Ex-DeepMind researcher's depthfirst essay: nuclear-age lessons for staying in control of AI
andreamichi · x · 2026-09-18
Andrea Michi, ex-DeepMind researcher and CTO of depthfirst, published an essay on why he left DeepMind two years ago to build AI security, examining the two main frontier-lab arguments for working at the frontier and what the nuclear age teaches about controlling AI.
- The two arguments: (1) you can only meaningfully influence AI safety at the frontier (echoing Demis's 'seat at the table' point); (2) if we slow down, less safety-conscious actors including authoritarian regimes will accelerate.
- He uses the Manhattan Project as an analogy: fear of Nazi Germany getting the bomb first drove the race, but the counterfactual is unknowable — and both arguments seem presumptuous.
- Unlike nukes, we don't yet know what forms AI will take, how risks emerge, or what containment requires; calls to slow down buy time, but we still need to build the means to stay in control.
- depthfirst builds an autonomous security command center: vulnerability discovery, supply chain security, secrets detection, agentic pentesting.
More from Companies & People
- RedMonk: Developers were the new kingmakers — agents are next in line — rseroter · 2026-09-18
- AI's PR Problem: More Jensen Huang, Less Sam and Dario, Says Commenter — NathanWilbanks_ · 2026-09-18
- Noam Brown's OpenAI team is hiring for long-horizon agents, alignment, and HCI — polynoamial · 2026-09-18
- Ex-AI exec: labs claim existential IP stakes yet operate in California where non-competes are void — suchenzang · 2026-09-18
- DeepMind vet vqctran leaves after 10 years to co-found biology AI startup Polyphron — infoxiao · 2026-09-18
- Pony model creators leave Google to launch Fictional, betting on AI relationships — Scobleizer · 2026-09-18