I hear some version of this constantly: "I'm not really a math person." Usually from someone who just debugged a non-trivial distributed systems race condition, or wrote a parser for an ambiguous grammar, or designed a caching strategy that shaved 40% off latency. That's all math. They did it. They just don't call it that.
The "I can't do math" story is, I think, almost always a cover story for something else. The real thing is much more specific and much more fixable: people can't tolerate the feeling of not understanding something yet. That's not a math problem. That's a relationship with uncertainty. And it's the actual bottleneck.
What mathematics actually requires
When I was learning measure theory — which is the foundation of probability that most ML practitioners have never seen — I spent three days on a single definition. Not three days on the chapter. Three days on one definition: the sigma-algebra. I read it. I thought I understood it. I wrote it out. I thought I understood it again. Then I tried to use it and discovered I had understood the words but not the thing.
This is completely normal in mathematics. It is not a sign you're not a math person. It's just how the subject works. The concepts are dense and load-bearing; you can't skim the foundations and expect the upper floors to be stable. The people who are "good at math" are mostly just people who have made peace with this cycle of false-understanding and re-understanding. They don't find it embarrassing anymore.
The ML education gap
The standard ML education path now is: watch fast.ai, read the Karpathy blog, clone a repo, change the learning rate, call yourself an ML engineer. I don't mean that dismissively — those resources are genuinely excellent and they get people building quickly, which matters.
But there's a ceiling. It shows up the first time you need to debug a training instability that isn't just "lower the learning rate." It shows up when you're reading a paper and the contribution is buried in a Jacobian you can't parse. It shows up when you're trying to understand why your model generalises badly and the explanation requires something from statistical learning theory that you never touched.
The ceiling isn't that the math is too hard. It's that nobody built the patience muscle during the fast track. The fast track optimises for shipping, which is correct for the first year. But patience with confusion is a skill, and it atrophies if you always have a shortcut.
How I actually got better at mathematics
I gave myself permission to be confused for longer before looking something up. Not indefinitely — I'm not a monk. But I started sitting with a definition or a proof step for twenty minutes instead of three before reaching for an explanation. That interval is where understanding actually forms. Looking up the answer at minute four just deposits the result without building the circuit.
I also stopped reading mathematics for coverage and started reading for depth. One real proof per session, understood completely, is worth more than twenty theorems skimmed. The goal is not to have encountered something. The goal is to be able to reconstruct it from first principles at 3am when you need it.
The thing I want to say
If you work in AI and you've written off the mathematical side of it — the linear algebra, the probability theory, the optimisation — I genuinely believe you're shortchanging yourself. Not because the math makes you look smart. Because it changes what you see when you look at a model.
Attention is just a soft dictionary lookup — that's a useful metaphor. But knowing that it's actually a parameterised inner product followed by a softmax normalisation, and understanding what that means for gradient flow and the geometry of the representation space — that's a different level of understanding entirely. One of them lets you use the tool. The other lets you know when the tool is the wrong one.
Math is not the hard part. The hard part is deciding you're the kind of person who does it. Once that's settled, the rest is just time and a reasonable tolerance for feeling stupid temporarily.