Hacker News Reader

Comment by adamddev1 | Hacker News Reader

> Have mathematicians become "less savvy" since the invention of the abacus or calculator? The invention of logarithms or analysis?

This is not a fair analogy. Those things are like compilers, languages, and type-checkers. They are deterministic, carefully proven, hand-crafted tools, that allow people to go safely to a higher level of abstraction and still make things in a deterministic way. The jump into the fuzzy world of LLMs is a completely different thing. It's different than the jump from assembly to a higher-level language.

View on HN

Not sure how much frontier math you follow, but probabilistic arguments are extremely popular (particularly in, weirdly enough, number theory), especially for fiddling with the bounds on things, or getting closer to full coverage of various theorems.

So this idea that doing math is some kind of fully deterministic & mechanistic process is a bit misleading.

Right, programmers also use randomly generated info from property based testing for example? But I was referencing the analogy to the abacus, calculator, logarithms, analysis, and higher level languages.

Is that kind of argument something like where you count the number of cranks claiming P=NP and the number of cranks claiming P!=NP and then if you get 95% papers claiming P!=NP that one wins?

What? Not at all. I'm referring to things like discussing distributions of certain things or probabilities of things as we go towards infinity, etc. These are proper papers, not surveys.

But if a LLM writes such a paper, how do you know there aren't any issues in it that cause wrong probabilities/distributions?

You don't, that's why you review it (just how you would handle a real human submitting a paper). The main issue with LLM papers is that they tend to be very dense/circuitous or—because LLMs don't truly understand what they're doing—have the wrong focus. An LLM might spend pages on a trivial result, but quickly gloss over a truly remarkable finding.

Terry Tao talks about this challenge in some of his blog posts and interviews, it's pretty interesting.

>You don't, that's why you review it

Review enough of those without doing the proving yourself, and you'll soon lose the ability to review. And have LLMs do the course work for you at math school, and you might never acquire such ability to begin with anyway.

That was adamddev1s point. It's a unreliable tool. You cannot copy & paste its results into a paper as proof. You could do this with e.g. Lean. To review what it writes you need to understand the topic to such an extend that you could have written it yourself (at least that is how it is for me when I review code).

Yes, for inspiration or search they are great tools.

Reply on news.ycombinator.com