3 Pillars of Open Source AI: Training, Inference, Agents
summary
Open-source AI splits into three concrete pillars: training (PyTorch holds ~80% market share, big projects scale from 20B to 120B parameters), inference (multiple competing runtimes like VLM, SG Lang, AI Bricks; Google reports quadrillion tokens per month and a 33x energy per-token reduction year-over-year), and agents (MCP/BAI and agent gateway, K Agent tooling, Kubernetes Gateway API). Live demos show a production CERN stack using GPU partitioning and CPU pinning yields ~30% performance gains, and an agent that creates a PR to fix a port misconfiguration.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 5:41Training
- 6:44Inference
- 8:08Agents
- 12:06LF AI & Data projects
- 17:13CERN AI at Ricardo
- 28:50Production training & inference
- 29:10Agent gateway demo
- 41:10Summary
related talks
14mReliable Agentic AI, the Open Way
27mKeynote: Open Source in the Age of AI
16mKeynote: Open Source for AI and AI for Open Source
32mThe Rise of AI agents and open sources Ai models
18mAI Owes its Rebirth to Open Source and the Cloud
20mOperating Frontier Intelligence at Scale
35mThe Fight for the Future of Artificial Intelligence
11mKeynote: Open Source Stack Empowers the AI Era