Headline

Demis Hassabis says AGI is just a few years away and wants a US-led group to test advanced models before they go live.

Summary

Demis Hassabis, who runs Google DeepMind, thinks AGI might show up soon and could shake things up more than earlier tech shifts. He's concerned that the race between companies and countries is moving faster than society can handle the risks. His fix is a US-initiated standards body, roughly modeled on FINRA, that would review top models for safety before deployment and set common benchmarks for dangerous capabilities.

Analysis

What stands out is a leading lab head pairing a short AGI timeline with an actual governance idea. Hassabis treats safety less like filling out compliance forms and more like running real technical tests that would need regular updates, serious compute access, and outside experts. The setup starts voluntary but could become mandatory for models used in the US, offering something between hands-off self-regulation and full licensing. Bringing in open-source voices and applying rules to both open and closed systems hits the core tension: how to control powerful AI without squeezing out startups, universities, or open research. Frontier labs would face tougher requirements on security, model documentation, staff screening, safety work, and capability testing, which means higher costs but also clearer lines on what counts as ready to ship.

Impact Assessment

Significance: High

Categories: AI Policy, AI Safety, Industry Trend