Great Minds

In 1938, after a mid-air collision over Wyoming killed everyone aboard both aircraft, the American airline industry did something that looked like commercial suicide and turned out to be the opposite. It asked to be regulated. The carriers had worked out that nobody would buy a ticket for a machine the public did not trust, and that trust was not something any single airline could manufacture on its own. On Tuesday this week, in an essay published to X, Demis Hassabis made the 2026 version of that argument. What happened next is the part worth paying attention to: within about six hours, four of his fiercest competitors publicly agreed with him.

The big conversation: everyone suddenly wants a referee

Hassabis's proposal is specific, which is what makes it interesting. He wants a Frontier AI Standards Body, modelled explicitly on FINRA, the industry-funded self-regulator that polices American stockbrokers. Frontier labs would voluntarily hand models over for review up to 30 days before release, and once the testing regime proved itself, that voluntary step would become a requirement to sell into the US market. Underneath it sits a claim he does not hedge: AGI is probably a few short years away, and advances on the frontier are outpacing our understanding of the technology. This is not a man talking down his own field. The same essay calls the technology closer to fire or electricity than to the internet, and includes the best sentence anyone wrote this week about what has actually been built.

Then the endorsements landed, and they landed fast. Sundar Pichai, whose company employs Hassabis, was the easy one: well said Demis, worth reading. Mustafa Suleyman went further, calling it important and adding, with the faint satisfaction of a man producing a receipt, that he had proposed an "IPCC for AI" with Eric Schmidt back in 2023. Satya Nadella framed it as systems risk, which is how Microsoft thinks about everything: the goal, he wrote, is a frontier ecosystem that promotes innovation and choice, while avoiding any one model drop that breaks the world. Even Sam Altman, who has spent two years being cast as the accelerationist in this drama, called it a thoughtful proposal.

Industries do not agree on rules because they have had a moral awakening. They agree when the rules become worth more than the thing the rules constrain. And that is exactly what the rest of the week was quietly demonstrating.

Where they disagree: FINRA or the FAA

The consensus is a mile wide and about an inch deep. Hassabis is proposing a self-regulatory organisation: industry-funded, industry-staffed at first, voluntary before it is binding. There is a competing proposal on the table, and its author said nothing at all this week. Five weeks ago, in an essay called Policy on the AI Exponential, Dario Amodei wrote that transparency was no longer enough and that frontier models should face mandatory third-party testing for cyber, bio, and autonomy risks, with the power to block or revoke a deployment outright. That is not a FINRA. That is closer to an FAA: a public agency that can ground the fleet. Amodei has been silent on his feed since 10 June and published nothing in this window, so he has not answered Hassabis. The gap between a body the labs fund and a body that can stop them is the entire argument, and it is currently being had by one side.

Chamath Palihapitiya said the quiet part out loud. He endorsed the framework, but pointedly, as the least-bad option: Hassabis is quite well reasoned especially compared to what he called the Pull Up The Ladder Framework from the other guys. By Sunday he had dropped the diplomacy: enough of the attempted reg capture, the genie is already out of the bottle, so let's just get on with it and compete on the field. Altman drew a similar line from the opposite direction, writing that OpenAI wants to do the right thing but does not want to scare people into doing our thing. Elon Musk, historically the loudest voice for slowing down, did not weigh in on his feed at all. He spent the week selling Grok Build and arguing that Grok has the best value for coding.

The quiet signals: the price of thinking fell through the floor

Here is the number that explains the sudden enthusiasm for rules. Palihapitiya, arguing that America must not close the door on open weights, pointed out that the same frontier-class intelligence can now be bought for either fifty cents or fifty-six dollars per million tokens, depending on whose you use. His framing was geopolitical: letting our adversaries attack us for $0.50 per 1MM tokens while we spend $26-56 per 1MM tokens to defend ourselves is equally ruinous, it's the Cold War Soviet collapse in reverse. Set that beside Altman boasting that GPT-5.6 Sol now costs a fraction of its rivals, and Musk competing on tokens-per-task, and you can see the shape of it. Raw intelligence is deflating faster than almost any input in economic history. When the product commoditises, the moat moves. It moves to distribution, to trust, and to the licence to operate. A standards body is not only a safety measure. It is also the last durable barrier to entry in a market where the thing being sold gets cheaper every quarter.

Jensen Huang, who does not post and therefore never argues, spent the week in Tokyo doing the most concrete thing anyone did: launching what NVIDIA calls the world's first national AI infrastructure with the Japanese government, and telling the country that it invented modern manufacturing and is now building the AI factories that will power the next industrial revolution. To a room of developers he put it in a register a small business can actually use: forty years ago was the beginning of the PC revolution, now, forty years later, instead of a personal computer, you can now have your own personal AI. That is the same democratisation story as the collapsing token price, told in hardware instead of dollars. The PC did not win because it was the best computer. It won because it was the first one an ordinary person could own outright.

Two more signals worth keeping. Palihapitiya's firm 8090 published a case study in which it decoded a fifty-year-old, 18 million line COBOL and Assembly repository governing billions in US healthcare payments, documenting more than 100,000 business rules in plain English, each traced to its source line, so that the people who own Medicare policy can finally read how their own system behaves. And Suleyman put out a Nature Health paper finding, across 1.7 million conversations in 109 countries, that a chatbot is most valuable to people with low confidence in their health systems. Both point the same way: the value is showing up where institutional knowledge was locked up or unavailable. Meanwhile Andrej Karpathy, two months into Anthropic, has still not posted a first-person word since June. The field's best explainer has gone quiet exactly when the arguments got interesting.

How a smaller business should read these minds

Strip out the geopolitics and this week hands an Australian small business two facts that point in opposite directions, which is why it is worth slowing down over. First, the intelligence you are buying is getting dramatically cheaper, and the gap between the expensive option and the cheap one is now roughly a hundred times for broadly comparable work. Anyone who signed a long, expensive AI contract last year should be re-reading it. Second, the industry is building a rulebook, and rulebooks always flow downhill. They start with the labs, then reach the vendors, then arrive in your quotes, your privacy policy and your customer disclosures. The EU's transparency rules begin biting from early August, and while that is a European law, it will shape what every serious vendor offers globally, the way European privacy law quietly rewrote Australian websites a decade ago. The businesses that get hurt are the ones who cannot say where their AI came from or where their customer data went. The ones who can will find that answer becoming a selling point, because when everyone has the same cheap intelligence, the only scarce thing left is trust.

That is the work we do at NextAura. We read the week so you do not have to, then help Australian small businesses act on the part that actually reaches them: paying the right price for intelligence that keeps getting cheaper, getting the messy knowledge out of people's heads and into systems they own, and being able to answer plainly where the AI in their business came from. The people building this technology spent the week asking to be held to a standard. That instinct is worth borrowing, because in a market where the intelligence is nearly free, being the business a customer can trust is the last thing that is genuinely yours. If you want a hand turning this week's signal into something running by next week, get in touch.