Who is speaking and why it matters

Forty-nine-year-old Londoner Demis Hassabis is not just another techno-optimist from Silicon Valley. Two years ago, he won the Nobel Prize in Chemistry for solving the mystery of protein folding using the AI system AlphaFold. Twelve years ago, he founded the DeepMind lab, which Google bought for 12 billion crowns — and it is precisely on this research that the key Gemini model now stands, powering Alphabet's entire AI offensive. It is precisely because of this that Hassabis is counted among the co-creators of the current AI boom. So when he talks about risks, it is not academic speculation — he speaks from the position of someone sitting in the front row. His essay, published on July 14, 2026, is resonating across the tech world and beyond. This is not the first warning we have heard from Hassabis, but this time it is exceptionally concrete. "Time is short, the hour is late. Something must be done, and now," he writes literally.

AGI as a new fire — not just another internet

In his essay, Hassabis talks about so-called artificial general intelligence (AGI) — a system that would handle all the cognitive capabilities of the human brain. He describes such technology as being just a few years away. He compares its impact more to the discovery of fire or electricity than to inventions like the internet or the mobile phone. "When we look back at this time in the coming decades, I think we will realize that we stood at the foot of the singularity — at nothing less than the dawn of a new age of humanity," he writes. The famous line about "sand that thinks" refers to the fact that today's AI models run on silicon chips — and silicon is a basic component of sand. A poetic metaphor, but behind it lies a hard reality: development is progressing faster than humanity can understand.

A concrete plan: The Office for Frontier AI

What sets Hassabis's text apart from previous warnings is the concrete proposed solution. He calls for the US government to establish a new specialized office that would test the safety of the most advanced models before they reach the market. As a model, he cites FINRA — the American organization for regulating financial markets. It operates as a partnership between the private sector and the state: it is largely funded by the industry itself but has powers delegated by the state. It is neither purely industry self-regulation nor a classic bureaucratic agency, which according to Hassabis "could not keep pace." The most advanced models would, according to his proposal, be labeled as "frontier" — based on meeting threshold values on selected capability tests, not based on the computational power used during training. This is a fundamental departure from earlier plans by the US and the European Union, which derived regulation precisely from the amount of computational power.

What would be tested

Testing would primarily focus on three areas:

• Cybersecurity — the model's ability to carry out or facilitate cyberattacks
• Biological risks — the model's potential to assist in the development of biological weapons
• The model's ability to bypass safety guardrails or intentionally deceive

Labs would initially submit their models to the office voluntarily, up to 30 days before release. Once the testing protocol proves itself, it would become mandatory — without a successful test, no model would be allowed on the US market. Creators of frontier models would also have to disclose technical details, maintain strong internal security, vet key employees, and allocate sufficient resources to safety research.

Altman and Hassabis: Two plans, one goal

Hassabis is not the only one coming forward with an AI regulation proposal these days. At the beginning of July, Sam Altman, CEO of OpenAI, published a similar piece in the Financial Times. According to The Economist magazine, both proposals were presented to world leaders at the recent G7 summit in France. Both men agree that the new oversight body should set standards, analyze risks, and ensure that only countries that join the initiative have access to the most advanced models. They differ, however, in approach: Altman calls for internationally coordinated effort under US leadership. Hassabis, on the other hand, argues that the United States must act first without waiting for broader consensus. Other countries, according to him, will eventually join the rules on their own to maintain access to American technology and markets — and in the long term, this could even apply to China.

What this means for Europe and the Czech Republic

For European readers, it is key that Hassabis's proposal takes a different path than current European regulation. The EU AI Act, which came into effect in 2024, classifies AI models by risk level and imposes additional obligations on the most advanced ones — referred to as GPAI (general-purpose AI). These, however, are derived precisely from computational power, which is exactly the approach Hassabis rejects. If the US were to adopt its own regulatory framework based on capability testing instead of computational power, it could become the global standard. For Czech companies and developers, this would mean that access to the latest models — whether from Google, OpenAI, or Anthropic — would be governed primarily by US rules. The European Union would have to find a way to align its approach, otherwise it risks European users being left with older models. Tension between the US and the EU in the area of AI access already exists. Recently, the United States restricted European access to some of the most advanced models for security reasons — and European officials sharply protested against it. Hassabis's proposal could at least codify these rules and make them more predictable. Gemini itself, the model powered by DeepMind, supports Czech and is commonly available in the Czech Republic. If new regulatory rules were to apply to it in the future, Czech users would likely not feel the change directly — but it could affect the pace at which Google releases new versions.