Rebuilding the AI Platform: Designing for Today’s Chaos and Tomorrow’s Unknowns
summary
Netflix is rebuilding its AI platform to serve half a billion users in 190 countries and claims a single system can handle classical ML, LLMs, foundational models and future agents. Concrete pieces: a capability-centric front door for personalization, conversational search and artwork personalization; a routing spine that picks models by latency cost and safety; unified telemetry and an InfraBill cost view. V1 works for CPU models but shows gaps for GPU inference, LLM eval and governance, so the rebuild is ongoing.
AI-written summary · about
chapters (16)
- 0:00Rebuilding Netflix’s AI platform
- 1:00AI use cases and platform mission
- 3:47Why rebuild now
- 5:35Evolution of Netflix’s AI platform
- 7:44Strengths and gaps in old platform
- 9:29Five core requirements
- 11:58From model-centric to capability-centric
- 14:24Unifying ML and LLM use cases
- 16:29North Star architecture
- 18:13Multi-model and multi-provider support
- 20:40Safety, observability, cost
- 22:28Measuring behavior and model quality
- 24:13Designing for agents
- 26:44Agent requirements today
- 29:54Agent-native future
- 31:39Q&A
related talks
29mEvolving the ML Platform Organisation at Netflix: A Case Study
29mBuilding Netflix - from movie magic to ML reality
03mInterview
38mNetflix's Ray Platform: From Deep Learning to GenAI
53mKeynote Interview
17mBuilding Resilient AI Pricing Platforms
28mAccelerating AI at Scale with Data
50mDemocratizing AI - Back-fitting end-to-end machine learning