pmaze

I ended up judging where to draw the line. Its initial suggestions were genuinely useful and focused on making the basic tool use more efficient. e.g. complaining about a missing CLI parameter that I'd neglected to add for a specific command, requesting to let it navigate the topic tree in ways I hadn't considered, or new definitions for related topics. After a couple iterations the low hanging fruit was exhausted, and its suggestions started spiralling out beyond what I thought would pay off (like training custom embeddings). As long as I kept asking it for new ideas, it would come up with something, but with rapidly diminishing returns.

The names & descriptions definitely have that distinct LLM flavour to them, regardless of which model I used. I decided to keep them, but as short as possible. In general, I find the recombination of human-written text to be the main interest.

There's two stages to the linking: first juxtaposing the excerpts, then finding and linking key phrases within them. I find the excerpts themselves often have interesting connections between them, but the key phrases can be a bit out there. The "fictions" to "internal motives" one does gel for me, given the theme of deceiving ourselves about our own motivations.

I did, there was a first round of UMAP to 50 dimensions. Running HDBSCAN on the full embeddings gave bad results, lots of singleton clusters.

The crash was indeed not intended - my mistake! Should be fixed now.

You've got the cluster semantics spot on, to be honest. Broad genres are grouped together, with a tendency for sub-genres to be grouped locally within those.

There is no interpretation of the overall shapes or the global structure, those are more a result of a particular UMAP run than inherent in the data.

Would love to provide different views on it and go more in depth next, thanks for the suggestion.

Hey, thanks for reporting - this is fixed now. I messed up the static build and some browsers freaked out. By law of showing things publicly, I of course only tested in a browser that didn't. Hope you can give it another chance!

My apologies for that! First time deploying Svelte Kit to Cloudflare Pages, and I messed up the static build. Should be fixed now, hope you can give it another shot.

Thanks!

The cluster memberships that come out of the first round are distributions over the different clusters, e.g. a given book is weighted 0.8 for cluster A and 0.2 for cluster B. The Hellinger distance is well-suited to quantify the difference between two distributions like that. Cosine similarity and Euclidean distance worked as well, but Hellinger gave subjectively nicer results.

Very interesting question, I'm not sure! While developing, I noticed that the systems thinking books were spread over different genres, which I found quite pleasing. However, I'm not sure if other books were even more diffuse. I'll have to dig back in and find out :)