A truly brilliant observation by Mr. Stephen Klein in a recent post on how China May Be The Only One With Mythos because they may have saved its responses (and thus figured out how to replicate it).

(Gen-) AI LLMs are just mega math models. Really big mega math models with probabilities being computed on top of probabilities being computed on top of probabilities in force-feedback loops that reinforce its computations (which could be brilliant deductions or hallucinations that equal the acid high of the most LSD addicted junkie on the planet).

And since the outputs are dependent on the equations that define the model and the training data, if someone can recreate the training data they can reverse engineer the equations from the output if they are sufficiently adept at mathematics, or at least a close proximity.

Which means that blocking off access to those who can evaluate and improve the model will not help if those you don’t want to have the model already have it.

But this isn’t a post about the blocking of AI models (because I personally think that’s great), but a post about what happens if you try to hide your capabilities behind “proprietary algorithms based on math” assuming that no on else can recreate it if you don’t talk about it and that the IP alone justifies an unreasonable price for your product or valuation for your company.

Every country has mathematical geniuses, and more than one can come up with the next iteration of a mathematical theory at about the same time. Maybe only one gets remembered (Newton vs. Leibniz), but it doesn’t mean they didn’t both invent the same concepts at about the same time (in calculus).

And the more you trump it up, the more it entices someone to tear it down. (And figure out how you built it.)

On its own, the math alone is not the advantage you think it is. There are a lot of free scientific papers with great math. In fact, more than enough to create your own Claude, DeepSeek, Gemini, Grok, etc. But most people can’t because it’s not just the equations, it’s the parameters, the training data, and the implementation.

With regards to the implementation, just because you have server racks that can do trillions of calculations per second, that doesn’t mean you can code inefficiently. For example, an average token output by Claude requires billions of operations, and an average question posed to Claude will require 1,000 to 10,000 tokens to answer, for 1 trillion to 1 quadrillion calculations for an output. A poorly designed model could require 10, 100, or 10,000 times that.

The same goes for classical machine learning or optimization algorithms. Good implementations will require billions of calculations. Bad, trillions to quadrillions to quintillions. Responses go from real-time to hours to days to the computations never end.

Then there is the training data. You can’t judge a model implementation unless you have good training sets that are representative of real-world problems. Optimizing for theoretical problems that don’t exist in the real world isn’t helpful and may, in fact, lead to a worse solution that not even testing it at all!

The real differentiator is the expertise both in the implementation of solutions based on math and deep knowledge of the domain the solution is for. That can’t be recreated by mathematical ability alone and requires experience. And that can be export controlled (and represents the real value).