Keynote: Tracking the Pulse of Generative AI
summary
Counts over 30 accessible models and argues that the winning edge is not training from scratch but post‑training competitive advantage through fine‑tuning, quantization and distillation; cites a healthcare survey where >33% plan to double GenAI budgets and ~50% increase ≥50%. Recommends routers, ensembles and cascades (llm blender, root llm, unifi.it), mentions Ray and VM for platforms, Sycamore, Tropic/Voyage AI embedding-with-context trick, TimeScale/Metaphor and Google Data Commons (240 billion data points) for graph RAG, and warns to build scaleable eval, incident plans and data tooling.
AI-written summary · about
chapters (20)
- 0:00Introduction to keynote
- 0:27Foundation models landscape
- 3:19Generative AI adoption trends
- 5:28Challenges in AI projects
- 6:52Early usage patterns for models
- 8:20Post-training and model optimization
- 9:27Model ensembles and cascades
- 11:17Data strategies for generative AI
- 15:35Retrieval augmented generation
- 18:03Graph RAG applications
- 22:42AI agents and software stack
- 25:34Analysts and mathematicians using AI
- 28:27Custom AI platforms and Ray
- 31:20Multimodality in AI models
- 33:07Open-source foundation models
- 34:34Hardware and software trends
- 36:21Evaluating AI applications
- 37:03Risk management in AI
- 39:54Building generative AI applications
- 41:18Key takeaways and conclusions
related talks
11mKeynote: Building an Interoperable Ecosystem for Generative AI
17mGenerative AI: Keeping it real
26mGenerative AI: Get a Grip … and a Look Under the Hood
41mGenerative AI: bliss or miss?
43mKeynote: Applied NLP in the age of Generative AI
41mKeynote
33mKeynote: Fine-tuning our way towards openness in AI
53mHow to Build and Scale Ultra-Lean Gen AI Startups