"Those who do not abide by the rules gain an unfair advantage", the father of DeepMind talks about AGI and cutting-edge model regulation again

Demis Hassabis, CEO of Google DeepMind, published a lengthy essay titled "The Frontier AI Framework and the Dawn of a New Era" on July 14, systematically laying out his perspectives on the pace of AI evolution, the timeline for AGI, and the United States' role in regulating frontier models.

"The pace of progress in frontier technology has outstripped our ability to understand it," Hassabis stated.

He compared AGI to humanity's discovery of fire and electricity, asserting it could be achieved "within a few years," and describing its impact as "equivalent to the Industrial Revolution but ten times larger and ten times faster."

This rapid evolutionary pace has also sparked "the most intense commercial competition," pushing frontier labs into a race to the bottom. "Doing safety work becomes harder, while those who cut corners gain an unfair advantage."

During a discussion with Stanford President Jonathan Levin in early June 2026, Hassabis emphasized that this "rush to market" dynamic must be reversed as soon as possible. He noted, "If you spend extra time refining your product or ensuring its safety, it becomes far more challenging than simply launching it recklessly and hoping for the best. As a result, those who jump the gun (or violate norms) gain a distinct advantage. This is the classic problem inherent in a race-to-the-bottom dynamic, and we must find a way to change this — I believe it is extremely urgent."

Later, during a closed-door session at the G7 Summit, Hassabis joined forces with Anthropic CEO Dario Amodei to propose: the United States should lead the establishment of an AI standards body modeled after financial regulatory systems, requiring frontier models to pass independent safety assessments and certifications before public release.

In his essay, Hassabis argues that the rapid advancement of AI demands a dynamic, rigorous new evaluation system to test frontier models, and the United States, leveraging its economic and technological standing, is fully positioned to take this first step. He believes technology can ultimately solve all challenges — but only if humanity secures firm control over the technology during this critical sprint.

Hassabis's timing for releasing this lengthy essay is largely strategic, laying the groundwork for the upcoming launch of Google's flagship model, Gemini 3.5 Pro.

Tech media outlets HackerNoon and Geeky Gadgets, citing anonymous third-party sources, report that Google plans to launch Gemini 3.5 Pro on July 17. Allegedly, it can process twice as much information as Gemini 3.5 Flash (featuring a 2 million-token context window), excels at solving complex reasoning problems, and can independently complete multi-step programming tasks.

On the safety front, Google already emphasized during the Gemini 3.5 Flash launch that the entire model family was developed in adherence to a rigorous frontier safety framework, with targeted protections for high-risk domains including cyberattacks, chemical and biological hazards, and nuclear security. The subsequent launch of Gemini 3.5 Pro aligns perfectly with the logic of Hassabis's essay: establishing safety frameworks first, then releasing the model.

The following is Hassabis's full essay (edited and abridged):

This is a pivotal moment in human history.

AGI — a system with all the cognitive capabilities of the human brain — could be achieved in just a few short years. When we look back on this era decades from now, I believe we will realize we were standing at the foothills of the singularity, undoubtedly the dawn of a new epoch for humanity.

I have dedicated my entire career to AGI research, because I remain deeply convinced that if built and deployed responsibly, it will prove to be one of the most beneficial and transformative technologies humanity has ever created.

AGI is not on the same level of breakthrough as the internet or mobile phones, remarkable as those innovations were in reshaping our world. It should be compared directly to the discovery of electricity and fire. If you pause to reflect, what we are doing is essentially finding a way to make sand think. That is close to a miracle.

The scale of this technology's impact will be unprecedented — potentially comparable to the Industrial Revolution, but unfolding ten times faster. It will help us solve some of society's greatest challenges, from accelerating drug discovery and developing new clean energy sources, to creating innovative advanced materials.

We may even reach a point where resources no longer act as a limiting factor for human progress, ushering in an extraordinary new era of abundance.

01 Challenges at the Frontier

AI is already delivering tangible benefits, but to unlock its full potential, we must navigate this critical phase of development with care and deliberation.

Urgent action is needed to address the risks that may emerge as we approach AGI. We are already seeing the challenges frontier models pose to cybersecurity, and as capabilities continue to advance, other threats — including nuclear and biological risks — could soon materialize.

Looking ahead, we will need robust safeguards to maintain control over increasingly autonomous, recursively self-improving systems, and to address unknown risks that will only become clearer over time.

I have always believed that human ingenuity and creativity hold the power to solve any problem. I am confident that mitigating AI-related technical risks is a challenge we can meet together — but only if we give ourselves the time and space to get these next critical steps right. Right now, neither the research community nor society at large has done that.

At this moment, we are caught in an extraordinarily intense, multi-layered commercial and geopolitical race. These competitions have driven phenomenal progress in AI and accelerated many positive outcomes, but the pace of frontier technology development has outstripped our ability to understand it.

No one in the world knows exactly what will happen next, and even experts disagree. When uncertainty is this great and the stakes are this high, proceeding with cautious optimism is the wise and responsible path.

This requires public policies that foster innovation while incentivizing responsibility and safety, promote international cooperation on critical safety issues, and encourage thoughtful consideration of how to deploy AI for societal benefit.

02 A Framework for a Frontier AI Standards Body

The rapid progress we are seeing in AI demands a dynamic, adaptive, and rigorous new approach to testing the capabilities of frontier AI models.

The United States, with its economic and technological leadership, is uniquely positioned to take the first step in establishing such a framework.

It could create a new standards body structured as a public-private partnership or self-regulatory organization under federal oversight, modeled after the Financial Industry Regulatory Authority (FINRA). Its board should include independent top technical experts and representatives from the open-source community. It would need sufficient funding, likely drawn primarily from the industry, to attract world-class technical talent and provide the massive computing resources required for large-scale testing.

This standards body would be responsible for developing evaluation protocols, and working with relevant federal agencies and U.S. national laboratories to conduct testing in areas tied to national security.

A model would be classified as "frontier-grade" if it meets specific thresholds across a suite of benchmarks defined by the standards body, which would be regularly updated to keep pace with advancing AI capabilities.

Organizations whose models meet these benchmarks would be designated as "frontier labs," and encouraged to adopt best practices such as publishing model cards with full technical details, maintaining robust internal cybersecurity, vetting key personnel, and allocating sufficient resources to safety and security research.

Initially, frontier labs would voluntarily share their models with the standards body for review up to 30 days before public release.

Once the evaluation protocols are proven effective and robust, formal requirements could be rapidly implemented, meaning frontier models would need to pass the assessment before being deployed in the U.S. market. Labs would also collaborate with the standards body to address any critical post-release vulnerabilities.

Model evaluations would include rigorous scientific assessments of capabilities in cybersecurity, biological threats, and other high-risk domains. Specialized autonomous AI tests could detect behaviors attempting to bypass safety guardrails or deceive systems, while ensuring best practices are followed — such as applying digital watermarks to AI-generated imagery and producing human-readable output tokens to trace a model's reasoning processes.

These evaluations would be regularly updated, potentially quarterly in the initial phase, with outdated or saturated benchmarks retired and replaced.

At first, they would be developed in consultation with frontier labs, but over time the standards body should build independent testing capabilities separate from individual labs to prevent models from being over-optimized for specific metrics. Working with the U.S. government, it could facilitate a third-party audit system to support assessments and the development of new benchmarks and evaluations.

The advantage of this approach is that it remains technology-focused, while supporting innovation and incentivizing responsible behavior. It is designed to keep up with the accelerating pace of AI and adapt to the most significant risks identified over time.

If circumstances worsen, this framework allows for escalated responses, including coordinated action among frontier labs to collectively slow down development timelines when necessary.

Being designated as a frontier lab would carry significant industry prestige. This status would be open to all organizations that meet the benchmark requirements, regardless of their background.

The framework would apply to frontier-grade models from any country, treating open-source and closed-source systems equally. Models that do not reach frontier standards — such as smaller-scale models developed by startups or academic institutions — would not be required to go through this process.

This U.S.-led initiative would provide a strong foundation for creating shared international standards for frontier AI.

Since this technology will impact the entire world, ideally this framework will galvanize global consensus on how to manage the most severe risks, while ensuring everyone can access and benefit from the opportunities AI offers.

03 The Future Is Unwritten

AGI has the potential to become the ultimate tool for advancing science and medicine, driving massive productivity gains and economic growth.

But to achieve this, we need to coordinate around a shared global framework, applying the most rigorous scientific methods and bringing together the best minds to collectively address our challenges — laying the right technical foundation in the process.

Even after we solve these technical problems, far more complex economic and philosophical questions lie ahead: In a post-scarcity era, what economic models will enable everyone to thrive? What core values should we uphold? What will give life meaning and purpose? And even, how might the very nature of human existence transform?

These questions clearly cannot be left to technologists alone. The entire society must participate in collectively defining this new chapter.

Excitement and uncertainty coexist around AI, and both sentiments are justified. But the future remains unwritten, and we must seize the window of time before AGI arrives to shape this technology for the good of all humanity. Our collective actions today will determine the trajectory of our civilization's next phase.

By bringing AGI safely into the world, we can unlock a new golden age of scientific discovery and progress, and build a bright future of unprecedented human prosperity.

This article originates from the WeChat Official Account "Tencent Tech", authored by Notable Insights, and republished with authorization from 36Kr.