I agree with your comment and the article. LLMs should be part of the answer, but the core of the progress should probably dive back into neural networks in general. Language is how we communicate as well as with other senses, but right now we're stuck at LLMs that just seem to be blown out elizas trained with other actual humans work. I remember early on, training of simple neural networks was done with rules in their environment and they evolved behavior according to criteria set, like genetic algorithms. I think the current LLMs are getting a "filtered" view of the environment, and that filter behaves like the average IQ of netizens lol