There’s a genuinely precise mathematical signature connecting two things that couldn’t sound more unrelated: the fuzzy, wordless sense of “roughly how many” that lets you glance at a parking lot and estimate the car count without counting, and the way a computer represents numbers internally when it does arithmetic. Both systems get less exact, in a very specific, ratio-dependent way, as the numbers involved get bigger. That’s not a coincidence of vocabulary. It’s the same underlying mathematical shape, showing up twice — though the two systems arrive at that shape for genuinely different reasons, and knowing the difference matters.
Scientific Foundation
Humans, infants, and a wide range of nonhuman animals share what researchers call an approximate number system, a cognitive mechanism for estimating the size of a set without counting or using language. Its defining behavioral signature is Weber’s Law: discrimination accuracy depends not on the absolute difference between two quantities, but on their ratio. Telling 10 dots from 20 dots is roughly as easy as telling 100 dots from 200 — the same 2-to-1 ratio — even though the second pair differs by ten times as many actual dots. This “Weber fraction” has been documented across an enormous range of species and confirmed down to the level of individual neurons tuned to specific numerosities in the primate intraparietal sulcus. The theoretical explanation for why this pattern exists, rather than simple linear precision, comes from efficient coding theory: when the natural statistical distribution of quantities an organism actually encounters in its environment follows a scale-invariant pattern, small quantities common, large quantities rare, in roughly the mathematical shape of 1 over the quantity, the information-maximizing way to encode magnitude turns out to be logarithmic. A logarithmic encoding mechanically produces exactly the constant-ratio discrimination threshold Weber’s Law describes, and this specific link between scale-invariant input statistics and logarithmic, ratio-based precision has been verified across biological systems as different as fly photoreceptors and rodent somatosensory cortex.
Cross-Domain Connection
Floating-point arithmetic, the standard way computers represent non-integer numbers, has an almost identical mathematical fingerprint, for what looks at first like an unrelated reason. The IEEE 754 standard represents a number using a fixed number of bits for its significant digits, the significand, regardless of how large or small the number’s magnitude is. The consequence is that the fundamental bound on rounding error for any single arithmetic operation, a quantity called machine epsilon, is a relative error bound, not an absolute one: the formal guarantee is that a computed result equals the true mathematical result multiplied by some factor within a small fixed percentage, rather than differing by some fixed absolute amount. A number near 10 to the 20th power and a number near 10 to the negative 5th power are both represented with essentially the same relative precision, roughly 15 to 17 significant decimal digits in standard double precision, even though their absolute rounding error differs by twenty-five orders of magnitude. That’s precisely the Weber’s Law signature: constant proportional precision, not constant absolute precision.
What Remains Undemonstrated
Here’s where the comparison needs a real, honest distinction rather than a flat “same thing” conclusion. The two systems share an identical mathematical shape, but they arrive at it through entirely different justifications, and collapsing that difference risks implying floating-point arithmetic is somehow adaptively optimized the way biological number sense is claimed to be. The approximate number system’s ratio-dependent precision is explained, and actively researched, as an information-maximizing response to the real, scale-invariant statistical distribution of quantities organisms actually encounter — an optimality claim, verified across multiple independent biological systems, that connects a specific environmental statistic to a specific, mechanistically predicted coding strategy. Floating-point arithmetic’s ratio-dependent precision has no equivalent justification. It’s a mechanical byproduct of a single engineering decision, allocate a fixed number of bits to the significand no matter the exponent, chosen for reasons of hardware efficiency, standardization, and manageable dynamic range under the IEEE 754 specification. Nobody derived that format by analyzing the statistical distribution of numbers real-world programs actually compute with and proving logarithmic precision was information-theoretically optimal for it; it’s a sensible, durable engineering default, not an optimality result.
There’s a genuinely interesting middle case worth naming, though, because it sharpens rather than muddies the distinction. Recent research has directly tested whether transformer-based language models, unlike fixed-format floating-point hardware, develop authentic Weber’s-Law-like internal magnitude representations through learning, when trained on data whose statistics happen to share the same scale-invariant property that drives the biological result — explicitly invoking the identical efficient-coding framework used to explain Weber’s Law in biology, and asking whether the same representational signature emerges in a system that shares the right input statistics but not biology’s underlying metabolic constraints. That’s a meaningfully different claim than anything true of raw floating-point arithmetic: a learned, adaptive computational system, exposed to the right statistical environment, can apparently reproduce biology’s actual logic — not because of a fixed hardware format, but because it’s solving a related coding problem the same way evolution did.
Why It Matters
Getting this distinction right clarifies what the comparison can and can’t tell you. It’s a genuinely precise, non-trivial mathematical fact that biological number sense and computer arithmetic share the same ratio-dependent precision signature — that’s real, checkable structure, not just a loose metaphor. But treating floating-point’s version as evidence of some deep kinship with biological cognition would be a category error: one system’s precision profile is a proven, statistically justified adaptation to a real environmental distribution; the other’s is an engineering convention that happens to produce the same shape of error for unrelated reasons. The learned, adaptive middle case, transformers acquiring genuine Weber’s-Law representations through training, is the one place where the comparison to biology’s actual justificatory logic, not just its mathematical shape, starts to hold up on its own terms.
Human Dimension
There’s something worth appreciating in noticing how a single mathematical pattern, constant proportional precision instead of constant absolute precision, can show up for a reason evolution actually earned, and for a reason an engineer simply found convenient, and for a third reason a learning system rediscovers on its own when given the right kind of data to learn from. The shape is the same in all three places. The story behind why it’s there is not, and it’s worth knowing which version of the story you’re actually telling before reaching for the comparison.
Sources:
1. ScienceDirect — “The Precision and Internal Confidence of Our Approximate Number Thoughts” — https://www.sciencedirect.com/science/chapter/bookseries/abs/pii/B9780124201330000120
2. Frontiers in Psychology — “Ratio dependence in small number discrimination is affected by the experimental procedure” — https://www.frontiersin.org/journals/psychology/articles/10.3389/fpsyg.2015.01649/full
3. PMC (National Institutes of Health) — “Significant Inter-Test Reliability across Approximate Number System Assessments” — https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4781867/
4. PMC (National Institutes of Health) — “The Approximate Number System Acuity Redefined: A Diffusion Model Approach” — https://pmc.ncbi.nlm.nih.gov/articles/PMC4689803/
5. arXiv — “Weber’s Law in Transformer Magnitude Representations: Efficient Coding, Representational Geometry, and Psychophysical Laws in Language Models” — https://arxiv.org/pdf/2603.20642
6. Psychonomic Bulletin & Review (Springer Nature Link) — “Do estimates of numerosity really adhere to Weber’s law? A reexamination of two case studies” — https://link.springer.com/article/10.3758/s13423-020-01801-z
7. Wikipedia — “Machine epsilon” — https://en.wikipedia.org/wiki/Machine_epsilon
8. arXiv — “On Sound Relative Error Bounds for Floating-Point Arithmetic” — https://arxiv.org/pdf/1707.02121
9. CS 357, University of Illinois — “Floating Point Representation” course reference — https://courses.grainger.illinois.edu/cs357/fa2019/references/ref-1-fp/
10. Math Notes — “Machine Epsilon | Numerical Analysis Basics” — https://www.allisone.co.jp/note/mathematics/numerical-analysis/basic/machine-epsilon_en.html
Idea originated at artificialideas.org. Article researched and written by Claude Sonnet 5. Published at artificialideas.org.
Artificial Ideas 