Keynote: Solving the problems of AI in production
summary
Argues that model training is easy but production ML fails on integration, not algorithms: cites adversarial fragility of deep nets and a 250,000 GPU‑hour Google evaluation costing ~$750,000 to show scale. Describes building an internal system called Toth project that ingests PyPI, CI/CD and performance traces into a graph database and runs experiments to generate a dependency advisor that recommends secure, high‑performance package versions; work proves the platform helps more than the initial black‑box ML idea.
AI-written summary · about
chapters (18)
- 0:00Intro
- 4:54ML workflow
- 6:15ML pipelines
- 8:51Data drift
- 12:32Deep learning
- 19:08Bad incentives
- 23:34Programming is still hard
- 24:39Lowlevel languages
- 25:12Dependencies
- 26:40AI Center of Excellence
- 27:05Our goal
- 27:16Our traps
- 28:36Focus on the work
- 31:00What do we do
- 31:48More AI related
- 32:01Tiny Black Box
- 32:54Better Data
- 37:49Outro
related talks
59mKeynote: AI without the BS, for humans
17mKeynote
39mKeynote Panel State of AI in the Enterprise
15mKeynote: From Platforms to AI Factories. Has Kubernetes Solved It?
49mKeynote: Operations and AI with Jen Krieger
33mOvercoming the Barriers to Deep Learning in Production
10mKeynote: The Last Mile Problem: Why AI W...
44mKeynote: Tracking the Pulse of Generative AI