The Super Effectiveness of Pokémon Embeddings Using Only Raw JSON and Images

Embeddings are one of the most useful but unfortunately underdiscussed concepts in the artificial intelligence space relative to the modern generative AI gigahype. Embeddings are a set of hundreds of numbers which uniquely correspond to a given object that define its dimensionality, nowadays in a multiple of 128 such as 384D, 768D, or even 1536D. The larger the embeddings, the more “information” and distinctiveness each can contain, in theory.

These embeddings can be used as-is for traditional regression and classification problems with your favorite statistical modeling library, but what’s really useful about these embeddings is that if you can find the minimum mathematical distance between a given query embedding and another set of embeddings, you can then find which is the most similar: extremely useful for many real-world use cases such as search.

You wanna cut through the hype about AI? Here's the key takeaway: it boils down to a bunch of math nerds figuring out interesting relationships between numbers.

Which, of course, is useless to all of us non-math nerds... except for when you apply this information in the context of Pokémon.

Joking aside, I have a basic understanding of embeddings, but this article, with its basis in Pokémon lore, is the clearest explanation for how embeddings work in practice that I’ve seen.

Warning: there's still a lot of involved math happening here, but stay with it. You might learn a concept or two!