Demis Hassabis is not calling for a pause. He is calling for structure.
In his recent essay, the DeepMind CEO states plainly what many still treat as distant: AGI is probably only a few years away. We are already standing in the foothills of a new age. The question is no longer whether the technology arrives — it's whether we meet it with clarity or with chaos.
His answer is practical. Create a U.S.-led standards body modelled on FINRA. Define "frontier" models through living benchmarks that update as capabilities advance. Require rigorous pre-release testing for the systems that matter most. Start voluntary, then formalise. And build in the ability to coordinate a measured slowdown if genuine high-stakes risks appear.
That separation is the whole idea. Most of what ships is ordinary AI: it moves on standard product rules and ordinary safety review, and the innovation engine stays unblocked. The rare, highest-capability models — the ones that could reshape security, biology, and decision-making itself — get the living benchmarks, the pre-release testing, and the shared technical guardrails. Rules match the risk instead of applying one blunt standard to everything.
The unique strength of the proposal is its realism. Hassabis accepts that the commercial and geopolitical race is already underway. Rather than pretend it can be stopped, he designs a mechanism that can keep pace with it — and escalate if needed. That combination of urgency and adaptability is rare in this debate.
For any industry adopting advanced AI, the value is straightforward. Clear upstream standards reduce uncertainty. Better-tested frontier models give downstream users stronger foundations to build on. Responsibility and progress stop being opposites.
Hassabis's central insight is simple and powerful, and it's the part worth carrying out of the essay. We still have room to decide how this goes.
The future is still unwritten.
The only question is whether we use the window to build systems that can actually keep up with the technology they're meant to govern.
The only question is whether we use the window.