Anthropic co-founder Jack Clark published his latest Import AI newsletter on May 4 arguing that all the engineering components required to automate AI research are now in place, that new research from Anthropic's Fellows Program demonstrates Claude can autonomously develop, test, and analyse alignment ideas with only minimal human calibration, and that he assigns approximately 60% probability to a frontier model autonomously training a successor version of itself before the end of 2028, a timeline that AI forecaster Ryan Greenblatt independently revised upward from 15% to 30% for full AI R&D automation by the same date.
Clark is careful about what he is and is not claiming. He does not expect a production demonstration of end-to-end AI research automation in 2026 and gives explicit reasons for the 2027 exclusion: frontier model training is expensive, the product of large numbers of humans working extremely hard, and the problem of proposing genuinely novel research directions remains unsolved. What he argues is that the distance between current capability and the point at which these bottlenecks are cleared is shorter than most outsiders believe, because the engineering scaffolding, the automated experiment execution, data generation, evaluation, and iterative fine-tuning, is already functional today. The Anthropic alignment research paper he anchors the newsletter around shows Claude proposing, running, and analysing its own alignment experiments autonomously. The researchers describe "automated research on outcome-gradable problems as already practical" and identify the remaining bottleneck as designing evaluations robust enough that AI systems can hill-climb without overfitting. That is a hard technical problem. It is also a finite one.
PostTrainBench, a benchmark Clark cites from earlier in his newsletter series, directly measures something adjacent to automated AI improvement: it tests whether frontier models can take smaller open-weight models and fine-tune them to exceed existing human-developed instruct-tuned baselines. The human baseline here is not a random amateur fine-tune. It is the best instruct model that a talented team at a frontier lab produced with significant resources and expertise. The fact that frontier models are now competitive with those baselines on some tasks in a controlled benchmark setting is the empirical foundation under Clark's broader claim. It is not proof that full automation is near. It is evidence that the capability gap between human AI researchers and frontier models on well-defined research subtasks is closing at a rate that warrants serious updating.
The incumbent advantage question is where founders and investors in AI infrastructure and tooling should focus. The standard intuition is that automation of AI research compounds advantages for whoever has the largest model, the most compute, and the best data. If OpenAI, Anthropic, Google DeepMind, and Meta can run AI research automation at scale while smaller competitors cannot, the frontier widens rather than narrows. That intuition is probably correct in aggregate. But Clark's framing suggests a more nuanced picture. He describes a research economy where given a small amount of expert human calibration, AI systems can autonomously conduct research end-to-end, producing improvements against specific problems. The phrase "small amount of expert human calibration" is doing significant work. If the leverage ratio between expert human researchers and automated research output is high enough, small technical teams with genuine research expertise could run automated research pipelines that produce improvements competitive with much larger teams running less efficient human-only workflows. That is not a certainty. It is the scenario that makes the question interesting.
The parts of AI research that are most automatable today are the most structured: hyperparameter search, architecture ablations, data augmentation experiments, evaluation harness construction, and iterative fine-tuning against measurable benchmarks. These are also the parts of AI research that are most expensive in researcher time at frontier labs and most accessible to automation with current model capabilities. The parts that remain bottlenecked by human judgment are the less structured ones: identifying which research directions are worth pursuing, recognising when a benchmark is misleading rather than informative, interpreting unexpected results, and making architectural bets that require synthesising intuitions that are difficult to formalise. Clark explicitly flags the "proposing research directions" problem as the last meaningful human role in the automated research pipeline, and notes that removing it is the threshold at which the dynamic changes qualitatively. His timeline suggests he thinks that threshold is two to three years out at the frontier level.
For startups building AI developer tools, research infrastructure, or model improvement pipelines, the near-term implication is more concrete than the AGI framing suggests. Automated fine-tuning, automated evaluation harness construction, and automated hyperparameter optimisation are all commercially available or commercially buildable today using current frontier models. The companies capturing value from these capabilities are not waiting for full AI research automation. They are selling the automatable fraction of the research workflow now, to the labs, startups, and enterprise teams running model improvement pipelines at scale. Sakana AI, a research lab building automated model training systems, has already published work on AI-generated research that was accepted at peer-reviewed conferences. The Atlantic's coverage in April described Silicon Valley as "in a frenzy over bots that build themselves." Clark's newsletter is a more measured version of the same signal, from someone with direct visibility into what is technically possible at the frontier, assigning specific probabilities rather than describing a frenzy. The specific probabilities are what make it worth taking seriously.
A 60% chance of frontier model self-training by end of 2028 is not a confident prediction. It is a statement that this outcome is more likely than not within a two-and-a-half year window, and that the evidence base for that estimate is public enough to examine. Founders building software on the assumption that model improvement requires large human research teams should be building optionality into their architectures. The leverage available to small technical teams in a world where AI can assist substantially with its own improvement is different in kind from the leverage available today, and the companies best positioned for that transition are the ones that understand the research workflow deeply enough to know which parts are already automated and which parts are not yet.
","excerpt":"Anthropic co-founder Jack Clark published Import AI 455 on May 4 arguing that all engineering components for AI research automation are now in place, anchoring the claim on new Anthropic research showing Claude can autonomously run alignment experiments end-to-end with minimal human calibration.
Also read: Image AI Is Outpacing Chatbot Upgrades as the Growth Driver That Actually Converts and Founders Are Repricing the Opportunity Accordingly • The EU Wants Anthropic to Test Its Banks for Mythos Vulnerabilities and That Negotiation Is Reshaping How Frontier AI Enters Regulated Markets • Long Lake Is Paying $6.3 Billion to Take Amex GBT Private and the AI Travel Thesis Behind It Is More Interesting Than the Price
