iThome Ironman 2026
iThome is a Taiwanese technology media platform and developer community. Its annual Ironman challenge invites participants to publish a technical article every day for 30 consecutive days. I wrote this series for the 2026 challenge as an accessible introduction to mechanistic interpretability and the broader questions it connects to.
Beginning with a small puzzle about factual recall, the 30 articles cover:
- representations, probing, and different levels of interpretability evidence;
- causal interventions, circuit discovery, and causal abstraction;
- sparse features, attribution graphs, and model diffing;
- activation steering and other ways of intervening in model behavior;
- Chain-of-Thought faithfulness, introspection, and model psychology; and
- cross-model representation comparison, NeuroAI, and Brain–LLM Alignment.
Throughout the series, I also emphasize what each method can and cannot establish. Hands-on companion notebooks let readers try the tools firsthand and build toward designing their own interpretability experiments.
