“There is no way for the [LLM] output x to acquire linguistic meaning without some further speech act in which someone capable of meaning something by x understands the potential meaning of x in language L.”

“While it is true that the outputs by LLMs do not have any meaning in themselves, the reason is not that the models are not acquainted with the objects referred to, but rather than they cannot carry out speech acts.”

“We are not mistaken by assigning [these outputs] meaning (…) but we are in error if we assume that the signs in question already have meaning merely due to their origin in the computational model.”

“An undead sign is a dead sign, but brought about with the intention to make it look like a living, meaningful sign. The difference is that not only is the original meaning suspended in order to be assigned a new one, but this very process is also designed not to be noticed.” 🧛

“Undead signs [are] the result of an intentional manipulation of meaningful language by the programmer of the model in order to generate the illusion of language in use in the outputs the model generates.”

“The construction of a semantic model of our language requires the machine learning mechanism to dissolve the semantic context in which the tokens are originally used.”🤔 Distributional semanticists will disagree, but in a certain sense of ‘semantic context’, yes.

There’s a lot in Lynch’s paper that I’m unable to evaluate, but at least this I thought unfair: “What is functional information? Wong et al. never tell us.” In fact, they do tell us, with enough maths and examples.

A ROSE by any other acronym?

“(…) jargon infused with technology industry hype, such as shown in Table 1, does not meaningfully explain. (…) We strive to remain critical of the vocabulary the technology industry coopts and deploys, and to remain respectful of scientific terminology.”

zenodo.org/records/1706...

8/🧵