That's also what he writes in the article, I.e. LLMs are large language models, so the approach is generally flawed. A sentiment I agree with.