Token prediction machines do not produce meaningful output.
It is much more likely that they will regurgitate, without attribution, something meaningful that a human wrote (in their plagiarism/training data).
Meaningful "output" is only meaningful if the author understood the meaning, instead of predicting words in sequence.
Just because a brain is made from organic material rather than silicon, doesn't make it less of a prediction machine.
The brain understands concepts. I can say, for example, "a penny saved is a penny earned" and understand that I'm talking about money and savings in general. I've felt the satisfaction of a full bank account and the panic of a near-empty one.
The AI knows that someone once wrote the words "a penny saved is a penny earned" in that order. It has ingested volumes of wise sayings and economic texts to facilitate the statistically accurate words which will follow previous words in the conversation.
AI is doing a lot more than just regurgitating facts. It can take abstract concepts and apply them reasonably well in novel scenarios.
To make things more concrete, how about you name a specific concept that you believe a frontier model can't grasp?
If AI was learning and understanding, this would have been solved in 2024. As it is, it's just predicting words that are likely to answer the question.
You're slurping up the excrement of corporations if you think AI is "learning" from reading others' experiences. All it is learning - all it can learn - is what words are likely to appear one after another, with no semantic understanding.
Second of all, not being able to properly spell a word doesn't mean one doesn't have a semantic understanding of the concepts the word represent. Do you think an illiterate person is incapable of intelligence?