woag - hmmmmm but that looks like a random probability distrobution of just a few tokens… hmmm…,. best case we have it walk thru an octree of all tokens of our own tokenizer and pick the correct ones progressively.

orrrr -… iduno… ummmm… hmm-…,…,.

see - we cant really make… agents happen with this, u see? if we wanted random samples, wed use som semi-random number generator and hook it up to the current time… LMs r time-independant and generate deterministic outcomes using a given text prompt and seed - but your model doesnt have that… iguess ur zero-electricity point is quite interesting tho… hmm-

we couuuuuld have it hooked up to a visual interpreter, shich sees the balls outcome and then picks from a set of predefined sentence blocks, to then create new sentences based off of that. now we r at the problem from earlier tho: we only got that many output states with the die inside.

so we could - theoretically - have this model navigate a given environment using code outout patterns instead of sentences - so… maybee this could be interesting…