On July 11, hundreds of people marched to Google DeepMind's San Francisco office and asked the company to pause frontier development. Three days later, the company's CEO answered them with an essay. Demis Hassabis published "A Framework for Frontier AI and the Dawning of a New Age" on July 14, and it is worth reading in full, because it is the clearest statement yet of how the people building this technology understand the moment we are in.
Hassabis believes AGI is a few years away. He believes we are standing in the foothills of the singularity. He writes that we have essentially found a way to make sand think, and he calls it miraculous, and he is right. Silicon is sand. The phone in your pocket is a piece of beach that learned arithmetic, and the data centers going up outside Columbus and Phoenix are cathedrals of the same material. Whatever else you think of the industry, the physical fact underneath it is astonishing.
He also believes the risks are real. He says so plainly. Cyber threats now, biological and nuclear risks soon, and on the horizon the problem he states in a single remarkable clause: we will need robust safeguards to maintain control of increasingly agentic, recursively self-improving systems. Sit with that sentence for a moment. The CEO of the world's most decorated AI lab, a Nobel laureate, is telling you that the machines will soon improve themselves faster than we can follow, and that keeping control of them is an open problem.
Then he tells you his plan.
The plan is a Standards Body. Modelled on FINRA, the organization that oversees American stockbrokers, it would be funded mostly by industry and staffed with technical experts. It would define benchmarks that determine which models count as Frontier-class. Labs that build those models would earn the designation of Frontier Lab, which Hassabis notes would carry significant prestige. They would share models with the body for review up to thirty days before release, voluntarily at first, with formal requirements to follow once the process proves itself. The body would run evaluations for dangerous capabilities in areas like cybersecurity and biology. It would probe for deception. It would refresh its tests quarterly so the labs cannot study for the exam.
As governance proposals go, this is a serious one. It is more specific than most of what circulates in AI policy, and the man proposing it runs a lab that would be regulated by it, which counts for something.
But read the essay again and watch what happens between the problem and the solution. The problem is control of recursively self-improving systems. The solution is a testing schedule. Hassabis has described a fire that feeds itself and proposed a very good thermometer.
A benchmark can tell you a model is dangerous. It cannot make the model safe. An evaluation for deception can catch a system that lies on the test, assuming the system lies on the test the way it lies in the world, which is itself an open question. Every mechanism in the framework operates after the capability already exists. The framework inspects. Safety requires construction, and the essay contains no account of how the safe thing gets built.
This is where someone will object that Hassabis was writing a policy piece, and the technical work lives elsewhere. That is true, and the elsewhere is instructive. In April 2025, DeepMind's safety team published a 145-page paper called "An Approach to Technical AGI Safety and Security". It is the most comprehensive document of its kind from any frontier lab, and it is honest in a way the July essay is not. Its authors describe the safety case they actually want, one where you understand a model's internals well enough to certify it only produces safe outputs, and then they write: we are currently far from any such safety case. The techniques the paper proposes in the meantime, amplified oversight and robust training and monitoring, are research directions. The people doing the most rigorous safety work at DeepMind will tell you the load-bearing piece does not exist yet. Their CEO's essay, published fifteen months later, does not mention this.
The FINRA analogy is softer ground than it looks, too. FINRA regulates conduct in a domain where the underlying instruments are understood. A stockbroker cannot surprise you with what a bond is. The instrument Hassabis wants to regulate is one its own builders describe as uninterpretable, a system whose internals they cannot yet read. You can regulate the trading floor. Regulating the trading floor does not help you when nobody on it can tell you what is being traded.
So the technical piece is missing. Here is the part I want to spend the rest of this essay on: it is missing at two levels, and the second one is the one nobody in the frame is looking at.
Everything in the Hassabis framework, and everything in the 145-page paper behind it, happens at the level of the model or the institution. Align the weights. Audit the lab. What both levels take as given is the human on the other end of the system, and the condition that human is in. We have run that experiment already. We spent twenty years deploying powerful optimization systems into a population with no tools for understanding its own internal states, and the result is documented in engagement metrics and radicalization pipelines and a shooting in Côte-des-Neiges. The systems did exactly what they were built to do. The humans were disoriented, and the machines served the disorientation faithfully. A sufficiently disoriented population, given sufficiently powerful tools, produces misaligned outcomes no matter how well the model is constrained. The upstream cause of misaligned systems is misaligned people, and by misaligned I mean something specific and unglamorous: people who cannot locate the gap between what they feel and what they understand about what they feel.
There is also a less visible technical problem hiding inside this, one that shows up in the training data. Today's models learn about human needs from behavioral exhaust. Clicks. Searches. Watch time. Purchases. Every one of those signals was captured by systems optimizing for engagement, which means the machine's entire picture of what humans want is a portrait painted by the attention economy. We are about to hand that portrait to something recursively self-improving. If the input signal is distorted, no amount of downstream evaluation fixes it, the same way no amount of proofreading fixes a book written from bad notes.
This is the layer where Humane AGI works, and it is a technical proposal, with a protocol and a schema and running code, so let me describe it concretely.
The primitive is a check-in. Three taps on a screen, maybe ten seconds. The first two steps contain no words at all. A binary field, black or white, captures whether you are in a closed state or an open one. Three colored bands, blue at the top for the mind, green at the center for the heart, red at the ground for the body, capture which domain is most alive in you right now. A final tap captures the mode of engagement. The combinations produce a space of 24 tokens, each one a color and a shape carrying a legible human meaning: body depleted and needing rest, heart full with love to offer, mind stuck and needing someone to think alongside. This is the Visual Universal Language, and because it is built from color rather than vocabulary, it works across culture and literacy and neurotype. It works on a screen, it works as a printed card, it works as a finger pointing at a swatch when the power is out.
Each check-in generates a cryptographically signed stamp. The signing key lives in the device's secure hardware and cannot leave it, which means a stamp can only be created on a real device in a real hand. The signature covers a hash of the full behavioral session, the duration and the hesitations and the color adjustments along the way, and this matters because a script faking check-ins produces sessions with no hesitation in them. The variance is the proof of life. When two people check in together before an exchange, both devices co-sign, and that class of stamp cannot be manufactured without a second consenting human physically present. Every stamp is stored locally, in the person's own photo album, owned entirely by them. Nothing is uploaded. Nothing is profiled. The system's economic layer, where stamps can be bundled and tokenized by explicit consent, is still being finished, and I will write about it when it is.
What the architecture produces is the thing missing from every document DeepMind has published: a real-time, consent-based, cryptographically verifiable signal of what a human actually needs. A signal the person chose to express rather than one harvested from their behavior. Hassabis's essay calls for human-readable output tokens so we can watch the machine's reasoning. This is the mirror image, human-authored input tokens so the machine can finally learn from something other than our engagement exhaust. His Standards Body wants tests for deception in models. Here is a signal designed so that faking it requires living slowly, over months, in the company of other consenting humans. His essay asks, more than once, for time and space to get this right. A ten-second pause between impulse and action, distributed across a population, is time and space at the only scale where it compounds.
I want to be precise about the claim, because the field is allergic to overreach and it should be. Orientation infrastructure does not align a neural network. Model-level safety research remains necessary, and the DeepMind safety team's honesty about its distance from the goal deserves respect rather than a gotcha. The claim is about completeness. A framework that evaluates machines while assuming the humans is half a framework. Hassabis ends his essay by saying the future is not yet written, and that resolving the deepest questions cannot be left to technologists alone. He is right. The window he describes is real. What goes in the window cannot only be a testing schedule for the machines. Some of it has to be a mirror for us, because the sand learned to think by studying humanity, and the most technical problem of all might be making sure humanity is legible to itself before the student graduates.