Notes on Act I Log Interpretation: Future Use of Decoder-only LLM

amplifiedamp · x · 2026-08-23

The author adds notes to the Act I log interpretation project. To avoid anomalies from using a small embedding model (SONAR) to generate activations and invert centroids into labels, future work may use a decoder-only LLM. Data was collected from Jul '24 to Apr '25, and a link to all 100 labels is provided.

Original post →

More from Research

Research channel →