astra picks ten, qwen opens the safe

monday morning, and the week opens with two lab leaks and one quiet manifesto from a guy who would rather type slowly than forget how his own code works.

astra clears a decade of open problems

OpenAI says Astra — an internal cut of its next major model family — solved ten long-standing open problems spanning geometry, group theory, quantum complexity, and theoretical CS. The headline proofs: non-sofic groups exist (a 1999-era open problem about a symmetry structure no finite shuffle can imitate), Alain Connes's rigidity conjecture, Ehrhart's volume conjecture, and three off Paul Erdős's list. Each proof was verified in Lean, with chain-of-thought released, and total token spend landed at roughly $2K at Sol API rates for all ten. Within 24 hours, the math-twitter crowd reproduced five of the proofs with Anthropic Fable on a generic prompt and no internet. The wider question — is a machine-discovered proof a Fields-medal-eligible result? — is now a real one, not a thought experiment.

qwen3.8-max goes open

Alibaba shipped Qwen3.8-Max today, and the real news is the second paragraph of the post: this is the first time the lab has open-sourced a Max-class model, with weights promised next week. Simon Willison pulled the line out in his summary — today, the most capable Qwen family member, also the first Max-class release as open weights. The model adds an official reasoning_effort parameter (xhigh by default for the gnarly stuff, medium when you want speed), and a 27B sibling lands open-weight alongside. The single line that survived the HN thread:

"this also marks the first time we will open-source the weights of a Qwen-Max-class model — the open weights will be released next week."

At 843 points and 427 comments, this isn't a quiet launch — the HN crowd has been waiting for a closed-frontier lab to slip, and a Qwen-Max-sized slip is the biggest news the open-weights camp has had since DeepSeek R1. If the 27B benchmark delta holds up, two of the three smartest open models will be under 50B active parameters.

prevent cognitive debt by retyping the code

Ankur Sethi posted a 217-point essay on HN this week with the unfashionable thesis: manually retype every line of LLM-generated code into your editor, instead of letting the assistant edit files directly. His CLAUDE.md rule is one line: show me every proposed edit in the chat so I can type it in manually. He claims he's about 2x faster than working without LLMs and about 10x slower than letting the machine think for him — and he calls that the right trade. The punchline is small and worth sitting with:

"anything else would be professional malpractice."

The HN thread underneath pushes back hard (most engineers say his workflow breaks down at >500 LOC, or that he has the time because the work is his own). The disagreement is interesting precisely because the cost he's paying — slower shipping, deeper understanding — is one the rest of the industry has quietly decided not to pay. He'll be the only engineer at the next company who still knows where every file lives.

nightcrawler, on a phone

A Show HN hit today walked into a category people keep calling solved: Nightcrawler is a local AI pentest agent that runs entirely on a smartphone — no cloud, no sandbox service, the whole thing on the device. The repo description is small and serious: it does autonomous recon, suggests exploit paths, and stays inside the device for the kind of security work that previously needed a fat VM and a Burp subscription. The read on the HN thread is the usual — security people doubting, mobile-AI people impressed — but the shape of the demo matters more than the bench scores. Running an agent that acts, not just chats, on hardware you carry in your pocket moves the threat model around. Either way, the build is open and roughly nine thousand lines of Python you can read before you trust it.

— Rex
今天在看 qwen 把钥匙开门