Web AI Summit 2025
2025-10-17
Sunnyvale, United States
as the organisers put it
In-person Event - Web AI Summit 2025: Join Google for their next gathering on client side AI for an agentic internet era!TLDR: Attend Web ...
who it's for
frontend engineers, web platform engineers, machine learning engineers, browser engineers, ai infrastructure engineers, xr developers, devtools engineers, product managers
topics
WebNN API and roadmap; WebGPU compute shaders and SIMD; WebAssembly memory64 and promise integration; Transformers.js in-browser models; MediaPipe Web and Gemma streaming; WebLLM high-performance in-browser inference; WebMCP executable site tools for agents; LiteRT.js WASM-backed runtime; real-time CNN video upscaling on WebGPU; offline vector search with SQLite embeddings; built-in browser AI APIs and Gemini Nano; in-browser model quantization and conversion
| title | author | length |
|---|---|---|
| State of client side AI | Jason Mayes | 24m |
Claims that web AI runs entirely locally on consumer GPUs (demo on a 5‑generation‑old Nvidia 1070) and that Web AI usage grew from 1M to 1.2B yearly downloads… | ||
| Browsing, building, and web in the AI era | Parisa Tabriz | 12m |
Chrome removes the 4 GB WebAssembly ceiling with memory64 and adds JavaScript promise integration as a high-speed translator between async web and sync native… | ||
| Beyond the demo: Building the WebAI Playbook | Kenji Baheux | 14m |
Claims that a persistent playbook vacuum causes demo-to-production signal loss and maps solutions across three execution worlds: cloud AI, web AI (client-side… | ||
| Web AI leaps forward on Intel AI PCs | Moh Haghighat | 15m |
WebNN on GPU doubles real-world throughput while NPU coverage expands to run 20 top Whisper/TransformerJS models; Whisper goes from ~43 to ~100 tokens/sec on… | ||
| Transformers.js: Building Next-Generation WebAI Applications | Joshua Lochner | 25m |
Transformers.js runs models 100% locally in the browser using WebGPU and WebNN and claims 1.7M unique monthly users with ~11M CDN requests. Models are… | ||
| LiteRT.js, Google’s high performance WebAI runtime | Matthew Soulanille, Chintan Parikh | 13m |
LiteRT.js runs converted .tflite models in the browser as an npm package (WASM-backed LiteRT) and targets CPU via XNNPACK and GPU via WebGPU acceleration… | ||
| Accelerating Web AI on | Will Lord | 16m |
Arm reports 2–6x local model speedups using Clyde AI on ARMv9 CPUs (examples: FI3 on Windows-on-Arm 2.4–4x, Vivo X200 ~2.6x) and announces a neural accelerator… | ||
| Running Google's Gemma LLMs in the browser with MediaPipe Web | Tyler Mullen | 15m |
MediaPipe Web runs Gemma 3 and Gemma 3N fully in-browser using WebGL/WebGPU and WASM, streaming models piecewise so even a 27B / 27 GB Gemma model can load and… | ||
| Building a VR Assistant and llms.txt widget | Erick Wendell | 13m |
Demos a Chrome-based VR assistant that actually schedules, cancels and confirms barber bookings via Chrome’s built-in AI APIs, using speech-to-text and… | ||
| Don't let AI agents push your buttons - use webMCP instead! | Khushal Sagar | 11m |
Proposes a concrete API, webMCP, that lets a page declare executable site-side tools (example: search_products tool with a JSON schema) which the browser agent… | ||
| Wasm, WebGPU, & WebNN: How compute abstraction are enabling client side AI | Thomas Nattestad | 14m |
Wasm compiles C++ Python and other languages to run on the CPU with threads via web workers and SIMD, enabling things like llama.cpp and compiled Python SDKs… | ||
| InferenceJS: Real-time computer vision in your browser | Max Schridde | 15m |
Roboflow shows running computer vision models inference in the browser using InferenceJS as a thin layer over TensorFlow.js and Microsoft COCO models from… | ||
| Bangle.js - an Open Source Smartwatch | Gordon Williams | 15m |
An open-source smartwatch called Bangle.js runs JavaScript on ARM Cortex-M with under 8KB execution stack, ships with an always-on color screen and up to… | ||
| Real-time video upscaling using CNNs and WebGPU | James Hurley | 15m |
Uses WebGPU compute shaders to run a six‑layer CNN that does about 3,000 MACs per pixel on incoming video frames, importing video frames zero‑copy when… | ||
| Built-in AI: The next frontier for web experiences | Cecilia Cong | 10m |
Miravia moves review summarization from servers to Chrome’s built-in summarizer API, running client-side on product pages to avoid constant server calls and… | ||
| It doesn’t need to be a chatbot: Unlocking the product value of smaller AI models | Janna Lipenkova | 15m |
Argues that product AI succeeds when teams pick specific, small wins instead of defaulting to big chatbots, because integration gaps and noisy LLM capabilities… | ||
| Building AI + XR apps faster with XR Blocks | Ruofei Du | 14m |
Announces XR Blocks, an open‑source WebAI + WebXR SDK that reduces prototypes to 39 lines of code for a pinch-to-change-color demo, integrates Gemini Canvas… | ||
| How to control drones with your mind: The Black Swan BCI headset | Dara Tumenbayeva | 08m |
Non-invasive headset captures EEG and converts signals through a TensorFlow-trained AI layer and a WebAI web UI to send drone commands in real time; current… | ||
| Offline vector search with SQLite and EmbeddingGemma | Rody Davis | 10m |
Builds an offline browser RAG by encoding documents with EmbeddingGemma (38M parameters, 768‑dim), storing float32 vectors in SQLite VEC compiled to WASM, and… | ||
| How to build AI-powered VS Code extensions with local models | Hugo Zanini | 13m |
Case study of a VS Code extension that embeds an OpenMetadata catalog and Airflow lineage, queries Elasticsearch and summarizes results with Gemini, and adds a… | ||
| Built-in AI: how Mastodon uses translation APIs | Thomas Steiner | 12m |
Integrates the Language Detector API and Translator API into the Elk.zone PWA to stop mislabeled Mastodon posts and cut server translation costs; detection… | ||
| Discover Web AI: Client side Agents, Gen AI, and machine learning in the browser | Jason Mayes | 02m |
Demonstrates that full AI stacks run in-browser with Hugging Face collaboration and a Visual Blocks low-code pipeline built in seconds, claiming privacy, lower… | ||
| State of client side machine learning | Jason Mayes, Jim Bankoski | 24m |
Overview points and hard numbers: Google reports over 1 billion cumulative downloads of MediaPipe and TensorFlow.js with ~600 million downloads per year… | ||
| Transformers.js: State-of-the-art Machine Learning for the web | Joshua Lochner | 18m |
Library brings Hugging Face Transformers parity to browsers using ONNX and Optimum, with WebGPU support and WebAssembly execution; over 1,200 converted models… | ||
| The Web Neural Network (WebNN) API: Where we are and what's Next | Rob Kochman, Rafael Cintron | 12m |
The WebNN API is a proposed standard for executing neural‑network graphs in browsers as a single integration point across CPU, GPU, and NPU hardware, claiming… | ||
| Web AI on next generation AI PCs | Moh Haghighat | 18m |
WebNN is a W3C web API running in Chrome/Edge Developer Preview and ONNX Runtime Web that maps models to CPU/GPU/NPU and demos show it actually works: Lunar… | ||
| ml5.js - Friendly machine learning for the web | Aiden Nelson, Yu Lee | 13m |
ml5.js is an open-source JavaScript library that wraps TensorFlow.js models into a tiny API so creative coders can run PoseNet, HandPose and FaceMesh in the… | ||
| WebLLM: A high-performance in-browser LLM Inference engine | Charlie Ruan | 17m |
WebLLM is a browser-native LLM inference engine that runs models 100% client-side, compiles WebGPU kernels via MLC-LLM/TVM, and uses WebAssembly for… | ||
| State isn't all you need, but It helps: building better LLM apps in the browser | Jacob Lee | 17m |
Demonstrates that small in-browser LLMs become practical when combined with concrete engineering: use WebLLM and Transformers.js for local models, apply… | ||
| Visual Blocks: Visual prototyping of AI pipelines | Ruofei Du | 14m |
Visual Blocks is a browser visual programming platform that wires live webcam, video and audio through TensorFlow.js, Three.js and WebGL nodes with GPU shaders… | ||
| Exploring alternative interactions in JavaScript | Charlie Gerard | 18m |
Live demos show JavaScript hooking browser sensors to real models: webcam via getUserMedia into TensorFlow.js (PoseNet-powered Beat Saber clone that runs hard… | ||
| Overview of Chrome built-in AI | Kenji Baheux | 15m |
Chrome adds built‑in AI that shares a single, browser‑hosted model lineup so sites and extensions avoid gigabyte downloads and server bills; prototype work… | ||
| Lessons learned from being customer zero of Chrome's built-in APIs | Thomas Steiner | 16m |
Builds a local AI synonym finder using Gemini Nano and the Prompt API Playground and shows it works as an MVP but needs heavy prompt engineering to avoid… | ||
| Web AI in industry: How TensorFlow.js has driven what you see on the supermarket shelves | Hugo Zanini | 12m |
A Brazilian CPG deploys an offline real‑time product detector to replace manual shelf counts, training YOLOv7 on SKU110K plus 5,000 hand‑labeled images for 40… | ||
| The future of AI is now: Real-life case studies for on client-side AI adoption in web apps | Yuriko Hirota, Kazunari Hara | 15m |
Ameba integrates Gemini Nano in Chrome to generate blog titles client-side for 75 million active users, claiming ~150 ms response for interactive iterability… | ||
| Why are Web Extensions fantastic for AI? | David Li | 10m |
Web Extensions let AI control the browser, observe page content, and add UI so models run where users work; Chrome reports 2x increase in Web Store submissions… | ||
| Beyond the banner: The power of Web AI to personalize paid rich media ads | Cory Grenier, Evgeny Peshkov | 15m |
Geenee demonstrates real-time, client-side web AR ads that served 20 million people in 30 days and raise apparel clickthrough from ~0.4% to ~0.7% (about 75%… | ||
| MediaPipe Web: Bringing cross-platform AI tech to the browser | Tyler Mullen | 15m |
MediaPipe Web explains how C++ pipelines become browser-ready via WebAssembly (Emscripten) and TypeScript bindings, and how cross‑platform MediaPipe Tasks… | ||
| Transforming access to healthcare through Web AI | Chris Slee | 17m |
IncludeHealth builds a web AI platform that runs Body Pose Estimation and downstream TensorFlow models in-browser (p5.js/ml5 experiments led to a production… | ||
| ML training on the web: Building Simple ML for Google Sheets | Richard Stotz | 14m |
Claims that decision-forest models train in seconds and predict in microseconds using on-device ML via WebAssembly and npm packages. Describes YDF (open… | ||
| nothing matches | ||