Hassabis wants an AI regulator like FINRA, and soon

Demis Hassabis has seen the future—or at least, he’s willing to guess it’s coming fast. The Google DeepMind CEO is out with a detailed framework for governing advanced AI, and the headline isn’t subtle: he thinks artificial general intelligence is just a few years away, a development he previously pegged as carrying ten times the impact of the Industrial Revolution at ten times the speed. That’s not a typo. He’s been consistent on this, and back in May he even placed us “in the foothills of the singularity,” a phrase that still makes a lot of researchers wince.

His proposal’s centerpiece is a new U.S. standards body, explicitly modeled on FINRA, the financial industry’s self-regulatory organization. It would start by developing voluntary evaluation protocols for frontier models, then shift to mandatory benchmarks funded by the industry itself. If things get dicey, this body could also coordinate a deliberate slowdown—an idea Anthropic has floated recently. Crucially, Hassabis wants startups and academic labs working on non-frontier systems left alone. That’s a smart carve-out that preempts the usual “regulatory capture” criticism, where giants like Google use red tape to kneecap smaller rivals before they can get off the ground.

The timing here isn’t accidental. The proposal landed just after a group of prominent AI researchers and economists fired off a warning letter about mass job displacement, a letter Hassabis notably didn’t sign. His own language is less alarmist but no less urgent. “Nobody in the world knows for sure what is going to happen from here, and even the experts disagree,” he writes. “When there is a large degree of uncertainty and the stakes are this high, proceeding with cautious optimism is the sensible and correct strategy.” That expert disagreement isn’t abstract—it blew up publicly last December when Meta’s Yann LeCun dismissed language-model-based AGI as “complete BS,” and Hassabis shot back that LeCun was “just plain incorrect.”

That feud is the real backdrop here. DeepMind’s own Gemini co-lead Oriol Vinyals offers a middle path, noting today’s models are strong in some areas but lack the spark of genuine innovation. Deep learning pioneer Richard Sutton just launched a startup to target that exact missing piece, while DeepMind co-founder Shane Legg thinks a “minimal AGI” could arrive by 2028. Hassabis’s regulatory push is essentially a bet that the LeCuns of the world are wrong, and that we need guardrails before we find out for sure.