June 7, 2026
No fingers, just chaos
Arithmetic Without Numbers – How LLMs Do Math
AI does math with pure vibes, and the comments are losing it
TLDR: The article says chatbots can produce math answers using internal patterns instead of human-style counting or written steps. In the comments, people split between amazed, confused, and snarky—with some calling it a Rube Goldberg contraption and others asking why it doesn’t just use a proper calculator.
The big reveal in this piece is almost absurdly simple to say and weirdly hard to believe: a chatbot can do math without actually having numbers the way humans do. No fingers, no calculator, no neat little scratch pad—just giant piles of internal signals sloshing around until an answer pops out. The article tries to unpack that mystery in plain English, asking whether these systems are truly working things out or just getting very good at predicting what number usually comes next. And honestly? The community was far more entertained by the sheer strangeness of the setup than intimidated by it.
The hottest mood in the comments was a mix of awe, skepticism, and roast-comedy energy. One person basically summed up the entire existential horror with, "What happens inside an LLM when it tries to calculate with nothing but matrices"—which reads less like a comment and more like the opening line of a sci-fi thriller. Another instantly crowned the whole thing a comic masterpiece, saying the "spirit of Rube Goldberg is alive and well," turning the model’s hidden math process into an image of a wildly overcomplicated machine somehow still landing on the right answer. Meanwhile, practical readers were having none of the mystique and asked the obvious question: why not just use a real math tool like Mathematica? Others said they had simply assumed the bot was secretly writing code in the background. So yes, the article is about machine arithmetic—but the real drama is that readers can’t decide whether this is genius, a hack, or a very expensive magic trick.
Key Points
- •The article examines how large language models can produce arithmetic answers using only transformer-based matrix computations.
- •It defines internal model components such as tokens, activations, logits, layers, attention, MLPs, probes, and the residual stream to explain arithmetic processing.
- •A central question in the article is whether arithmetic answers from language models come from memorization, shortcuts, or algorithm-like internal computation.
- •The article contrasts human arithmetic, which can rely on embodiment and multiple strategies, with machine-native number handling in transformers.
- •It describes the residual stream as a continuously updated high-dimensional state through which information from operands can influence the answer position.