[jqpabc123]

...large language models, rather than large math solvers or large fact models.

And why not?

Math is the most logical and precise language ever invented.

If LLMs can truly think and reason and understand, I would expect them to excel at math problems. Or at least admit that it can't do math and logic.

[elpocko]

>If LLMs can truly think and reason and understand

But they can't, which was kind of my point. They are clever token predictors, they know language, which makes them really good text generators ("stochastic parrots"), but even trivial tasks like counting the letters in a word is hit-or-miss, especially if the solution is not found in their training data.

I don't understand why people find this surprising. It's remarkable that LLMs can solve some problems at all, not the other way around.

[jqpabc123]

I posed this exact same test problem to a selected LLM.

It produced the correct answer --- including pointing out the irrelevance of the "smaller" kiwis.

Then I changed the person's name (John instead of Oliver) and jiggled the numbers a bit. It confidently produced an answer including an explanation with the irrelevance noted --- but it still did the simple addition wrong.

It obviously doesn't *understand* anything. It just regurgitates what it finds posted somewhere on the internet.

Frankly, I wouldn't trust anything it says. It's hard to imagine what this would be truly useful for --- propaganda maybe?

Billions and billions of dollars are being wasted on this.