Demis Hassabis, co-founder and CEO of Google DeepMind and a recent Nobel Prize winner, published an essay on his Substack in which he makes a bold claim. In his view, artificial general intelligence—a system possessing all the cognitive capabilities of the human brain—is probably only a few years away. He describes it as a breakthrough that cannot be compared to the internet or mobile phones. Instead, he likens it to the discovery of fire or electricity. And he immediately adds that right now, it is being decided whether humanity will make the most of it or set itself up for trouble.
His essay offers two perspectives. One enthusiastic and the other concerned. Let us examine both.
Enthusiasm: a technology that he says will surpass the Industrial Revolution
Hassabis estimates that this technology will have roughly ten times the impact on the world as the Industrial Revolution, in one-tenth the time. He famously summed it up by saying that we have essentially figured out how to make sand think, calling it a miracle.
He sees specific benefits in science and medicine. Artificial intelligence should accelerate drug development, help create new sources of clean energy, and deliver advanced materials. He also envisions a world where resources cease to constrain human progress—a kind of age of abundance. He has spent his entire life working on artificial general intelligence precisely because he believes it will be beneficial if it is built and deployed responsibly. But that “if” is a very big one for him.
What concerns him: a race that no one fully understands
On the other hand, Hassabis writes that the world is caught in an extraordinarily intense commercial and geopolitical race. And although this pressure is driving progress forward, there is a catch. Development at the cutting edge of the field is outpacing our understanding of the technology itself. He openly admits that no one in the world knows for certain what comes next. Even experts disagree. And when uncertainty is this great and so much is at stake, he recommends the only sensible approach: cautious optimism.
He is primarily concerned about security threats that may emerge as models become more capable. In his view, we are already seeing cybersecurity problems, while risks related to the nuclear and biological fields may appear on the horizon. He is particularly preoccupied by the prospect of systems that act autonomously and can improve themselves. With such systems, it will be necessary to maintain control and address problems that only become apparent over time.
His main criticism is directed at all of us. As a field and as a society, he says, we are not yet giving ourselves the time and space needed to take the next crucial step properly.
His solution: a new regulator modeled on securities oversight
Hassabis puts forward a proposal. He wants a new approach to testing the capabilities of cutting-edge models that would be both flexible and rigorous. In his view, the United States should take the first step because it has the economic and technical capacity to do so.
He proposes establishing a standards body. As a model, he cites FINRA, the U.S. organization that oversees the financial sector—a public-private partnership under federal supervision. Its leadership would include independent technical experts as well as representatives of the open-model community. Funding would have to be generous and would come predominantly from companies so that the body could attract top technical talent and pay for the enormous computing capacity required for extensive testing.
How would such a body work in practice? A model would be designated “frontier-class” if it met certain thresholds across a suite of tests that the body would update regularly. Companies with such models would become “frontier laboratories” and would be expected to follow best practices, including publishing model cards, maintaining strong internal cybersecurity, vetting key personnel, and providing sufficient resources for safety research.
Participation would initially be voluntary. Laboratories would provide their models to the body for evaluation no later than thirty days before release. Once the process had proven effective, it could quickly gain authority, meaning that a frontier model would have to pass the tests before it could enter the U.S. market at all.
What exactly would be tested
Evaluations should include rigorous scientific assessments of capabilities in cybersecurity, biological threats, and other high-risk areas. For autonomous systems, the tests would look for attempts to circumvent safety safeguards or signs of deception.
Among the best practices he mentions are digital watermarks for AI-generated images and outputs in a form humans can understand, making it possible to track how a model reasons. The tests would be updated—perhaps quarterly at first—with outdated or compromised ones replaced by new tests.
Over time, the body should be able to create its own confidential tests independently of the laboratories so that models could not become accustomed to them in advance. An ecosystem of independent third-party auditors could develop around this process.
A safeguard for the worst-case scenario
Hassabis writes that the entire system could take a tougher approach if necessary. If the seriousness of the situation required it, the body could even coordinate a slowdown in development among frontier laboratories. That is an unusual statement from the head of one of the fastest-moving laboratories.
The rules would apply fairly to everyone. They would cover frontier models regardless of their country of origin and regardless of whether they were open or closed. Smaller models from startups or academia would be exempt from the process. Hassabis sees the U.S. effort as a starting point. He hopes it will encourage an international consensus on how to manage the most serious risks.
Questions that engineers cannot answer alone
At the end of the essay, Hassabis acknowledges that solving the technical challenges will not be enough. He says we face difficult economic and philosophical questions. What new economic models will the world need? What principles will guide our lives, where will we find meaning, and how might the human condition itself change?
According to him, these questions must not be answered by technologists alone. They require the involvement of every part of society. And he adds a sentence that captures the entire point of his essay. The future, he says, has not yet been written, and we must use the small window of time remaining before the arrival of AGI to steer the technology toward benefiting everyone.
