"We made sand think" — Google DeepMind chief Demis Hassabis sounds the alarm on AI

Google DeepMind chief Demis Hassabis — one of the key architects behind today's AI boom — has published a lengthy essay comparing the current technological leap to the discovery of fire or electricity, while at the same time warning that humanity is failing to keep pace with what it is creating.

"In essence, we've found a way to make sand think. It's a marvel," Hassabis writes, referring to the fact that today's AI models run on silicon chips whose capabilities are growing faster than the world can adjust to them.

The 49-year-old London-born scientist won the Nobel Prize in Chemistry two years ago for cracking the problem of protein folding, and twelve years ago founded the DeepMind lab, which was later acquired by Google for roughly 12 billion CZK. Today, DeepMind's work underpins the Gemini model, which forms the backbone of the entire AI strategy at Alphabet, Google's parent company.

In the essay, Hassabis discusses "artificial general intelligence" (AGI) — a system capable of performing all the cognitive functions of the human brain. By his estimate, only a few years remain before such technology emerges, and he compares its impact not to the internet or mobile phones, but to the discovery of fire or electricity. "Looking back on this period decades from now, we'll realize we stood at the threshold of a singularity — at the dawn of a new age for humanity," he writes.

Developed responsibly, Hassabis argues, AGI could help solve humanity's biggest problems — from accelerating drug development to clean energy and new materials. He compares its potential effect to speeding up the industrial revolution tenfold. In the most optimistic scenario, humanity could enter an "age of abundance," in which resources cease to be a limiting factor for progress.

At the same time, however, Hassabis warns that the world is currently caught up in a commercial and geopolitical race that is accelerating progress — while our understanding of the technology lags behind its actual capabilities. Cybersecurity risks are already apparent, and as models grow more capable, nuclear and biological threats could follow. He is particularly concerned about increasingly autonomous systems capable of "recursive self-improvement" — improving themselves without human involvement.

Hassabis also proposes a concrete solution: creating a new government body in the US that would test the safety of the most powerful models before they reach the market. In his view, this should be neither pure industry self-regulation nor a classic bureaucratic agency, which "won't be able to keep up with the pace and won't have the necessary resources."

As a model, he points to FINRA, the US financial markets regulator — a body funded largely by the industry itself but granted governmental authority. Under his proposal, the most advanced models would be designated "frontier" once they cross certain threshold capabilities.

"Time is running out, the hour is late. Something needs to be done, and done now," the DeepMind founder concludes.