Stanford CS25: Transformers United V6 I From Language Models to Native Multimodal Intelligence

Short summary

Stanford's CS25 seminar explores the evolution from language models to native multimodal systems, covering architectural principles, training approaches, and emerging directions like sparsity and modality specialization. Led by Christopher Manning and colleagues, with guest Victoria Lin from Thinking Machines.

• •Deep dive into multimodal LLM architecture and training principles inherited from language models
• •Covers modality representations, autoregressive modeling, and reasoning capabilities
• •Discusses emerging directions including sparsity and modality specialization

Generated with AI, which can make mistakes.