Two DeepMind exits in one week, both citing extinction risk — this isn't a rogue whistleblower anymore, it's a pattern inside the companies building AGI.
The Summary
- Bilal Chughtai resigned from Google DeepMind, stating "I earnestly believe that AI has the potential to kill us all" and that humanity "might be running out of time"
- His exit follows former Anthropic researcher Jacob Coxon's viral warning that AI lab employees privately fear the technology could wipe out humanity
- Former DeepMind scientist Alex Turner also left over safety concerns, arguing AI progress is moving too fast for governments or even superpowers to control
- Turner suggests tracking compute could be the key to real regulation, noting Big Tech's "slow down" calls might be both sincere and strategic
The Signal
The departures are piling up. Chughtai worked on AGI safety and alignment research at DeepMind, the division explicitly tasked with preventing catastrophic outcomes. His resignation statement posted across social platforms echoes Coxon's warning from days earlier, but with one crucial difference: Chughtai was inside Google's operation, watching "AI development first hand" and concluding the "default trajectory of this technology" is headed somewhere dark.
This isn't outside critics anymore. These are people who were in the rooms where capabilities are measured, where model behavior gets debugged, where the gap between "we think this is safe" and "we can prove this is safe" becomes obvious. Turner's interview reveals the internal tension: AI progress is outpacing not just regulation, but the cognitive capacity of institutions to even understand what they're regulating.
"AI progress is moving too fast for governments, companies, and even superpowers like the US and China to control."
The timing matters. Anthropic CEO Dario Amodei said over the weekend that frontier AI labs need "to pace" their development. That's the head of a leading AI company saying pump the brakes, the same week two researchers from competing labs resign over safety. When the people building the technology and the people leaving the companies both say the same thing, that's not agenda, that's consensus forming in real time.
Turner's compute tracking proposal is the first specific mechanism anyone's offered that isn't pure hand-waving. If you can measure the computational resources being deployed, you can detect when someone's training a frontier model. It's verifiable. It's enforceable. And Turner suggests it might be the only lever that works across jurisdictions, because compute doesn't lie about its scale.
But here's the darker read: Turner also notes that Big Tech's sudden interest in slowing down "might be both sincere and a power play." Translation: if you're Google or Anthropic and you've already built massive compute infrastructure, calling for a pause locks in your lead. New entrants can't catch up if the race stops. So when researchers who worked on alignment at these companies walk away citing existential risk, you have to ask: are they leaving because the risk is real, or because they see the regulatory capture coming and want no part of it?
The Implication
Watch the exits. When safety researchers resign from the labs with the most resources and the best talent, that's a leading indicator. Either the internal culture around caution is breaking down, or the people tasked with making AGI safe have concluded it can't be done under current conditions. Either way, the companies building this technology are losing the people who understand the risks best.
If you're building in the agent space, the compute tracking idea is worth understanding. Regulatory frameworks that actually work will likely center on measurable inputs like compute, not vague promises about "responsible AI." The companies that get ahead of that shift, building transparency and auditability into their infrastructure now, won't be scrambling when governments finally figure out what levers to pull.