When Yann LeCun said LLMs are doomed, he was taken out of context; LLMs are only meant to work with languages because language is a symbolic system with much less noise, and if you want to truly scale beyond language (especially for natural or real-life data), his point is valid.