AI is actually deterministic, a random input is usually included to let you get multiple outputs for generative tasks. And anyway, you could just save the “random” output when you get a good one.
Maybe deterministic wasn’t quite the correct word but basically it only gives you a result that resembles your previous result if you change absolutely nothing, not the training data for the model, not the model, not the random seed, not the prompt,… which makes it useless for iteratively approaching a usable result. I guess the output space is not contiguous might be a better way to describe it.
Oh, I said that as a programmer all right. And that’s how I’ve approached AI - I ran it locally, and kept poking it until I began to get a feel for it. Until I could see patterns. Until I could put together a methodology
They exist. Word choice matters greatly. Shorter is better. Varied word choice is better. Less “orders” is better. Strange combinations of tokens can convey something in non-obvious ways. They all seem to have a very strong attachment to the name “Luna”
They’re as deterministic as any software is, if you run it in the same state with the same input you’ll get the same result, sometimes with minor wording changes
And software isn’t as deterministic as we pretend it is. Programming doesn’t require it either, luckily. Every program you’ll ever write is interacting with complex systems no one fully understands, and it will sometimes act unpredictably
Programming is about finding patterns in the chaos, then using them to get the result you want. You need consistency - not deterministic outcomes. You can program with anything you can find the patterns in - even human behavior or the physical world. You can program yourself.
You can treat AI like something unknowable, or you can find the patterns and put them in your toolbox
As a programmer I can tell you that AI is nothing like programming because programming is deterministic and repeatable and AI is anything but.
AI is actually deterministic, a random input is usually included to let you get multiple outputs for generative tasks. And anyway, you could just save the “random” output when you get a good one.
Maybe deterministic wasn’t quite the correct word but basically it only gives you a result that resembles your previous result if you change absolutely nothing, not the training data for the model, not the model, not the random seed, not the prompt,… which makes it useless for iteratively approaching a usable result. I guess the output space is not contiguous might be a better way to describe it.
Oh, I said that as a programmer all right. And that’s how I’ve approached AI - I ran it locally, and kept poking it until I began to get a feel for it. Until I could see patterns. Until I could put together a methodology
They exist. Word choice matters greatly. Shorter is better. Varied word choice is better. Less “orders” is better. Strange combinations of tokens can convey something in non-obvious ways. They all seem to have a very strong attachment to the name “Luna”
They’re as deterministic as any software is, if you run it in the same state with the same input you’ll get the same result, sometimes with minor wording changes
And software isn’t as deterministic as we pretend it is. Programming doesn’t require it either, luckily. Every program you’ll ever write is interacting with complex systems no one fully understands, and it will sometimes act unpredictably
Programming is about finding patterns in the chaos, then using them to get the result you want. You need consistency - not deterministic outcomes. You can program with anything you can find the patterns in - even human behavior or the physical world. You can program yourself.
You can treat AI like something unknowable, or you can find the patterns and put them in your toolbox