Semantic Galaxy
What a small model sees when you hand it a name.
What is a model looking at when it reads a word
Anthropic's piece on the Jacobian Lens showed Claude building something like a quiet sense of itself, just from the data it had seen. It got me wondering something smaller. When I type "Hi," what is a model actually looking at? We know the textbook answer, a vector sitting in a cloud of everything semantically near it. I wanted to see that cloud, not just know it's there.
Handing a small model my own name
Ran it through a small model on Hugging Face and gave it "hari." Then the full "Hariprasad." Two pretty different neighborhoods came back.
It's a small model on a small dataset, so plenty of names are probably missing from its world entirely. Not rigorous, just fun, and a little strange to see your own name plotted next to strangers a model thinks live nearby.
Other words have their own neighborhoods too
Kept feeding it words after that. "Prasad" pulls in a cluster of names that end the same way. "Tamil" sits near other South Indian languages, and near a few odd near-matches. "Raga" lands closer to music than I expected. And "fart," because I had to check, mostly keeps company with words you'd rather not see plotted next to your name.
The space is up on Hugging Face if you want to hand it a word of your own and see where it lands.