Days of Miracle and Wonder

September has been a busy media month for those connected to the AI frontier. Most of the focus has centered on the headline grabbing, “It’s Dangerously Good” theme, engineered by Anthropic et al. This has been so successful, it managed to become a subplot at the recent UN General Assembly and a core topic for a Papal visit to France. All the fuss about AI joining the long list of End-of-Days things to be fearful of, distracted attention away from another significant AI moment that happened at the start of the month.

On September 8, OpenAI announced that they had solved the Navier Stokes problem, one of Mathematics’ Clay Institute’s Millenium Problems. For those of you who are unfamiliar, the Millenium Problems are seven great mathematical problems, for which the institute, in 2000, offered a prize of $1M for the solving of any of these problems. Prior to this month, only, one of the seven had been solved, the Poincaré Conjecture, by Russian mathematician Grigori Perelmann in 2010. To be clear, this is not the first mathematical problem to be solved using AI but, in this case, OpenAI was loudly touting its bragging rights at having solved a Millenium Problem.

So What?

This all might be interesting to mathematicians, but how you might ask is this of interest or relevant to investors?

Scale and efficiency. As investors are starting to getting nervous about how the AI trade might unwind, they are beginning to consider more closely how the big AI firms’ capex spend is going to yield dividends, particularly in light of a potential fall in price across the sector. How valuable are these companies going to end up being relative to what we think they are worth today?

To return to the solving of the mathematical problem. The scale of the endeavor is vast. According to OpenAI the resources used were 10,000 agents running in parallel for 88 hours finding the solution and 17 additional hours proofing the work. The solution itself required the exchange of 2.7 million messages and use of 130 billion output tokens. This compute is truly massive. The efficiency of the process, well that is much more difficult to quantify. An unsolved 90-year problem solved in less than 5 days—that appears efficient. The computational, energy resources deployed to do so—the efficiency is hard to assess. Then what about it being financially efficient? The cost of this as you can imagine is being kept deliberately vague but it is thought to have been in the region of $15M (with estimates ranging from $10M to $40M).

Counting the Cost

In this case, unusually, a hard dollar figure can be put on the value of the Navier Stokes solution, $1M albeit a somewhat arbitrary number given that it is its value as a prize. Is a $1M return on $15M expenditure sensible? OpenAI clearly thinks so, with money to burn, this expense is just a line item in the advertising budget. However, one would not do much bragging for winning a $1M lottery having bought $15M worth of tickets. This of course is an unfair parody, but it indicates something that should be worrying investors.

There is a tricky balance between monetizing knowledge and innovation and they are not necessarily the same thing. Solving a mathematical problem may be invaluable, in the real sense that it is cannot be valued. It can also add enormous economic value through application in the world. Say the worth of Pythagoras’ theorem or Newton’s laws of motion. Key to this value creation is that such discoveries that expand human knowledge generally, manifest themselves through the free sharing of that knowledge. It would be of very limited use to OpenAI or anyone else if the Navier Stokes or any other similar solution needs to have an OpenAI trademark or its can only be used by paying some licensing fee.

Prohibition on AI Distilleries Won’t Work

Knowledge naturally disseminates and becomes more accessible. This does not bode well for investors who may unwittingly be subsidizing the creation of AI systems, that inevitably end up having copies that are cheaper, more basic, maybe not as good, but good enough to do the tasks the end user needs to get done. This can clearly be observed in AI distillation, meaning the training of AI on someone else’s AI models. This issue is so concerning to the US administration that it prompted US Treasury Secretary Scott Bessant recently to call AI distillation, theft.

Put this in context, anyone can go and have their AI of choice access OpenAI’s 165-page Navier Stokes proof without having to spend a dime of the $15M bill to solve it and possibly use it creatively to brew up their own AI moonshine! The speed with which this type of progress accelerates is what most threatens Big AI’s long-term value and whether the investment will end up disappointing.

Beware a Corporate Request for More Regulation

It is little wonder that Big AI’s corporate leaders are signaling the need for regulation—under the auspices of protecting humanity (of course). Suddenly we are getting a whole new vocabulary of fear, “Rogue Agents Escaping” etc. Silicon Valley sounds like an out-of-control zoo with wild AI animals or penitentiary with dangerous AI agents, ready to take over the world. The general public cannot be blamed for being scared, given the self-serving narrative at play. The reality is that humanity has spent 75 years or so with the technology to severely damage/destroy itself, a reality that is not lessened by adding AI agentic destruction to the mix. The regulatory spin should be interpreted for what it is, spin, designed to maintain power and status threatened by improving tech and diminishing value. There is nothing like a bit of regulation to crush or at least slow down the less established competition. ■

Image Credit: Funkyart-images.