New here?
Create an account to submit posts, participate in discussions and chat with people.
Sign up
46
Every single time (media.scored.co)
posted 1 year ago by RealWildRanter on scored.co (+0 / -0 / +46Score on mirror )
9 comments
Permalink
Mirrored from scored.co
You are viewing a single comment's thread. View all
0
IGOexiled on scored.co
1 year ago 0 points (+0 / -0 ) 1 child
Yeah. So, a prompt, the thing you type and hit Enter... those can be thought of as a complicated shape. And the LLM can be thought of as a Pachinko machine. The prompt bounces down the board of pegs and eventually lands in a cup. That cup is the answer it gives you.

The illusion is that because answers are long, it seems like there couldn't be that many cups and it's doing something more complicated. Nope. Pachinko.

The shape of your question causes it to bounce one direction preferentially vs. others.

Kike Meta may block off sections of Pachinko track, or cover up some cups, or tilt the board at an angle.

But it's all just shaking the dice.
Permalink Reply
Mirrored from scored.co
0
RealWildRanter on scored.co
1 year ago 0 points (+0 / -0 ) 1 child
It's all a trick to make you think it's a roll of dice when in practice computer programs are deterministic, and have extreme difficulty generating true randomness. It's the gas lighting and the profuse use of circular logic that gives that up. But normal users are too dumb to recognize these tricks. They give it some input, and after a quarter of a second an output comes up, and they become impressed or happy that can carry on being lazy because they realized the machine can do the work for them. Let's not talk about coding as AI is barely batter than a horde of pajeet monkeys typing on keyboards until the compiler stops complaining.
Permalink Reply
Mirrored from scored.co
1
IGOexiled on scored.co
1 year ago 1 point (+0 / -0 / +1Score on mirror )
It's interesting to set the seed and see the differences in responses due to to the RNG functions. (Which you are correct, are not truly random)

You can really get a feel for the topography of the model. Either try same prompt with lots of seeds or many prompts with the same seed. Compare responses.




Permalink Reply
Mirrored from scored.co
Toast message