Blog
Research notes.
What we measured inside language models, and what didn't work.
November 15, 2026 · 12 min read
NeuroSpeak: a language for reading and rewriting what a model thinks.
A script language where every line runs one operation on a model's internals and prints what it did in English. Words are defined by examples, edits carry promises the model has to keep, and someone else's weight edit can be read back into words. Here is how it works, and which parts failed.
November 14, 2026 · 14 min read
We found where a small model doubts itself. Deleting it got 8× more math problems right.
One direction inside Qwen3.5-0.8B, found from four example pairs, holds its habit of second-guessing. Removing it took GSM8K from 4 of 64 to 32, and beat MIT's Astra weight edit 61 to 31 under their own settings. Here is what we found across 30 studies, and what we didn't.