Firstly, non-determinism isn't nearly the gotcha you think it is - it's trivial to make an LLM deterministic by just passing one parameter to the inference engine.

Secondly, nothing is non-deterministic in a strict sense on any traditional computer architecture (including the ones LLMs run on). Given all the inputs, and all the algorithms I can recreate the output; what makes the output 'non deterministic' is the complexity of the input and the feasibility of doing that. Frankly, compilers come pretty close to reaching the necessary complexity to be practically non-deterministic; anyone who had never spent days banging their head against code that works perfectly with debug enabled but fails utterly when compiled with release optimisations, but only when two logically entirely unrelated pieces of code happen to be written in a particular order, and then only if interrupts are enabled on the target... Well, anyone who hasn't been through that has no business calling themselves a software engineer - and certainly no business telling other software engineers what tools they are allowed to use.

Honestly, a bunch of wets brought up on PHP and Javascript and managed runtimes now thinking they're qualified to dictate what Real Programming entails is just laughable.