Large language models can draft legal briefs, compose sonnets in the style of Keats, and explain quantum mechanics to a curious child. Ask one to multiply 47 by 89, and there is a reasonable chance it will confidently produce the wrong answer. This is not a temporary limitation awaiting the next model release. It is a window into the fundamental architecture of these systems — and a reminder that intelligence comes in radically different flavors.

The disconnect baffles users who assume that a system capable of passing bar exams should find third-grade arithmetic trivial. The opposite is true. For a language model, multiplying two numbers is genuinely harder than writing a persuasive essay, because the two tasks require entirely different cognitive machinery.

How language models actually see numbers

When you type "47 × 89" into a chat interface, the model does not see two numbers waiting to be multiplied. It sees a sequence of tokens — discrete chunks that might be individual digits, partial words, or common character combinations. The number 47 might become one token; 89 another. The model then predicts what tokens are statistically likely to follow this sequence, based on patterns absorbed from trillions of words of training text.

This is pattern matching at cosmic scale, not calculation. The model has seen many multiplication problems and their answers. When the numbers are small or common, it can often retrieve the correct answer from memory, the way a human might instantly recall that 7 × 8 = 56. But novel combinations — especially larger numbers — require actual computation, which the architecture simply was not designed to perform.

A calculator follows explicit rules: take these digits, apply this algorithm, produce this output. A language model has no such algorithm. It is, at its core, a prediction engine optimized to produce plausible-sounding continuations of text. "Plausible-sounding" and "mathematically correct" are related but distinct properties.

The training data paradox

The irony runs deeper. Language models trained on more data become better at most tasks — but not proportionally better at arithmetic. A model might encounter millions of correct multiplication tables during training, yet this does not teach it the underlying procedure. It teaches the model that certain digit sequences tend to follow certain other digit sequences in mathematical contexts.

This is roughly analogous to a student who memorizes thousands of solved equations without ever learning algebra. They might recognize familiar problems and produce correct answers, but novel problems expose the absence of genuine understanding. The model is interpolating from examples, not deriving from principles.

Recent models have improved at arithmetic, partly through chain-of-thought prompting that encourages step-by-step reasoning, and partly through hybrid systems that route mathematical queries to actual calculators. But these are workarounds, not solutions. The underlying architecture remains fundamentally unsuited to precise symbolic manipulation.

What this tells us about intelligence

The arithmetic limitation is clarifying rather than damning. It reveals that language models are not general-purpose reasoning engines but specialized systems exquisitely tuned for a particular kind of task: understanding and generating human language. They excel at ambiguity, context, nuance, and the fuzzy pattern-matching that characterizes most human communication.

Human brains, notably, are also not optimized for arithmetic. We invented calculators, abacuses, and written numerals precisely because mental math is cognitively expensive and error-prone. The difference is that humans can learn procedures and apply them reliably; language models, as currently architected, cannot.

This distinction matters as these systems become embedded in consequential applications. A language model can draft a financial report beautifully while making elementary numerical errors that a spreadsheet would never commit. The eloquence masks the brittleness.

Our take

The arithmetic problem is not a scandal but a lesson in epistemic humility. Language models are staggeringly capable at what they were built to do and predictably weak at what they were not. The hype cycle encourages us to see them as nascent general intelligences, a few iterations from superintelligence. The reality is more interesting: they are a genuinely new kind of cognitive tool, powerful and limited in ways we are only beginning to map. Knowing where the edges are is not pessimism. It is the beginning of using these systems wisely.