Why you should care about AI interpretability
summary
Mechanistic interpretability demonstrates programmable control of neurons: the Golden Gate demo locates a bridge feature, flipping it forces the model to mention the bridge; Ember platform shows clicking token attributions and boosting a “sensitive information” feature by ~60% prevents PII leaks in a Llama chat. Ember also enables Ember platform neural programming and real‑time Paint with Ember image canvases that paint concepts and actions, supports dynamic prompting (Coca‑Cola example), model diffs after training, and uses sparse autoencoders to find features.
AI-written summary · about
chapters (8)
- 0:00Introduction to interpretability
- 2:48Practical AI use cases
- 4:08Debugging and neural programming
- 7:07Text model steering demo
- 9:22Guardrails and dynamic prompting
- 12:28Image model canvas demo
- 16:18Advanced research applications
- 18:33The future of AI engineering
related talks
20mMechanistic Interpretability
41mIf We Want AI to be Interpretable, We Need to Measure Interpretability
41mThe Explainability Problem: Towards Understanding AI
59mBetween the Layers– Interpreting Large Language Models
01h 29mNavigating ML Explainability: From Theory to Implementation
01h 01mBetween the Layers– Interpreting Large Language Models
32mOn the Importance of Explainability
47mML: Explainable Artificial Intelligence explained