Google DeepMind CEO warns AGI is coming, wants frontier AI models checked by US standards body before launch
AI has evolved from answering questions to writing code, generating videos and acting as autonomous agents. But what’s the end goal? For many frontier AI labs, the next big breakthrough is developing Artificial General Intelligence (AGI), AI that can match human cognitive abilities. So, when will we get there? Google DeepMind CEO Demis Hassabis believes AGI could arrive within the next few years. But before it does, he warns the world’s most powerful AI models should undergo independent safety checks before being released to the public.
In a detailed essay titled A Framework for Frontier AI and the Dawning of a New Age on X, Hassabis says “This is a pivotal moment in human history. Artificial General Intelligence (AGI), a system that exhibits all the cognitive capabilities the brain has, is probably only a few short years away.”
Hassabis argues that AI is advancing so rapidly that governments and the industry need a new way to assess the risks posed by frontier AI models. To address this, he proposes creating a new standards body that would independently evaluate the world’s most powerful AI systems before they are deployed.
AGI will be like electricity and fire
The DeepMind chief believes AGI will be far more transformative than the internet or smartphones. Instead, he compares its impact with discoveries such as electricity and fire. If developed responsibly, he says, AGI could dramatically speed up drug discovery, help develop cleaner energy technologies, create advanced materials and unlock significant productivity gains across industries.
Hassabis writes that humans have “essentially found a way to make sand think.” He believes AGI could have an impact “10x the Industrial Revolution at 10x the speed”, helping solve some of the world’s biggest challenges and creating a future where resources are less of a limitation.
But he also cautions that the same technology could create new security risks. According to him, today’s frontier AI models are already challenging cybersecurity, while future systems could pose biological and nuclear risks. Hence, he argues that strong safeguards will be needed to ensure humans remain in control in the era of AGI.
“The field is now locked in an extremely intense commercial and geopolitical race,” he writes, adding that AI capabilities are advancing faster than society’s understanding of them. “Nobody in the world knows for sure what is going to happen from here,” he says, calling “cautious optimism” the right approach.
How can AGI be controlled?
To prepare for AGI, Hassabis proposes a US-led Frontier AI Standards Body that would independently test the world’s most powerful AI models before they are released. Under the proposal, models that cross certain capability thresholds would be classified as “Frontier Models”, while the companies developing them would become “Frontier Labs”. These companies would be expected to strengthen cybersecurity, publish model cards and invest more in AI safety research.
Initially, companies would voluntarily submit their models for testing up to 30 days before launch. If the system proves effective, Hassabis says these checks could eventually become mandatory before frontier AI models are deployed in the US.
The assessments he suggests would also examine cybersecurity, biological and other high-risk threats, while also checking whether models attempt to bypass safety guardrails or show deceptive behaviour. He also recommends safeguards such as digital watermarking for AI-generated content.
Hassabis believes the proposed framework could eventually become the basis for global AI standards ,as the technology will affect the entire world. “The future is not yet written,” he writes, arguing that the decisions made today will determine how AGI shapes humanity.
– Ends
