LLM Navigator

I had an idea about creating a browser that doesn't use the internet.

What does that even mean? The idea is seeing whether you could get an AI to dream up a webpage given the URL, and then make that operate like a real web browser.

At the Tokyo Mistral AI Worldwide Hackathon, I revisited this idea and built a version of this. It's now on my vibe-coding sandbox v1 as LLM Navigator.

How it works

When a URL is entered, a prompt is given to the LLM to generate the HTML and CSS for the web page. It is also fed some context like whether the new URL is coming from a click from a previous page (in which the previous page is also fed in as context).

One other little trick is images. I had tried using the image generation models but it soon blew through all my credits. So instead, it does a separate image generation pass but mainly sticks to SVG for the images to keep speed high and costs low.

Examples

If you give it a fictional twitter account, it will try to generate what it would look like. Imagine if William Shakespeare had a twitter account:

Rather than be asking the LLM "what do you know about this?", things can get an bit interesting if it become free form in HTML. It could also make up the year it wants to respond in.