Table of Contents
Welcome back to AI News Friday. 📰🤖
Some weeks feel like a natural continuation of the story from the week before. And some weeks feel like someone grabbed the chessboard and flipped it over.
This was the second kind of week.
Last Friday I wrote about the moment Washington switched off Anthropic’s frontier models. I thought that was the story for the month. Turns out, it was just the opening chapter. This week we got the backstory — who lit the fuse, and why. We got a Chinese open-weight model that quietly out-coded every Claude Opus on the leaderboard. We got Midjourney — yes, the image generator company — announcing they’re entering medicine with a radiation-free full-body scanner. And we got a $41 billion bet on physical AI from Jeff Bezos.
Let’s get into it.
1. Midjourney Wants to Scan Your Whole Body in 60 Seconds
Let’s start with the story that has nothing to do with regulation, nothing to do with Washington, and everything to do with the fact that AI companies are no longer just AI companies.
Midjourney — the image generation platform that defined the aesthetic of the 2023-2024 AI era — unveiled a radiation-free, full-body ultrasound scanner this week. It takes 60 seconds. Zero radiation. And they’re launching a whole new division called Midjourney Medical to go with it.
Yes, Midjourney. The company that made your weird textured 4K images of wizards in cyberpunk cities is now building medical hardware.
Kenny’s Take: Of all the directions I expected Midjourney to go, medical imaging was not on the list. But honestly? It makes a weird kind of sense. They’ve spent years perfecting the skill of turning noise into coherent visual output. Ultrasound is essentially the same problem — reconstructing a signal into a readable image — with much higher stakes. The question is whether they can navigate FDA approvals, hospital procurement cycles, and liability frameworks, which are a very different kind of hard than building a better diffusion model. I’m skeptical but fascinated. This is either genius or the most expensive pivot in AI history.
2. The Amazon Paper Trail
Remember last week’s bombshell — Washington forced Anthropic to disable Fable 5 and Mythos 5 globally? This week we got the backstory, and it pins the start of it on Amazon.
New reporting reveals that Amazon — Anthropic’s biggest investor and primary cloud partner — flagged concerns about the capabilities of Anthropic’s frontier models to Washington. That warning triggered a review that ultimately led to the export-control letter that shut down the models.
Not a government agency. Not a competitor. Anthropic’s own partner.
Kenny’s Take: This changes the story completely. Last week I asked whether Amazon was a partner, an investor, or a leash-holder. Now we have the answer. Amazon saw a national security risk in the very models it was helping fund and train — and instead of dealing with it internally, went to Washington. That is a profound breach of trust, regardless of whether you think the shutdown was justified. Anthropic is now in an impossible position: their cloud provider and investor has shown they will go to the government when they get uncomfortable. How do you build a partnership on that foundation?
3. GLM-5.2 Cracks the Coding Top Two
While all the Anthropic drama was happening, a Chinese open-weight model quietly walked onto the leaderboard and started breaking benchmarks.
GLM-5.2, developed by the team behind the GLM series, now ranks in the top two for coding on the Chatbot Arena leaderboard — ahead of every Claude Opus model. And it’s open-weight. You can download it. Run it yourself. Fine-tune it.
Meanwhile, America’s best models — Opus, Fable, Gemini Ultra — are either locked behind APIs or, in Anthropic’s case, literally switched off by government letter.
Kenny’s Take: Read that again. An open-weight model from China out-coded every Claude Opus while Washington is busy making sure America’s best models can’t be used anywhere. The irony is almost too perfect. This is exactly the dynamic that export controls are supposed to prevent — instead of keeping the US ahead, they’re creating an environment where closed American models are restricted while open Chinese models get better in the open. If Washington’s goal was to accelerate Chinese AI competitiveness, they could not have picked a better strategy.
4. Bezos Bets Big on Physical AI — Again
Project Prometheus emerged from stealth this week at a $41 billion valuation. Jeff Bezos is co-CEO. The mission: use AI to automate the design and manufacturing of physical machines — robots, industrial equipment, hardware systems.
This is the biggest bet on “physical AI” we’ve seen yet, and it comes from someone who has already bet big on AI through Amazon’s internal work and his investment in Anthropic. Bezos is now building AI that doesn’t just write code — it designs forklifts and builds factories.
Kenny’s Take: I said it last week and I’ll say it again: the real AI money is in the physical world. Chatbots are a $200 billion market. Manufacturing is a $12 trillion market. Bezos sees this clearly — he always sees the infrastructure play before everyone else. $41B is a staggering valuation for a company with no public product, but Bezos has the track record and the capital to make it real. The interesting question is whether Prometheus builds the factory of the future or the factory that replaces the factory workers. Washington is going to have opinions.
5. OpenAI’s $38 Billion Reality Check
Leaked financials for 2025 show OpenAI tripled revenue to ~$13 billion while spending ~$34 billion. The reported net loss is around $38 billion — though most of that is non-cash (stock-based compensation, mostly).
The headline numbers look catastrophic. The underlying story is more nuanced. Revenue is growing fast. Compute costs are the biggest drain. And the company is moving toward a for-profit restructure that would give it access to the kind of capital it needs.
Kenny’s Take: Three things can be true at once. OpenAI is growing revenue faster than almost any company in history. OpenAI is burning cash at a rate that would have been unthinkable even five years ago. And the for-profit restructure is a lifeline, not a strategy — it buys time, it doesn’t buy profitability. The real question is whether the market’s patience for “invest now, profit later” holds when the model landscape is getting more competitive, not less. The GLM-5.2 news from earlier in this post? That is a direct threat to OpenAI’s pricing power.
6. Dario Amodei’s Policy Counter-Punch
One day after shipping Claude Fable 5 to everyone, Anthropic CEO Dario Amodei published a five-part policy blueprint that reads like a direct response to the shutdown.
The demands: mandatory third-party testing for frontier models, government authority to block unsafe models from deployment, a major jobs and retraining program, public science access, and a national security framework. He is essentially daring Washington to turn the ad-hoc power they just exercised into a transparent, rule-bound system.
Kenny’s Take: Amodei is playing 4D chess here. By calling for the very oversight that just shut his models down, he accomplishes three things at once: (1) he looks like the responsible adult in the room, (2) he forces Washington to either codify their power (which means limits and transparency) or admit they prefer opaque control, and (3) he aligns Anthropic with the “safe AI” narrative at exactly the moment when the alternative — unconstrained frontier development — looks scarier than ever. It is a masterful piece of positioning. Whether it works depends entirely on whether Washington wants rules or power.
⚡ Quick Hits
- Midjourney’s pivot proves every AI company is now a science company: image generation was never the endgame — it was the training ground for something bigger.
- Amazon’s role in the Anthropic shutdown will have consequences: every other frontier lab is now reassessing who they let inside their cloud infrastructure.
- GLM-5.2 is a warning shot: open-weight models are catching closed models, and export controls are not stopping it — they may be accelerating it.
- Physical AI is the $41B bet that nobody is talking about enough: Prometheus, Tesla Optimus, Figure — the race to automate the physical world is on, and it is bigger than all the LLM drama combined.
- OpenAI’s burn rate is survivable, but the math is getting tighter: $13B revenue on $34B spend is not a sustainable ratio, and the competition is getting cheaper by the week.
Bottom line: This week had everything — a company you thought was about making art suddenly building medical devices, a partnership drama that reads like a Silicon Valley thriller, and a Chinese open-weight model that reminds us the world does not stop because Washington sent a letter. The throughline is clear: the AI industry is moving faster than any single government, company, or investor can control. And the ones who think they are in control? They are probably wrong.
— Kenny