Cloud security feels a little like alphabet soup sometimes. You hear terms like CSPM , CNAPP , CWPP , CIEM  — and it can all start to...

Move over, PayPal mafia: There’s a new tech mafia in Silicon Valley. As the startup behind ChatGPT, OpenAI is arguably the biggest AI player in town. Its meteoric rise to a $300 billion valuation has spurred many employees to leave the AI giant to create startups of their own. The hype around OpenAI is so […]

On this week’s GeekWire Podcast: Microsoft envisions a future where humans manage fleets of AI agents — fundamentally changing the way companies are run and work gets done.

Plus, we dive into the story of Chainguard, the $3.5 billion cybersecurity startup showing how a fully remote company can scale fast without a physical office.

And we explore how AI is reshaping everyday life, from travel planning to home repairs.

Related stories

Meet your new AI teammate: Microsoft sees humans as ‘agent bosses,’ upending the workplace: Microsoft’s 2025 Work Trend Index envisions a future where employees manage AI agents, fundamentally reshaping organizational structures.

Chainguard doesn’t have an office. Here’s how the $3.5B cybersecurity startup makes remote work: A look at how Chainguard successfully operates as a fully remote company, including intentional communication and periodic in-person gatherings.​

Subscribe to GeekWire in Apple Podcasts, Spotify, or wherever you listen.

With GeekWire co-founder Todd Bishop and editor Taylor Soper. Audio editing by Curt Milton.

React added experimental support for two new techniques this week: View Transitions and Activity.

View Transitions makes it easier to add animations to the UI transitions in an app.

“Under-the-hood, these animations use the new start ViewTransition API available in most modern browsers,” the React Labs blog post read. “This new component lets you declaratively define ‘what’ to animate when an animation is activated.”

For the very curious, the React team created an explanation of how View Transition works, but the post focuses on how developers can use it with three triggers:

1. startTransition
2. useDeferredValue
3. Suspense

When the DOM updates due to one of these animation triggers, React will use declarative heuristics to automatically determine which components to activate for the animation, the post explained. The browser will then run the animation that’s defined in CSS. That said, the post noted this is not to replace all the animation in an app.

“View Transitions are meant to be used for UI transitions such as navigation, expanding, opening, or re-ordering,” the post stated.

It explores ways developers can use the feature, including animating Suspense boundaries, navigations and lists.

Response from the dev community included this from Josh W. Comeau, a trainer and developer, on Bluesky:

Well this is exciting! React just added experimental support for View Transitions 🤩

react.dev/blog/2025/04…

[image or embed]

— Josh W. Comeau (@joshwcomeau.com), April 24, 2025

Activity is an API that hides and shows part of the UI. The goal is to allow developers to deprioritize components, “preserving UI state with reduced performance costs relative to unmounting or hiding with CSS.”

As the React team wrote in its blog post, “When an Activity is visible, it’s rendered as normal. When an Activity is hidden, it is unmounted, but will save its state and continue to render at a lower priority than anything visible on screen.”

Developers can use it to save state for the parts of the UI that the user isn’t using, the post read, or to pre-render parts that a user is likely to use next.

In other news, this past Monday, the React team issued the release candidate for React Compiler, a build-time tool that optimizes React apps through automatic memoization.

React also provided two case studies showing how content creation application Sanity Studio and content and curation and collaboration platform Wakelet used React Compiler to improve performance.

Angular Team Lead’s LLM-First Web Framework

Minko Gechev is known among the Angular community as Google’s team lead for the framework, but of course, he’s also a developer in his own right. He’s been tinkering independently and he shared the results this week: a large language model (LLM)-first web framework.

In a post on his blog, Gechev wrote that he sees two problems with the current landscape of platforms for vibe coding:

1. A mismatch of API versions: “Often the LLM will generate code that uses deprecated or missing APIs from previous versions,” he wrote. “I see that across the board with all frameworks since they all use versions and deprecate APIs.”
2. Lack of substantial training data: If an LLM is not familiar with an API because it’s new or just not very popular, the results are “dissatisfactory,” he stated.

There are several ways to solve the problem, including building a framework. In his blog post, he recounted how he designed the framework with fine-grained reactivity, using very basic syntax and minimal versus expressive syntax.

“It has orthogonal APIs that complement each other and there’s a single way of doing things,” he wrote. “This could potentially make building apps with it more verbose, but there’s less for the LLM to ‘learn’ and ‘know.’”

It’s based on a toy framework he created called Revolt, with one major change: The framework gets all text node and attribute values by invoking a getter function, he wrote. That allows the framework to access reactive and static values in the same way, he added.

The project is open sourced and available on GitHub.

If you’re interested in other options people are exploring, we recommend checking out the comments on his LinkedIn post.

RedwoodJS Releases RedwoodSDK

The RedwoodJS blog has been quiet since September, but this month it updated with a surprise: The team has created a RedwoodSDK. It’s a full-stack, composable React framework built for Cloudflare.

“RedwoodSDK is the natural evolution of everything we learned while building and supporting RedwoodJS,” wrote Peter Pistorius, who co-created RedwoodJS, on the community’s blog. “Since the early days of RedwoodJS, our mission has been clear: To help developers build and deploy full-stack apps faster, smarter, and with confidence. That mission hasn’t changed — but the platform has.”

There’s a demo available on GitHub.

Pistorius also addresses those who have wondered about the framework, writing that “RedwoodJS will remain fully supported, with active maintenance and updates.”

Anthropic’s New Agentic Coding Assistant

Anthropic released a command line tool for its agentic coding assistant called Claude Code earlier this month.

“Developed as a research project, Claude Code gives Anthropic engineers and researchers a more native way to integrate Claude into their coding workflows,” the Anthropic engineering blog stated.

It’s designed to be low-level and unopinionated, but that also means a learning curve for those new to the tool, so the AI company has created a list of suggestions for using it effectively.

In a company website post, Anthropic told users how to customize the setup for Claude Code, including an explanation of four ways to manage the allowed tools and how to give Claude access to more tools.

Claude doesn’t impose specific workflows, so the post also covers what internal and external teams have identified as successful patterns for using Claude. As part of that, it explained how to use Claude to interact with git and Jupyter notebooks. You can also use it as a linter, the post noted.

Finally, the post explained how to “up level” with multi-Claude workflows, such as having one Claude write code and another Claude to verify.

IBM Open Sources Speech-To-Text Model

IBM open sourced its new speech-to-text model Granite Speech 3.3 8B last week. The lightweight model automates speech-to-text transcription and translation for enterprise use cases. It supports eight languages: English, French, German, Italian, Japanese, Mandarin, Portuguese and Spanish. It can also process large audio files with its 128k context window.

Developers can also use the model to create voice-enabled AI agents that transcribe user voice input and trigger additional actions, enabling a more interactive experience for end users. For instance, in a customer support use case, the model makes it possible for customers to interact with a chatbot using voice commands.

IBM added that Granite Speech, available on Hugging Face, consistently delivers greater accuracy than leading open- and closed-model competitors in testing across several prominent public datasets, surpassing comparably sized models like OpenAI’s Whisper and a larger frontier model, Gemini 2.0 Flash.

The new audio model is built on top of Granite 3.3 8B Instruct, the latest update to IBM’s LLM family.

The post React Adds New Experimental Animation Feature appeared first on The New Stack.

---

To prevent prompt injection attacks when working with untrusted sources, Google DeepMind researchers have proposed CaMeL, a defense layer around LLMs that blocks malicious inputs by extracting the control and data flows from the query. According to their results, CaMeL can neutralize 67% of attacks in the AgentDojo security benchmark.

By Sergio De Simone