State of client side machine learning
summary
Overview points and hard numbers: Google reports over 1 billion cumulative downloads of MediaPipe and TensorFlow.js with ~600 million downloads per year, body-pose and segmentation models hitting ~120 FPS on a mid‑range laptop GPU, and a back‑of‑the‑envelope example where 1 million 30‑minute meetings would require ~54 billion segmentations/day (roughly $2B/year if done server‑side). Demos show on‑device LLMs for email and PDF Q&A, video frame search, and real products using WebGPU and emerging WebNN; Chrome ships smaller built‑in models and recommends a hybrid client/server fallback when devices lack capacity.
AI-written summary · about
chapters (8)
- 0:00Web AI Summit welcome
- 2:13Client-side ML defined
- 3:59Why client-side AI now
- 10:23Community demos
- 14:43Generative AI in browser
- 16:53Call to action
- 17:34Google's web AI investment
- 23:23Closing remarks
related talks
24mWeb Apps of the Future With Web AI
19mML at Client Side
31mNode.js <=> Machine Learning
46mFrom ML to LLM: on-device AI in the browser
23mNode.js = Machine Learning
15mMachine learning for next gen web apps with TensorFlow.js
20mTensorFlow.js Bringing Machine Learning to the Web and Beyond
31mFrom ML to LLM on device AI in the browser