the White House's reported pre-release vetting framework by establishing that the capability gap to "AI that's a serious national-security concern" is short.

The skeptical view is also worth flagging. Clark works at Anthropic, which commercially benefits from AGI urgency narratives. The lab's investment thesis depends partly on AGI timelines being short; underweighting this institutional incentive is dangerous. Independent researchers like Stuart Russell, Yoshua Bengio, and Geoff Hinton have similar timelines but different commercial incentives, which strengthens the credibility of the claim. Researchers like Yann LeCun and Gary Marcus argue significantly longer timelines or fundamental architecture limits that would push recursive self-improvement years or decades further.

My Take

Clark's prediction is in the credible range but not the consensus range. The mainstream AI research community's current 50th-percentile estimate for "AI builds successors" is closer to 2030-2032 based on aggregated forecasting markets and surveys. Clark's 2028 estimate is on the aggressive end of credible forecasts but not implausible given current capability trajectories.

The right frame is probability distributions rather than point estimates. By 2028, there's likely a 30-50% chance of meaningful recursive self-improvement capability; by 2030, that rises to 60-80%; by 2033, it's near-certain. Clark's specific 60%+ at 2028 falls within the upper range of credible distributions but not absurdly so. Importantly, the policy implications hold across the credible range — even at 30% probability for 2028, the asymmetric downside risk justifies robust safety oversight.

The deeper question is what "AI builds successors" actually means commercially. If recursive self-improvement produces models that are 10x more capable than current frontier models, the entire AI commercial landscape gets reshuffled — frontier labs that hit this capability transition first capture massive economic value, and labs that miss it face structural disadvantage. That's why labs are investing so aggressively in research-acceleration tooling right now. Whether the timing is 2028 or 2032 doesn't change the strategic picture; it just changes the urgency.

What this means for AI safety and policy

Three implications. First, expect AI safety frameworks to formalize "recursive self-improvement" as a specific capability threshold requiring oversight — UK AISI, US AISI, and the EU's AI Act enforcement bodies will likely converge on similar language. Second, expect frontier AI labs to publish more concrete safety-evaluation methodologies for recursive self-improvement scenarios; these are currently underdeveloped. Third, expect increased policy focus on compute-cluster oversight — if AI is autonomously running training pipelines, the visibility and control points become the underlying compute infrastructure rather than the model itself.

For investors and operators, the practical implication is that the AI capability landscape may be meaningfully different by 2028 than current extrapolations suggest. Strategic decisions made today should bake in optionality for fast-than-expected capability progress, not just continued linear scaling.

Frequently Asked Questions

Who is Jack Clark?
Co-founder and head of policy at Anthropic. Previously worked at OpenAI on policy and at Bloomberg as a journalist covering AI. Author of the widely-read Import AI newsletter.

What is "recursive self-improvement"?
The capability for AI systems to autonomously design and train their successor models — running the entire research-and-engineering pipeline without human researchers in the loop for core technical decisions.

Is 2028 a consensus timeline?
No. Most aggregate AI capability forecasts put the 50th-percentile estimate for recursive self-improvement closer to 2030-2032. Clark's 2028 estimate is aggressive but within the upper range of credible forecasts.

What are the safety implications?
Significant. Recursive self-improvement is a natural threshold for heightened safety oversight — once AI systems can autonomously improve themselves, the pace of capability progress could accelerate beyond human ability to monitor. Most AI safety frameworks identify this as a critical inflection point.

The Bottom Line

Jack Clark's 60%+ probability estimate for recursive self-improvement by 2028 is one of the most concrete public timelines from a frontier AI lab leader. The prediction is aggressive but credible, and it amplifies the case for the White House's reported pre-release AI vetting framework. Whether the actual timing is 2028 or 2032, the policy and commercial implications are substantial and require strategic preparation today.