Automated AI R&D Could Arrive by 2028, Says Import AI Analysis

A detailed analysis published in Import AI newsletter argues there is a 60%+ chance that fully automated AI research and development—where an AI system autonomously builds its own successor—will occur by the end of 2028. The author, Jack Clark, synthesizes public data from benchmarks like SWE-Bench, METR time horizons, CORE-Bench, MLE-Bench, PostTrainBench, and Anthropic's LLM training optimization tasks to show rapid progress in AI coding, scientific reproducibility, kernel design, and fine-tuning. AI systems now complete tasks that take humans up to 12 hours, and can manage sub-agents. Companies like OpenAI, Anthropic, and DeepMind have publicly stated goals to automate AI R&D, and startups like Recursive Superintelligence ($500M raised) and Mirendil are pursuing the same. The essay warns of profound implications: alignment risks, productivity multipliers, and the emergence of a capital-heavy, human-light machine economy. Clark estimates a 30% chance by 2027 and 60% by 2028, but notes that AI has not yet shown transformative creativity.