yessss im going to be digesting this idea for a little while this is a great reminder thank you

cause like then it'd be even more about the geometry of the space and you can re-center / rotate the state some how to show that it is the right shape... um i dont have a real world analog for this that doesn't make me sound really really schizophrenic

in attached paper instead of determining the origin of "a text" like i am it determines the origin of "the answer to a prompt" and the generator generates an answer to the prompt

last night though i was toying with Fourier transforms and thinking maybe we should try to manufacture specific attractors instead of reach specific states... maybe that's "ignore till we get there" though

i hope u see what i'm saying n don't think i'm on "we should train an agent to navigate a space," we should train an agent to navigate any space while only seeing a low dimensional projection of the space, only applying forces to the environment in a low dimensional action space

yeah okay i'll put a pot of coffee on and try to get it running cause now i'm worked up about it again

obv with a traditional reservoir it's different than "reached" vs didnt, we'd have some threshold

i think though nearly Everything in the real world is reservoir-y right so navigating reservoir spaces would be super useful

you would or wouldn't give the navigator the exact state if you wanted it to be the most useful, like it sees how close it is to the right state half of the time and has to deduce that itself from seeing the current state and the target the other half?

i think im overthinking it because this seems like something i could just implement right now and test in like google colab
i think an intelligent agent like this could also be used to modulate specific forces to confine plasma in new ways

you'd make sure the states _are reachable_ but _hard to reach_ so maybe +1 for proposer for "navigator eventually reached" and then proportional reward to the length of the trajectory that the navigator needed? navigator gets rewards for reaching and inverse other term?

im still trying to develop this idea outside of code as much as i can before going into the code (which could be the wrong approach but whatever..) like do we just try for novelty? i think it'd have to be a self-play with proposing a state to reach and then reaching it

discriminator's output gets assigned positive reward when it correctly determines the origin of a text (generated / real)

generator's output gets assigned positive reward when it fools the discriminator

increase likelihood of high rwd sequences, decrease likelihood of low rwd sequences

text gan with reasoning

for obvious reasons i am not making this dataset public but it is incredibly easy to recreate, yikes! or cool! whatever you think is the right thing for me to say here

if you keep a tight bound on the distance from the original embedding this looks like changes in specific word choice or quirks of grammar

a model trained on this set, along with my text autoencoder "essence" can be used to create a system that autonomously edits a post to maximize the likelihood of a reply from a user and tune specific qualities of their likely reply

we did some scraping and data engineering and have 2 billion tokens to use to model posts that users reply to and their replies to them given their recent activity