A clay laptop on a kitchen table showing long streams of pebbles, a small figure sitting beside it for a long time while a clock turns

Thought bubbles above the laptop show three simple games, a paddle and ball, a maze and a brick wall, and a fourth bubble is empty

A large robot and a small laptop stand at the same table looking at the same pebbles, both reading

The small figure puts a closed padlocked toolbox aside untested, looking thoughtful

Illustration, AI-generated from the story text. Tap a panel to read the story through it.

Afternoon Brief · November 3rd 2026

Updated

An open 27-billion-parameter model on a laptop repeats part of the Astra weight-reading test

Simon Willison gave the paper's raw-weight game test to Qwen 3.8 27B, running locally in about 17GB, and it named Pac-Man, Pong and Breakout from the numbers alone and failed on Freeway, taking 46 to 65 minutes each. Kimi K3.1 Pro got Breakout in under ten minutes. The paper presented reading weights as something GPT-6 Astra could do; an open model on a laptop now does most of it too, given an hour. Willison did not test the editing result, and says he is "not ready to declare that the age of source code is over." SourcesC

Sources

CQwen 3.8 27B on my laptop can interpret model weights, too (Simon Willison)simonwillison.net

This story ran in the Afternoon Brief of November 3rd 2026.

Western startups are selling Chinese open models with the censorship edited out

Netflix tests an AI Discover tab on US televisions