A New Stack for the Gen AI Era
summary
Claims Gemini is natively multimodal and is rumored to have 5× the parameters and 20× the tokens of GPT‑4, producing SOTA benchmarks; the recommended stack fragments into five concrete workflows: model training/fine‑tuning, data processing and embeddings, prompt construction, prompt execution, and multi‑agent orchestration. Concrete tools named are LangChain, LlamaIndex, Ray, Spark, Cassandra and vector databases, plus LangStream, Weights & Biases and OpenAI/Vertex orchestration. A production example uses Cassandra + vector search to generate SQL via dynamic few‑shot prompting and shows measurable accuracy gains for complex joins.
AI-written summary · about
chapters (16)
- 0:00Introduction
- 1:05AI stack evolution
- 2:08Gemini model capabilities
- 3:57Brain-inspired AI architecture
- 7:49Skypoint agent workflows
- 9:58Generative AI workflows
- 12:22System of context concept
- 14:29Data processing for AI
- 15:55Prompt construction
- 17:20Chain of Thought prompting
- 21:20Dynamic few-shot prompting
- 22:24Prompt execution monitoring
- 24:52Multi-agent orchestration
- 26:17Orchestration layer battle
- 29:49Conclusion
- 30:38Q&A
related talks
35mGenAI Stack for Developers
24mScalable GenAI in Production
23mScaling GenAI Applications and Agentic AI for Business Competitive Differentiation
06mThe Building Blocks of Agentic Al
31mExploring LLMs across clouds – A Year in the Life
39mBuild the new generation of AI Agents
05h 56mFrom the main stage to the terminal
02h 44mUsing Generative AI with Java: Beyond Simple Chatbots