AI research automation
reinforcing loop · 5 steps · Trace on the map →
The cycle
- AI speeds up AI researchAI R&D uplift
- Superhuman AI researcherSuperhuman AI researcher
- Better algorithms and trainingAlgorithmic efficiency
- Stronger modelsReasoning models
- Longer, more reliable agent runsLong-horizon agents
What drives it
Each model generation speeds up the research that builds the next one.
What breaks it
Forethought puts returns to software research at roughly 0.5 to 2 once hardware is held fixed. Above 1, progress accelerates; below 1, it fizzles as ideas get harder to find. Limited compute for experiments and months-long training runs are the brakes.
What to watch
AI R&D uplift: 1.04 to 2x in 2026, against about 25x for a superhuman researcher. METR time horizons on AI research tasks.