Roll of weighted dice
"AI is a roll of the dice" - people, all the time
I don't like this trope and I hear it a few times a week.
Is using AI, specifically autoregressive models, like a roll of the dice?
Not really. That would imply an evenly distributed randomness in outputs given a common input.
Stochastic? Yes. Random? No.
Using AI is more like "a roll of weighted dice." We set AI up for an outcome we'd like but there's always a bit of chance in there.
We add weight through training data selection, fine tuning, context engineering, and prompting techniques. All of that will increase the odds "our number" comes up. Doesn't guarantee it. But it does systematically affect the outcomes over time.