Paper-shaped blog post claims GPT-6 Astra can edit raw weights

single source· 1 articles · confidence: medium · first seen 2026-11-03 16:29 UTC

What this means for you

Nothing to act on. The method is described in a blog post, not shipped as a tool, and the strongest numbers come from a single unreviewed write-up with no evaluation date. If you fine-tune small models, the claim to track is that weights can be altered without gradients — but none of it is usable today.

A write-up by Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab — paper-shaped, though the authors call it a blog post — reports that GPT-6 Astra can read and edit the raw weights of a neural network without running training code. It says the model cut loops in Qwen3.5-0.8B's reasoning by editing 0.016% of its weights, raising MATH-500 from 7.9% to 48.1%; no evaluation date is given. Simon Willison says one example also worked with Qwen 3.8 27B on his laptop.

Models in this story

Key facts

  • ·The write-up is by Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab, who classify it as a blog post rather than a paper. source
  • ·It reports that GPT-6 Astra raised Qwen3.5-0.8B's MATH-500 score from 7.9% to 48.1% by editing 0.016% of its weights, with no evaluation date given. source
  • ·It reports the same weight edit tripled IFEval, using the same sampling settings. source
  • ·It claims small game-playing networks written from scratch by GPT-6 Astra outplayed networks trained with PPO for 100 million steps, without environment feedback. source
  • ·Simon Willison reports that Qwen 3.8 27B, running on his laptop, handled one of the write-up's examples. source

What the sources say

  • Simon Willison — Repeats one example on a local 27B model and separates the two claims bundled together.

Sources

The original reporting. Follow these — they did the work.

← the wire