Automating AI Research

a day ago

AI research is rapidly advancing towards full automation, with a 60%+ chance of achieving no-human-involved AI R&D by the end of 2028, where an AI system could autonomously build its successor.
Key evidence includes dramatic improvements in AI coding capabilities (e.g., SWE-Bench scores rising from ~2% to 93.9%), extended task time horizons (from seconds to hours), and progress in automating scientific tasks like reproducing research (CORE-Bench) and optimizing AI training (e.g., 52x speedup in language model training).
AI systems are now capable of managing other AIs, forming synthetic teams to tackle complex problems, and showing early signs of creativity in fields like mathematics (e.g., solving Erdős problems), though their ability to generate novel insights remains uncertain.
Automating AI R&D raises critical challenges: alignment risks may escalate with recursive self-improvement, economic productivity multipliers could exacerbate inequality, and a shift toward a capital-heavy, human-light economy may emerge, with potential for fully autonomous AI-run corporations.
Major AI companies (e.g., OpenAI, Anthropic) and startups are investing heavily in automating AI research, indicating a strong industry push toward this goal, which could accelerate progress despite uncertainties about AI creativity and scalability.

Hasty Briefsbeta