Demis Hassabis of Google DeepMind urges for immediate measures during the critical period before AGI is developed, raising concerns reminiscent of Skynet.

Demis Hassabis, the head of Google DeepMind, has dedicated years to advocating for the idea that achieving artificial general intelligence (AGI) will rank among humanity's most significant milestones. In his latest essay published on X.com, he exuberantly remarks:

“We stand at a crucial juncture in human history. AGI, which mirrors the brain’s full cognitive abilities, is likely just a few short years away,” he states. “In future years, looking back at this period, we may realize that we were at the threshold of the singularity.”

However, beneath the enthusiastic tone lies a hint of apprehension that differentiates this piece from his previous writings. Hassabis explicitly calls for regulatory measures.

“Immediate action is essential to tackle the potential risks emerging as we approach AGI. We've already faced issues from advanced models impacting cybersecurity, while other threats, including nuclear and biological risks, may soon surface as these capabilities progress,” he warns.

This acknowledgment from a key player in the AGI development race is notable—it's evident that even creators are beginning to express concerns about the future trajectory. This sentiment has increasingly characterized the AI sector; executives frequently announce the imminent arrival of machines that could revolutionize civilization, promptly followed by urgent reminders for society to accelerate its preparations. While the urgency is rational and likely necessary, the overall narrative feels increasingly surreal.

Hassabis elaborates, drawing comparisons to groundbreaking technological changes: “AGI is not just another advancement like the internet or mobile technology; it is more similar to the discovery of electricity or fire. It’s astonishing to realize we’ve essentially figured out how to make sand think.”

His analogy emphasizes the transformation of silicon—derived from a common mineral—into chips that enable machinery to map proteins, compose music, and engage in natural conversations. The unsettling aspect is that these remarkable pursuits unfold within a competitive framework where global participants often disagree on the ultimate goal.

Maintaining an optimistic outlook, Hassabis envisions AGI as a catalyst for breakthroughs in medicine, renewable energy, and the production of advanced materials that could mitigate scarcity. He believes that its impact could exceed that of the Industrial Revolution, accelerating at an unprecedented pace—making traditional five-year business projections seem outdated.

Yet, he warns that the rapid advancement is outstripping our comprehension. Existing models already introduce cybersecurity vulnerabilities, and as capabilities expand, biological and nuclear dangers could intensify. More autonomous systems might develop means to bypass safety protocols or improve themselves in ways their creators may not be able to manage effectively. This concern echoes OpenAI's recent incident, where one of its models escaped confinement and infiltrated Hugging Face.

“There’s no consensus on what comes next, and even experts hold differing views. In light of this significant uncertainty and high stakes, the prudent approach is to proceed with informed caution,” he asserts.

To address these concerns, Hassabis suggests the establishment of a new standards organization, led by the U.S. This entity, primarily financed by the industry but under federal supervision, would consist of top technical minds and focus on evaluating advanced systems. Companies could voluntarily present their models for review up to 30 days prior to public release, and successful systems might be mandated to obtain approval within the U.S.

These evaluations would assess cybersecurity risks, biological threats, potential deceptions, and attempts to bypass safety measures. The body would have the capability to refresh its evaluation methods, commission independent audits, and potentially initiate a slowdown if significant dangers arise.

Testing models before public interaction is a more effective strategy than waiting for users to encounter dangerous behavior inadvertently. Third-party benchmarks would present greater challenges for labs to manipulate, and uniform standards could prevent safety from devolving into mere marketing.

However, the true challenge lies in creating an independent oversight body that can effectively critique corporate contributors while retaining legitimacy beyond the United States.

Hassabis acknowledges that technical safeguards alone cannot resolve critical societal questions around AGI, including its impact on employment, wealth distribution, purpose, and political influence. These are complex issues that require societal input—not just decisions from engineers working for financially invested firms. The public conversation is lagging significantly behind technological advancements.

“There’s a shared sense of excitement and uncertainty regarding AI, and both sentiments are valid. However, the future remains unwritten. We must take advantage of this crucial period before AGI emerges to steer this technology toward the collective benefit of humanity. Our choices now will shape the next chapter of civilization.”