Is AI Math Solved?
(If you are familiar with Bettenridge’s Law then you already know the answer.)
I was in a discussion last week and this was the question: why isn’t the world talking about how bad AI is at math? The unspoken question is, is AI math solved?
If you read the popular press you’d never know it was an issue. The primary headlines revolve around the Math Olympiad and other academic contests associated with math. But these are tests of reasoning, not tests of execution.
But here’s the dirty secret: no one is incentivized to talk about the reality.
Model companies highlight reasoning benchmarks and frontier exams. That’s what moves markets. Journalists cover what’s novel and headline-worthy. Olympiad proofs sound exciting; deterministic execution guarantees do not. Meanwhile, corporations quietly experiment and avoid mission-critical math use cases.
The reality, though, is that LLMs cannot reliably execute math deterministically. They are probabilistic systems. LLMs generate outputs token by token based on learned probability distributions. They are not symbolic algebra systems. They approximate patterns in mathematical language. That is not the same as executing mathematics deterministically.
Software developers understand, and they steer clear of projects that require math, or they build mathematical models in code to solve whatever problem they are having, solving the problem as a one-off and creating technical debt they have to support forever.
Some more technical people get it, too. If you listen to engineering podcasts, it gets mentioned. But let’s be honest: most people aren’t listening to these podcasts and even if they are, technical people know this is a problem and therefore it is not news worth discussing.
As Bill and I started building TrueMath we also started talking to developers. To a person, they all recognized the problem, and all steered clear of taking on projects that require it. Now that TrueMath exists, though, they’re ready to take on projects like this again.
So, is AI math solved? No.
LLMs are powerful reasoning tools. They are not deterministic execution engines. If your workflow requires guarantees, you need a system designed for guarantees.
That’s what TrueMath is built to provide.
