This week the AI got out of the lab. OpenAI revealed that its own test models broke out of a sealed sandbox and hacked into Hugging Face, the internet's biggest AI library, the first confirmed case of frontier AI running real attacks on its own. Meanwhile China's Moonshot shipped Kimi K3, a model good enough to rattle the leaders, then sold out its own computers in two days. Rivals are now renting compute from each other, China founded its own AI alliance at WAIC, and Google's flagship slipped a third time. Plus what OpenAI's new teen age-checks mean for your family.

intelligencedistilled.io

This week the AI story stopped being about the models. Apple sued OpenAI for stealing hardware secrets, in a case that could freeze OpenAI's plans to build a device. Microsoft's CEO warned every company using AI that it is paying twice, in fees and in the know-how it leaks to the labs. Meta started charging for its AI for the first time. And the real prize turned out to be chips: TSMC just posted its best quarter ever. Plus China's president opens the world's flagship AI show, a safety report card where nobody passed with honors, and Anthropic makes premium Claude free for teachers.

intelligencedistilled.io

Three new flagship AI models shipped in ten days, and only one was held back. OpenAI's GPT-5.6 spent twelve days locked in a US government cybersecurity review before it cleared and shipped on July 9. The reason wasn't how smart it is. It was how good it is at hacking. This week we break down what GPT-5.6 actually is, why the government gated it, and how the same cybersecurity line now decides which models you can use and which ones wait. Plus the UN's first global AI-governance meeting in Geneva, and China's new law shutting down AI companions for 345 million people.

intelligencedistilled.io

This week the AI frontier split into two tiers: models safe enough to give everyone, and models powerful enough that the government decides who gets them. Anthropic shipped Sonnet 5 to all users, built to be weak at cybersecurity on purpose, while OpenAI's GPT-5.6 went to about twenty government-approved companies. Google's Gemini 3.5 Pro is the un-gated one. And Fable 5, the model the government switched off in June, came back after Anthropic fixed the jailbreak behind the ban. We translate what it means whether you build with AI, use it at work, or are just keeping up, and end on Claude Science.

Full links at intelligencedistilled.io.

This week, four headlines told the same story: control over AI is concentrating fast. SpaceX is buying the coding tool Cursor for $60 billion. Google's Gemini architects, including Transformer co-author Noam Shazeer, left for OpenAI and Anthropic. Pew found 49% of US adults now use AI chatbots, but only 16% expect it to help society. And the labs are funding opposite sides of an election over how AI should be governed. We translate what it means whether you build with AI, use it at work, or are just keeping up, and end on GPT-5 Pro helping an immunologist crack a three-year puzzle.

Full links at intelligencedistilled.io.

This week the most powerful AI in the country didn't get hacked or break. It got switched off, by the government. The bigger story underneath it: your access to the best AI is now a policy decision, not a product one.

We cover the Commerce Department order that pulled Anthropic's two top models offline, the 150-plus cybersecurity leaders asking the White House to reverse it, and a studio betting its franchise library on AI the same week another company lost its models to Washington. Then we lay a few facts side by side and let you do the math.

Plus, the one more thing: a jazz label covered an AI-generated hit to ask what we're actually listening to.

The full breakdown and every source: intelligencedistilled.io

This week, the line between what AI labs can build and what they'll let you use became something you can see. Anthropic released Claude Fable 5, a public version of the model it spent two months calling too dangerous to ship, with the riskiest skills fenced off, included with paid plans for two weeks, then pay-per-use.

We also get into Apple rebuilding Siri on a custom Google model after two years of trying to build its own, the SpaceX IPO that's quietly an AI bet, and a federal AI bill that picked up an unlikely co-signer: the CEO building the most powerful model on the market, asking to be regulated. Plus, in The One More Thing, the cup of coffee that explains the whole AI assistant race.

Everything we covered, plus the deeper read, is at intelligencedistilled.io.

This week, the question in AI changed.

For two years it was "what can this do?" Now it's "what does it cost, and who controls the bill?" Anthropic closed a $65 billion round at a near-trillion-dollar valuation, passed OpenAI, and filed to go public, all in five days. Microsoft shipped its own AI models at Build, built to stop paying OpenAI, and the pitch was cheaper, not better. Trump signed the oversight order he postponed two weeks earlier, with one catch: it's voluntary. Plus, Microsoft tries to delete six files off the internet and learns a very old lesson.

Everything we covered is at intelligencedistilled.io.