Today, you’re in for a mess. This is the first time I’ve felt really bad about the state of tech and academia. While I’ve been hating on generative AI for two years, most of my complaints were about potential risks (i.e., things I could foresee going wrong). Now, we have real issues hitting the math community, and I’m certain it’s just the start.
Improving Matrix Multiplication
You may have heard recently that all the frontier AI models are being used to solve famous math problems (e.g., the Navier-Stokes equations). I don’t keep up with this, and I frankly don’t care for reasons that I won’t bore you with today.
However, what I’ve found interesting is that a lot of academics have been using LLMs to aid in their research. I am not strictly opposed to this, but I find it troubling (again, for reasons I won’t bore you with but are nicely captured by one of my colleagues in the “Research Costs” section here).
In fact, some academics at my university have used LLMs to—as far as I understand—improve the upper bound on matrix multiplication. While that’s cool, I suppose, they had a much more interesting takeaway from the process. You can read it here.
In short, the solutions that these agents provide are slop, plain and simple. They may solve the problem, but the results are incomprehensible. And funnily enough, just a week later, OpenAI dropped their own solution to the matrix multiplication problem that offered an even better upper bound, which you can find on GitHub.
This led to quite the comical exchange between faculty over email. I won’t share too much for the sake of privacy, but I’ll share a couple snippets (not something you couldn’t get by making a FOIA request anyway):
- “I have a 108 page slop proving a problem I deeply care about to read and I’m not cited at all even though I’ve written papers related to it”
- “I doubt OpenAI’s employees understand more than %5 of their AI slop nuke. But they seem to enjoy flexing their muscles.”
- “OpenAI didn’t cite Andrew and Oliver. Tssk tssk tssk.”
In other words, I have a sense that academics are starting to turn on the AI companies. After all, I haven’t seen the word “slop” thrown around so casually, and I’m a huge fan of it.
But Wait, There’s More!
After wrapping this piece up on October 11th, I went to bed and woke up to a series of angry emails from faculty in my department to the associate dean of research and graduate studies. If you’re wondering what the associate dean could have done to invite such vitriol, consider that they had previously asked faculty on September 24th to fill out a survey indicating our interest in purchasing access to frontier AI models (you can imagine how I filled out that survey).
Since then, OpenAI has clearly disrupted theory research, so now faculty don’t want anything to do with them. I won’t share the exact emails, but their sentiments basically boiled down to the following: 1) OpenAI threatens the livelihoods of researchers, 2) OpenAI threatens the role of a university, which is to “educate students, train researchers, and sustain an environment in which the next generation of scholars can build meaningful careers,” 3) OpenAI threatens to become “Academia’s Landlord” as they set the “price of admission” even in applied research, and 4) OpenAI threatens to destroy the academic community “that produced it without actually delivering any meaningful benefits to the community or the society at large.”
That last point is of particular interest to me because it’s something I ranted about way back when I was working at GE: industry steals the work of academics for profit, and it’s finally destroying its host like a cancer.
With that said, while I’m excited to see a disdain for AI popping up in an unlikely place, I’m actually more interested in talking about the implications of LLM usage in math and beyond. Specifically, I’m hoping to segue into how it’s fascinating that LLMs are producing solutions that are incomprehensible while costing us the byproducts of failure. Then, I’ll talk about how these ideas transfer to other domains, like software development. So, let’s get into it!
Incomprehensible Solutions
Something I talk about a lot on this site is that “innovation” always seems to come at a cost. Here, I’m using “innovation” in scare quotes because I’m deeply skeptical of words like “innovation” and “progress.” Just because we’ve built something new doesn’t mean it’s a net positive for society, and I strongly believe that generative AI is a net negative on society.
So, you can probably imagine how excited I am to see the math community turning on these AI companies. One of the frustrations seems to be around how incomprehensible the solutions are. I find this absolutely fascinating.
After all, I often joke that these agents function somewhat like a sophisticated version of Monte Carlo Simulation: just randomly throwing darts at a wall until a solution emerges. In this case, the darts are every possible method across every possible math domain. The result is a complete mess of a solution that can only be verified by consulting experts from each domain.
Personally—and perhaps I’m not alone here—I can’t imagine there is much meaning or enjoyment in playing the role of “mathematician who verifies LLM outputs.” In fact, I know there is little enjoyment there because this technology has only been around for a few years and software developers have already shifted away from reading agent outputs.
Not only that but what value is in information that we don’t understand? It’s not an issue of “LLMs are super intelligent, and our dumb human brains don’t get it.” It’s an issue of “LLMs stumble into a working solution by brute force and can’t even begin to explain it.” It’s like when new programmers shuffle lines of code around until their tests pass. Like, congratulations. I guess.
And like who knows, maybe they’ll build an agent that can do all the verification too, so you don’t even need a human in the loop. How cool is that? Knowledge by bots for bots.
Failure Is Important
Snark aside, I actually think there’s an even bigger cost to relying on generative AI for mathematical discovery, and that’s losing the byproducts of failure. After all, in our efforts to solve math problems, we more often than not try things that don’t end up working out, and those failures actually lead to new discoveries.
Unfortunately, we’ve been trending in this direction in the STEM disciplines for a long time because of this obsession with efficiency. In fact, I’ve ranted about that several times on this site, such as here and here. In an effort to cut out all “waste” in a process, we inadvertently lose meaningful pieces of the process for the sake of efficiency.
And, you don’t have to think too hard to demonstrate this. If a company could run more efficiently without a portion of its labor force, the CEO might choose to let a lot of people go. For example, anybody in charge of compliance is going to slow down your output, so you might lobby for deregulation to be able to run more efficiently. Of course, deregulation is bad. It’s how you end up with dangerous chemicals in your water supply, increases in greenhouse gas emissions, and unethical labor practices.
In the case of math, we’re trading all of the tangential discoveries that might be made through failure for the ability to arrive at solutions quicker. In the short term, this might seem good; we’re rapidly solving problems that have been around for decades. But, we’re losing all the accidental discoveries that could have been made had humans been allowed to continue failing.
Likewise, if my discipline is any indication, I suspect a lot of folks are going to find other work that is more fulfilling. As a result, we’ll have fewer people trained in discovery—people who know what questions to ask or what problems to explore. It’s ultimately very short-sighted.
My Most Charitable Take
I’ve been mulling this idea over for a few days since I wrote this piece, and I’m really trying to imagine a scenario where we’re not just “trading all of the tangential discoveries that might be made through failure for the ability to arrive at solutions quicker.”
The most positive outlook I can imagine is what happened to chess. Chess engines haven’t “solved” chess, at least not yet, but they have allowed for a deeper understanding of the game. For instance, pro players are able to use the engine to plan out refutations for certain openings and end games. Personally, I think this makes the game of chess less enjoyable because it basically becomes a game of memorization (though, still very impressive), but chess engines have certainly elevated the game of chess. Of course, they’ve also hurt the game of chess by making cheating much easier to do.
Regardless, I think one angle we could take on LLMs in theory research is that we could potentially still learn new things from the approaches that the agents take. The phrase “that’s not a human move” in chess could also apply here, where agents do things that are unexpected, leading to new discoveries. That’s my most charitable outlook.
In practice, I really don’t see this happening. Like anything that LLMs touch, you’re basically incentivizing laziness. When I hear that someone is using an LLM to write their emails, I have to wonder where else they’re using it. It completely puts into question the authenticity of anything that person does. In fact, I’ve already made this argument to other academics who “only” use LLMs in certain parts of their work. Why should I trust you to have self-control? First, it’s to help with wording a sentence. Then, it’s editing paragraphs of text. Suddenly, you’re just having it write your papers outright.
That same snowball effect is why you’re already—a reminder that ChatGPT launched only 4 years ago—seeing people advocate for never looking at a line of code again. We’re talking about the eradication of fundamental programming skills in four years. Why would anyone learn to code if this was the case? Why would anyone study math for that matter? These skills will be relegated to hobbyists, which means no one will have the expertise to actually leverage these bots effectively, which I wouldn’t advocate for anyway.
Essentially, you can’t have computers do all the thinking and expect people to continue thinking for themselves. You’re already seeing this with social media. We latch onto creators and let them form our opinions and beliefs for us. Now, we can latch onto a bot. What a wonderful time to be alive. I have never been more interested in running off to a cabin in the woods.
Like, let’s be so real for a second. Do you think there will be trained mathematicians who can review OpenAI slop papers and admire the methodology? Who aren’t just going to prompt the bot harder? Who aren’t just going to say, “hey chat, what should you study next?,” and then have it write a slop paper on its “work?” Oh wait! That’s what OpenAI is doing now.
I really do not understand the end goal of this technology. It’s hardly a tool, when it’s actively sabotaging the folks with the current ability to use it. Though, I suppose the end goal could be an uneducated underclass of consumers, which is something I think the CEO of Nvidia actually wants. After all, I swear I just saw a clip of him saying we don’t need to teach kids algebra or calculus anymore. What an absurd take.
Chasing Efficiency in Computer Science
For a long time, I’ve been frustrated with how much the field of computer science obsesses over efficiency. Almost every class you take makes a point to prioritize efficiency when building software. In fact, most programs have a few classes that are just dedicated to computing runtimes. I took one of those as recently as grad school in 2018 where we learned about several classes of algorithms, such as divide-and-conquer and dynamic programming.
Hell, there’s an entire website dedicated to writing the most efficient algorithm for a problem called LeetCode that plagues every aspiring developer. You can’t even get into the industry without being able to pass some sort of LeetCode-style assessment.
Ironically, despite how hard efficiency is pushed, anyone who has built software can tell you that it’s only a piece of the puzzle. In fact, I often reference Knuth’s premature optimization quote in pushback to this mindset. In contrast, software developers are actually juggling a ton of metrics, such as correctness, readability, maintainability, scope, product-market fit, size, usability, etc.
Are there times when efficiency is important? Absolutely. Video games should be optimized to ensure the experience is smooth for the player. Embedded systems should be optimized because they’re on a small footprint. Often, data analytics and machine learning applications are dealing with massive datasets, so they should be optimized.
These are exceptions and are certainly not the rule, and even in these cases, you only really need to optimize bottlenecks. There is no reason to get upset if a teammate uses a suboptimal algorithm if the time saved from a better solution is negligible compared to the size of the bottleneck. Likewise, who cares if your code is optimized if whatever you’re building sucks?
What’s really interesting is that despite everything I’ve said up this point, I don’t even need to prove to you that our field is obsessed with efficiency. We literally identified ourselves as the bottleneck and are happily automating ourselves out of jobs with LLMs.
Perhaps that’s why I find this situation in math so interesting. After all, unlike (most) software developers, mathematicians are openly upset about what OpenAI is doing. They understand that discovery cannot be optimized. They understand that when theorems are seen simply as outcomes, we lose all the value created by the traditional process. So, maybe we (e.g., software developers, computer scientists, etc.) should follow their lead and resist the AI industry’s desire to “disrupt” all of our institutions.
Marc Andreesen’s Wet Dream
Apologies for all the tangents, but I want to sneak one more thought down here because I’ve really been thinking about these topics a lot. Specifically, I want to take this idea of efficiency in computer science a step further. Previously, I made the argument that arriving at solutions quickly in fields like mathematics and theoretical computer science is bad because it closes the book on the possibility of further exploration. I’ve been thinking about how this applies to computer science—especially software development—and I think I’ve figured it out.
By being able to build software rapidly, there is never space for reflection. There’s never going to be a moment where you’re halfway into creating a feature and you think: “do we even need this feature?” or “why are we even making this?” While that might be the wet dream of someone like Marc Andreesen, reflection is a fundamentally human trait. What do we do when we’re not considering the implications of what we’re building? That’s not a world I want to live in.
What’s up? Thanks for reading all the way through this one. It took me a bit longer to write than my usual pieces because I felt I had something a little more reflective or introspective to say today. I feel like the message from this piece is a bit more important than my usual rants. It’s not just “AI bad” or “efficiency bad.” It’s something more along the lines of “consider what we’re losing when we replace everything with AI.”
I’ve been ranting in this series since March 8th, 2024, and while I feel like we have made significant progress in terms of public sentiment around AI, the world feels like a worse place to be in than back then. There just doesn’t feel like there is anything untouched by AI at this point, and I don’t feel great about the next couple of years.
With that said, I enjoy writing these pieces, and I’ll continue writing even if I’m the only one left doing it myself. That goes for every facet of my life that the AI industry seeks to replace. It’s a core principle of mine at this point.
Can you tell that the latest stories in the math field have gotten to me? Anyway, enough doomer-posting. Check out some of these related pieces:
- No, Generative AI Is Not Just Another Innovation
- Why Generative AI Makes the Future of Software Development Worse
- Is Anyone Else Bothered by How Quickly We Adopted Generative AI?
Likewise, you’re welcome to take your support a bit further by checking out my list of ways to grow the site. Otherwise, take care!
Recent Blog Posts
It's 2026, and I'm picking up DnD for the first time. Needless to say, I haven't had this much fun in ages.
Lately, I've noticed a lot of people obsessing over their own relevancy. I'm not sure that's healthy, nor should it matter.

