The Minter is a 2.86M-parameter model (11 MB) that invents Pokémon that never existed: a coherent typing, weaknesses that follow from it, moves, a colour, a habitat, for an IRI it has never seen and cannot look up.
The Memoriser was trained on 901 real, stable IRIs, so memorising id-to-facts was the optimal thing for it to do, and it snaps any new name to the nearest stored row. This model was trained on species whose IRIs are minted fresh and never repeated. Memorisation buys nothing when the key is new every time, so the only thing that lowers its loss is learning what a Pokémon is: it has one or two types, its weaknesses follow from them, a Fire-ish thing knows Fire-ish moves.
It is the same decoder, 15× smaller, and it does the thing the big one can't. The architecture was never the problem. The incentive was.
You'll notice the model rarely keeps the name you typed. Ask for zapfrog and it
might dream up pyroeon instead. It is the same snap-to-what-I-know reflex the
Memoriser has, except here it snaps to the shape of a plausible name rather than to a stored
row, then invents a fresh animal to go with it. You suggest, it dreams. So the demo keeps
whatever creature the model settled on and tells you both names.
Unlike recall or derivation, invention needs temperature. Greedy decoding would emit the single likeliest move over and over and never close the entry. The wildness slider is that temperature: low for a plain, typical creature, high for stranger typings and rarer moves. Each press reseeds, so you can reroll the same name into a different creature.
weakTo set actually follows from the hasType lines.