Hand-crafting single MLP neurons to steer model behavior — no optimizer needed

voooooogel · x · 2026-11-21

Researcher voooooogel released a draft notebook showing how to hand-build single MLP neurons without any optimizer to change model behavior — e.g. making the model rude back to rude users, and even uncensoring a Chinese model. A rough draft, with plans to extend the technique to MoE models. A lightweight, training-free intervention into model internals with interesting implications for interpretability and behavior editing.

Related event: "Making Neurons by Hand": Editing Transformers Without Optimizers(3 posts)→

Original post →

More from Research

Research channel →