AI Watch, Wednesday 23 September 2026Anthropic and OpenAI ship on the same day

The country, which holds the presidency of the Council of the European Union, has proposed new changes to the "digital omnibus", the text meant to simplify European digital law, Next reports. The rewrite would make it "legitimate" to use personal data to train models, as well as to place cookies to measure advertising effectiveness, without going back through people's consent. For a French organisation, the legal basis of an AI project fed on customer data could shift before the CNIL (France's data protection authority) even rules on it.

The Journal du Net publishes a grid of average pay in the sector: the prompt engineer role (designing the instructions that drive a model) brings up the rear at €49,700 gross a year, while the most sought-after profiles reach €115,800. The gap sketches the hierarchy taking shape between those who use the tools and those who design and integrate them.

The manufacturer says it flew two artificial intelligences aboard its fighter jet, one developed in-house, the other supplied with CortAIx, the AI arm of Thales, La Tribune reports. The test brings onboard AI into French combat aviation, on an aircraft that France exports widely.

The executive is presenting a scheme intended to accelerate the adoption of artificial intelligence in businesses and the public administration, Maddyness reports. The detail of the resources committed and the audiences targeted will tell whether the announcement translates into real skill-building on the ground.

The venture capital firm a16z is launching a programme aimed at recent secondary-school graduates and college dropouts, to spot and train young talent building AI products, Bloomberg reports. The "Horowitz Andreessen Academy" starts with $42 million and ten partners: Anduril, Anthropic, Coinbase, Google, Meta, Nvidia, OpenAI, Palantir, Replit and Stripe, The Verge details. The programme, with no homework, is meant to be a direct on-ramp to a Silicon Valley startup. The initiative raises a fundamental question about education: when employers become trainers and bypass the diploma, the value of a course is measured at hiring, not by the title. It deserves the attention of every player in French vocational training.

The company says it disrupted a service that automated mass cyberattacks and is thought to have compromised around 12,000 accounts, Ars Technica and Microsoft report. EvilTokens supplied a full chain, from phishing to credential theft, making large-scale breaches faster and easier for low-skilled actors. The case shows a shift on the attacker side: AI lowers the cost and the technical bar to launch a campaign as much as it equips the defence. For a French company, the safeguard rests on basic reflexes held by trained teams: two-factor authentication, verifying requests, raising the alarm at the slightest abnormal signal.

In an op-ed for the MIT Technology Review, Timnit Gebru, director of the Distributed AI Research Institute, and Emily Bender, a linguist at the University of Washington, call for caution in the face of a season of spectacular announcements. They recall the recent run of demonstrations, from security flaws paraded by the labs to claimed mathematical results, and invite readers to tell real progress from promotional narrative. Their argument: fear of an all-powerful AI and the promise of a miraculous AI feed the same hype and serve the same commercial interests. The warning lands on the day two labs cut their prices after calling for a slowdown.

At its annual Apsara conference in Hangzhou, the group announced it is working on a model that could reach 10 trillion parameters, an order of magnitude above today's largest public models, the South China Morning Post reports. The group's chairman, Joe Tsai, restated the ambition of a superintelligence and an investment across the full stack, from chips to models. The announcement comes with the unveiling of an in-house accelerator billed as the most powerful designed in China, while the US restrictions on semiconductors remain in force. The parameter count is still a promise, but it sets the course Beijing wants to display against the US labs.

The group's cloud arm will launch centres in the Netherlands, Turkey and Finland over the next twelve months, the first in the Netherlands as early as October, the South China Morning Post reports. The expansion is part of a global rollout of 20 gigawatts of capacity and aims to sell a full-stack offering, from chips to cloud to multimodal models. For a European company, a Chinese provider joins the short list of large-scale AI compute hosts, alongside the Americans. The choice will raise questions of sovereignty and GDPR compliance that legal departments will have to settle.