Keynote: Fine-tuning our way towards openness in AI
summary
Open source models are now claimed to match or beat proprietary ones and the session shows working demos: local vision and text inference with Rama Llama, synthetic training-data generation that increases output quality and cuts cost using Instruct Lab, and PDF/table extraction that preserves layout with Dockling. The stack runs on VLM with quantization, Llama Stack agent tooling, Podman AI Lab and MCP APIs for function calling, with monitoring via Prometheus/Grafana and Jaeger — demos run live or as pre-recorded proofs of concept.
AI-written summary · about
chapters (11)
- 0:00Welcome and introductions
- 2:07Technological history
- 6:00Evolution of user interfaces
- 9:24AI demo and local models
- 12:11The case for open source AI
- 15:28Education and learning
- 17:33Customizing LLMs
- 21:06Data preparation with Docling
- 24:36Inference and optimization
- 28:32Building agentic applications
- 31:26Future of AI and community
related talks
40mKeynote: Open-source Multimodal AI in the Wild
16mKeynote: Open Source for AI and AI for Open Source
16mKeynote: Open Source AI Everywhere
51mKeynote Panel: Open Innovation in AI - Present an...
20mKeynote: Open-Sourced Inspiration - The Present and Future of Tech - AI
27mKeynote: Open Source in the Age of AI
11mKeynote: Open Source Stack Empowers the AI Era
16mKeynote: An Open-Source Stack for AI Compute: Kubernetes + Ray + PyTorch + vLLM