WebNN, Built-in AI, WebMCP: What’s New in Web AI?
summary
Updates on browser AI standards and working demos: WebNN gives near‑native NPU/GPU inference (often ~90% of native), shown with object, landmark and depth models via WebGPU; a cross‑origin storage proposal prevents repeated multi‑GB model downloads; Built-in AI exposes use‑case APIs (prompt, writer, translator, embedder) that run offline but are Chromium‑only and show cold starts and browser‑dependent model behavior (Gemini/Edge LLMs); WebMCP registers site tools so agents call endpoints directly — demos include a pizza builder and maze solver that work but expose latency and policy limits.
AI-written summary · about
chapters (11)
- 0:00Introduction to Web AI
- 8:00Bring your own AI approach
- 13:06WebNN and local inference demos
- 27:38Cross origin storage proposal
- 28:00Built-in AI APIs overview
- 32:57Prompt API demo
- 41:31Multimodal models in browsers
- 45:03AI browsing with WebMCP
- 50:50WebMCP demos and tools
- 55:33Pros and cons of Web AI
- 58:47Q&A and closing remarks
