I am Alan and I work at Google on A2UI. AI can generate bespoke UI in your design system. A2UI 1.0 just launched.
New features like Express DSL and Macros (templates+code), SDKs and the skill generator make it faster, safer and easier.
It works with system one models, Gemma fine tuned, and any frontier LLM. It works without any model, because it's just declarative UI, made standard across web, mobile, app, wearables and more.
From the demo I saw, I actually really like this idea. I've thought about what generating UI on the fly for users with LLMs could look like, but my concern would always be wondering if the LLM was going to create a UI that's good, let alone usable?
So a few things that's got me a little fuzzy here with this protocol: how do I know the user is getting a UI that works? How do I test this? What kind of undertaking is it for a service to implement this protocol in their system (i.e., OpenAI, Anthropic, etc)? How do I know the model won't accidentally assign some critical functionality to the wrong button press in a generated UI?
And finally... my intuition is telling me, "oh like skills or mcp servers for embedding UI in the chat" - how far off is that thinking?
It seems models are really good at outputing MDX, it's Markdown + JSX after all. You just tell model which Components exist (e.g. Column, Row, Card, Form), show it some examples and it does really good job, even smaller/older models.
OpenAI's new GPT-6 Intelligent UI uses streaming MDX with custom inline language for React style hooks.
New features like Express DSL and Macros (templates+code), SDKs and the skill generator make it faster, safer and easier.
It works with system one models, Gemma fine tuned, and any frontier LLM. It works without any model, because it's just declarative UI, made standard across web, mobile, app, wearables and more.
I'm pretty excited about it.
So a few things that's got me a little fuzzy here with this protocol: how do I know the user is getting a UI that works? How do I test this? What kind of undertaking is it for a service to implement this protocol in their system (i.e., OpenAI, Anthropic, etc)? How do I know the model won't accidentally assign some critical functionality to the wrong button press in a generated UI?
And finally... my intuition is telling me, "oh like skills or mcp servers for embedding UI in the chat" - how far off is that thinking?
It seems models are really good at outputing MDX, it's Markdown + JSX after all. You just tell model which Components exist (e.g. Column, Row, Card, Form), show it some examples and it does really good job, even smaller/older models.
OpenAI's new GPT-6 Intelligent UI uses streaming MDX with custom inline language for React style hooks.
1) A2UI is purely JSON, while MDX embedding JSX creates some security issues.
2) A2UI is a protocol that's implemented natively for the platform, while MDX is ultimately bound to some JS runtime.