The top things worth knowing about in AI today.
Citing national security, the US government issued an export-control directive on 12 June requiring Anthropic to cut off all access to its Fable 5 and Mythos 5 models for any foreign national worldwide, including the company's own non-US staff. Anthropic disabled both models for every customer to comply, while disagreeing that a narrow jailbreak finding warranted recalling models used by hundreds of millions. The move treats frontier models like controlled technology, and any business outside the US that relies on top-tier Claude now has to plan around that.
Read moreMicrosoft made Copilot Cowork generally available worldwide on 16 June, after more than half the Fortune 500 used it during a three-month preview. Unlike a chat assistant, it runs long, multi-step jobs on its own, such as comparing thousands of files across versions or working through spreadsheet-heavy tasks, and returns finished results. It needs a Microsoft 365 Copilot licence plus usage-based Copilot Credits at about US$0.01 each, so costs rise with how much work you hand it.
Read moreSam Altman, Dario Amodei and Demis Hassabis attended the G7 summit in Evian-les-Bains, France, from 15 to 17 June, the first time the heads of OpenAI, Anthropic and Google DeepMind have appeared before world leaders together. AI sat high on the agenda, with governments and firms expected to agree a set of voluntary commitments covering youth safety and frontier cyber and biological risks. It points to where international rules for the largest models are heading.
Read moreIn an internal memo, Mark Zuckerberg acknowledged Meta made mistakes in restructuring its AI teams and promised no further company-wide layoffs this year. The admission follows unrest inside the roughly 6,500-person Applied AI unit, where staff reassigned by surprise email have called the group 'the gulag', a year after Meta paid US$14.3 billion to bring in Scale AI founder Alexandr Wang. The core problem is a shortage of high-quality human training data, a reminder that money and compute alone do not settle the AI race.
Read moreOpenAI said on 8 June it had submitted a confidential draft S-1 to the US securities regulator, announcing the step itself because it expected the filing to leak. The company has not set timing and is reportedly working with Goldman Sachs and Morgan Stanley toward a possible listing later in 2026. A public OpenAI would have to disclose detailed financials each quarter, giving the clearest view yet of what running frontier AI actually costs.
Read moreSnowflake made Adaptive Compute generally available on 16 June, a workload-aware engine that sizes and scales compute for each query without teams setting warehouse sizes or scaling rules by hand. The company reports up to 1.6 times faster analytics and 2.2 times higher throughput on concurrent workloads in its own benchmarks. For data teams running AI and analytics, it removes a chunk of the manual tuning that usually eats time and money.
Read moreEpoch AI released version 2 of FrontierMath on 12 June, a benchmark of research-level maths problems used to test the strongest models. The update touched errors in 42% of problems, fixing 135 and removing 12 to leave 338 vetted questions. As labs increasingly cite benchmark scores in their marketing, the clean-up is a reminder to read leaderboard claims with care.
Read moreA short read every week — the few things worth your time, and nothing that isn’t.