The article further discusses the integration of Visual Blocks, which allows for client-side deep estimation without coding, and the introduction of 16 new custom nodes in collaboration with Hugging Face. Eight of these nodes run entirely on the client-side using Web AI. The article also mentions the large-scale use of JavaScript for Web AI in Chrome, with updates to WebGPU and the addition of the Memory64 scheme in WebAssembly. The article concludes by introducing the use of Headless Chrome for testing Web AI models, leveraging server-side GPU acceleration.
Key takeaways:
- Web AI is a set of technologies designed to use machine learning models on the client side through a web browser running on device CPU or GPU.
- Gemma Web is a new open model launched by Google that can run in the browser on user devices, which can significantly save costs, enhance user privacy protection and shorten latency compared to running inference on cloud servers.
- Visual Blocks allows you to run deep estimation on the client side without writing code, and 16 new custom nodes have been created in cooperation with Hugging Face.
- Chrome is working on built-in device-side AI, allowing you to access models using standardized JavaScript API for specific tasks, and has updated WebGPU to support 16-bit floating point values.