A Santa Claus Machine is something that can take any input and give you any output.

Something I’m thinking of is that although LLMs are just text right now, we should be training them on everything! Bring back Gato, but for more interesting things.

I really liked this talk on UI Transducers and how LLMs are almost text santa claus machines. If this is true, then we should be trying to expand their capabilities by training them on more esoteric things, but we should also be coming up with ways to integrate them into the UI in more cohesive ways.