A rant on the current AI math cultural moment

When I heard that the Riemann hypothesis was proven today, I felt strangely relieved. I’m not entirely sure why, but I suppose it felt like a sword of Damocles hanging over the heads of the math community, and now it’s (probably) been taken down.

There are apparently two purported proofs out now, one by Anthropic employee Levent Alpöge using Claude and one by mathematician Alessandro Masullo who apparently didn’t use any AI assistance.

I’m still holding my horses on accepting the RH as true tho, since these proofs are both long and will take some time to be independently verified. I don’t really have much to say about Masullo’s work, other than that I find it way more impressive and wholeheartedly hope for its successful verification.

In contrast, Alpöge, or rather Anthropic, or rather the whole AI sector and the hype around it, have massively pissed me off. Alpöge’s statement can be found at https://alpo.ge/remark.pdf and I must say I am offended at some of the language he chose to use, treating Claude like a sentient collaborator.

This is indicative of a larger issue I have noticed for some time now. I used to enjoy watching math videos on YouTube,1 but now I have to click off half of them2 because instead of “look at this cool math!!!” it’s all just “look at how amazing AI is!”, “AI has revolutionized mathematics!”, “How to engage with math in the age of AI”, etc. etc.

Are you people fucking listening to yourself???!!! You sound like you’ve been secretly hired by the AI bubble companies as propagandists! The whole thing is a massive PR stunt to begin with! As if these AI companies care even a little bit about the math they’re pushing.

“But it’s true that AI is helping advance math”

Yeah, it probably is. But you also need to be aware of the many criticisms that can be levied against it. There are environmental issues from data centers, mental health issues for the users, exploitation of low-wage foreign workers to generate training data, and many other things unrelated to math, but I feel they at least deserved a mention here.3

The most relevant issue to bring up here is truth-irrelevance: An LLM generates tokens based on a statistical distribution built from a gigantic set of training data. By its very mechanism, there is no point in the process where truth ever enters the picture. It’s just a very very very very advanced relative of your phone’s autocomplete. It doesn’t care about truth, which is why “hallucination” is the wrong word for when it gets something wrong, as it never even tried to get it right in the firstplace. In fact it’s worse than that: What it tries to do is generate convincing-looking text, which means that even the false statements will look convincing!

The chatbot interface is also a lie by the way; Behind the scenes you’re collaborating with the LLM program on something like the script for a play, where you write the parts said by your character and the LLM predicts the parts said by its character.

Luckily for the math community, there are external tools that are used to control the LLM’s output; Since formal mathematical proofs are quite structured and mechanical, a computer can be used to check them. Then, I assume, such a checker (in this case a software called Lean) is integrated into the screenplay as a sort of narrator or third character, which responds to the LLM’s character’s statements with whether they worked or not, and the LLM can then predict that its character will respond to any failures by adjusting the code to fix the issues.

Of course, there’s still the issue of “who checks the checker”, i.e. the checker itself may have bugs that let it incorrectly say a proof is correct when it in fact isn’t, and this has happened before. Personally I especially find this concerning when the all-important kernel, the thing that must be trusted for the checker’s output to be of any worth, itself contains LLM-generated code.

But I admit that these concerns are mostly based on principle and in practice it seems pretty unlikely for such an issue to have occured in Alpöge’s RH proof.

The bubble must pop

My main issue is still this: There’s currently a massive economic bubble around AI, despite the fact that the technology is just not nearly as impressive as it is made out to be. It has its uses, but it’s not gonna revolutionize literally everything to the point that we can talk about an “AI age” (🤮). Let’s face it, we’re not in the “VR age”, we’re not in the “cryptocurrency age”, we’re not in the “NFT age”, and we’re not in the “AI age”. We’re still in the internet age, or information age if you will.

The bubble will pop, the bubble needs to pop, but the AI companies are clutching their pearls, desperate to hold onto the wealth and power they currently have, and in an economy dominated by investors, they are absolutely dependent on the idea that AI is incredibly, unfathomably powerful and will generate yet unforeseen gigantic mountains of profits.

And they do this by taking hold of the culture, of our culture. AI is an infestation in the math sphere, feeding on our enthusiasm and destroying our metaphorical homes, just to keep itself alive.

It’s fine to mention that AI was used in the process of arriving at some mathematical result. Necessary, even, given all the other issues around it I touched on. But PLEASE, don’t feed the parasites. Don’t dwell on it, just say something like “Alpöge proved the RH with LLM assistance” and move on to talk about the details of the proof or the mathematical implications or whatever.

  1. Of course this isn’t just happening on YouTube, but I don’t go on social media (assuming you don’t count YouTube), so that’s where I notice it. ↩︎
  2. The other half are fortunately still great! Shoutout to Another Roof in particular; He also has a similar take on AI in math. ↩︎
  3. Tho not “art theft”, since I don’t believe in intellectual property. ↩︎