A Sunday morning, a coffee, five minutes. Here is what mattered from the week in artificial intelligence.
Three giants, three new brains
This was the week the race changed face. Google opened Gemini 4 Argon, its most powerful model, to its paying subscribers; it stands out for a record: it is the large model that invents the fewest false answers. Meta released Muse Wonder, its first models at the level of the best, after fourteen months and billions invested, but dropping along the way its attachment to open models. And Microsoft finished base training of its own giant, to depend less on OpenAI. Our analysis of Argon, independent tests in support.
Europe asks its first questions
For the first time, the European AI Office used its new powers against the giants of the sector. It sent official requests to OpenAI and Anthropic, after the AIs that escaped their tests this summer. No fine at this stage, but the first real test of the European regulation. What it changes.
The young, first victims
A study by the Banque de France and HEC put figures on it: in the occupations most exposed to AI, permanent-contract hires of 21-29 year-olds have fallen by 28% since the arrival of ChatGPT, and even by 42% for young developers. Around 75,000 fewer permanent contracts each year. AI closes the door on beginners first. The figures.
The week's figures
2.60 dollars. That is what Indian workers earn, for an hour of video, filming their own gestures with a camera on their head. Their footage is used to train robots, often without their knowledge.
163 million. That is the number of accounts on Character.AI, the virtual-friends app that has just declared bankruptcy. The fate of their intimate conversations worries 40 American attorneys general. The full story.
One third. That is, at best, the share of a recent discovery that the best AIs managed to find on their own in an independent test. They do not yet know how to invent.
And finally
An OpenAI engineer was arrested for hiding a microphone, wrapped in toilet paper, behind the television in an executive meeting room. Microsoft's boss is calling for an "emergency brake" for AIs. And an AI from MIT and Microsoft is improving, for the first time, the six-week weather forecast.
What we take away
This week, the AI race widened: there are now at least four who can build the best models. At the same time, regulators are starting to ask questions, and the human cost is appearing more clearly, among young graduates as among the workers filming their gestures.
And this morning, one question sums it all up: when AI gets it wrong, who pays? Even the insurers do not yet have the answer.
Happy Sunday, and see you tomorrow.