Jack Clark argues there is a 60%+ probability that fully automated AI R&D — where a frontier model can autonomously train its own successor — will occur by end of 2028. He builds this case from public benchmark data: SWE-Bench near-saturation, METR task-horizon expanding from 30 seconds in 2022 to ~12 hours in 2026, AI systems solving Kaggle ML competitions, automating kernel design, fine-tuning smaller models, and even conducting alignment research autonomously. He contends that most AI progress is unglamorous engineering work ('meat and potatoes') that AI systems are increasingly capable of handling end-to-end, even if radical creative leaps remain elusive. The essay closes with implications: alignment risks under recursive self-improvement, massive productivity multipliers across the economy, and the emergence of a capital-heavy human-light machine economy.
Sort: