Today in our AI evals series, we’re building an alignment dataset to ensure our LLM judge actually agrees with human reasoning. Plus, learn how to use a statistical hack called Bootstrapping to prove your high scores aren't just a lucky draw → https://goo.gle/4cQCsJH
Chrome for Developers
Technology, Information and Internet
Helping you build, grow, and innovate on the web.
About us
The official Chrome for Developers LinkedIn account from Google. We want to help you build beautiful, accessible, fast, and secure websites that work cross-browser, and for all of your users.
- Website
-
https://developer.chrome.com/
External link for Chrome for Developers
- Industry
- Technology, Information and Internet
- Company size
- 5,001-10,000 employees
Updates
-
Sometimes you just need a little inspiration from the Chrome Dino to power through #GoogleIO session prep. Join us to discover how advanced AI tools running directly in the browser will enable better agentic capabilities and streamline your daily workflows → https://lnkd.in/g922Pj_n
-
We are checking in on the WebMCP Early Preview to hear your ideas on building the agentic web → https://goo.gle/41QD9MQ Previously, we opened up access to this proposed web standard so you can hand AI agents an API to understand your site. Instead of forcing agents to scrape screens and waste valuable tokens, WebMCP lets you define your capabilities as tools. So whether you're using the Declarative API to translate existing HTML forms or the Imperative API to grant direct JavaScript access, we want to know what you're building!
-
Learn how to build sites that are ready for the next generation of web agents → https://goo.gle/3Obk1pn The foundations of the web are more important than ever. By focusing on well-structured, accessible, and semantic HTML, you create experiences that are easy for both humans and AI agents to navigate. Recommit to these core principles to ensure your content remains discoverable, inclusive, and effective in an evolving digital landscape.
-
-
Chrome for Developers reposted this
Can we resolve the latency of multi-step AI agents by changing where the logic executes? 🔌 Currently, WebMCP interactions often rely on "Atomic Tools". Discrete actions like move(), look(), or use(). For complex tasks, this back-and-forth creates a significant bottleneck. In a maze-solving scenario, 40 sequential moves can translate to over 60 seconds of model round-trips and considerable context bloat. I've been investigating an alternative to this sequential tool calling: client-side code execution. Instead of the model deciding every discrete step in a back-and-forth loop, the agent generates a complete JavaScript algorithm that executes locally in the browser. In my maze experiment, the model produces the logic once, and a sandboxed Web Worker handles the execution loop. This partitions the logic differently: 🧠 The Model: Handles high-level strategy and code generation. 🏗️ The Client: Handles high-frequency execution and state management. This approach shifts the bottleneck from model latency to the browser’s own execution speed. In the maze demo, it reduced the time to solution from minutes of model "thinking" to seconds of local execution. The primary trade-off is the current browser security model. While the performance gains are clear, executing agent-generated code on the client side involves significant security risks and environment constraints that make it unsuitable for production today. These hurdles suggest that a robust, secure execution environment may eventually need to be a platform-level feature of the agentic web. I've documented the technical implementation of the eval_code tool and the performance data from the experiment here: https://lnkd.in/eWzDQFwM #WebMCP #CodeExecution #WebDev #AI #AgenticWeb
-
-
Prepare your Single Page Apps (SPAs) for the future by testing the Soft Navigations API in our final origin trial → https://goo.gle/4ttf9eP If you are building with React, Angular, Vue etc, this API is the missing piece to help you accurately measure Core Web Vitals for your SPAs. We have been experimenting with this API for a while, and thanks to your continuous feedback, we are finally ready to ship it later this year. Before the official launch, we are offering one last origin trial starting in Chrome 147 and running through Chrome 149. This is the perfect opportunity to conduct a final test of the API's expected final shape.
-
-
Working hard to bring you the best web updates at #GoogleIO → https://goo.gle/3Qn6nQW Expect highly technical sessions on how to use Chrome DevTools for agents, HTML in Canvas, AI APIs and more.
-
Improve your LLM-based application by building a robust AI testing pipeline → https://goo.gle/4uDaaZl Part 4 and 5 of our AI evals series explore how to combine objective rule checks with a custom LLM judge to build your testing pipeline.
-
In his new article, Sam Richard explores why the real work of a software engineer lies in architecting solutions, building consensus, and navigating the "people problems" of software, not just writing syntax ⬇️
What is the role of software engineering in the age of #ai? I've had a number of conversations recently, and they all seem to converge on the same thought: LLMs can't replace software engineers, because coding was never the job. Let's start having a conversation about what the job actually is, and what the actual limits of LLMs and coding agents are. Only then can we start to think about how to move our industry forward. This is the first of what will likely be a multi-part series on my evolving thoughts on the subject. Stay tuned for more! #webdev #software #engineering
-
Before you write any code, you need to define exactly what success and failure looks like for your AI application → https://goo.gle/4tOVshb Today, we hit the drafting board to map out objective failures, like broken JSON, and subjective ones, like a toxic output or bad brand fit. Watch this video for a quick rundown and map out your own eval strategy.