The BBC never published a digital edition of the Battle of Waterloo, but in the latent space of a Large Language Model (LLM), that page does exist.
LLMs are famously trained to predict the next word in a sentence. Somehow, as a side effect, they also learn facts, concepts, and relationships. Among those relationships is the link between a URL and its content. This results in something like a compressed model of the web content the model was trained on. A vast, invisible architecture where every URL, real or imagined, exists in a latent space, waiting to be rendered. The URL serves as the coordinate when exploring this space. For every address on the web, the AI has a preconceived notion of what you should find there. If the URL is well-known, the model stays grounded; if it is obscure, the results become fantastic; and if the URL is entirely fictional, the the model generates something entirely new. This echoes Borges’ Library of Babel. The library in the story contains all possible books, you just need to know where in the infinite space it sits. In the story, the order of the books is random, so practically speaking you’ll never find a reasonable book, it’s almost entirely gibberish. The Web of Babel improves upon this notion since web pages already have a URL. It’s a browser that allows you to visit any possible URL in a seemingly infinite World Wide Web, encoded in the mind of a large language model. It’s not fast or cheap. The results stream in, but since modern web pages usually start with a lot of CSS and JS, it still takes quite a while for the first thing readable to appear. And you’ll need an API key for one of the supported providers to play with it. The key is stored on your local browser, so it is safe, but depending on the model and the size of the page, it costs you anywhere from 1 to 25 cents. Web of Babel comes with a number of pre-generated pages that allows you to explore the concept. For example, to see what’s new in the world of tech, visit the hallucinated hacker news page. Planning a trip to Berlin, check out the imagined lonely planet page. Check out the generic google calendar to see what’s happening this week - don’t forget to call Mom. Generic Elon Musk on X is hard to tell from the real product. Leaning more into hallucinations, GeoCities is still alive on the Web of Babel and ready to answer all your crystal related questions, complete with blink tag. We can even do some form of time travel. As that Waterloo example suggests, the system can be surprisingly specific. Because BBC News URLs encode the date, we can browse the edition from June 18, 1815. An interesting detail is the weather widget showing heavy rain, which according to some helped defeat Napoleon. Remarkably, the version of Google embedded in the Web of Babel works reasonably well. You can search for the Library of Babel and get results quite comparable with what the real Google would show you. No AI summary on top yet though. Taking self reference one step further, you can visit the Web of Babel page inside the Web of Babel. It returns a page vaguely fitting with the rest of my projects, possibly more oriented towards the library than the web, but still, it’s a pretty good fit: “part wonder, part joke, and part meditation on how much of the web is really indexing rather than reading.” I couldn’t have said it better than my virtual self did.
