GPT-6 Astra reads raw weights only on tiny networks, and Postman installs itself into Codex

On Tuesday's Hacker News front page, the lead story was GPT-6 Astra, and how much of a network it can really edit by reading its weights. A paper called Sparks of Weight-Space Programming has Astra look at a network's raw numbers, with no tools, and say what it was trained to do. Matthew Honnibal, who wrote spaCy, sorted every experiment in it. The no-tools setting only covers small networks, up to about 170,000 weights. Every test on a real language model has Astra writing Python.

The headline result is a fix for Qwen3.5 0.8B, which got stuck in reasoning loops. Astra edited a tiny fraction of its weights, and the model went from solving about one math problem in thirteen to nearly half. To get there it wrote code and ran Qwen on up to fifty batches of practice problems. Honnibal calls that work a human machine learning engineer could do. x313, who works in interpretability research, puts it at four to eight hours by hand.

The no-tools test is one you can run yourself. Simon Willison fed the paper's game prompts to Qwen 3.8 27B, a 17 gigabyte model, in LM Studio on his laptop. The prompts are on GitHub, and they need a long context window. From the weights alone, Qwen named Pac-Man, Pong and Breakout, taking close to an hour for each one, and it missed Freeway. Willison thinks a patient person with a pencil and paper could get there too.

Nicholas Carlini ported Stockfish to a 200 megahertz Pentium Pro from 1995, with 64 megabytes of RAM. Tuned for that chip, it plays far above Kasparov, by a margin where it should score better than nine games in ten. So a desktop of that era could have beaten him without Deep Blue's custom chess chips, and most of the gain since then came from better software. The machine is online and you can play it. k2052, an expert-level club player, reports finding a line that beats it.

If you turned off training on your Codex chats, check that it is still off. One company had everyone disable data sharing to keep its code away from the labs, and several of its engineers found the setting, Improve the model for everyone, switched back on in ChatGPT's data controls. chriskanan had asked twice for training to be disabled, and both requests were granted, and found it on again. OpenAI's privacy portal has a second switch for the same thing.

The Postman desktop app installs plugins into Codex and Claude Code when an existing, signed-in user opens it, along with the Postman CLI. indicava found a startup hook that tells Codex to load a Postman skill before any non-trivial API task, and installer code that tries to approve that hook on its own. A second user saw the same install on a newer build. If I needed an API client this week, I'd use Bruno, because it keeps your requests as plain files in your git repo.