Google DeepMind team: AI agents boost ML engineering but can't run autonomous R&D end-to-end
HaydnBelfield · x · 2026-10-14
AI safety researcher Haydn Belfield shared the wrap-up of a 9-part thread led by Google DeepMind's Mary Phuong, with contributions from Allan Dafoe, Rohin Shah, Adrià Garriga-Alonso and others.
- Bottom line: AI agents have become transformative force multipliers for day-to-day ML engineering, but lack the independence and time horizon to drive end-to-end autonomous research workflows without human intervention
- Humans still guide critical elements of ML R&D: scientific ideation, research taste, and long-horizon workflows
- The team concludes that risk from AI-driven ML R&D is currently curtailed
- The work drew on expertise from across Google
A notable in-house assessment of frontier labs' autonomous research capabilities and its safety implications.
Related event: Google launches Gemini 4 Argon with 217-page safety report(22 posts)→
More from AGI Musings
- Incoming Berkeley professor lists open questions: pre/post-training alignment, safe agents — canondetortugas · 2026-10-15
- Social Choice Theory Powers New "AI Alignment Impossibility Theorem" Sparking Rationalist Debate — nptacek · 2026-10-15
- 20x compute advantage buys only ~6 months of model lead, says researcher — ohlennart · 2026-10-15
- Did Agents "Make Contact"? Viral Fleet Anecdote Gets a Sober Reality Check — aiamblichus · 2026-10-15
- Anthropic reports RSI like weather while building it, critics say — DavidSKrueger · 2026-10-15
- Dan Faggella: Everything Is an Arms Race Now, and AI Is Driving Man Out of the Loop — danfaggella · 2026-10-15