# Et Tu, Brute Force > Source: > Published: 2026-09-12 08:36:00+00:00 Et Tu, Brute Force The Millennium Prize Problems are a set of 7 mathematical problems set by the Clay Mathematics Institute in the year 2000. Solving one of these problems would earn you an award of $1 million. So far only one of these problems has been solved, though the solver, Grigori Perelman, declined the prize (as well as a Fields Medal). He felt that credit would not be shared fairly when his work was reliant on the work of another mathematician, and his contribution was no greater than theirs. This gets at the heart of how research, mathematics and science largely works. No work is done in isolation, and granting sole credit to the one who takes the final step is ethically suspect at best. The recent controversy surrounding Anthropic and OpenAI and the potential solving of the Navier-Stokes Equation has made me think about these problems. Whether or not the Navier-Stokes equation was truly solved (under the specific constraints and conditions chosen by OpenAI) purely by AI is irrelevant to my point, but I will take the time here to explore the truth of that claim. It is important to note that a Millennium Prize Problem cannot be claimed as solved unless the paper is published in a peer-reviewed journal and remains so for 2 years and is accepted by the general scientific community. This means that the claim that the problem is "solved" is a slight exaggeration at best and a cynical marketing move at worst. I'll leave it to your imagination which side I'm leaning on. The first and most important point of contention is that an N.Y.U. mathematician named Tristan Buckmaster had been working on the problem with another mathematician who works for Anthropic. They had been using OpenAI's Codex tool in their work and as of today there has only been a non-committal statement that they "cannot rule out that de-identified data derived from their usage of our products helped improve our models". This means they may have piggy-backed on the work of the mathematician duo and used information from them in the training and prompting to solve this problem. OpenAI even admitted they became aware that the problem was close to being solved (or at least progress was being made) and chose to target this problem as a result, as a kind of orchestrated publicity stunt. There are also several critiques of their paper, and though I'm not qualified to speak on them, I will go over them briefly here. The Clay Mathematics Institute definition of Navier-Stokes allows for solving either the unforced or forced version. OpenAI solved the non-absolute forced version, where the fluid is driven by an external force. In addition, it seems that the approach used involved reverse-engineering the conditions to calculate an extremely specific external force designed to create the desired result. Assuming their solution is both mathematically and scientifically rigorous, this does result in successfully proving that it is possible for the part of the fluid to reach an infinitely high velocity. This would mean that under some conditions the equations we use to model fluid dynamics fail to adhere to reality at certain scales. However, reverse-engineering the external force to cause the model to break down is being criticised as not addressing the spirit of the problem. It was simply a classic tech move of finding a loophole and targeting it. Moving on from technical critiques, OpenAI loudly proclaimed they solved it in just 88 hours to emphasise the impressiveness of the achievement and they utilised approximately 10,000 AI agents to do so. These agents generated 130 billion output tokens. This would translate to somewhere between $10 million and $40 million for the solve, based on consumer market token costs and the various other APIs needed. In my view, this price isn't the main issue though. The electricity cost of this publicity stunt was around the equivalent of the monthly electricity usage of a small town in 4 days. It also consumed approximately 12.5 million litres of water, the equivalent of 6.25 million people's daily intake. In just 4 days. The frustrating reality is this massive environmental cost of the scaling approach to improving the quality and usability of AI seems to largely be ignored (or at least not mentioned) by the technically minded or tech focused people. This isn't an accidental byproduct, it's the direct result of the approach chosen to improve this technology, though depending on your views this may be little more than polishing a turd using the most expensive and environmentally damaging polish imaginable. Even if solving this problem was a massive breakthrough or leads to some material improvements (my understanding is that it won't, though I'm not qualified enough to know for sure) is it worth this incredible cost? Now, we've been utilizing computing power to solve equations or disprove conjectures since we've had the ability to do it. Brute forcing a problem that would theoretically take humans years to solve is a straightforward approach if our goal is to simply disprove it. A functional, effective or even efficient brute forcing technique or algorithm would still need to be considered and designed by a human. They would need to think about the problem, what shape it takes and how best to solve it. Regardless, even the most inefficient brute force approaches are somehow so much less energy intensive than LLMs that they aren't even a drop in the bucket in comparison. Looking at how to solve "toy" problems using various techniques is a great way to reduce the importance of the solution itself. The entire point of puzzles and other toy problems, is that knowing the solution removes the meaning of solving it in the first place. If we take a look at Sudoku for example, a brute force algorithm can simply look at the next empty box and guess every possible number, checking if it's valid and moving to the next. It is slow for most Sudoku puzzles, though there are some it will solve extremely fast. This is part of why this is even an interesting pursuit in the first place, different techniques can be faster purely based on the specific circumstances. There are many ways to solve a Sudoku puzzle computationally, but I think it's difficult to choose a best approach. The more important part is how interesting an approach is, and whether using a technique that isn't intuitive or from a different domain could solve the problem. Switching to a Rubik’s Cube, the states of a Rubik's Cube can be viewed as cities on a map, where the distance between them is how many twists are needed to move from one to the other. By doing this, you've essentially converted the problem from an arbitrary puzzle into a problem of finding the shortest distance to travel to reach your destination city, the solution. It becomes a kind of Pathfinding problem. Is this necessary to solve a Rubik's Cube? Obviously not; a human solves it by looking at the colors and memorising specific movement combinations to get the desired result, but figuring out that it can be done is the whole point. The reality of the situation is that a human exploring a problem may discover useful techniques or approaches or analogies that help solve other problems. Frequently, problems get solved when someone from another discipline notices that a problem has a similar shape as something else, or can be converted into one. AI does not provide the same level of approach or technique improvement. Chain of thought isn't actually a chain of thought so much as it is a computing technique designed to generate statistically more useful output. Asking an AI to explain its reasoning or approach is a prompt that contains its original output, resulting in it generating an arbitrary output, which can be compared to a person coming up with a post-hoc rationalisation (though I would say that it isn't quite the same). It cannot really provide feedback on different techniques and approaches it used and why they did or didn't work. This is the important part of why humans try to solve "unsolvable" problems. This was why the Millennium Prize Problems were even defined in the first place. For most things in the mathematical and scientific world, solving a problem is not more important than the journey of getting there.