About
neuronengineering.dev is a personal research site devoted to one question: how much of a frozen model's behavior can be elicited by editing activations, and how systematically can we find the edits that work?
The domain is the brand, and the brand is the engineering position: neuron-level control of neural networks, implemented with the seriousness of systems software. The Rust implementation of the MAS-Steer harness is the reason the TLD fits — interventions are data structures, hooks are zero-cost, and every claim on the paper page is backed by something you could replay.
The site hosts papers first. MAS-Steer is the founding entry in /papers/; the structure is built to accumulate more. A research log (/blog/) is planned once there is something to log beyond the paper itself — the first post will likely trace the line from single-neuron personality flips to population-based search.
This project sits between mechanistic interpretability and inference engineering. The interpretability side supplies the hypothesis — individual neurons and small sets carry behaviorally decisive signal. The engineering side supplies the discipline — interventions that cannot be logged, replayed, and stress-tested do not count as results here. Spectacle demos are fun, and they are credited where they inspired this work; but the standard on this site is the protocol.
For questions, corrections, or collaboration: [email protected]. The GitHub repository and arXiv entry are linked from the footer once public.