Big news in the world of money, as billionaire Alex Gerko is offering $5 million to whoever makes “the first publicly-shared AI model to enter an AIMO (Artificial Intelligence Mathematical Olympiad) approved competition and perform at a standard equivalent to a gold medal in the IMO (International Mathematical Olympiad),” as the XTX Markets website puts it. There is an additional $5 million up for grabs in smaller chunks for people making steps towards this goal. Basically, $10 million can be yours if you can make and publicly share an AI that is really, really good at performing in math competitions. It isn’t an easy task, given that the IMO is highly competitive and tests critical thinking, logical reasoning, and creativity, all of which AI currently struggles with. XTX Markets claim that their fund “intends to spur the development of AI models that can reason mathematically” very well. Though their intentions are probably noble, I doubt that their goals will be achieved.
I applaud their requirement that the model be open source. The current leading player in AI, OpenAI, does not open source their models, despite the misleading name. Neither does Google, Microsoft, or X.Ai. Meta graciously made their Llama model open upon request for AI researchers, and then it pretty quickly got leaked on 4chan and is now accessible to everyone. There are still some great open source models out there, but the big players have made it clear that they intend to keep their sandboxes locked up. What a shame.
The problem is simply that this hypothetical incredible AI model doesn’t exist because we just don’t know how to make it yet, not because of some lack of motivation that $10 million dollars would solve. It is evident that this hypothetical model would be worth much, much more than $10 million. Let alone the model, anybody able to create the model would be worth more than $10 million to any company working on AI.
I am reminded of the Clay Mathematics Institute’s Millennium Prize Problems, seven unsolved math problems, each with a million dollar bounty for solving. These aren’t like 15*348, or even 24357657*7815263/(123712.2137632 + 12356.2 – pi1/e). As an example, the Poincaré conjecture states that “Any three-dimensional topological manifold which is closed and simply-connected must be homeomorphic to the 3-sphere.” This is the only one of the problems to be solved, and it took almost 100 years to do so. And also, Grigori Perelman, the man awarded the $1 million for solving the conjecture, rejected the prize. It’s not about the money, despite what these new-money STEM-geniuses (that made all their wealth on technologies that are at best net neutral for society) may believe. (XTX Markets, the source of Mr. Gerko’s enormous $10 billion wealth, is an algorithmic trading company. Presumably, they’ve made money for some people, lost money for others, pocketed the difference, and increased the average global temperature by 0.001 degrees Celsius.)
It seems likely that a model able to perform this task will exist one day, perhaps even soon. This model would probably share very few similarities to the current state-of-the-art AIs that have captured the hearts and minds of the public, Large Language Models that work by predicting subsequent tokens from text. This can make it difficult to generate the insights that form the basis of IMO answers, or even perform general high-school level math. In fact, it is shocking that models like GPT4 are as good at math (not very) as they are.
Leave a Reply