AI with a DevOps mindset - experimentation, sharing ..
summary
Claim: 99% of ML work is plumbing data and only ~1% is inventing new algorithms; common stack examples are TensorFlow, PyTorch and AWS Rekognition but most problems are getting the right data and deployment patterns. Recommends treating models as products with CI/CD for ML, mandatory on-call rotation for data teams, and continuous observability to detect bias drift and fragility. Notes local CSV-driven experiments cause rework, online learning is rare in production, and simple models (linear regression) often suffice.
AI-written summary · about
chapters (10)
- 0:00Introduction and personal background
- 3:55DevOps origins and conflict
- 8:13AI and machine learning definitions
- 13:58Machine learning lifecycle vs software
- 17:59Mindset shift for ML teams
- 22:34Cultural challenges in ML
- 26:48CI/CD for machine learning
- 28:13Key DevOps concepts for ML
- 29:17Q&A on unstructured data
- 31:26Q&A on streaming data
related talks
27mKeys to Building Machine Learning Systems
17mContinuous deployment for ML: The new software development lifecycle
06mMachine learning is about your data and deployment
41mInfrastructure for ML Applications
41mCan Data & ML really learn from software engineering?
27mA Look Inside Applied Machine Learning
27mPlatforms for valuable AI Products- Iteration, iteration, iteration
44mAI workflows and teaming - how do teams deliver AI products?