Talk by Zachary Lipton
summary
Argues that academic distribution-shift theory rarely helps in practice and shows a concrete counterexample: a deployed audio→note pipeline that strings together language-recognition, speech transcription, diarization, alignment, generative summarization and guardrails, plus human-in-the-loop evaluation and staged rollouts. The system runs in approaching 100 hospital systems with tens of thousands of clinicians and millions of notes generated regularly, and it succeeds only because of product processes, challenge datasets and continual feedback rather than neat theoretical assumptions.
AI-written summary · about
chapters (6)
- 0:00The state of machine learning
- 2:38The evolution of ML research
- 10:18Tackling distribution shift
- 17:40Paradigm battle in research
- 27:48Real-world AI deployment
- 42:10Final reflections and closing
related talks
13mClosing Address
18mReality is not an end-to-end prediction problem
45mArchitecture of an NLP Deployment
31mKeynote Q&A – Barbara Plank
15mKeynote: Empowering Accessibility Through Kubernetes: The Future of Real-Time Sign Language Interpretation
46mContext is all you need: FUNdamental linguistics for NLP
24mTowards Large-scale Speech Analytics Systems
27mWhat happens when AI starts grocking streaming audio directly