Is Your LLM Evaluation Missing the Point
summary
A concrete mismatch emerges: evaluation suites report 93% accuracy while deployed systems produce 48,000 flagged cases that a review finds 93% wrong and tools like PromptFoo, DeepEval, RAGAS prove necessary but insufficient. Case studies from federal benefit automation, child welfare triage, and histopathology show which features (immutable criminal history, stale 60‑month poverty counts) cause harm, which thresholds and aggregation rules fail, and which red‑teaming, ethnography, model cards, and co‑designed tests actually catch errors before rollout.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 1:51Domain experts in LLM evaluation
- 5:44Real-world harms from ignoring expertise
- 12:12Distributional justice in AI
- 15:44Bias from feedback loops
- 17:09Medical AI use cases
- 20:00Communication gaps in AI
- 23:10Invisible labor in teams
- 25:01Benchmarks vs domain expertise
- 26:26Participatory AI models
- 29:17Including domain experts
- 34:10Model cards for collaboration
- 34:52Expertise fosters trust
- 35:17Q&A
related talks
35mEvaluating LLM systems: the nuts and bolts
46mEnhancing Agent Reliability Through Evaluation-Driven Development
25mEngineering Better Evals: Scalable LLM Evaluation Pipelines That Work
38mMaking LLMs reliable - A practical framework
50mBeyond the Prompt: Evaluating, Testing, and Securing LLM Applications
39mBeyond the Prompt: Evaluating, Testing, and Securing LLM Applications
28mTrustworthy LLMs: Vibe checks are not all you need
48mBeyond the Prompt: Evaluating, Testing, and Securing LLM Applications