Inference
The act of running a trained model to produce output — as opposed to training, which is the earlier one-time process of creating the model.
In AI, "inference" is the runtime step: your input goes in, the model processes it, output comes out. Training is what happened before — the slow, expensive process (often months of compute) that produced the model's parameters. Inference happens every time you send a chat message; training happens once, then again months later for the next model version.
Inference is expensive at scale because each token requires the model to run all its parameters through the input. For a 400-billion-parameter model producing a 500-token response, that's 200 trillion parameter reads. This is why AI companies invest heavily in specialized inference hardware (GPUs, TPUs, custom chips) — even at the per-query level, inference is what determines the unit economics of a subscription like LADLE.
The word gets used loosely in some contexts to mean "the AI's reasoning" (as in "the inference stage where the model concludes X"). That usage is less technical; the strict definition is the mechanical one above.