Re: [spam][ot][spam][ignore] Enlrge Ur Pen1s

[okay i'm imagining like a small transformer that could learn a feature if given data to do so. how do we bind it to the utility for predicting?] [so maybe 2 small transformers? one to predict other data, and one to produce a feature useful for predicting it] [maybe! you could start with really small transformers so more use structure emerges.] [you'd want something about the location?] [this could start having a lot of models. i do like the idea of taking apart a deep transformer model and turning it from a single stack into a useful graph.]