Transformers are plenty deterministic. We just decide to run an rng against their output to get the final token in most cases.

when concept A applies to concept B in practice, a lack of theoretical connection between A and B is basically irrelevant to me as a lay person. this perspective applies to the other responders argument too.

Non-associativeness of floating point arithmetic plays a major role as well.

This is not inherently a property of either LLMs or "classical" ML.

I can run a linear regression and get the same result every time or I can go to chat.com and get A/B tested upon. yes these are cherry picked examples. they're also the ones that apply to me personally. for a higher level version of this argument please see my response to the other responder.

linear regression on commodity hardware is generally not going to be truly deterministic delta implementation details