OpenAI recently released a ton of math problems that they spent, on average, three hours each solving with one of their latest (obviously unreleased) models. These are problems that humans have been unable to make progress on, or whose solutions they thought could not be improved upon, for decades. For instance, subquadratic 3SUM—while the gains are tiny, improving on what people thought was an unimprovable lower bound is extremely impressive. It’s on the order of voodoo magic, as 3SUM intuitively seems unsolvable in subquadratic time. OpenAI claims that they spent under three hours per problem, and that seems about right.
And academia is not reacting well. Imagine you’re a PhD student working on these problems, one of the brightest minds in the world. You’ve spent years trying to solve a problem, and you learn that an AI model took just three hours to one-shot it and surpass all the progress you’ve made on it. You learn that not only were you on the wrong track, but you had previously been on the right track and deviated from the correct path. Soul-crushing, really.
Which is why the math community is so strongly against it. While some of the broader math community’s concerns have merit (see Terence Tao’s concern that there’s too much math to consume), its reaction has largely been one of protectionism. Thousands of researchers, professors, and academics are concerned that their fields of research are being rendered useless and that their life’s work is meaningless. It’s understandable! You’ve spent years of your life toiling away at numbers and letters, graphs and equations! It’s frustrating.
But progress is progress! You cannot stop progress. And policing it is futile. This is the world you live in and the world that you will continue to experience. And it’s a good thing: who knows what future innovations a more efficient multiplication method could lead to? Suppose one day we discover a significantly more efficient algorithm for something that’s widely adopted. Think of how much energy we would save every time that routine runs anywhere in the world.
It’s really about jobs. Academics are concerned that they are losing their jobs, and they should be. But they should not mask that concern by treating the field of mathematics like some kind of hobby. If we are using public funding and other research grants to pay academics to discover things for the good of the public, we should expect some kind of return on that funding. We are not paying them to dilly-dally and indulge their own egos. We are paying them to make discoveries that benefit us all.
This pattern of AI replacing white-collar work is all too common. First, it came for the software engineers, as software engineers are paid very well in today’s industry. The cost of developing some crappy internal tool has essentially dropped to zero, when it used to cost a company a few junior software engineers’ salaries (which were still quite high). Now it’s coming for academics, because academics are also paid very well. And it makes sense! It’s the highest ROI you can get; automating a software engineer’s work or a professor’s work saves much more money than automating a barista’s.
There’s a disconnect between the kinds of jobs most people think will be AI’d away first and those that actually will be. The usual concern is that we’ll replace artists, Uber drivers, and truck drivers. But why do that? Self-driving cars, despite the progress we have made, are still in their infancy. And even if they do replace human drivers, is the token cost worth it? As of now, maybe not, but certainly in the future. But why deal with the real world, which has all sorts of obstacles popping up in the road, when we can deal in the abstract, in the realm of words or tokens, and automate away the high-paying jobs first?
But there’s no sympathy for white-collar jobs. Nobody cares if some rich software engineer gets laid off (unless it’s a game developer, in which case their whining on Twitter somehow gets amplified and treated differently from that of the thousands working on Microsoft Excel). They make $200k a year. Who cares? They’re fine. There’s nobody to really go to bat for them. And academics—what do the pesky nerds in the math department really do? Sure, people will fight to preserve jobs researching some endangered fish in the Nevada desert, but nobody cares about math. It’s boring nerd shit.
The real way to prevent yourself from being automated away is either to be on top of the AI (sell the AI and its product) or to be on the bottom, doing something that is not worth AI’ing away. Plumbing is one example; you cannot easily AI away plumbing because it requires hardware, and you are also dealing with unpredictable customers who don’t really know what they want. You can automate everything else away, but plumbing is something you would have to do manually. You can create better tools for plumbers, and perhaps in the future we will have robots, but in the end, that root will need to be cut away with an auger. It’s a much harder problem than theorizing about 3SUM.
I don’t think this is a bad thing. Plumbers can make good money. In the future, they may make more money. And that’s good! They deserve it.