Distributed Mechanistic Interpretability at Scale: Activation Streaming, Split-Layer Inference, and Distributed Sparse Autoencoder Training

J. Melton · American Code Labs · Preprint, August 2026

Abstract

ActStream splits inference and analysis across two nodes, streams intermediate activations over a purpose-built binary wire protocol, and trains sparse autoencoders on the streaming output without staging activations to disk. The Thunderbolt link sustains 2.150 GB/s, 86% of its negotiated rate and 26x the LAN path, holding p99 latency within 1.5x of median. Includes a feature-based safety classifier evaluated against an LLM-judge baseline on PKU-SafeRLHF.

Code and data

github.com/american-code/ResearchPapers — data/distributed-capture/

Licensed CC BY 4.0.