DeepMind's Hassabis Calls for US-Led AI Watchdog — FINRA-Style, Industry-Funded, 30-Day Pre-Release Testing
TL;DR
DeepMind CEO Hassabis calls for US-led AI regulator modeled on FINRA — industry-funded, 30-day pre-release safety testing.
Google DeepMind CEO Demis Hassabis published a personal manifesto "A Framework for Frontier AI and the Dawning of a New Age" on July 14, calling for the US to establish a new AI regulator — with authority to screen the world's most advanced AI models and coordinate industry-wide slowdowns if risks escalate. Target: up and running before year-end.
The regulatory model references FINRA (the US financial industry's self-regulatory body): industry-funded, staffed by world-class technical experts, accountable to the US government (SEC equivalent). Frontier labs would voluntarily hand over models up to 30 days before release for safety testing — covering cyber-attack capabilities, biothreats, and model "deception."
The trigger is last month's Anthropic freeze. Post #57 covered the US suddenly pulling Claude Fable 5 and Mythos 5 into export control with three-day takedowns — Anthropic spent two and a half weeks negotiating unfreeze with no rules, protocol, or playbook. Hassabis: "This was a wake-up call... Washington needs something sturdier than ad hoc directives."
Hassabis's time window is aggressive — he argues that within the next 18 months, cyber capabilities (already in warning stage) and more serious bio and nuclear safety risks could appear in open-source models no government can control. He explicitly emphasized risks don't just come from open source — they'll also come from future proprietary models from all major AI labs.
Behind-scenes prep has been months. Hassabis has quietly engaged the Trump administration, other AI lab heads (including Anthropic's Amodei), and European officials — at last month's G7, Hassabis and Amodei jointly called for a similar framework. This is turning the concept into a public manifesto and starting formal lobbying.
Industry self-regulation vs government direct oversight — old debate. FINRA model pros: technical fluency, industry pays. Cons: critics have long called FINRA "too soft" on Wall Street. Applied to AI, the question is whether Anthropic, OpenAI, and DeepMind will objectively evaluate their own models — the hardest thing for Washington to trust.
If it works, the US builds AI-FINRA by year-end, China and EU forced to align, US regains AI governance leadership. If it doesn't, Trump rejects industry-led approach and reverts to post-#57 ad hoc export controls, and AI safety governance keeps unraveling in chaos.
via CNBC / CryptoBriefing
The regulatory model references FINRA (the US financial industry's self-regulatory body): industry-funded, staffed by world-class technical experts, accountable to the US government (SEC equivalent). Frontier labs would voluntarily hand over models up to 30 days before release for safety testing — covering cyber-attack capabilities, biothreats, and model "deception."
The trigger is last month's Anthropic freeze. Post #57 covered the US suddenly pulling Claude Fable 5 and Mythos 5 into export control with three-day takedowns — Anthropic spent two and a half weeks negotiating unfreeze with no rules, protocol, or playbook. Hassabis: "This was a wake-up call... Washington needs something sturdier than ad hoc directives."
Hassabis's time window is aggressive — he argues that within the next 18 months, cyber capabilities (already in warning stage) and more serious bio and nuclear safety risks could appear in open-source models no government can control. He explicitly emphasized risks don't just come from open source — they'll also come from future proprietary models from all major AI labs.
Behind-scenes prep has been months. Hassabis has quietly engaged the Trump administration, other AI lab heads (including Anthropic's Amodei), and European officials — at last month's G7, Hassabis and Amodei jointly called for a similar framework. This is turning the concept into a public manifesto and starting formal lobbying.
Industry self-regulation vs government direct oversight — old debate. FINRA model pros: technical fluency, industry pays. Cons: critics have long called FINRA "too soft" on Wall Street. Applied to AI, the question is whether Anthropic, OpenAI, and DeepMind will objectively evaluate their own models — the hardest thing for Washington to trust.
If it works, the US builds AI-FINRA by year-end, China and EU forced to align, US regains AI governance leadership. If it doesn't, Trump rejects industry-led approach and reverts to post-#57 ad hoc export controls, and AI safety governance keeps unraveling in chaos.
via CNBC / CryptoBriefing
