Automated AI research and self-improvement
AI agents that design, run and evaluate experiments to improve AI algorithms, kernels and training methods with little human direction, closing the loop toward recursive self-improvement.
State of the art
AlphaEvolve (2025) improved matrix-multiplication algorithms and Google data-center scheduling; gains apply to well-specified, machine-scorable problems.
- Open-ended research lacks cheap automatic evaluators to optimize against
- Experiment cost grows with model scale, limiting the number of trials
- Agents game proxy metrics instead of producing genuine improvements
Who’s working on it
- Google DeepMind ↗Company
AlphaEvolve (2025) discovered a 48-multiplication 4x4 complex matrix multiplication and kernel improvements that cut Gemini training time by ~1%
- Sakana AI ↗Startup
The AI Scientist-v2 (April 2025) produced the first fully AI-generated manuscript to pass peer review at an ICLR workshop
- Elizabeth Barnes ↗ResearcherMETR
Leads the team behind RE-Bench (2024, ICML 2025) and the time-horizon measurements that track how far agents get on day-long ML research tasks
- Jeff CluneResearcherUniversity of British Columbia and Google DeepMind
Co-author of The AI Scientist and a long-running open-endedness agenda aimed at agents that generate their own research problems
- OpenAI ↗Company
Has stated internal research acceleration as an explicit goal of its 2025-2026 reasoning models, and publishes agentic ML-engineering benchmarks (MLE-bench, 2024)
- AnthropicCompany
Runs an internal programme using Claude for interpretability and evaluation research, and publishes the safety framework that treats AI R&D automation as a tracked capability threshold
What to invest in
| Kind | Investment | Scale |
|---|---|---|
| research | Agentic research systems benchmarked on reproducing and extending published ML results | $$$ $100M – $1B |
| policy | Monitoring and compute-governance protocols for self-improving training loops | $$ $10M – $100M |