You’ve probably tried asking an AI to solve a complex algebraic equation, only to watch it confidently hallucinate a completely wrong answer. We laugh at these blunders. We screenshot them. We use them as proof that AI isn’t really that smart. But we’re missing the most fascinating—and slightly unsettling—truth about how these machines actually think.
We built a machine out of algebra, only to discover it hates algebra.
It sounds like a paradox, but it’s the reality of Large Language Models. Under the hood, LLMs are massive, sprawling webs of matrix multiplications and continuous weights. They are, quite literally, made of math. Yet, if you ask them to perform explicit, symbolic algebra—the kind of rigid, step-by-step rule-following we learned in high school—they often stumble without extensive, targeted training.
But rather than seeing this as a bug, we need to recognize it for what it is: a mirror. The AI isn’t broken; it’s acting exactly like you do.
Think about it. If you throw a baseball to a friend, your brain is instantly calculating a complex trajectory involving velocity, gravity, and wind resistance. Your brain is doing calculus and linear algebra in real-time, completely unconsciously. But if I hand you a piece of paper and ask you to formally solve a differential equation for that exact same trajectory? You break out in a cold sweat.
The true magic of intelligence—human or artificial—isn’t in following rules, but in the silent, continuous math happening in the background.
As mathematician Tim Gowers recently pointed out, we are fundamentally evaluating AI’s mathematical intelligence by testing it on the exact tasks humans find cognitively difficult: formal symbolic manipulation. We judge AI by its ability to pass the exams we struggle to pass. But in doing so, we completely ignore the AI’s actual superpower.
LLMs excel at intuitive, continuous mathematics. They are incredible pattern-matching engines that navigate high-dimensional spaces with an uncanny, almost biological grace. They can write poetry, code software, and predict the next logical thought because they are effortlessly executing the continuous math of language and context. We just don’t recognize it as math because we can’t easily write it on a chalkboard.
This realization should completely recalibrate how we use these tools. If you’re forcing an LLM to act as a rigid, symbolic calculator, you’re wasting its potential and setting yourself up for frustration. Stop trying to make it follow discrete rules it inherently resists. Instead, leverage it for what it does best: intuitive pattern-matching, contextual analysis, and continuous reasoning.
We are asking machines to pass the exams we failed, while ignoring the calculus they do just to breathe.
The uncanny truth is that our artificial minds are bound by the exact same cognitive walls as our biological ones. They excel at the unconscious, continuous flow of intuition, and they struggle with the rigid, discrete steps of conscious logic. AI isn’t an alien intelligence. It’s a reflection of us, mathematical warts and all.
FAQ
Q: If AI is made of math, shouldn't it be a perfect calculator by default?
A: No. Being made of a material doesn't mean you excel at manipulating it symbolically. Your brain is made of neurons and electrical impulses, but you can't consciously rewire your synapses on command.
Q: How should I change my AI prompts based on this?
A: Stop forcing AI to do step-by-step symbolic math. Use it for pattern recognition, data synthesis, and intuitive tasks. Leave the rigid algebraic manipulation to traditional calculators and deterministic code.
Q: Is AI actually bad at math?
A: AI isn't bad at math; it's bad at human math. It's actually doing incredibly complex calculus continuously just to generate a single word. We're just measuring it with the wrong yardstick.