Thanks for the link, that’s a fascinating article.

The thought experiment (I really liked it) has convinced me that an LLM cannot learn meaning the same way a human does.

But two counter arguments arise: first - does it matter? If the only interaction is through thai text, how could the difference between understanding the meaning of Thai text and simple text completion through infinite studying of Thai books be asserted? And second - how is this changed by multimodal models? The author explicitly states that all images are removed from the library and, when asking others on their opinions of the thought experiment, “I’d look for an encyclopedia with images” is considered cheating. That means the author considers images as a weak point of the thought experiment. If the presence of other media did not change the outcome they would not have to be excluded. And if multi modal models change your opinion is that not simply because you underestimated how much you can do with infinite time in a Thai library? What is the fundamental difference between text and an image?