Quantity of training material doesn’t confer new abilities. It makes the resulting weights more representative of the language of the materials, but it doesn’t give something the text doesn’t have.

You left out half of my sentence though. Size of the model does in fact confer new abilities. See, for example, this paper: https://arxiv.org/abs/2206.07682

All of the problems you describe that LLMs have - it's true, and it holds for current LLMs. But it is not necessarily a fundamental limitation. Time will show how much better LLMs will become. I think the time of super fast progress is probably over, but there will still be improvements.

That is a damning verdict for a machine literally invented for computing. If there is one thing a computer should be good at, it should be the thing it was built for.

An LLM is not a computer. An LLM runs on a computer. You're mixing up two things here.

I think you vastly overestimate human abilities. It's a phenomenon I come across here on Lemmy all the time. It's like people believe their brains are capable of some sort of "analytical logic" as opposed to "numerical logic". As if "understanding" something was some sort of godly ability. I don't believe in a supernatural spirit or anything like that. We learn from external influences, we are statistical parrots as well. There is no absolute truth mechanism in our brains. We're conscious, yes, but just because understanding something feels so absolutely logical and true to you doesn't mean it's not just a result of the statistics your brain has learned from.