NEW paper from Microsoft Research.
(bookmark it)
The entire interpretability literature is built around human readers. As more analysis gets delegated to agents, the right target of interpretability shifts. This paper is a recipe for designing tools that agents can actually
For anyone questioning whether this post is overhyped, I think they are actually *underselling* the approach.
We did similar experiments and found that Claude models can write entire, functional neural net weights from scratch! csinva.io/weight-space-p…
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming:
One exciting thing: when LLMs write weights directly, they can usually *explain* the changes they're making. Huge win for interpretability!
Especially true for scientific modeling (e.g. constructing an interpretable model to predict brain responses)
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming:
An obvious way to improve over Jev:
1. open-source/cheap embedding model
2. sklearn model on top, fit to your labels
Almost always cheaper + way more accurate
blog.voyageai.com/2026/10/27/how…@VoyageAI
Excited to share TokenEconomizer, a plugin that learns from sequential tasks to make coding agents significantly cheaper (w/ no perf drop). Works on codex, claude, & more!
Great work led by Aaron Li github.com/microsoft/Toke…
