Keynote Interview
summary
Wikimedia runs about 100 ORES models for edit quality, topic and damage detection and serves them at Wikipedia scale from on‑prem hardware rather than public cloud. The team deploys Kubernetes with Kubeflow and KFServing (standalone serving first), uses AMD GPUs (not NVIDIA/CUDA), pre‑caches predictions for every edit, and plans a feature store plus continual retraining pipelines. Hosting includes Cloud VPS and a one‑click Jupyter service called Pause; almost all code, models and data are published openly for reuse and inspection.
AI-written summary · about
chapters (9)
- 0:00Introduction and welcome
- 2:08Chris Albin’s background
- 8:25Wikimedia and machine learning
- 15:05Wikimedia’s open philosophy
- 21:44Professionalizing ML at Wikimedia
- 35:22Research community adoption
- 42:02Q&A and community impact
- 50:01Future roadmap and goals
- 55:16Closing and open culture
related talks
12mKeynote: AI Infrastructure
34mKeynote: Building for AI at Scale: Infra, Integrity, & Innovation
21mKeynote: A Tale of Using Kubeflow to Make the Electricity Smarter in China
37mKeynote – Carol Willing
16mKeynote: An Open-Source Stack for AI Compute: Kubernetes + Ray + PyTorch + vLLM
08mKeynote: Plug in and Scale: Serving LLM Models on Kubernetes Ma...
16mKeynote: Marin: An Open Lab for Frontier AI
14mKeynote: Building the Infrastructure that Powers the Future of AI