Thinking specifically about AI here: if a process does not give a consistent or predictable output (and cannot reliably replace work done by humans) then can it really be considered “automation”?
Thinking specifically about AI here: if a process does not give a consistent or predictable output (and cannot reliably replace work done by humans) then can it really be considered “automation”?
See, that’s the crux of it for me. Something with stochastic elements could totally count as automation, but it has to actually replaces some manual work.
LLMs could be made deterministic, by the way. They just produce the nth best token sometimes instead of the 1st best for the sake of naturalness.
Good reminder that LLM output could be made deterministic!
Though correct me if I’m wrong, their training is, with few exceptions, very much going to be stochastic? Ofc it’s not an actual requirement, but under real world efficiency & resource constraints, it’s very very often going to be stochastic?
Personally, I’m not sure I’d argue automation can’t be stochastic. But either way, OP asks a good question for us to ponder! The short answer imo: it depends what you mean by “automation” :)
Uhh, that actually raises some questions about the definition of determinism. If the order in which it sees training material is generated by a single known seed, for example, does that count? What if it’s a really bad RNG algorithm, or literally just a complex but obvious pattern?
They are a black box, so in a sense the way they’re constructed might as well be totally random.
good points on the training order!
i was mostly thinking of intentionally introduced stochastic processes during training, eg. quantisation noise which is pretty broadband when uncorrelated, and even correlated from real-world datasets will inevitably contain non-determinism, though some contraints re. language “rules” could possibly shape that in interesting ways for LLMs.
and especially the use of stochastic functions for convergence & stochastic rounding in quantisation etc. not to mention intentionally introduced randomisation in training set augmentation. so i think for most purposes, and with few exceptions they are mathematically definable as stochastic processes.
where that overlaps with true theoretical determinism certainly becomes fuzzy without an exact context. afaict most kernel backed random seeds on x86 since 2015 with the RDSEED instruction, will have an asynchronous thermal noise based NIST 800-90B approved entropy source within the silicon and a NIST 800-90C Non-deterministic Random Bit Generator (NRBG).
on other more probable architectures (GPU/TPU) I think that is going to be alot rarer and from a cryptographic perspective hardware implementations of even stochastic rounding are going to be a deterministic circuit under the hood for a while yet.
but given the combination of overwhelming complexity, trade secrets and classical high entropy sources, I think most serious attempts at formal proofs would have to resign to stochastic terms in their formulation for some time yet.
there may be some very specific and non-general exceptions, and i do believe this is going to change in the future as both extremes (highly formal AI models, and non-deterministic hardware backed instructions) are further developed. and ofc overcoming the computational resource hurdles for training could lead to relaxing some of the current practical requirements for stochastic processes during training.
this is ofc only afaict, i don’t work in LLM field.
In practice there really is no incentive to avoid stochastic or pseudorandom elements, so don’t hold your breath, haha. It’s a pretty academic question if you could theoretically train an LLM without any randomness.
Thanks for writing that up, I learned a few things.