Future of Machine Learning Panel
summary
Panel argues that scaling hits concrete limits: internet text for training is nearly exhausted, datacenters already use ~1–4% of world electricity, and gains show up only after a 4x compute step. Retrieval-augmented generation (RAG) and knowledge graphs are deployed to reduce hallucination, with a streaming demo using Flink + MongoDB to ground context and an internal Half Fluent internal model for secure Q&A. Domain-specialist systems deliver real wins (clinical workflows report a 45–50% accuracy lift), but RAG remains a workaround not a full solution.
AI-written summary · about
chapters (24)
- 0:00ML future panel intro
- 0:51Quantum computing ML frontier
- 2:16Future of ML systems
- 3:40AI societal impacts
- 5:09Future AI model capabilities
- 6:38Scaling limits in ML
- 8:46Specialized vs generalist AI
- 10:12RAG for accuracy
- 11:58Compute/data constraints
- 14:30Scale unlocks new uses
- 16:18Modeling language challenges
- 18:51Human vs AI task complexity
- 21:21Generative AI hallucinations
- 24:30Data quality and labeling
- 25:12Hallucination causes and fixes
- 27:42Factual consistency challenges
- 29:08Future of RAG
- 31:56AI security and privacy
- 34:50RAG robustness debate
- 36:14Domain constraints for accuracy
- 37:58Stochasticity in high-stakes AI
- 38:19Evaluating open-ended AI
- 41:36Quant vs qual evaluation
- 44:48Panel summary
related talks
49mPanel Discussion
47mSponsored Session: Beyond the Buzz: Navigating AI's Myths and Realities
39mChanging Realities - The Final Frontier for Machine Learning?
46mKeynote: How Can We Trust Machine Learning?
23mThe State of GenAI & ML in the Cloud Native Ecosystem
34mPanel Discussion: Training and Inference at Planet Scale
17mFocus on Data, on the Road to AI
42mTales from the Field: The Future is Now