Willison replicates weight-reading trick on his laptop
Simon Willison says a 27B model running locally did in 45 minutes what the buzzy GPT-6 Astra paper showed. Meanwhile Anthropic softened its game-copying rule and got told off anyway.
The big disagreement
Ball says watching the agent use GitLab beats a unit test
Thorsten Ball asked an agent to stand up a GitLab instance and build an integration against it, then checked the commits, MRs and issues it actually pushed instead of reading tests. Thariq called the shrinking need for self-checking the bitter lesson at work: the smartest models will need fewer tokens and will just be right. Ball agreed in one line.
Why have training wheels if you never fall over?
Receipts
Willison got the same weights trick out of Qwen 3.8 27B on a laptop
Simon Willison tried the GPT-6 Astra weights paper on his own machine and says Qwen 3.8 27B reached the same result, after 45 plus minutes. The paper's claim is that a model can directly interpret model weights fed in as prompts full of floating point numbers. He frames it as deflating the buzz, not confirming it.
I tried it with Qwen 3.8 27B on my laptop and found that, after 45+ minutes, that model could achieve the same thing
Model and agent watch
Anthropic loosened its game-copying rule, Theo and Florence still object
Addy Osmani says the rule that had Claude Code refusing game projects was updated in 2.1.317: you can name a game to describe mechanics and get an original one in that genre, and asking for a copy gets an offer to build something similar.
Imagine a pencil that wouldn't let you draw a picture of iron man.