SIGNAL//DESK
AI/MLsrc: Roost FA glossary v2

Non-determinism

Same input → DIFFERENT output. That is the AI half of the 98/2. When you ask Roost the same question twice you can get two different replies — because the next word is sampled from a distribution, not computed from a rule. This is a feature, not a bug. It is also why AI cannot live alone on the regulated paths.

A characteristic of generative models where the inference process involves sampling from a probability distribution rather than selecting the single highest-probability token. This variability is intentionally preserved to ensure output diversity, distinguishing it from the deterministic logic found in traditional software stacks.

A property of a stochastic computational system wherein the mapping from input to output is not a fixed function, but a conditional probability distribution. By sampling from the model's output logits—often modulated by temperature or top-k/p parameters—the system produces non-identical outputs for identical inputs, a fundamental requirement for generative fluidity that necessitates architectural separation from deterministic execution layers.

evolution

  1. 1956 · history
    Dartmouth Workshop

    Early AI research established the use of probabilistic models and random search strategies to simulate intelligent behavior.

  2. 1988 · history
    Probabilistic Reasoning in Intelligent Systems

    Judea Pearl formalized Bayesian networks, shifting AI from rigid symbolic logic to uncertainty-based inference.

  3. 2014 · history
    Generative Adversarial Networks (GANs)

    Ian Goodfellow introduced a framework where stochastic sampling became central to generating novel, non-deterministic data outputs.

  4. 2017 · history
    Transformer Architecture

    The 'Attention Is All You Need' paper popularized temperature-based sampling, making non-deterministic text generation a standard feature of LLMs.


← all terms