Live Q&A: Day 1
summary
Live Q&A covers concrete systems and numbers: a video generator runs at 24 fps native generation (vs prior 4 fps) and ships 1,000+ non‑cherry‑picked samples on GitHub; captioning uses Llama models to produce detailed training captions; an AI‑twin flow enforces live scripted recordings plus verification; lip‑sync works from 2s input; moderation runs in production with LLMs and human review; mobile upscaling targets 720p–1080p and ABR optimizes for watch time; Netflix uses elastic reserved instances since 2017.
AI-written summary · about
chapters (19)
- 0:00Day 1 Live Q&A
- 0:28Video generation model motion
- 3:39AI twin impersonation safeguards
- 4:42Lip dub variable video lengths
- 6:06Content moderation deployment
- 7:30AI ads for promotion
- 8:30Future of machine translations
- 10:36Real-time translation challenges
- 11:59Media processing in AI translation
- 13:50ML in video quality pipelines
- 16:23Microservice vs monolith tradeoffs
- 19:37Video scaling resolution formats
- 21:29Mobile data savings vs quality
- 22:32Video input enhancement
- 23:56Validating captions with LLMs
- 25:02Dynamic load handling in Netflix
- 26:30Validation in streaming
- 27:35Netflix Cosmos implementation
- 28:41Closing remarks
