The Cutting Edge of AI: From LLMs to Multimodal Intelligence
summary
Overview of concrete claims and tools: argues small LLMs are more practical than enormous models, names Llama, Qwen v3, DeepSeek 3.1 as open options, and claims Gemini 2.5 leads text while newer vision models (GPT5/GP5) lead vision. Describes systems built with embeddings stored in a vector database, RAG vs agentic RAG with MCP and A2A protocols, video RAG using ASR OCR and object detection, and an AR pipeline built on TensorFlow and nnx/ONNX. Notes cost grows with dialogue length and error accumulates across multi-step chains, and recommends model-as-judge plus golden datasets for evaluation.
AI-written summary · about
chapters (14)
- 0:00AI trends and Net Company
- 1:07Large language models overview
- 2:54Multimodal and agentic AI
- 4:42RAG vs agentic RAG
- 7:09AI deployment challenges
- 8:55Cost vs accuracy trade-offs
- 10:22AI model evaluation
- 12:57AI in daily life
- 15:23Context engineering in AI
- 17:52Video RAG frameworks
- 19:38AR frameworks for AI
- 20:00AI trends 2025
- 21:44New AI industry roles
- 22:48Q&A on AI evaluation
related talks
49mThe Building Blocks of Today’s and Tomorrow’s Language Models
40mKeynote: Open-source Multimodal AI in the Wild
18mLLMs and GenAI: A Guide to Building Your Own RAG Model for Document Interaction
32mSponsored Session: Beyond the Buzz: Navigating AIs Myths and Realities
27mAI Demystified Unveiling the Intelligence Behind the Machine
36mLLMs and AI agents demystified
26mHow Does LLM and Generative AI Really Work?
36mInsights from The State of AI Report 2023