One of the problems when building applications with ML, Deep Learning, LLM is not much the stochasticity. Once trained most models are not stochastic, think of convNets, even if you set the temperature parameter of a LLM to 0 it should be deterministic, some models are intrinsically stochastic, the simple one being Variational Auto Encoder where you sample from a latent probability distribution in order to generate content. One issue comes form the non linearity of these models, given a similar input (with a tiny difference) there are some chances that the output will differ a lot (think of Adversarial examples). The second issue is related to the training data vs the space of all possible inputs that might come in. The training data is always never representative enough of the use cases…