Keynote
summary
Estimates put a full foundational LLM training run at $65M training cost, so organizations favor hosted models and RAG (retrieval augmented generation) to constrain hallucinations and preserve provenance. Observability shifts from uptime to semantic fitness tests and continuous evals; LLM Ops must track model, data and CI artifacts for reproducibility. Real incidents show risk — a chatbot created a bogus bereavement policy; only single‑digit percent of PoCs reach production. Coding assistants scale output but slip on refactoring (≈33% accuracy) while firms report 25% of code generated by copilot.
AI-written summary · about
chapters (19)
- 0:00Introduction to AI architecture
- 1:08Training large language models
- 4:44AI in drug discovery
- 6:10Serving AI models securely
- 11:26AI architecture stack explained
- 13:34Choosing the right AI model
- 16:01AI applications for non-technical users
- 18:12Document answering with LLMs
- 19:18AI for social impact in India
- 21:08Emerging AI patterns and RAG
- 23:13Multi-agent AI systems
- 25:44Observability and privacy in AI
- 26:47Challenges of AI in production
- 28:59Role of architects in AI systems
- 31:56Enterprise coding standards with AI
- 33:23Testing AI applications
- 35:09MLOps and AI reproducibility
- 37:38AI in software delivery lifecycle
- 38:42Future of AI evolution
related talks
33mOpening Keynote: LLMs, between fantasy and reality - the poin...
44mProduction-Grade LLMs & AI Agents
13mNew Challenges in Software Quality Brought by GenAI and LLMs
10mKeynote: Enterprise Value with Generative AI
12mKeynote: How LLMs Have Changed the Way Organizations Do AI
05mKeynote: Teaching LLMs New Knowledge
17mArchitecting LLM-Based Chatbots using free open-source LLMs
45mPanel Talk