AI Daily: AI Agents Finally Get Their Own UI Language? Google A2UI and Anthropic Bloom Lead a New Development Wave

The AI landscape has been buzzing lately, with both underlying protocols and everyday tools undergoing a transformation. If you’ve felt that AI Agents have been stuck—unable to do much beyond typing in a chat box—Google’s new A2UI protocol might be a game-changer. On another front, Anthropic has open-sourced Bloom, a tool designed to take over the tedious “bug-hunting” work that previously required massive human effort.

These developments suggest one thing: we are one step closer to a future where we can get everything done just by speaking.

No Longer Just “Chatting”: Google A2UI Reshapes Interaction Logic

Interacting with AI can sometimes be frustrating. You want a button to check out or a form to fill in, but the AI just spits out a long paragraph of text, forcing you to go elsewhere to complete the task. This is highly inefficient.

The Google development team has addressed this “all talk, no action” mode with A2UI (Agent-to-User Interface). This open-source project aims to set the industry standard for Agent-Driven Interfaces.

Simply put, A2UI gives AI Agents the ability to provide the most appropriate UI interface based on the chat context. This isn’t just about throwing crude HTML code; it uses a declarative format. This means the same AI-generated interface can run on the web, in a Flutter app, or even on future new devices, all while maintaining a native feel. In the current v0.8 version, Web Components, Angular, and Flutter are already supported.

I believe the brilliance of this technology lies in two areas:

First is the trust issue. In a future web where multiple AIs collaborate, if an external AI directly passes JavaScript code to your main program for execution, it’s like giving your house keys to a stranger. A2UI cleverly chooses to pass pure data (JSON). The main program is only responsible for rendering the data and never executes foreign code. This directly solves the biggest security headache in cross-organizational collaboration.

Second is Incrementally Updateable. This is crucial for the user experience. Imagine you’re filling out a form, and the AI notices you’ve changed your budget; it only needs to quietly update that price field. This real-time fluidity, achieved through Server-Sent Events (SSE), is what makes an AI application feel like “proper software.”

For those interested in the code, check out their GitHub or read the Google Developers Blog.

Anthropic Bloom: Fighting Magic with Magic

AI safety testing is, frankly, a chore. Researchers have to think of all sorts of tricky questions to test the model’s limits. But in 2025, models are evolving faster than humans, and relying solely on human brains to find these “traps” is no longer enough.

Bloom functions like a rigorous forensic team. You only need to provide a “Seed Configuration”—the DNA of the test case—and Bloom will automatically grow various mutations based on it. Its workflow is as follows:

1. Understanding: Figure out what flaws we need to test.
2. Ideation: Design various conversational traps that are hard to defend against.
3. Execution: Interestingly, it doesn’t just test dialogue; it also supports a Simulated Environment. This means it can observe if an Agent does anything malicious while writing code, using tools, or performing long-term tasks.
4. Judgment: Finally, another model is used to score the results.

To prove this isn’t just “self-praise,” Anthropic even developed a Meta-judgment mechanism, using AI to monitor the quality of the AI’s scoring. More interestingly, they intentionally created “Model Organisms”—somewhat like mice in medical experiments—to verify if Bloom can truly catch flaws. This scientific rigor fits Anthropic’s established persona. More details can be found on their official blog.

Gemma Scope 2: Opening the Black Box

Google DeepMind has taken another bite out of the “interpretability” challenge. They released Gemma Scope 2, which is essentially a high-powered microscope for the Gemma 3 model family, covering the full parameter range from 270M to 27B.

We often say neural networks are black boxes—we know the input and output but not what happens in between. Gemma Scope 2 uses Sparse Autoencoders (SAEs) and transcoders to try and turn this black box into a transparent glass box.

This update is technically significant, with two highlights:

First, the introduction of Matryoshka training technology, which helps the model detect more precise and useful concepts.

Second, the addition of Skip-transcoders and Cross-layer transcoders. This allows researchers to go beyond single-layer slices and track how information jumps and flows across layers in complex neural networks.

This scale is said to be the largest in the open-source community, processing up to 110 PB of data. If you’re interested in “what AI is really thinking,” this DeepMind article is worth reading.

NotebookLM Heart Transplant: Gemini 3 is Live

This is probably the best news for note-taking enthusiasts. Google’s note-taking tool, NotebookLM, has finally switched its engine to Gemini 3.

The official X account had previously hinted that this was the most requested feature. With the new engine, the most noticeable improvements should be in reasoning and the ability to “read the room.” When processing complex documents of several hundred pages or performing cross-document correlation analysis, there should be fewer instances of “hallucinations.” Official announcement here.

Developer’s Toolbox: New Toys for Codex and Qwen

In addition to the big news, there are two interesting small tools:

• OpenAI Codex CLI supports Skills: Coding is annoying when you’re constantly reinventing the wheel. OpenAI added a Skills feature to the Codex CLI. It’s designed thoughtfully, using Progressive Disclosure—it only shows a directory at startup and loads details when needed. This helps save precious Context Window space. Documentation link.

• Qwen-Image-Layered Model: The Alibaba Cloud Qwen team developed an image model that can “peel an onion.” It doesn’t just generate images; it can decompose them into multiple independent RGBA layers, enabling physical-level isolated editing. Even more impressive is the support for Recursive Decomposition. Imagine taking a person out of a photo and then continuing to separate their clothes and hair, theoretically infinitely. This Matryoshka-style editing capability offers many possibilities. Try it on HuggingFace Space.

Industry Dynamics Full of Gunpowder

Finally, two more serious news items.

Google vs. SerpApi: This lawsuit was bound to happen. Google has officially sued SerpApi, accusing the scraping company of using Cloaking technology and deceiving servers with constantly changing fake names and IPs to bypass protections. Google is truly furious because SerpApi doesn’t just scrape public data; it even resells content that Google has paid for under license (such as Knowledge Panel data). This is no longer simple “data scraping” but a direct hit on commercial interests. The outcome of this lawsuit could rewrite the rules for the scraping industry. Google statement.

METR’s Stress Test on Claude Opus 4.5: METR Evals released data estimating that Claude Opus 4.5 has about a 50/50 success rate for a complex task taking nearly 5 hours. But the devil is in the details: the 95% confidence interval they provided is startlingly wide—ranging from less than 2 hours to over 20 hours. What does this mean? It means that for such super-models, we don’t yet have a precise enough ruler to measure their limits. METR data.

FAQ

Q: How is A2UI different from just outputting a piece of HTML code? A: There’s a big difference. Besides being safer by transmitting pure data, A2UI’s strongest feature is incremental updates. Imagine the AI just flipping a switch or changing a number, and the interface reacts instantly, rather than clumsily re-rendering the entire page. This native app fluidity is something traditional HTML output can’t provide.

Q: Is a tool like Bloom useful for ordinary developers? A: Honestly, it’s mainly for those doing AI safety research. You have to write a Seed Configuration to define the “genes” of the test. The barrier to entry is high, but if your team needs to ensure a model absolutely cannot exhibit a specific bad behavior, it’s a powerful automated tool.

Q: Does upgrading NotebookLM to Gemini 3 cost extra? A: Google hasn’t mentioned money. Usually, these underlying model upgrades are platform optimizations. Just think of it as a free performance boost and use it with peace of mind.

Q: Why is Google so intent on suing SerpApi? Isn’t scraping common? A: The nature of this case is different. Google accuses SerpApi of using cloaking technology to deceive servers and reselling licensed data that Google paid for. This has crossed the line of “public data scraping” into malicious circumvention of security mechanisms and copyright infringement. If Google wins, life might get much harder for AI data collection companies.