About two years ago, I was at a coffee shop and two maths students were trying to understand a concept in the book. I suggested they might ask a language model to help them understand it. They scoffed at me, called a stochastic parrot, and said they can't do math at all.

I responded with, "Well, you don't even need it to do math. It can explain the concept." They scoffed again. "Those things are useless toys"

▲

georgemcbay 4 hours ago | parent [-]

IMO they weren't exactly wrong... 2 years ago.

2 years ago LLMs were a great tech demo but were mostly awful for actual practical purposes and hallucinated almost as much as they got right.

They'd be wrong today, but that's a different thing.

▲

preommr 2 hours ago | parent [-]

Reading comments like this is so depressing because it was extremely obvious that better results were a matter of time/hardware. It evokes the seem feeling as that trump quote about "who knew healthcare could be so complicated".

▲

skew-aberration 2 hours ago | parent | next [-]

Also, even GPT-3 (6 years ago) was a very powerful search tool for exactly these kinds of problems. You needed some result that would give a bound on some obscure property, and it instantly gave you five different applicable theorems from three different domains of math. Usually, the theorems even existed.

I was taken aback by the progress in 2025/2026 because late 2024 did seem to be a plateau, but 5-10 years seemed inevitable.

Even when I first studied automated theorem provers in around 2013, the 10-20 years to automating proofs (not theory building) seemed inevitable. They were already superhumanly fast and generating real results, and neural networks were making massive progress on translation, image processing, and games. Someone just had to work out how to stick a neural net inside an ATP. The surprise was that it happened the other way around, and that it solved even more than just mathematical proofs.

AI has only just begun to expose the depth of willful human ignorance.

	▲	bananaflag an hour ago \| parent [-]
		> late 2024 did seem to be a plateau Why? o1 was released in late 2024

▲

georgemcbay an hour ago | parent | prev [-]

> Reading comments like this is so depressing because it was extremely obvious that better results were a matter of time/hardware. It evokes the seem feeling as that trump quote about "who knew healthcare could be so complicated".

I honestly don't see how this has anything to do with my reply to the original comment.

Even if you were absolutely sure LLMs were going to get as good as they are today 2 years ago, that didn't make them any more of a practical solution to the issues you were trying to solve 2 years ago.

It is likely that the SpaceX Starship will eventually be a reliable rocket platform. That doesn't mean that given the chance I would hop on the next flight.