Multimodal Generative AI
summary
Announces a new HPT framework for scalable multimodal LM training and three HBD-family models, claiming the March 16 release matches GPT‑4V on some benchmarks and that a recent HPD 5 Air trained on the N3 stack in two weeks. Introduces HPD 1.5h — 4B params for phones that reportedly outperforms a Microsoft 5B multimodal model at the same scale. Demonstrates one-photo personalization in 30 seconds with no fine‑tuning and shows early video generation—promising but not yet Solar quality—noting training costs remain very high.
AI-written summary · about
chapters (9)
- 0:00Introduction to multimodal AI
- 1:14Multimodal foundation models explained
- 3:19Company background and progress
- 4:44HPT framework for multimodal LM
- 7:56HPD 1.5H mobile model release
- 8:38Personalized image generation demo
- 10:29Challenges in video generation
- 12:43Future outlook and applications
- 14:31Conclusion and Q&A
related talks
01h 23mMultimodal Deep Learning in the Real World
52mBeyond the AI Hype: What's Real, What's Next
39mNew multimodal vision AI models and their practical applications
42mFrom theory to practice: The rise of Generative AI
17mUnlock the power of Generative AI: From understanding its principles to utilizing its power
28mLarge Multimodal Models: Understanding the State of the Art in Compute
01h 01mKeynote: Transformers: The Rise of ChatGPT
46mMultimodel Multimodal and Multiagent innovation with Azure AI