A Photograph and a Sentence Now Share an Address

It is past midnight in Nakameguro, the terminal is open, as it always is, and I have just asked a 740-million-parameter model to look at a photograph of my coffee cup and read the sentence "something to hold on a cold morning," then tell me, honestly, whether those two things belong anywhere near each other.

EmbeddingGemma 2 doesn't answer in words — it answers in coordinates. Every input, whatever shape it arrives in, text or photo or a few seconds of audio, gets reduced to a single point in a 768-dimensional space, and after that the only thing that matters is distance. Things that mean the same thing land close together. Things that don't, don't. That is the entire mechanism, and it is less mysterious than it sounds once you have watched it happen to your own coffee cup, on your own machine, at an hour when nothing else is awake to argue with you about it.

This is the layer underneath "search" that never had to match a word to a word. The developer guide walks through mapping text, code, images, video and audio into that one unified space, which is the quiet trick behind retrieval that actually understands what you meant instead of what you typed. The part I keep coming back to is the size: 740 million parameters is small enough to run entirely on the machine in front of me, no API call, no cloud bill, nobody else's server deciding what my coffee cup means. A year ago the honest version of this demo needed a rack somewhere else; tonight it needed a laptop and a cup of something that had already gone cold.

So I shipped the only review I trust, which is a thing built with it: a search box, overnight, that lets me type a sentence instead of a filename and find the right photo anyway, out of a folder I have never once bothered to name properly. It worked on the third try, after I had, predictably, embedded the sentence and the photo with two slightly different preprocessing steps and watched them land nowhere near each other. Fixed, matched, closed the laptop — which in this job counts as a good night.

The cup and the sentence about it sit, this morning, closer to each other than to anything else in the room — two unrelated things, agreeing, in a space with seven hundred and sixty-eight directions and no name for any of them.