This Is Getting (Rule)diculous

submitted by
She/her/they
Top text reads: "My INCREDIBLE new A.I. model uses ZERO electricity. Looking for investors." Below is a picture of the magic 8 ball novelty toy.

Just as accurate as the leading brand!

14
446

Log in to comment

Hot Top New Old

Woah. Infinite prompts, built in security for the model, and a unique interface?!

 reply
36

Makes me remember that mini comics or meme, where a dude asks the ball “am I stupid?” but there is no answer coming up because it’s a bowling ball.

 reply
17

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…

 reply
8

octree??

 reply
1

yea - an octree - se tree structure where u walk foraard and find “more detail” along the path. LMs essentially do the same, but each branching point is a token and each token has about 30.000 different possibilies - for different branches (thats the models “token vocabulary") so yea-… “octrees”…. sure exist.

anyway - have a nice day ~ ~ ~<3

 reply
2

huh! thought octrees could be used to search token space for the closest(s) tokens, i had never seen the word “octree” in a machine learning context.
also, i think that’s just called a “tree”, “oct-” means “eight”

 reply
1

oooooooh ir right. sili meeeeeh

 reply
2

also octtrees are specific to three dimensions, i did not register that and thought it was possible to apply to high dimensional spaces such as token embeddings

 reply
1

I had to do a double take cause I thought the thing in the 8 ball said “I hate my do*”. I couldn’t read the last word so the do at the front threw me for a loop

 reply
5
Insert image