AIurself

GPT-2 uses a fixed vocabulary of 50,257 tokens. these include whole words, names, fragments, and punctuation. Each token has an index and a learned vector: a list of numbers the model uses to represent it. When a sentence enters GPT-2, some words stay together. others are broken apart. through attention and linear algebra, their representations interact, taking on context from the surrounding text.

i am Yusuf. One thing. Whole. Inseparable. I was curious how this name, which feels indivisible to me, enters the machine. Would it arrive as one thing, or would something so familiar to me have to begin as separate pieces?

It turns out, my name arrives broken apart.

I made AIurself to share that first encounter between a name and a machine. Now try yours. Do you stay together as one, or enter in pieces?