Band-Aids don’t fix bullet holes
summary
Argues that production ML pain comes from pipeline entanglement and vendor-driven point solutions, not lack of models: deep learning excels at vision/speech but is brittle to adversarial noise and costly to tune (one study used 250,000 GPU hours). Shows glue-code failure modes, hidden technical debt, and a working alternative: prioritize feature engineering, summaries and sketches (example: min hash sketch for recommender problems) plus stronger observability on Kubernetes and better types to make models interpretable and maintainable. Choose this if you want concrete fixes, not hype.
AI-written summary · about
chapters (8)
- 0:00Start
- 1:12Taylor Swift
- 7:20Hidden Technical Debt in Machine Learning Systems
- 16:48Network Architectures and Hyper Parameters for Machine Translation
- 18:07Why Do We Use the Wrong Tools
- 19:13Incentives of Vendors and Even of Open Source Communities
- 30:44Type Signature
- 34:00Conclusion
related talks
26mCompleting the machine learning loop
48mPrinciples and Practices of Building Machine Learning Systems
21mWhy is Production ML so Hard?
30mBuilding Teams and Culture that Support ML Innovation
28mUnravelling Hidden Technical Debt in ML: A Pythonic Approach to Robust Systems
08mConsiderations for producing the best AI and Machine Learning models
34mForget about AI and do mathematical modeling instead!
02mInterview