In the last three years of writing and researching generative AI, I have seen many respected people predict that AGI is close. Most of those predictions come with vague timelines, dramatic warnings, or promises that everything is about to change.
But this one feels different because it comes from Demis Hassabis, someone who has spent most of his life trying to build artificial general intelligence. He co-founded DeepMind, led the development of systems such as AlphaGo and AlphaFold, and now runs one of the few labs with the talent, compute, and resources needed to seriously push the frontier.
So when he says AGI is probably only a few years away, it is difficult to treat it as another prediction on social media.
In a new essay titled A Framework for Frontier AI and the Dawning of a New Age, Hassabis argues that humanity is standing in the “foothills of the singularity.” He believes future generations may look back on this period as the beginning of a completely new age.
He does not compare AGI to the internet, smartphones, or another major software platform. He compares it to electricity and fire.
I know. It sounds poetic, but it is also surprisingly close to what modern AI is. We take silicon, electricity, enormous amounts of data, and carefully designed software, then turn them into systems that can reason, write code, discover patterns, and increasingly act on their own.
Demis is surprisingly optimistic about what comes next
The way Demis describes it, he is very optimistic about the benefits of generative AI. He believes AGI could accelerate drug discovery, help develop new clean-energy systems, create advanced materials, and dramatically speed up scientific research.
He even suggests that its impact could be ten times larger than the Industrial Revolution and happen ten times faster. That is an almost impossible scale to imagine.

Excerpt from Demis Hassabis’s A Framework for Frontier AI and the Dawning of a New Age
The Industrial Revolution changed where people lived, how they worked, how goods were produced, and how wealth was created. If AGI produces a bigger transformation in a fraction of the time, nearly every institution we rely on may have to adapt while the technology is still moving.
Demis also talks about a future where resources may no longer be the main limit on human progress. In the most optimistic version, AI helps us reach a form of abundance where energy, medicine, materials, and knowledge become much cheaper to produce.
This is not a doom-focused essay. He clearly believes AGI could lead to a golden age of scientific progress and human flourishing, but his optimism comes with a warning.
The race to build these systems is moving much faster than our ability to understand or govern them.
The race is becoming the problem
Right now, frontier AI is shaped by two kinds of competition.
- The first is commercial. Google, OpenAI, Anthropic, Meta, xAI, and other labs are trying to release more capable models, win users, attract developers, and secure enterprise customers.
- The second is geopolitical. The United States and China both see advanced AI as a strategic technology that could shape national security, economic power, and military advantage.
Competition has clear benefits. It pushes models forward, lowers prices, improves tools, and forces companies to move quickly.
But it also creates a dangerous incentive. Even if one lab believes it should slow down, it has to assume that its competitors may continue.
We are expecting individual companies to act carefully inside a system that punishes anyone who falls behind.

Commercial and geopolitical AI race
Demis argues that this race is pushing capabilities forward faster than researchers can evaluate them. Cybersecurity risks are already visible, while biological, nuclear, and other dangerous capabilities may become more serious as the models improve.
Agentic systems introduce another layer of risk. Future models may be able to take long sequences of actions, use tools, write and run code, copy themselves across systems, or improve parts of their own workflows.
We are still struggling to evaluate today’s coding agents properly. It is not hard to see why testing a recursively improving system would be much more difficult.
His proposed solution is a Frontier AI Standards Body
Demis does not stop at warning that AGI is close. He proposes a specific institution for testing the most capable models before they are released.
His idea is a U.S.-led Frontier AI Standards Body. It would operate as a public-private organization with federal oversight, independent technical experts, open-source representatives, and funding largely provided by the AI industry.
The closest comparison he gives is FINRA, the organization that oversees parts of the U.S. financial industry.
The body would define what counts as a Frontier Model based on capability thresholds. Any organization building models above those thresholds would be classified as a Frontier Lab.
Those labs would be expected to follow stronger safety and security practices. This would include better cybersecurity, more detailed model cards, internal personnel vetting, and dedicated safety research.
The most consequential part is the proposed review process. Initially, labs would voluntarily provide models for evaluation up to 30 days before release.
Once the process proved reliable, passing the assessment could become a requirement for deploying frontier models in the U.S. market.

Frontier AI Standards body
The tests would focus on areas such as cybersecurity, biological threats, deception, guardrail bypassing, and other dangerous agentic behavior. The benchmarks would be updated regularly as older evaluations become saturated.
Demis also wants the standards body to develop its own private tests over time. This is important because public benchmarks eventually leak into training data, get optimized against, or stop measuring meaningful progress.
In the most serious cases, the organization could coordinate a slowdown between frontier labs. That is a much bigger proposal than asking companies to publish better safety reports. It creates a path toward an institution that could delay or block the release of the most capable AI systems.
This is where the idea becomes controversial
The proposal sounds reasonable when framed as independent testing for dangerous models. The harder questions begin when we ask who controls the process.
A standards body funded by frontier AI companies could easily become shaped by those same companies. Google, OpenAI, Anthropic, Meta, and other large labs already have the money, policy teams, and infrastructure needed to participate in a complex regulatory system, while smaller companies may not.
Demis says startups, academic groups, and non-frontier models would be exempt. That helps, but someone still has to decide where the frontier begins.
A threshold based on model size or compute could miss smaller systems with dangerous capabilities. A threshold based on benchmark scores could be gamed, leaked, or designed in ways that favor certain architectures and labs.
This is why some of the public reaction focused on regulatory capture. A system designed to control frontier labs could also protect them from new competitors.

I share part of that concern. I do not think the answer is to let companies regulate themselves, but I also do not think handing full control to governments automatically makes the system safer.
A credible standards body would need independent leadership, transparent rules, clear appeals, strict conflict-of-interest policies, and enough technical talent to challenge the labs it oversees. Without those protections, it could become either too weak to matter or powerful enough to be abused.
The international problem is even harder
The largest weakness in a U.S.-led framework is that AGI development is not limited to the United States.
American labs could agree to stronger testing while foreign companies, governments, or covert projects continue building systems outside the framework. Advanced models can also be accessed through APIs, copied through leaked weights, or deployed across borders.
This is the concern behind many of the negative reactions to Demis’s proposal. What happens if democratic countries slow down while authoritarian states or hostile groups keep moving?
Demis treats the U.S. body as a starting point. His longer-term goal is a shared international framework for the most serious risks.
That is probably the correct direction, but international cooperation is much harder than creating a domestic standards organization. Countries may agree that AGI is dangerous while still refusing to reveal their compute infrastructure, model capabilities, or military research.
The framework only works if competitors believe everyone else is following the same rules. Otherwise, every participant has an incentive to quietly continue preparing in case the agreement falls apart.
AI 2040 arrives at a similar conclusion
This is where the timing becomes interesting.
The AI 2040 report appeared only days before Demis published his framework. The two proposals are different, but they start from the same basic concern.

Both believe that advanced AI may arrive before governments and institutions are ready. Both argue that the current race is moving too quickly, and both believe humanity needs more time to prepare.
The difference is how far they are willing to go.
Demis wants a standards body that tests frontier models and can gradually increase restrictions as the risks become clearer. AI 2040 proposes a much broader international slowdown, mainly coordinated between the United States and China.
Its preferred scenario would delay superintelligence until around 2040. The goal is to buy time for alignment research, governance, transparency, verification, and a wider distribution of advanced AI capabilities.
AI 2040 also focuses heavily on tracking compute. Major training clusters would be declared and monitored so countries could verify that no one was secretly running a frontier training project outside the agreement.
That is far more aggressive than Demis’s proposal. Demis is trying to build a gate around the frontier, while AI 2040 is trying to slow down the entire road leading to it.
The similarities are more important than the timing
I do not think the publication dates mean Demis copied AI 2040. His proposal is too developed to have been written in a few days.
The more important point is that two very different groups reached a similar conclusion. One proposal came from an independent organization focused on long-term AI risk, while the other came from the CEO of Google DeepMind, one of the companies actively racing toward AGI.
Both are effectively saying that the current incentives are not enough. They also share the belief that the future should not be decided by whichever company or country reaches superintelligence first.
That idea goes much further than safety testing. It raises questions about who owns the models, who controls the infrastructure, who benefits from the productivity gains, and whether ordinary people have any meaningful role in shaping the transition.

Side-by-side comparison of Demis Hassabis’s framework and AI 2040
I still wonder what made him write this now
One of the most common reactions to the post was simple.
What did Demis see?
It is tempting to imagine that Google has an internal Gemini model showing capabilities the public has never seen. There is no evidence of that yet, and the less dramatic explanation is probably enough.
Demis can see internal evaluations, scaling trends, security tests, agent behavior, and research progress that most people will not encounter until much later. He also understands the pressure inside a frontier lab better than almost anyone.
You do not need a secret AGI system to become worried. You only need to believe that current trends will continue while safety research, regulation, and international coordination move at their usual pace.
Still, the tone of the essay is hard to ignore. This does not read like a distant thought experiment from someone speculating about what might happen decades from now.
We need a believable path to the optimistic future
What I like about Demis’s essay is that he does not end with an apocalypse.
He asks what new economic systems may be needed in a post-scarcity world. He also asks how people will find meaning and purpose when intelligence and productivity are no longer scarce in the same way.
Those questions should not be answered only by AI companies.
The same is true for AI 2040. Its proposal may be politically unrealistic, technically incomplete, or too aggressive in some areas, but it at least tries to describe a future where humanity does not simply lose control or hand permanent power to the first lab that reaches superintelligence.
We need more positive visions for the future, but optimism alone is not enough. A serious plan has to explain how power will be distributed, how dangerous capabilities will be tested, how countries will cooperate, and how the public will have a voice in decisions that may affect every part of society.
Demis believes we still have a small window to shape that outcome. He may be too early, or he may be much later than we think.
Either way, if AGI really is only a few years away, the institutions that govern it cannot be designed after it arrives.
Sources
- A Framework for Frontier AI and the Dawning of a New Agex.com
- FINRAfinra.org
- public reactionx.com
- AI 2040 reportai-2040.com
