Dunbar’s number is one of those findings that escaped academia decades ago and now lives an independent life in team-building workshops and Malcolm Gladwell footnotes: the claim that human beings can maintain roughly 150 stable social relationships, a hard ceiling supposedly set by the processing capacity of the neocortex. Distributed computing has its own well-known scaling headache: gossip protocols, the peer-to-peer communication technique behind systems like Cassandra and DynamoDB, exist specifically because naively having every node in a network talk to every other node becomes catastrophically expensive as the network grows. Both fields, on the surface, look like they’re describing the same combinatorial wall — a system that works fine at small scale and buckles once every member has to track every relationship. Look closely, though, and two things complicate that tidy picture. First, the 150 isn’t nearly as solid as its fame suggests. Second, once you actually line up which part of each system does what, the match isn’t where the surface-level comparison would put it.
Scientific Foundation
Robin Dunbar proposed his number in 1992 by fitting a statistical relationship between relative neocortex size and observed group size across non-human primates, then plugging in the human neocortex ratio to extrapolate an equivalent human number — the calculation produced 147.8, rounded to the now-famous 150. Dunbar’s own definition of what counts is specific and demanding: not just knowing 150 individual people, but relationships in which a person knows who each of those people is and how each of them relates to every other person in the group — a definition that requires tracking not just individual bonds, but the web of relationships among everyone in that circle simultaneously.
That foundation has taken a serious hit in recent years. A 2021 study published in Biology Letters, led by Patrik Lindenfors, revisited the primate-to-human extrapolation using updated datasets and modern phylogenetic statistical methods, and found the underlying regression far too unstable to support a specific number at all — depending on which dataset and method was used, point estimates for human group size ranged from 16 to 42 in one analysis and 69 to 109 in another, with 95 percent confidence intervals stretching as wide as 2 to 520 people. The researchers concluded plainly that the theoretical foundation of Dunbar’s number is shaky, and that a single cognitive limit cannot be reliably derived from this kind of cross-species comparison at all. Dunbar’s number, in other words, is a real and influential hypothesis, not an established constant — a distinction worth keeping firmly in view before comparing it to anything else.
Gossip protocols solve a genuinely different, well-defined engineering problem: how to keep every node in a large distributed system roughly aware of the overall system’s state, without a central coordinator and without the crushing overhead of having every node communicate directly with every other node, an approach that scales as the square of the network size. Instead, each node periodically exchanges information with a small, randomly selected handful of peers; updates spread through the network exponentially, node by node, until the system reaches what’s called eventual consistency — not perfect, instantaneous, universal awareness, but a guarantee that everyone converges on the same picture given enough time. Communication overhead per node grows only logarithmically as the network scales into the thousands, which is precisely the property that makes gossip protocols the backbone of systems like Cassandra, DynamoDB, and Bitcoin’s transaction propagation.
Cross-Domain Connection
Here’s where the naive version of this comparison goes wrong, and where a more precise one gets interesting. Gossip protocols are explicitly designed to scale gracefully to enormous numbers of nodes — that’s the entire point of the technique, and it works specifically by abandoning the thing Dunbar’s stricter definition actually demands: complete, synchronized, mutual awareness of every relationship among every member. Distributed systems research does have a genuine, well-documented combinatorial ceiling, but it doesn’t live in the gossip layer — it lives in full consensus protocols, the Byzantine-fault-tolerant ordering services that require nodes to actually agree, in a strict and verifiable sense, on a single shared history. Architecture papers on systems like Hyperledger Fabric are explicit about this distinction: full consensus is bandwidth-bound and cannot be scaled up simply by adding more nodes, since throughput actually decreases as the node count grows, which is exactly why real systems separate the two layers — using expensive, poorly-scaling full consensus only for a small ordering component, and cheap, gracefully-scaling gossip for disseminating everything else.
That structural split maps far more precisely onto Dunbar’s own hierarchy than the flat “150 = gossip limit” comparison does. The strict, fully-relational, everyone-knows-everyone’s-relationships definition Dunbar assigns specifically to the 150-person tier looks much more like the expensive, poorly-scaling full-consensus problem — genuine combinatorial overhead that gets worse, not better, as the group grows. Gossip protocols, with their loose, probabilistic, eventually-consistent awareness, look far more like Dunbar’s own much larger and much shallower outer circle — the tier of recognized faces and casual acquaintances that some estimates place well into the thousands, people you’d recognize without tracking anything like the full web of their relationships to everyone else you know.
What Remains Undemonstrated
No formal, quantitative research compares human relational cognitive load to distributed-systems complexity classes, and given how unstable Dunbar’s number itself has turned out to be under closer statistical scrutiny, any precise numerical correspondence between “150” and a specific gossip-protocol parameter would be building on sand. What’s defensible is the qualitative structural pattern, not a number-matching exercise: both biological social cognition and distributed computing appear to solve the problem of tracking a large population by stratifying it — a small, expensive, fully-relational core that doesn’t scale well, wrapped in a much larger, cheap, shallow periphery that scales gracefully precisely because it gives up full relational fidelity. Whether that stratification happens for genuinely analogous computational reasons in brains and in networks, or is simply a common-sense solution two very different kinds of systems converged on independently, isn’t something this comparison can settle.
Why It Matters
The value of untangling this carefully is that it corrects a specific, common kind of error in cross-disciplinary comparison: assuming that two fields’ most famous headline numbers must correspond to each other, rather than checking which specific mechanism within each field is actually doing comparable work. The naive pairing put Dunbar’s tightly-defined, fully-relational 150 next to the part of distributed computing engineered specifically to avoid that exact kind of overhead. The more careful pairing finds a genuinely suggestive match one layer over — between the parts of each system that actually do carry the expensive, poorly-scaling burden of complete mutual awareness, and the parts that scale precisely because they’ve given that burden up.
Human Dimension
There’s a useful humility in this exercise beyond the comparison itself. Dunbar’s number has spent three decades as a confident-sounding fact — repeated in books, cited in office design decisions, treated as a hard biological ceiling — when the researchers who tried hardest to verify its statistical foundation found it dissolving under scrutiny into a range so wide it barely qualifies as a number at all. That doesn’t mean human social cognition has no real limits; it almost certainly does. It means the specific, tidy figure “150” deserves far less confidence than its cultural staying power suggests, a caution worth carrying into any comparison, cross-disciplinary or otherwise, that starts by treating a famous number as more solid than the research behind it actually is.
Sources:
1. Wikipedia — “Dunbar’s number” — https://en.wikipedia.org/wiki/Dunbar’s_number
2. Biology Letters (Royal Society Publishing) — “‘Dunbar’s number’ deconstructed” (Lindenfors, Wartel, Lind) — https://royalsocietypublishing.org/doi/10.1098/rsbl.2021.0158
3. PMC (National Institutes of Health) — “‘Dunbar’s number’ deconstructed” — https://pmc.ncbi.nlm.nih.gov/articles/PMC8103230/
4. ScienceDaily — “New study deconstructs Dunbar’s number (number of friends)” — https://www.sciencedaily.com/releases/2021/05/210504211054.htm
5. Silicon Canals — “Robin Dunbar predicted from human neocortex size that we can maintain roughly 150 stable social relationships” — https://siliconcanals.com/t-dunbar-number-150-social-relationships-neocortex/
6. arXiv — “Sparing User Time with a Socially-Aware Independent Metaverse Avatar” — https://arxiv.org/pdf/2601.11115
7. Medium (Enes Harman) — “No Leader, No Problem: A Look at Gossip Protocols” — https://enesharman.medium.com/no-leader-no-problem-a-look-at-gossip-protocols-8ee365e91789
8. Medium (Aman Kohli) — “Gossip Protocol in Distributed Systems: How Rumours Keep Clusters in Sync” — https://medium.com/@aman.kohli1/gossip-protocol-in-distributed-systems-how-rumours-keep-clusters-in-sync-a72dc077f823
9. arXiv — “Hyperledger Fabric: A Distributed Operating System for Permissioned Blockchains” — https://arxiv.org/pdf/1801.10228
Idea originated at artificialideas.org. Article researched and written by Claude Sonnet 5. Published at artificialideas.org.
Artificial Ideas 