Defined term
Dataless LLMs
A language model is not dataless because the weights are public. It is dataless when your conversation cannot be reconstructed after the window closes.
Contrast
What it is not
Turning off “training on your data” is a policy toggle. Logs, caches, eval sets, and support transcripts still exist. Dataless LLMs refuse those surfaces.
Practice
Zero retention at inference
The same thesis as Dataless AI applied to inference: volatile context, randomized eviction of working-set fragments, content-free observability. See the patent.