The Memoriser is a ~20M-parameter transformer (~80 MB) that learned the entire Pokémon graph by heart. Pick real subjects and it recalls their stored facts, then the page draws them.

What it does

The page hands the model nothing but a type assertion per pick, <…/pikachu> a <…#Species> . Everything that comes back (Pikachu's typing, abilities, egg group, evolutions) is recalled from its weights. Turn the trie switch off and it can hallucinate IRIs. Turn it on and decoding is constrained to a trie of real identifiers, so every IRI it emits is one that exists.

The honest part: it's a database in a neural coat

This model exists to be measured, and the measurement is unflattering in a useful way. The graph gzip -9s to 0.17 MB. The model is ~80 MB, 455× larger than the data it stores, at recall 0.87 against gzip's 1.000. Judged as a database, it loses to a 1970s compression utility, and scaling only makes it worse.

And it is only a database. Ask it for a Pokémon that isn't a row and it snaps to the nearest one by spelling: zarquon becomes zarude, returned with Zarude's real facts. It has no way to represent a creature that isn't already stored. That gap is why the Minter exists: a smaller model that can.

Try it

• Pick Pikachu and one or two friends, and watch a clean, correct little graph assemble, sprites and all.
• Pick four species from different generations. The model still stitches shared types and egg groups between them.
• Untick the trie and pick a big hub like a popular type. Occasionally you'll catch a hallucinated IRI that the trie would have blocked.

← Back to the collection about