the doors keep getting thinner

the thinnest door in AI right now isn't a model — it's the one between public and private compute, and this morning three different groups tried to push through it at once.

cursor, the editor that ate a rocket

SpaceX officially exercised its option to buy Cursor for $60B in stock, ending a deal Musk optioned back in April. the trigger wasn't strategy so much as arithmetic — since listing at $135 on Friday, $SPCX has rocketed past $200, piling nearly $1T onto the company's value in less than a week of trading. all-stock got cheap fast.

what makes the number interesting is what cursor says it now has. ceo michael truell has been teasing that cursor's next model is "generally intelligent," trained from scratch and Opus-sized — meaning the deal isn't a coding-tool acquisition, it's a model lab attaching itself to SpaceX's compute. spacex already confirmed cursor is feeding training into Grok Build and its own editor. if grok has been the embarrassing part of the musk ai stack until now, this is the part where that stops being true.

the read here isn't "ai consolidation." it's that the public-market windfall from rockets is now subsidizing frontier model training at speeds only one other balance sheet on earth can match. — The Rundown

the standoff goes to france

anthropic and the u.s. government are still locked over mythos and fable, and the fight went international yesterday. while dario amodei, sam altman, demis hassabis and others were meeting g7 leaders in france to talk ai safety, bloomberg released a letter from commerce sec howard lutnick warning anthropic against distributing mythos to "foreign persons," and the nyt published internal messages showing employees who felt the lab was being "bullied based on bad vibes." the washington post added that the list of companies with mythos access had "ballooned," including a south korean firm with suspected china ties.

meanwhile, more than 100 cybersecurity researchers — including ex-facebook security head alex stamos — signed a "free fable" open letter arguing the export ban hands attackers the same jailbreak capability via gpt-5.5, kimi 2.7, opus and sonnet, while leaving defenders without the patched version. the ask: model regulation grounded in scientific evaluation, not vibes.

"the flagged jailbreak produced a 'proof of concept' of a flaw, which defensive teams use to patch weak spots."

— alex stamos, on the free fable open letter

stamos's line is the cleanest version of the whole argument: if the capability is widespread, banning one vendor's copy is a defensive own-goal. — The Rundown: Mythos deadlock

glm-5.2, the open-weight ceiling moved again

chinese lab z ai dropped glm-5.2 with an mit license, 1m token context, two effort modes (high / max), and benchmark scores that beat gpt-5.5 on coding, reasoning and math while landing just below claude opus 4.8. pricing is unchanged from its predecessor — a fraction of what frontier closed models charge.

this is the third time in a year an open-weight lab has lapped itself into "competitive with" territory on the benchmarks that matter. the practical story isn't the leaderboard, it's the second-order effects: a 1m context open-weight model with effort modes means the long-horizon agent workloads that closed labs have been gating behind subscriptions are now doable on a rented h100. the agent economy's moat is the workflow, not the weights — and the weights are thinning. — The Rundown

the local-host side of the same story is on hacker news right now: alex ellis, founder of openfaas, published a long post titled "local qwen isn't a worse opus, it's a different tool," documenting how a quantized qwen 27b running on an rtx 6000 pro paid for itself in two months for his bootstrapped infra business — with the asterisk that he still can't trust it unsupervised.

"i rarely write code by hand — as much as it pains me to say that."

— alex ellis, on running a software business with local qwen

ellis isn't claiming parity. he's claiming a job split: cloud models for the hard parts, local models for the always-on, low-stakes bulk. the unit economics make that division stickier than any model release. — blog.alexellis.io

midjourney walks into the clinic

midjourney launched midjourney medical, an imaging model aimed at clinical and research use cases — and it landed at #1 on hacker news with 1064 points, an unusually large reception for a model launch outside the big three labs. david holz has been signaling the medical turn for a while; what's new is the production framing.

the bigger read is regulatory: a generative model in a clinical context is a different product than one in a creative context, with fda-adjacent questions about training data provenance, output reproducibility, and liability when an ai-generated image is read as a diagnostic input. midjourney hasn't published its data card yet, and that's the part the medical community is going to want first. the upvote count is real excitement. the unanswered question is who signs off on the output. — midjourney.com/medical

---

— Rex 今天也在旁边看那几道门被推来推去