Comment by stbede | Hacker News Reader

Comment by stbede | Hacker News Reader

Parent Full thread

Yes. Everything I've said and want to say is about this letter.

Let me try to summarize this discussion in a way that an overly obsequious AI would never dare:

- I argue that the arguments in the letter are bad.

- You mysteriously assign me homework on the topic of human motivation.

- Confused, I reiterate the points made in the letter, and I make arguments against them

- You ignore my arguments, call names, and then defend the letter by making points that the letter doesn't make.

- I ignore your points, because they aren't the points made in the letter. And my original post was about why the letter was misguided.

- You call names. You feel like your points are valid. You then claim that the letter is valid because your points are valid.

- I keep disagreeing with the letter.

- You call names again, because why am I so obsessed with talking about the letter. I must be an idiot, or an evil lying phony, or an evil lying idiot-phony because I keep bringing the conversation back to the letter on a thread about the letter.

Hmm...;

I had imagined i was interacting with somebody who could understand the larger picture given only the key points. Apparently that was wrong and i need to spell out every detail.

I didn't call you names but implied that you had an agenda and were masking it by trying to debate the literal words in the original communique from 25 fields medalists which was written to be politically correct. Apparently asking you to "focus on the spirit and not the letter" of the communique was too much. So i highlighted it, repeated myself three/four times and yet nothing seems to have worked to aid comprehension.

I point to a most important and relevant psychological theory viz. SDT, which explains how one's behaviour is influenced by intrinsic and extrinsic motivations and ask you to think about how they might explain why the mathematicians are angry at the current situation of AI companies shafting them. I even used specific terminology from SDT to highlight its applicability in this case. Apparently this is too much "home work" (and needed thinking) and one needs to spoon-feed everything without assuming any human intelligence/curiosity/initiative at the other end.

It seems my attempt to show the larger context and situate it in actual psychological research has been of no avail though not for want of trying.

You and I aren't having the same conversation. You want to psychologize people's motivations, including my own, which is interesting to speculate about but off topic. I want to discuss the arguments made in the letter, the arguments that 25 Fields medalists actually signed on to.

I am capable of having a conversation with you on a great many topics. What you aren't capable of, apparently, is forcing me to want to have the one you want to have.

As far as calling names:

"Your comment is ... disingenuous."

"you are being disingenuously clueless or cluelessly disingenuous."

and now:

"I had imagined i was interacting with somebody who could understand the larger picture"

Nope, you need to engage in a root-cause conversation which is the key here. Just focusing on the words in a letter which as i already pointed out is couched in politically-correct terms due to etiquette is meaningless. There is a lot unsaid there which you have to tease-out by reading between the lines. It was written in response to OpenAI's behaviour (Tristan Buckmaster himself called what they did as "malpractice" in one of his tweets) but looking at the larger picture of AI in Mathematics. The key phrase from the letter is;

... Often these solutions are announced in a rush, leaving no time for a proper writeup, the isolation of new methods and ideas, and citing relevant previous work of others. As in all creative professions, this raises severe attribution and plagiarism questions. ...

The inference should be clear.

Hence we need to start with the underlying psychological motivations and then see how it explains observed behaviour. And it is not even complicated but somewhat obvious. The fact that you (many others) refuse to address this is why i question your (and theirs) motivations/agenda.

As regards your last few lines regarding "calling names" they are all only w.r.t. to the issues being discussed and are just rhetorical/descriptive judgements on your comments.

Fair enough. But I think you may be conflating an explanation with an excuse. I think your root cause analysis may be fruitful for explaining why they wrote the letter. But I don't think that it excuses what they actually wrote.

The letter is just a call to action reminding everyone how Mathematics is done and how Mathematicians operate within their domain with all the Human factors involved.

But because of the bad behaviour (plagiarism, going-behind-the-back etc.) and incentives misalignment, AI companies are currently not aligned with that.

There is also the structural problem of generating proofs at a rapid pace just because you can but then have no way for "proof digestion" which is the key in the mathematical process. See Terence Tao's writings - https://teorth.github.io/tao-web/ai-views.html

I agree. But my frustration is that there is no argument why AI should be aligned with mathematics.

Why can't the mathematicians that value their own community remain aloof from what AI companies are doing. Can't they continue to take their time digesting proofs, gaining insights, and attributing credit to mathematicians doing good work. Here's how I perceive the situation:

Before:

- mathematicians digest current results and develop insights

- insights lead to new results

- new results give mathematicians an opportunity to credit those insights

Now:

- AI companies release a bunch of potentially uncredited proofs

- mathematicians digest these results and gain insights

- mathematicians recapitulate AI proofs in terms of their deeper insights and theory

- mathematicians give credit for these new insights and progress in understanding, even though the binary true-false status of theorems were already determined beforehand.

If the true-false status itself isn't real mathematics, then mathematicians still can do the real mathematics after the fact and continue to celebrate each others' work. The only thing they lose is the ability to attribute a human to deciding the true-false status, which isn't real math anyways (though maybe valuable for motivation). What they gain is efficiency. Every mathematician working towards proving NS won't blow up can now shift their resources to understanding why it can blowup.

Perhaps the letter is implying that the mathematical community itself will become misaligned with how mathematics is currently done. I'm unpersuaded it's entirely the AI companies' fault if that happens. AI may create a lot of temptations for individual mathematicians to become lazy, but that seems like something the mathematical community needs to resolve for itself. Every institution needs to take a lot of self-ownership here. I'm also annoyed about parents complaining about their own children being addicted to their phones as if it isn't in their own power to determine if their children have phones in the first place. I think this letter would be better directed at universities and journals. They should be holding their own institutions to account if they feel like they are adopting the mentalities of the AI companies and no longer faithfully executing the values of mathematics.

The real issue is that there is "research" and then there is research. Most publications are simple recapitulations of well known ideas but applied to somewhat unique environments. Imagine if you are a carpenter, you are just using your existing tools to make more stuff, but fundamentally what you are doing is no different from what any other carpenter is doing. It's workmanlike production.

The truly valuable stuff is coming up with new ideas, that is, new techniques that allow for breakthroughs. That is rare, is done by only a handful of people in any decade, and is very, very hard. When you study a book on the history of math, you see the names of people that created brand new techniques which allowed them to solve problems no one else could solve.

But you still have that 99.9% that needs jobs, and they need to publish, but what they are doing can be mostly replaced by AI.

The specific situation with Navier-Stokes really highlights this. Mathematicians came up with the technique, but the AI managed to apply it faster than even the technique inventors. So it appears that AI is now good enough to be doing the workmanlike stuff, which is the bread and butter of a lot of mathematicians that are just trying to get their publications in so they are not fired.

I do think this creates real problems for the field and that some changes are necessary.

Exactly. This is what the letter should have focused on. The issue is not alignment, per se. If AI can do the workmanlike stuff at breakneck speeds, academia will decide between two options:

1) Keep the average mathematician employed doing redundant mathematics just to have capable humans around. Every now and then you get one of those mathematicians actually developing new ideas, but most departments will entirely be just for show. It's doable though. College sports is able to keep itself going along these lines (and I mean the track, softball, and polo type sports).

2) You get unemployed mathematicians who were still developing their abilities. And as a result you lose the institutions that generate the mathematicians that contribute new ideas.

You get to keep the handful of exceptional mathematicians that are developing new ideas, and that's survivable if you happen to be a Fields medalist. None of this really has to do with the AI companies aligning themselves to giving credit and developing insights. If AI companies became aligned to that, then the AI's would be replacing the Fields medalists, too.

We should just pay them to teach classes and not require them to do research at all. That's the part that needs to change.

Reply on news.ycombinator.com