Transcription

• Speaker #0

  Next time, the ones we let in. That was last week's promise. This week, we saw who walked in on Wednesday in New York, in the chamber of the United Nations Security Council. Four guests came to brief governments on the risks of AI, and three of them run the very companies that governments are trying to rein in. They were invited to talk about the dangers. of the technology they are building themselves. The day before, the American president had rejected any global rule, and on that same Tuesday, two of those labs were fighting a price war. Four days earlier, a federal court had been asked whether they were allowed to agree with each other at all. So who writes the rule, and who holds the door? Welcome to the week of September 21st, 2026, Synth AI Cast, the AI that listens for you. Episode 26, the 4th of Season 2. This week, we are not asking who is right. We are asking who benefits from which rule, and why now. First signal. On Monday the 21st, OpenAI published a proposal. Common technical standards for the most advanced AI, with the United States in the lead. The text calls for shared measurements, one common way to evaluate what a model can do, and common rules for reporting incidents. But it explicitly rules out two things, licenses and approval before a model goes to market. Each government would then decide how to use the standards. Keep that distinction in mind. It will matter later. A standard tells you how to measure. A license tells you whether you are allowed to sell. On Tuesday, at the General Assembly, Donald Trump took the podium. He rejected what he called a globalist scheme to control artificial intelligence and promised to encourage the technology, not to rein it in. On Wednesday, the Security Council you chaired this month by France, held its first ever high-level meeting, devoted to the security risks posed by AI. Sam Altman, from OpenAI, was in the room. Dario Amadei, from Anthropic, spoke remotely. And Yoshua Bengio, the Turing Award winner, who co-chairs the UN's Scientific Panel on AI, described the dangers as real and imminent. A quick aside. As always, when that name comes up, Claude, Anthropix Model, produces this episode. So we hold it to exactly the same standard as everyone else. Amadé said that, if managed poorly, AI could even be a risk to humanity as a whole. And he argued for coordination between countries, including between democracies and authoritarian governments. Now put those three days side by side. The labs are happy to have rules. Rules for measuring, not permits, and the country where they are based. Wants no rule at all that comes from the whole world. Second signal. And this is where the picture flips. On Friday the 18th, four paying subscribers of ChatGPT, Claude, Grok, or Gemini filed a lawsuit. in a federal court in California, and asked to represent every paying subscriber in the country. They are suing Anthropic, OpenAI, Google, and SpaceX AI, the company behind Grok, and their accusation is unusual. Not that the companies agreed to speed up, but that they agreed to slow down. The complaint goes back to September the 12th. That day, Amadé published an essay, calling on the whole industry to slow down together, and, according to the complaint, his rivals endorsed it, publicly, that very same day. They invoke the Sherman Act, America's antitrust law, which is more than a century old, and their argument fits in a single sentence. Antitrust law does not allow competitors. to decide among themselves that competition is too dangerous. The harm, they say, is a subscription that is worth less, because the products would improve more slowly. A word of caution. This is a lawsuit, not a ruling, and the companies had not responded to the press that weekend. Now, look at Tuesday the 22nd. Anthropic released Opus 5.5, 20% cheaper at list price than the model before it. The same day, OpenAI released GPT-6 Sol, at half the price of the model it replaces. For a developer, this is very concrete. And it came just four days after the lawsuit was filed. A $100 bill last week now costs $80. With Anthropic. And $50 with OpenAI at posted prices. A secret pack to slow down? On pricing, at least, nobody is slowing down. The French podcast Silicon Carnet summed it up on Friday. One foot on the brake and one on the accelerator. Perhaps more a communication war than a technology war, they suggest. That proves nothing, one way or the other. It is the backdrop. And the backdrop matters for what comes next. Third signal. And this one needs to be handled with care. What follows is not confirmed. None of the three labs has announced it. Several outlets attribute it to the information. a subscription news site covering the tech industry, and we have not been able to read the original article. According to those reports, Google, OpenAI, and Anthropit would be preparing their own standards body, with no government oversight, modeled on the way Wall Street polices itself. There is a detail worth knowing here. On Wall Street, the industry's own watchdog. Is it self-supervised by a government regulator? Self-regulation there does not mean no state, with a launch mentioned for the end of this year, or the beginning of the next. On the agenda, still, according to those reports, outside testing of models before they are released, rules for reporting incidents, and criteria for choosing the auditors. If it turns out to be true, What should we watch? Just one thing. Who would be able to sanction whom? A body with no legal power cannot exclude anyone. To exclude, it would need the state. Let me take you back to 17th century France for a moment. When the French Academy was founded, it was given a mission. To give the French language fixed and certain rules. And to produce a dictionary. In 1674, the king granted it the exclusive right to print one. In 1685, it expelled one of its own members, a writer called Antoine Furetiere, for writing his own dictionary in spite of that privilege. The Academy's own dictionary finally came out in 1694, 60 years after the mission began, holding the monopoly on writing the rule. and being able to deliver it, are two very different things, and that monopoly was granted by the king. This week, the labs asked for standards, they turned down permits, and the power that could grant them a privilege, had just turned down any global rule. Four days, that is the gap, between a lawsuit for colluding, and a price war between the accused. The number proves nothing, but it puts everything in place. So each player has an interest in a different rule. The labs, in measuring rules they help to write. The White House, in no rule coming from abroad. And some subscribers, in competition that never stops. And while everyone fights over the pen, Silicon Carne had asked a simpler question. On Thursday. Will we even be able to check? That a model is not just pretending to obey? Last week, we asked who would get to look. This week, the ones we were supposed to be watching were invited in. And if the reports are confirmed, they would be the ones drawing the room. Three things to try this week. Five minutes. Read the summary of OpenAI's proposal. And note what it leaves out, not just what it puts forward. 30 minutes. Compare the price. of your favorite AI tool. Before and after this week's price cuts. One hour for one tool you use at work. Find out who checks its safety and who pays that checker. Is a referee chosen by the players still a referee? Next time, who pays the referee? See you next week.