CLIP isn’t an llm either. point being, the pattern recognition argument doesn’t hold. they’re feature detectors, which is basically a giant grid of classifiers. that’s not what llms do.