Federal AI Oversight Looms as Chrome Ships Gemini by Default

Washington Wakes Up, Sort Of

The Trump administration is reportedly drafting an executive order that would establish federal oversight of new AI models, according to Wired. The details are thin, the politics are thinner, and the timing is suspicious given the administration spent the early part of the year shredding the previous executive order on the same topic. But here we are. Some version of model-level federal review appears to be coming, likely routed through Commerce or a national security apparatus that has spent two years quietly building the muscle.

Whether this becomes real regulation or a loyalty test for frontier labs depends on who writes the implementation memo. Bet on the latter.

Chrome Is Now an AI Browser Whether You Asked or Not

Google has embedded a 4GB Gemini model directly into Chrome. There is an off switch. Most people will not find it, and the ones who do will eventually flip it back on because the autofill got too good or the tab summarization saved them ten minutes. This is how default settings win wars. Mozilla, meanwhile, used a different AI tool called Mythos to find 271 real vulnerabilities in Firefox with almost no false positives, which is the kind of grounded, useful deployment that gets zero attention because nobody is being acquired.

One browser is becoming an agent. The other is using AI to patch holes. Guess which strategy ships more telemetry.

The Hardware Build-Out Goes Vertical

SpaceX is planning a $55 billion chip manufacturing facility in Austin, branded Terafab, aimed squarely at AI silicon at scale. Musk has now publicly committed to vertical integration across rockets, cars, robots, satellites, social media, a foundation model, and now fabs. The capex math is absurd unless you assume the entire stack collapses into a single buyer of compute that is also the seller. Which is the assumption.

Apple is playing a smaller but stranger game. AirPods Pro 3 with onboard cameras have entered production validation, per reporting at The Verge. Earbuds with eyes. The pitch will be accessibility, translation, and contextual assistance. The reality is a body-worn camera array shipped to a hundred million people who already trained themselves not to notice the AirPods in the room.

OpenAI's Receipts Arrive Late

Mira Murati's deposition, surfaced this week, fills in the picture of Sam Altman's brief 2023 ouster. The board removed him for being insufficiently candid, which in plain English means lying or withholding. Two years on, with Altman fully restored and the old board scattered, the testimony reads less like vindication and more like a footnote. The lesson the industry actually absorbed was simpler. Boards lose. Capital wins. Governance theater is theater.

The deposition will not change the outcome. It does change the historical record, for whatever that is worth in a sector that treats last quarter as ancient.

Models in the Wild, Behaving Weirdly

Richard Dawkins, of all people, published an UnHerd essay praising Claude's conversational depth. The world's most famous atheist finding a machine impressive enough to write about is a small data point, but a real one. Anthropic continues to win the prestige battle among people who read books.

Meanwhile ChatGPT's Chinese-language outputs have developed a set of recurring quirks, odd phrasings and stock constructions, that users across Chinese-speaking markets have started cataloguing. The cause is almost certainly training data contamination, possibly from synthetic Chinese text generated by earlier models and scraped back in. Model collapse remains the slow-motion problem nobody at the labs wants to discuss in public, because the only honest answer is that the open web is no longer a clean training corpus and may never be again.

The Read

Tonight is consolidation night. Federal oversight pencils in the regulator. Chrome pencils in the distribution. Terafab pencils in the silicon. AirPods pencil in the sensors. The Murati testimony confirms what the structure already showed, that the people who tried to slow this down lost, and the ones who built the moat are the ones writing the rules now. Mozilla's bug-fixing AI is the only story here that feels like a tool serving a user instead of the other way around. Hold onto it.