Keynote + Conversation
summary
Announces concrete numbers and engineering choices: 1.2 billion downloads for Llama 4 variants; a sparse mixture-of-experts design distilled from a two‑trillion‑parameter teacher into a Maverick 17B (distributed across eight GPUs) and a tiny Scout on one H100 for single‑host deployment. Shows an API with “one line of code,” downloadable fine‑tunes, speculative decoding (1.5–2.5x speedup), page‑KV for long context, partnerships for Cerebras/Grok inference, multimodal progress (SAM object segmentation, upcoming SAM 3) and real deployments from ISS to Crisis Text Line.
AI-written summary · about
chapters (12)
- 0:00Llama open source journey
- 3:05Why open source AI matters
- 8:25Llama 4 features and performance
- 14:48Llama API and capabilities
- 23:09Fine-tuning models workflow
- 26:58Meta AI app innovations
- 31:14Visual AI tools and SAM 3
- 35:12Conversation with Ali Ghodsi
- 47:05Voice interfaces and AI future
- 53:29Open source impact on AI
- 1:02:13Advice for AI developers
- 1:06:43Closing remarks
