Super cool! I admit my ignorance regarding state space models, but could this architecture be adapted for "standard" masked language modeling and zero-shot fitness prediction?