September 2025

Reichstadter asked Anthropic to stop frontier AI development. Trazzi asked DeepMind’s chief executive to commit to releasing no further frontier models if the other leading labs also stopped. The protests took public concern about advanced AI from open letters to the companies’ doors.

Apertus treated openness as more than released weights: researchers and developers could examine how the system was built and adapt it under a permissive license.

The case showed how a language model could write malicious code on demand, and why a sample’s origin matters when judging a new threat.

The plans tied AI adoption to hiring and training, a field long served by established recruitment platforms.

The meeting joined federal agencies and technology companies in a push for AI literacy in schools and the workforce.

The suit followed similar cases brought by Disney and Universal earlier in the summer, making Warner Bros. Discovery the third major studio to sue.

The reported aim was to reduce OpenAI’s reliance on Nvidia’s processors for part of its computing needs.

The analysis made the handling of uncertainty part of how a model is judged. Improving a benchmark score and reducing the harm of a wrong answer can call for different incentives.

The claims concerned how the books were obtained, not whether training AI models on them was lawful.

The project was conceived as a public demonstration of generative tools in feature animation.

The expansion brought AI-generated search answers to speakers of five more languages.

The law made disclosure, incident reporting and whistleblower protection the core of California’s approach to frontier AI.

The orders gathered information under the FTC’s study authority, separate from any enforcement action.

The remarks placed adaptability, rather than any single technical skill, at the center of education for an AI economy.

The announcements placed generation tools inside a major platform’s publishing workflow, some available at launch and others to follow.

The peer-reviewed paper gave a fuller account of how a model released in January had been trained.

The result measured performance on demanding programming problems under contest conditions. It came from a research version of the model, not the one offered to users.

The design paired a small in-lens display with an input method that leaves the hands largely free.

Its developers described it as a proof of concept, not yet ready for clinical use; its estimates describe risk across populations, not individual diagnoses.

The publishers used the findings to press for stronger transparency rules, including under the EU’s AI Act.

The browser became a place to read, compare and act across the web with an assistant built in.

The partnership joined two long-standing competitors’ technologies in planned products for data centers and PCs.

The announcement tied staged investment to the construction of computing capacity.

The lab extended the institute’s language-model research into robotics.

The arrangement gave agencies a common route to an open-weight model family; each agency still decides which tasks it suits.

The round reflected investor interest in AI agents built for a single business function.

The release concentrated on reliability within an existing model family, refinements that matter in daily use.

The figures described an infrastructure program under development, dependent on construction, power supply and equipment arriving over time.

The estimate was a forecast, not a measurement, and it gave a public marker against which later progress can be compared.