Language models and latent spaces AI
An autoregressive language model learns to predict a next token from preceding tokens. Tokens are units such as words, word fragments, or punctuation. Repeated prediction can produce a passage responsive to the supplied context; subsequent training can further shape how a deployed assistant responds.
Parameters are learned numerical values, while representations are patterns of activation used in processing particular inputs. “Latent space” is a useful broad term for learned representations, but it should not suggest a single hidden library containing every potential sentence. A generated passage is constructed through computation, rather than ordinarily retrieved as a complete, pre-existing document.
Models can nevertheless memorize and reproduce training material. Generation and memorization are therefore compatible; neither a fluent answer nor an apparent quotation identifies a reliable source by itself. Systems that retrieve documents introduce an additional mechanism, whose sources can be inspected separately from the model’s learned parameters.
In Latent Writing, the phrase names a philosophical way of thinking about inherited language becoming generative. Technically, hallucination usually refers to unsupported or erroneous output, rather than all generated language. Calling the whole process hallucinatory is a broader metaphor; it should not remove the distinction between a checkable claim and an invented one.
Further reading
- Brown et al., Language Models are Few-Shot Learners, a primary account of autoregressive language modeling and prompting.
- Carlini et al., Extracting Training Data from Large Language Models, evidence of training-text memorization and extraction.