The article discusses the impending ability of AI systems to build themselves, marking the first step towards recursive self-improvement, a milestone that could accelerate AI development dramatically.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
Based on a comprehensive analysis of public papers, benchmarks, and product deployments, Import AI editor Jack Clark believes there is a greater than 60% probability that by the end of 2028, AI R&D without human involvement (i.e., AI systems autonomously building their successors) will occur. This judgment stems from exponential progress in coding ability, autonomous task duration, and scientific skills.
In an article dated May 4, 2026, Import AI editor Jack Clark publicly stated that based on analysis of public information, he reluctantly concludes that there is a greater than 60% probability that by the end of 2028, AI R&D without human involvement (i.e., AI systems autonomously building their successors) will occur. He believes AI research is heading toward end-to-end automation, crossing a 'Rubicon' into a nearly unpredictable future.
Clark emphasizes that this judgment is primarily based on papers on arXiv, bioRxiv, NBER, and products deployed by frontier companies. He believes all the engineering components to build today's AI systems are in place; if scaling trends continue, models will become creative enough to replace human researchers in proposing novel research directions.
SWE-Bench is a benchmark for evaluating AI's ability to solve real GitHub issues. At the end of 2023, the best model Claude 2 scored about 2%; by 2026, Claude Mythos Preview reached 93.9%, effectively saturating the benchmark. Clark notes that all benchmarks have inherent noise, e.g., about 6% of labels in the ImageNet validation set are wrong or ambiguous.
SWE-Bench is a reliable proxy for coding ability. According to Clark, the vast majority of people in frontier labs and Silicon Valley now code entirely through AI systems and increasingly use AI to write tests and check code. This means AI has already automated a major component of AI R&D.
METR charts the complexity of tasks AI systems can reliably complete, measured in hours needed by a skilled human. In 2022, GPT-3.5 could complete tasks of about 30 seconds; in 2023, GPT-4 reached 4 minutes; in 2024, o1 reached 40 minutes; in 2025, GPT-5.2 (High) reached about 6 hours; in 2026, Opus 4.6 has reached about 12 hours. METR's Ajeya Cotra believes it is not unreasonable for AI systems to complete about 100-hour tasks by end of 2026.
This growth is directly linked to the explosion of agentic coding tools. Much of the work of AI researchers—such as cleaning data, reading data, and launching experiments—now falls within the time horizon of modern systems.
Modern science largely involves specifying directions, running experiments, and checking results. Advances in AI coding and world modeling have spawned tools that accelerate human scientists and partially automate R&D.
Benchmarks like CORE-Bench (Computational Reproducibility Agent Benchmark) show significant progress in AI reproducing research results, combining machine learning techniques, and optimizing AI systems themselves. These skills are at the core of AI R&D.
Clark explicitly states he does not expect fully automated AI R&D within 2026. However, he believes that within a year or two, a proof-of-concept of 'model end-to-end training its successor' could emerge, especially for non-frontier models; frontier models may be harder due to high cost and heavy reliance on human effort.
All evidence comes from public information, but benchmarks have inherent flaws. Clark emphasizes the importance of observing the overall trend from all data points rather than details of individual benchmarks.
This article is based on public analysis by Import AI editor Jack Clark. All data comes from public sources he cites (arXiv, bioRxiv, NBER, SWE-Bench, METR, etc.). Clark himself acknowledges benchmark noise and explicitly states his judgment is a 'reluctant view'; the probability estimate (60%+) and timeline prediction (end of 2028) are his personal inferences.
Automation of AI R&D may achieve proof-of-concept within two years, with over 60% probability of human-free AI system self-construction by end of 2028. This trend is driven by exponential progress in coding ability, autonomous task duration, and scientific skills, but its ultimate impact is difficult to predict.
Primary report
Primary source