A conference talk covering explainability and interpretability in AI, from classical machine learning techniques to LLMs. It explains why explainable AI (XAI) matters for transparency, fairness, and regulatory compliance (EU AI Act), then walks through techniques like LIME, feature importance, and interpretable models. The talk then addresses the unique challenges LLMs pose for explainability due to their scale, nondeterminism, and tendency to humanize outputs. Practical techniques covered include source attribution with RAG, scope detection, prompt chaining, chain-of-thought prompting, and tree-of-thought reasoning. The core message is that LLM outputs should never be used without human validation, and that explainability tools help make AI usage more trustworthy and transparent.
Sort: