Hacker News Reader

determinism isn’t really enough, we want “predictable”. Most of these AI wavefunctions are “chaotic” - tiny changes in state can cause wildly divergent outcomes

A part of my question that you didn't go into was, can new knowledge be added in a new version without making the answers with knowledge learned in previous versions non-deterministic?

that’s not really how training works.

changing the input (data) means you get a different output (model).

source data has nothing to do with model determinism.

as an end-user of AI products, your perspective might be that the models are non-deterministic, but really it’s just different models returning different results … because they are different models.

“end-user non-determinism” is only really solved by repeatedly using the same version of a trained model (like a normal software dependency), potentially needing a bunch of work to upgrade the (model) dependency version later on.

This requires an exact lock-down of things like the hardware and driver version, doesn't it? Is that sustainable?

It shouldn't. It didn't used to, at least.

But that won't survive an upgrade, will it?