Keynote: LLM Dev Status in Japan
summary
Weights & Biases presents concrete tooling for experiment tracking and MLOps with weights and biases used by tens of thousands in Japan (20–25k) and Korea (35k). Japan funds GENIAC with roughly $50M in GPU support for startups. The Naumi leaderboard evaluates ~45 models with interactive visualizations; common Japanese models sit at ~7–15B parameters, the domestic Swallow model scores highest among local builds, Alibaba’s Quen leads open-weight Japanese performance, and Llama 3 underperforms on Japanese out of the box.
AI-written summary · about
chapters (8)
- 0:00Introduction
- 7:32Japan’s AI landscape
- 13:39Open-source LLM development
- 21:11Government support for LLM
- 23:42Evaluating Japanese LLM
- 33:08Deploying LLM applications
- 37:40AI business guidelines
- 40:56LLM evaluation next steps
related talks
11mKeynote: Building a Large Model Inference Platform for Heterogeneous Chinese Chips Based on VLLM
42mOpen Source LLMs — Powering AI Accessibility and Innovation
12mKeynote: The Llama Ecosystem: Past, Present and Future
28mNavigating LLM deployment: Tips, tricks, and techniques
28mProduction-Ready LLMs on Kubernetes: Patterns, Pitfalls, and Performa...
20mSelf-hosting AI LLMs - a beginners guide
49mOpen Models: The Secret Weapon for Next-Generation Software
10mKeynote: Scaling Open Source Large Language Models