Robert Nishihara, Anyscale | AWS Startup Showcase S3 E1
summary
Claims that Ray is the open source runtime behind large LLM work and that Anyscale is a managed Ray service that hides infra are backed by specifics: demo runs OPT training across 32 GPUs with a single-line scale change to 64 or 128, cited trend of compute growing ~35x every 18 months, and customer-reported moves from ~20% to 95% GPU utilization (≈5x cost efficiency). Tooling includes Ray Train, DeepSpeed, VS Code workspaces, dashboards, spot instance support and auto‑terminate to cut waste.
AI-written summary · about
