How Google Is Shifting AI From the Cloud to Your Browser

The Google Web AI summit was held earlier this month as an invite-only event in Sunnyvale, Calif. After the event, I caught up with the organizer, Jason Mayes, who leads Web AI initiatives at Google. The last time I’d interviewed Mayes was February 2023, when our main topic of conversation was TensorFlow.js — a JavaScript library for machine learning (ML). At that time, just a couple of months after the launch of ChatGPT, the term Mayes used was Web ML. But now the term du jour is “Web AI.” I asked Mayes when that switch happened. “Yeah, so I pivoted when I realized web developers were searching the term ‘Web AI’ because they don’t know the [term] machine learning,” he chuckled. In his view, Web ML is more technically correct, but the entire industry has chosen to use “AI” over “ML.” In any case, the Web AI trend — and indeed the focus of the event Mayes just ran — is primarily about running client-side AI in the browser. And now that’s about much more than TensorFlow.js; other technologies discussed at the event were MediaPipe LLM, WebMCP, Transformers.js, Chrome’s Web AI API, WebAssembly (Wasm), WebGPU, WebNN and more. From TensorFlow.js to LiteRT.js for In-Browser Inference In fact, TensorFlow.js is less widely applicable now and is on the path to being replaced by a new library, called LiteRT.js (LT is short for Lite Runtime). That’s because the focus of Mayes’ team has shifted from supporting both training and inference — which is what TensorFlow.js does — to focusing solely on inference, which LiteRT.js was designed for. “So TensorFlow.js was inference and training,” Mayes said. “You could even do training in the browser. We decided to drop the training there to focus purely on inference, because I think Python’s basically the way to go for things on the training side, but the deployability of the web is still the winning factor here, and the reach and the privacy and all that kind of stuff. So, we wanted to make a library that’s really focused on the fastest inference in town in that kind of situation.” “We wanted to make a library [LiteRT.js] that’s really focused on the fastest inference in town [for browsers].” — Jason Mayes, Google Web AI lead In other words, Google has ceded the training of AI models to Python, rather than continuing to try and push a JavaScript solution for that. The company seems to have made the bet that inference — the process of making a decision or a prediction based on an AI model — is something that can be done more effectively in the browser, which means it’s ripe for JavaScript. To further explain: LiteRT.js is Google’s Web AI runtime, targeting production web applications. It builds on LiteRT, which is designed to run ML models directly on devices (mobile, embedded, or edge) rather than relying on cloud inference. So LiteRT.js is the browser-specific implementation of the LiteRT runtime, enabling in-browser inference via WebAssembly and WebGPU. I asked Mayes whether JavaScript will ever compete with Python as the primary language to use with AI. He replied that due to academia’s bias towards Python, “you’re going to see the latest research models come out in Python first” and that the training will happen on the backend. However, he thinks there will be abstractions that allow Node.js and JavaScript developers to talk to the models. “So I think then that enables all the JavaScript developers to fine-tune models or retrain them for their business needs, and then convert those models to a web AI form,” he said, “and that’s what actually gets deployed, especially if they care about privacy or cost.” Understanding the ‘Agentic Internet’ As with everyone else, Mayes has also become enamored with AI agents this year. At the Web AI Summit, he proposed what he calls “an agentic internet.” “I believe what we actually will have is an internet designed for humans and agents alike. So in the same way we had to go from desktop to mobile when the smartphone was invented, I think we need to go from human to human-and-agentic now that agents are invented, essentially.” For JavaScript developers, WebMCP “might be a better thing than the original MCP [Model Context Protocol].” — Jason Mayes He points to the new WebMCP standard that Google and Microsoft are working on as an indicator of where things are going. He suggests that for JavaScript developers, WebMCP “might be a better thing than the original MCP [Model Context Protocol].” WebMCP, as I discovered when I spoke to Microsoft’s Kyle Pflug at the end of September, is like setting up an MCP server for your website or app, except that the functionality WebMCP gives is client-side rather than from the server. “The original MCP is very much designed by people who have never touched the web,” Mayes said. “It’s a command-line kind of ecosystem on the backend, and it’s almost like they just discovered the REST protocol for the first time. We’ve been using it for years, obviously, on the website. So maybe there’s something better we can do here for the communication standards of the web and the frontend and JavaScript; and there [are] better ways to annotate your website, such that browsers and other agents can discover the tools that you’re exposing on your website, versus a command-line tool.” “Imagine every website is annotated for the tools it provides.” — Jason Mayes Mayes thinks WebMCP and similar agentic tools will lead to AI applications that go way beyond the chatbots that defined the first wave of generative AI (GenAI) during 2023 and 2024. “Imagine every website is annotated for the tools it provides,” he offered as a thought exercise. “I believe that instead of adverts in the future — I mean, adverts will still exist, but instead of only adverts — companies that offer any kind of service … will offer a WebMCP interface (or something like this) that can be understood by the browser, such that they can bid to do work when users ask to do something.” The Browser as the Human Interface for AI Other than agents and MCP, 2025 has also been the year of AI browsers. Microsoft Edge gained a “Copilot Mode” in July, then in September, Google announced a raft of new AI features in Chrome (with further “agentic capabilities” to be added in the coming months), and of course last month OpenAI announced its new browser, Atlas, which joins similar chat-based browsers on the market, like Perplexity’s Comet and The Browser Company’s Dia (now owned by Atlassian). Given the new AI focus of browsers, I asked Mayes whether he expects web browsers to be the human interface for AI agents. “There’s a reason all this GenAI stuff started with a web browser.” — Jason Mayes “That is the entry point to all of this,” he agreed. “The browser is the only place that you’re signed in to all of those things you might want to use WebMCP from. So if they [website owners] just annotate their sites properly, as you browse the internet, the browser discovers tools that are relevant to you — versus having a centralized repository with, like, a billion tools that makes it then a paradox of choice, [where] you can’t choose which ones to use. … So you can build up a personalized repository of tools that you probably care about.” “There’s a reason all this GenAI stuff started with a web browser, right?” he added, referencing ChatGPT. “It didn’t start as application-native; it was a web interface.” The Coming ‘Aha Moment’ for Web AI Local AI seems to be the big trend driving the larger Web AI trend. Towards the end of our conversation, Mayes noted that “at some point, we’re going to have a model in the future that’s as good as today’s cloud models, [but] that fits on-device.” He referenced the recent advancements in power in Google’s Gemma models, which are lightweight enough to run on laptops and even phones. Mayes believes that when we get to the point where sophisticated models can run on-device, then organizations will begin asking whether they even need the cloud anymore for AI. “Maybe for 95% of use cases, you won’t need to delegate to the cloud.” — Jason Mayes “Yeah, there might be some new advances,” Mayes said. “There’s always going to be something new and shiny to run on the cloud that doesn’t fit on-device, or some state-of-the-art thing, but maybe for 95% of use cases, you won’t need to delegate to the cloud. And that’s when you’re going to start seeing a lot of people have this aha moment, and discover Web AI.” The post How Google Is Shifting AI From the Cloud to Your Browser appeared first on The New Stack.