Contrary to some popular views, the performance of general empirical approaches (statistical, neural) to many NLP tasks remains inferior to that of linguistic (expert) approaches, at least when specific sublanguages are concerned. This is demonstrated in the case of Machine Translation. But is that still true for LLMs? Is it true that expert, linguistics-based methods are, or might soon be, beaten by LLM-based tools, which would surpass the performance levels of expert humans in any task? Using as an example the title of a recent talk by Mathieu Lafourcade at EGS in Montpellier and the task of high-quality translation from French into three languages, we show that this is probably not the case. However, a synergy between LLMs and expert resources and tools (like NooJ-based dictionaries and grammars) might be used to “inflate” the data (corpora) fed to the LLM and improve performance.