Dehumanizing Agents: Why Explainability is Crucial in the LLM Era
summary
Built a churn prototype that flags customers above 90% and then runs feature importance plus clustering to replace costly manual calls, which worked and produced actionable segments; created an AutoML pipeline prioritizing compactness and stability to pick interpretable models for a thrombosis risk project but found those approaches break when LLMs arrive; adopted retrieval augmented generation with inline citations and a scope detection classifier (adds a scope related class) to cut hallucinations—out‑of‑scope queries cause ~10% of enterprise errors.
AI-written summary · about
chapters (21)
- 0:00Introduction
- 1:31What explainable AI is
- 5:43Challenges in explainability
- 8:55AI failures without explainability
- 10:43AI Act requirements
- 12:07Global vs local explainability
- 15:47Interpretable AI solutions
- 19:18Limitations of LLM explainability
- 23:57Humanization in LLMs
- 27:53Dangers of LLM overreliance
- 31:03LLMs for retrieval vs prediction
- 32:06RAG and source attribution
- 37:03Scope detection for answers
- 41:45Reducing hallucinations
- 43:11Self-citations reliability
- 45:22Prompting techniques
- 47:53Prompt chaining
- 50:42Chain of thought techniques
- 52:27Validating LLM outputs
- 54:39Human oversight for AI
- 55:02Q&A
related talks
50mDehumanizing Agents: Why Explainability is Crucial in the LLM Era
28mExplaining Machine Learning Models with LLMs
24mWhy and How to Build Explainability into your ML Workflow
15mHarnessing LLMs for Scalable Data Minimization
33mDesigning LLM Applications for Massive Datasets
55mLessons Learned From Implementing Generative AI Solutions
29mAutomating Privacy Reviews with LLMs: Challenges & Opportunities
43mAI Development Lifecycle: Learnings of What Changed with LLMs