{"id":837,"date":"2026-01-13T09:00:00","date_gmt":"2026-01-13T08:00:00","guid":{"rendered":"https:\/\/neurostratum.com\/?p=837"},"modified":"2026-05-14T11:03:42","modified_gmt":"2026-05-14T09:03:42","slug":"when-machines-count-without-calculating","status":"publish","type":"post","link":"https:\/\/neurostratum.com\/index.php\/en\/2026\/01\/13\/when-machines-count-without-calculating\/","title":{"rendered":"When Machines Count Without Calculating: A Journey to the Heart of a Misunderstanding"},"content":{"rendered":"\n
\n \n
\n

MATHEMATICS & AI<\/a><\/p>\n

By Jp@NeuroStratum \u2014 Originally published on January 13, 2026<\/p>\n <\/div>\n\n \n

\n

In brief<\/em> \u2014 LLMs can produce \u00ab\u00a07 \u00d7 8 = 56\u00a0\u00bb while doing \u00ab\u00a0nothing more\u00a0\u00bb than predicting the next word. A paradox? No \u2014 a misunderstanding. The statistical mechanism doesn’t determine the sometimes uncanny behavior. These models have learned to mimic the arithmetic patterns of their training data. It often works, but it guarantees nothing. The moment the numbers grow or the formats stray from familiar ground, the house of cards starts to wobble. The rule of thumb: if accuracy matters, hand it off to a real calculator. If the stakes are low, an approximation will do.<\/p>\n

\u23f1 Estimated reading time: 7 minutes<\/p>\n <\/div>\n\n \n

\n

How LLMs produce arithmetic results \u2014 and why they don’t always deserve your trust.<\/p>\n <\/blockquote>\n\n \n

The Objection That Stings<\/h2>\n\n

\u00ab\u00a0Large language models don’t think. They predict the next token.\u00a0\u00bb<\/p>\n\n

The claim lands like a verdict. And without fail, the counterattack fires back: \u00ab\u00a0Oh really? Then how do they handle 347 \u00d7 28?\u00a0\u00bb<\/p>\n\n

You have to admit, the question is sly. If these machines do nothing but guess the next likely word in a sentence, how on earth could they handle a multiplication? Nobody’s ever seen a parrot do long division.<\/p>\n\n

Except a delicious misunderstanding hides behind the exchange. \u00ab\u00a0Predicting the next token\u00a0\u00bb describes the technical how<\/em> \u2014 not the observable what<\/em>. It’s confusing the mechanics of the piano with the music that comes out of it. A pianist produces sound by striking keys; that doesn’t reduce Chopin to random percussion.<\/p>\n\n

Let me try to clear this up: untangle the paradox, see when LLMs get their calculations right (and why they sometimes fail so spectacularly), and above all give you a practical compass for knowing when to trust them.<\/p>\n\n

Ready for the journey?<\/p>\n\n \n

First, Let’s Get Our Terms Straight<\/h2>\n\n

\u00ab\u00a0Predicting the next token\u00a0\u00bb is the heart of the engine. The model looks at what came before, estimates the probability of each possible word, and picks one. Then it starts again. That’s the mechanism \u2014 not the purpose.<\/p>\n\n

\u00ab\u00a0Reasoning,\u00a0\u00bb in the strong sense, means applying logical rules systematically, with guarantees of validity. A formal computation program reasons: it unfolds axioms and inference rules without flinching.<\/p>\n\n

The distinction that changes everything has three layers: the mechanism<\/em> \u2014 how the output gets made (statistical prediction); the behavior<\/em> \u2014 what the output appears to show (solving a problem); and the guarantee<\/em> \u2014 what’s formally assured (spoiler: nothing, on the LLM side).<\/p>\n\n

What the research tells us is that the behavior can look strikingly like reasoning even when the underlying mechanism offers none of its guarantees. It’s not human reasoning in the strict sense. But it’s not \u00ab\u00a0nothing\u00a0\u00bb either. It’s something else \u2014 and that something is fascinating.<\/p>\n\n \n

Try It Yourself<\/h2>\n\n

Open your favorite LLM \u2014 Claude, GPT, Gemini, it doesn’t matter \u2014 and run these three exercises by it.<\/p>\n\n

First test: the warm-up.<\/strong> Calculate: 234 + 567. Expected answer: 801. Most models handle it without flinching.<\/p>\n\n

Second test: turning up the heat.<\/strong> Calculate: 987654321 \u00d7 123456789. Now things get murky. The exact answer is eighteen digits long, and the models often spin their wheels.<\/p>\n\n

Third test: the nasty trap.<\/strong> Which is bigger: 9.11 or 9.9?<\/p>\n\n

You know the answer: 9.9 is bigger. But some models reply \u00ab\u00a09.11\u00a0\u00bb \u2014 because they treat the decimal portion as a separate integer (11 > 9).<\/p>\n\n

Why these differences? The key is called tokenization. To you, \u00ab\u00a0987654321\u00a0\u00bb is a number. To the model, it’s a string of chunks sliced up according to rules that have nothing to do with mathematics. See the gap?<\/p>\n\n \n

How It Works (When It Works)<\/h2>\n\n

Memory of regularities.<\/strong> An LLM trained on billions of texts has run into \u00ab\u00a07 \u00d7 8 = 56\u00a0\u00bb a staggering number of times. It’s absorbed these patterns like a sponge. For small numbers, this implicit memorization does the heavy lifting. No need to \u00ab\u00a0calculate\u00a0\u00bb: recognizing the pattern and completing it is enough.<\/p>\n\n

The limits of generalization.<\/strong> The GSM-Symbolic study made waves in 2024. Changing the numerical values of a problem (without touching its logic) shifts performance. More troubling: adding irrelevant information can drop performance by 65%. The noise pulls the model off course.<\/p>\n\n

Exact or plausible?<\/strong> That’s the knot at the heart of the problem. An LLM doesn’t intrinsically distinguish between \u00ab\u00a0correct answer\u00a0\u00bb and \u00ab\u00a0answer that sounds right.\u00a0\u00bb To the model, \u00ab\u00a07 \u00d7 8 = 56\u00a0\u00bb and \u00ab\u00a07 \u00d7 8 = 54\u00a0\u00bb are two strings of tokens; one simply appears far more often in the training data.<\/p>\n\n \n

The Distinction That Changes Everything<\/h2>\n\n

Empirical performance:<\/strong> \u00ab\u00a0The model answers correctly in X% of cases.\u00a0\u00bb That’s a statistical measure. Useful, but partial.<\/p>\n\n

Formal guarantee:<\/strong> \u00ab\u00a0The system always produces the right answer.\u00a0\u00bb That’s what your pocket calculator offers.<\/p>\n\n

The difference isn’t one of degree. It’s one of nature. It’s the difference between \u00ab\u00a0often works\u00a0\u00bb and \u00ab\u00a0always works.\u00a0\u00bb Between talent and certainty.<\/p>\n\n \n

The Pianist and the Mathematician<\/h2>\n\n

A virtuoso can play a sonata sublimely. But a wrong note can slip in. The mathematician proving a theorem doesn’t get that luxury: either the proof holds, or it collapses.<\/p>\n\n

The LLM is on the pianist’s side. Brilliant, often right, sometimes off-key.<\/p>\n\n \n

The Real Solution: Hybrid Systems<\/h2>\n\n

The naked LLM.<\/strong> Default mode. You ask a question, the model generates a text answer. No guarantee of arithmetic accuracy. It’s text about calculation \u2014 not calculation itself.<\/p>\n\n

The tooled LLM.<\/strong> This is the approach that reshuffles the deck. GPT-4 with Code Interpreter, Claude with its tool-use capabilities: the LLM understands the question, generates the code, and an external interpreter runs the actual computation. Adding tools cuts errors by a factor of 5 to 13. That’s not a marginal improvement \u2014 it’s a paradigm shift.<\/p>\n\n \n

What to Take Away<\/h2>\n\n

LLMs can produce correct arithmetic results while being \u00ab\u00a0only\u00a0\u00bb token predictors. These models have learned regularities that work well in common cases. And they fail in predictable ways at the edge cases<\/em>.<\/p>\n\n \n

The Rule of Thumb<\/h2>\n\n

If accuracy is critical \u2014 accounting, engineering, medical work \u2014 offload to a dedicated tool and check the result. Always.<\/p>\n\n

If the stakes are low \u2014 estimation, orders of magnitude \u2014 an approximation will do.<\/p>\n\n

In between? Use your judgment. And when in doubt, a calculator costs nothing.<\/p>\n\n

As the old saying goes: \u00ab\u00a0Trust, but verify.\u00a0\u00bb LLMs didn’t invent that advice. They’ve just made it more relevant than ever.<\/p>\n\n \n

And You \u2014 What Do You Think?<\/h2>\n\n

Have you ever been surprised \u2014 pleasantly or not \u2014 by an LLM’s arithmetic?<\/p>\n\n

Share your experience in the comments: I’m curious whether you’re in the verify-everything camp, or the trust-your-gut camp.<\/p>\n\n \n

\n\n \n

Written with the support of AI to help organize the thinking and shape the language.<\/p>\n

Jp@NeuroStratum<\/p>\n\n \n

For Further Reading<\/h2>\n