Managing Machines
summary
Claims that shipping ML is hard because labeled data and post-deploy operations dominate effort, not model code; gives concrete figures and practices: labels are often the bottleneck, uses CAPTCHAs as an example of cheap labeling at scale, and says 80% of engineering pain happens after deployment. Notes models are probabilistic and fail some percentage so monitoring and runbook automation are required, warns off relying on plug‑and‑play cloud APIs, and highlights feature engineering, data transformation and labeling tools as the missing pieces.
AI-written summary · about
chapters (4)
related talks
30mChallenges of machine learning: The new ML lifecycle
25mAutomating Operations with ML
04mThe age of machine learning
21mWhy is Production ML so Hard?
06mMachine learning is about your data and deployment
26mMLOps for the rest of us- A poor man's guide to putting models in production
13mMLOps at Scale
01h 05mMachine Learning Models and Productionization