The last month has been “Radical Speed Month” at Automattic. Unlimited AI token budget, no bureaucracies, and no constraints. Pick something to work on, pick a partner, build and ship something in one month.
I paired up with James Grierson, a long-time coworker and someone I trust to help with big ideas. And our idea was big! Choosing to help envision the foundation of the future of the open web is no small task, especially within a month. But I’m happy to announce we’ve done just that! Let explain the idea, the philosophy, and what we shipped.
Idea and philosophy
The open web philosophy has always been about a decentralized, accessible, and user-empowered Internet. I can’t pretend to know what the future of the open web will actually be, how it will function, or what it will look like.
But I do know a couple things. It will be about trust. It will be about controlling your own algorithm. It will be about privacy. And we must demand transparency from the big-boxes before we give them our data.
As an open source community, we have to get going on this! We have to actually get started somewhere. The ideas we’ve explored here may or may not be the actual future. But it’s a starting point. Who better than Automattic and WordPress to help establish what the future of the open web should look like?
A bold assertion
I hereby declare that the website of the future should work a lot like the website of today works! A user will build a site on a web host of their choice, create or use a design (perhaps all AI empowered!), and they’ll publish content on pages and posts.
The radical change comes in at the interface to that website. As a consumer of the content on the web, augmented by AI, users will demand both control and privacy. That control will likely be enacted in many ways, but I’ll introduce you to the first way.
The “User Model”
Are you a text skimmer who likes to click on links and go down rabbit-holes? Do you prefer humor when you read posts, or do you like long form no-nonsense content? Why is it up the publisher to demand that you consume content in their style of writing?
Side note – but doesn’t this kill soul of the art of writing somehow? Yes. I think so! That’s something we have to figure out. But ultimately, the control should be in the hands of the user, the consumer, the reader. Maybe we give them the option to consume certain brands natively. I certainly don’t have all the answers.
Imagine a small, portable, machine interpretable model – a description of you, your interests, your geographical location, and your content consumption preferences. You generate this model on-device, and provide it (or certain parts) only to the services you trust – and/or – it gets used for local inference on content served to you by the website.
The user model is both binary-packable in a tight specification, or JSON-packable for web portability. It contains three main sections
- The fingerprint – metadata, checksum bits, model versioning.
- Style vectors – 8-bit integer packed spectrums for preferences about how you consume content – things like “serious vs humorous“, “skimmable vs immersive“, or “simple vs technical“.
- Interest embedding strands – a list of floating point numbers in n-dimensions used to describe your interests, packed with a header that describes both generating model and length, so that the user model can be portable in strict binary.
The model is open-specification, open-source, uniform, easy to generate and easy to consume and apply. I’ll open-source the specification we developed soon, likely as part of the Aurora project detailed below.
User Model Applications
We spent a lot of time brainstorming uses for the user model. I’ll list out a few that we tinkered with here.
Aurora
Aurora is the codename for a “web-browser-like” OSX native application that I built, which can generate the user-model on device, and provide it to services selected by the user. The model also gets used to interpret the user’s behavior on sites they visit. The model would theoretically be self-updating, so that it learns your preferences over time and tailors the experience of using the web to your liking to make you more productive or entertained.
Again, the model is generated and lives on-device. LLM inference can even take place directly on-device. The app is currently capable of loading both local LLM models and Nomic embedding models directly on the machine. No dependency on OpenAI, Anthropic, or Google for inference for private data and preferences.

Just imagine for a moment visiting a site like Facebook or (insert your “news” site preference here) and having the site actually deliver you the content you want to see, in a format that’s best for you. Maybe you’re on the move with your earbuds in, and you want that content delivered in podcast style. Maybe the AIs of the future can turn that content into shortform video in near realtime and cast to your TV. These are some of the ideas and goals here.
Future web host compatibility
Is it possible that in the future, your user model will be binary-packed and delivered to websites you trust, and then used for on-the-fly inference against the content that the site is delivering to you? That gives the user control and flexibility with their preferences and content interests, but also allows web hosts or other web sites the flexibility to do things like deliver advertisements, or special features tailored for AI users.
I thought deeply about implementing this into Jetpack as a possible service, but with only a month to work on this, I didn’t have time to materialize it. But this is a possibility for the future – providing this functionality to websites without their host needing to be compatible. This could be done now.
Content selection and readability
For content aggregators, like the WordPress.com Reader (one of my favorite Automattic creations!), imagine finding highly relevant blog posts without having to dig so hard for the right blogs. Imagine seeing relevant content without having to subscribe to a bunch of blogs. And then imagine that content being delivered to you summarized to your liking. That’s the future of the web. And I do believe that this concept of a “user model” will be a cornerstone of it. It’s technical and complex right now, but mark my words, AI will make it transparent and easy over the coming years.

What we shipped
- reada.blog – The advertising/marketing site for the Reada.blog app.
- my.reada.blog – A web app for interacting with the WordPress.com Reader, but with user-model integration and a Smart Feed which selects and tailors content for you.
- Reada.Blog App – We built Reada.Blog as a native app for iPhone, but unfortunately the Apple store is very backed up with submissions, so we aren’t shipping this yet. Instead, we shipped the web app, which is less ideal, but embodies the proof of concept we aimed for.
- Aurora – I’m not ready to open-source this yet, but will be doing so as soon as I’m ready. My gut tells me this is the best example of the future of the web, so I want to get it right. I’m going to continue to work on this in my free time, and will aim to release it as an Automattic product by the end of the year.
Conclusion
This project was ambitious, and I’m not going to lie – getting something shipped was way harder than I thought it would be. But I’m proud of what we did ship, and I’m so excited to continue this work in some capacity.
I believe in the open web. It made me who I am. And I simply cannot fathom a future for my children where they are only distracted, entertained, and sold to – instead of educated and empowered, like my generation was. This is something we all should be fighting for.







































































































