Why 2031 Might Be the Last Year Humans Do AI Research
This title could be clearer and more informative.Try out Clickbait Shieldfor free (5 uses left this month).
A speculative forecast, built on claims from Redwood Research's Ryan Greenblatt in a Dwarkesh Podcast debate, argues that by 2030-2031 AI systems will fully automate AI research and development, triggering recursive self-improvement that compresses roughly five years of historical AI progress into a single year. The piece lays out a hypothetical multi-tier training pipeline for turning models into AI researchers, argues that data availability won't bottleneck this process since compute spending dwarfs data spending 10:1 to 20:1, and extends the argument into an 'industrial explosion' where AI-designed chips and robotics reshape the physical economy, culminating in projected artificial superintelligence by 2033.
Table of contents
1. Why AI R&D is the Ultimate AI-Native DomainFast, Containerized Feedback LoopsAdditive InnovationsHigh Empirical Visibility2. How Labs Will Train AI Researchers: The GPT-7.5 to GPT-9 PipelineLayer 1: Small-Scale Pre-Training EnvironmentsLayer 2: Mid-Scale Fine-Tuning ExperimentsLayer 3: Production Research & Bug Hunting3. The Math of Compressed Progress: 5 Years in a Single YearGet Ezekiel Njuguna ’s stories in your inboxOvercoming the Compute Deficit4. Debunking the Data Bottleneck Myth1. Compute vs. Data Economics2. Pre-Training Progress is Algorithmic, Not Human-Typed5. From Software to Physical World: The “Industrial Explosion”6. The Post-Human Research Era (2032 and Beyond)17.6K Impressions4 Comments