This is my blog drawn as a network, with a light coat of paint to make it look like a video game map. Each city is a post, bigger cities have more incoming links, and roads are connections between posts. It updates semi-automatically every time I publish. (I used an LLM to write the code and drew about 90% of the art myself in Inkscape.)
For a long time I've been unsatisfied with the linear feeds and tag soups that most blogs use, so I decided to experiment with a non-linear view. I built some CLI tools that combine embedding similarity, LLM suggestions and human review to find the links between posts and quantify their strength. Now, every time I write a new piece, it takes me about 5 minutes to generate its links to/from older posts and grow the map accordingly.
The map is meant to look playful more than precise. For me, the key feature is how it enables making your own "journey" through the blog. The tricky part was finding a layout where closely related posts end up near each other AND the landscape doesn't change dramatically with every new post. I'm not a graph expert, but t-SNE on shortest-path distances over the link graph turned out to work well in my tests. Each update starts from the previous positions with early exaggeration off, and a Procrustes alignment keeps the map from rotating or drifting.
I hope it helps people get lost in there and find things they want to read. I'd like to see something like this on other blogs, so I'm sharing it here. I wonder if anyone else has tried similar ways to un-linearize their blogs.
This is my blog drawn as a network, with a light coat of paint to make it look like a video game map. Each city is a post, bigger cities have more incoming links, and roads are connections between posts. It updates semi-automatically every time I publish. (I used an LLM to write the code and drew about 90% of the art myself in Inkscape.)
For a long time I've been unsatisfied with the linear feeds and tag soups that most blogs use, so I decided to experiment with a non-linear view. I built some CLI tools that combine embedding similarity, LLM suggestions and human review to find the links between posts and quantify their strength. Now, every time I write a new piece, it takes me about 5 minutes to generate its links to/from older posts and grow the map accordingly.
The map is meant to look playful more than precise. For me, the key feature is how it enables making your own "journey" through the blog. The tricky part was finding a layout where closely related posts end up near each other AND the landscape doesn't change dramatically with every new post. I'm not a graph expert, but t-SNE on shortest-path distances over the link graph turned out to work well in my tests. Each update starts from the previous positions with early exaggeration off, and a Procrustes alignment keeps the map from rotating or drifting.
I hope it helps people get lost in there and find things they want to read. I'd like to see something like this on other blogs, so I'm sharing it here. I wonder if anyone else has tried similar ways to un-linearize their blogs.
Very cool. Is the terrain (coastlines, roads) procedurally generated from the t-SNE positions, or hand-drawn on top of the layout?
It's all procedural. I get some parameters I can tune, like thresholds and how crooked to make the road, but the algorithm does the rest.