westurner
12 hours ago
> Circuit Tracer on Gemma-2-2b
decoderesearch/circuit-tracer: https://github.com/decoderesearch/circuit-tracer
ScholarlyArticle: "Dissecting the Black Box: Circuit-Level Analysis of LLM Vulnerability Detection" (2026-05) https://arxiv.org/abs/2605.29901v1
westurner
11 hours ago
Explainable AI: https://en.wikipedia.org/wiki/Explainable_artificial_intelli...
"Harmonic Loss Trains Interpretable AI Models" (2025) https://news.ycombinator.com/item?id=42941954 :
> Harmonic loss enables improved interpretability and faster convergence, owing to its scale invariance and finite convergence point by design,