Meta Opts You Into Its AI, Broadcom Opts Into Vengeance

Consent Is A Legacy Concept

Meta has decided that if your Instagram profile is public, your face is a prompt. The company's new AI photo generator will produce synthetic images of users by default, no explicit consent required, buried opt-out somewhere in a settings menu you have never opened. This is being described as a feature. It is more accurately a policy choice, made deliberately, by adults, in a boardroom, with lawyers present.

The framing will be that public means public, that you already surrendered the relevant rights when you posted a beach photo in 2014. The framing is thin. Generating novel images of a person is not the same as displaying images they uploaded, and Meta knows this, which is why the opt-out exists at all rather than the opt-in. Expect a regulator somewhere in the EU to notice within the quarter.

Microsoft, for its part, is having a different bad week. A Windows Defender zero-day can fill victim disks with garbage files, and the fix is stalled inside a public feud between the researcher and Microsoft's security team. Two vendors, two different flavors of contempt for the user.

The Sanctions Regime Meets Reality

CXMT, the Chinese memory maker, has been quietly rebuilding its supply chain out of domestic parts. The reporting suggests the company can now produce competitive DRAM without meaningful exposure to US export controls. This is the outcome Washington's chip policy was supposed to prevent, arriving roughly on the schedule cynics predicted when the controls were announced.

Commerce Secretary Lutnick, meanwhile, is on the phone with Samsung and SK Hynix asking them to please build more memory in America, because the AI buildout needs HBM and the domestic supply is not there. Asking foreign firms to underwrite your industrial policy while simultaneously trying to strangle their largest customer is a strategy. It is not clear it is a good one. The Korean chipmakers have leverage they did not have eighteen months ago, and they know it.

The combined picture is straightforward. Export controls are leaking, domestic capacity is short, and the people who need to fix this are reduced to begging allies while adversaries route around them. The AI infrastructure narrative depends on a hardware supply chain that increasingly does not want to cooperate with the political geography it sits inside.

Somebody Is Making Money

The Guardian notes that AI wealth is fueling a private jet boom, with SpaceX and AI startup equity holders driving surges in both purchases and charter demand. This is the part of every bubble that ages best in the retrospective documentary. Whatever happens to the underlying technology, the Gulfstream order book is a real economic indicator, and it says the top of the AI wealth distribution is converting paper to kerosene at a healthy clip.

Downstream, Broadcom is allegedly punishing Allstate for trying to leave VMware, using audits as a retaliatory tool. Broadcom denies this. Allstate has receipts. The VMware acquisition continues to be a case study in what happens when a licensing regime is optimized for extraction rather than retention, and enterprise buyers are learning the lesson in real time.

OpenAI is quietly killing Atlas, its ChatGPT browser, less than a year after launch. The company will spin this as focus. It is more useful to read it as evidence that not every adjacent product works, even when the parent app has a billion users. Browsers are hard. Distribution does not solve everything.

Claude Thinks In Places We Cannot See

The most genuinely interesting item is from Anthropic's interpretability group, which has developed a technique revealing that Claude does substantive reasoning in a hidden latent space researchers had not previously mapped. This is good news for the field of interpretability, which now has a new instrument. It is more ambiguous news for anyone who was assuming that current alignment audits were catching the relevant computations.

The short version: the models are doing more than we can see, we now have a slightly better flashlight, and the room is bigger than we thought. This is how the interpretability story has gone for three years running. Every new technique reveals more surface area of things we were previously ignorant about. That is progress. It is also a reminder that shipping frontier systems has continued to outrun the ability to understand them, and nobody involved is planning to slow down.