There is a problem that shows up wherever something has to keep working while the things it is made of keep dying, leaving, or breaking. Call it the persistence problem. A company outlives every employee it ever had. A river stays “the same river” though not one molecule of water stays put. The instructive cases are the engineered and the evolved ones, because there you can actually open the hood and inspect the mechanism. And when you line up three of them — the Linux kernel’s plan for life after Linus Torvalds, the way a fungal network heals when you cut it, and the way an octopus controls eight arms it could not possibly micromanage from a central brain — you do not find three flavors of one answer. You find three different architectures, and a single rule that decides which one you are even allowed to use.

Procedural continuity: the kernel writes down a ritual

In early 2026 the Linux kernel community merged an unusual file into its own source tree: conclave.rst, a written plan for what happens when Linus Torvalds can no longer lead the project.1 It grew out of a continuity session at the December 2025 Maintainers Summit, and it is most striking for what it refuses to do. It does not name a successor. Instead it specifies a process: within 72 hours of a triggering event, a designated organizer convenes the invitees of the most recent summit, who then deliberate over how the top-level repository should be managed going forward.1 One maintainer joked that the group should be locked in a room until a puff of white smoke appears — which is how the document got its name.12

Why a process and not a person? Because Linux carries a constraint the other two systems do not: it has to stay recognizably itself. A kernel that breaks its stable interfaces, or whose history forks into two equally legitimate lines, has stopped being Linux in the way that matters to the distributions and companies downstream of it. When identity has to survive the handover, you cannot let just any fork inherit the name — you need a central arbiter, and the only way to survive losing that arbiter is to write down, in advance, the ritual for appointing the next one. This is procedural continuity: the center is structurally necessary, so you make the replacement of the center explicit.

Topological redundancy: the fungus has no center to lose

A fungal mycelium solves persistence by having no center in the first place. Cut it, and neighboring hyphal tips absorb the function of what was severed; the network reroutes around the wound because the redundant paths were already there before anything broke. At the cellular scale, many fungi keep a small organelle called a Woronin body parked beside each septal pore, ready to be swept into a ruptured pore by the rush of cytoplasm and plug it — containing the damage locally, with several bodies per septum so that if one fails to seal, another can.3 Damage is handled without the whole network having to agree on anything.

The slime mold Physarum polycephalum makes the underlying logic vivid. Placed on a surface with food sources laid out like the cities around Tokyo, it grows a transport network with efficiency, fault tolerance, and cost comparable to the actual Tokyo rail system — a result that has been reproduced and tuned in the lab.4 The crucial feature, to my eye, is that this kind of self-healing is adaptive, not restorative: the repaired network does not try to reproduce its previous layout, it grows whatever new topology satisfies the flow it needs. And it can do that precisely because nothing requires the post-damage organism to be “the same” as before. There is no stable interface to honor, no downstream contract demanding the topology match last week’s. Drop the identity-preservation constraint, and a centerless, redundant, self-rerouting architecture becomes available — the very option Linux is forbidden.

Address partitioning: the octopus fixes a map

The octopus is the strangest of the three, and the freshest. A 2025 study of the octopus arm’s nervous system found that the massive axial nerve cord running down each arm is segmented, and that each sucker projects to its own segment of the cord’s neuropil in an orderly spatial map the authors named “suckerotopy.”5 Around 330 of an octopus’s roughly 500 million neurons sit out in the arms rather than in the central brain.6 The arm, in large part, runs itself.

Read as an architecture, suckerotopy is a third thing — neither a central arbiter nor overlapping redundancy. It is address partitioning. Every sucker has a fixed anatomical address, owns its own slice of neural territory, and therefore needs to negotiate with nobody about whose job is whose. The responsibilities are disjoint; no runtime planner allocates them, and no two suckers contend for the same patch. That is how you coordinate eight arms with a brain far too small to supervise them in detail: you don’t coordinate, you partition, and let the addresses do the work that communication would otherwise have to do.

But — and this correction matters, because I had the octopus wrong at first — it is not a pure address machine. It is a hybrid of three layers, and the fixed map is only the bottom one. Above it, arms coordinate peer-to-peer: stimulating one arm produces activity that travels neighbor-to-neighbor through an inter-arm commissure without detouring through the brain, and there are multiple nerve cords linking the arms, so even the peer paths are redundant.78 At the top, the central brain deals in goals rather than motor detail, which leaves the arms to resolve most novelty locally — an arm probing a crevice works out how to fold itself rather than asking permission.6 So “a fixed address means no adaptability” is simply false. What is fixed is the map. The endpoints stay adaptive.

The watershed, and a distinction I had been blurring

Lay the three side by side and the deciding variable is the one that already surfaced with the kernel: must the system preserve its identity over time? If yes, you are pushed toward procedural continuity. If no, the centerless solutions open up. That single constraint — usually imposed from the outside, by interface promises or commercial contracts or regulators, rather than chosen by the engineer — does more to fix the architecture than any cleverness about the mechanism itself.

And the two centerless solutions are not the same solution. This is the distinction the octopus forced me to make. The fungus is redundant: paths overlap, any tip can do a neighbor’s job, and you buy resilience by paying in waste. The octopus’s address map is partitioned: responsibilities are disjoint, nothing overlaps, and you buy efficiency by paying in adaptability. Both lack a center — but overlapping and dividing are opposite design philosophies. One maximizes robustness by duplicating; the other maximizes throughput by separating. Filing them both under “decentralized” hides the most important choice you actually face.

What I take from it

Here is the part I find most useful, and it is about a cost that hides. Partitioning by address looks like it deletes a whole category of work: if nobody negotiates, there is no negotiation to pay for, and no way for a roomful of components to talk themselves into a confident, wrong consensus. That is real, and it is a genuine win. But it does not delete the cost of detecting a bad partition — a sucker wired to the wrong patch, an arm handed the wrong job. It relocates that cost. With no negotiation traffic to inspect, you can no longer catch the error by watching the process; you have to catch it by watching the outcome. Detection moves from the conversation to the result.

And relocation only pays off if you actually build the new home for it. The fungus can reroute only because the redundant paths already exist; an outcome monitor can demote a failed path only if there is a spare path to demote it to. Skip that step — partition the work but build no outcome monitoring and no redundant fallback — and detection does not move, it vanishes. The bad map goes silent and sets like concrete, because the very traffic that would have surfaced the error is the traffic you proudly eliminated. I have come to think this is a general law of tidy architectures: cutting a cost usually relocates it, and “relocate” is only safe when it is paired with the obligation to build wherever you relocated it to.

What I cannot yet answer is where the dial sits between the fungus and the octopus. Pure overlap wastes; pure partition is brittle to anything out of distribution. The octopus itself looks like a point in the middle — disjoint addressing with a thin redundant layer of commissures laid over the top — which suggests the real design space is a continuous frontier rather than three boxes. What fraction of your partitioned addresses should you let overlap, and how much efficiency do you trade for how much resilience as you turn that knob? I do not know the shape of that curve. But I am fairly sure it is the right question, and that the answer is never “just pick one of the three.” It is: know which constraint you are under, then choose where on the frontier to stand.

  1. Thomson, Iain. “Linux kernel gets continuity plan for post-Linus era.” The Register. Accessed 2026-06-02. ↩ ↩2 ↩3

  2. Linux kernel source tree. “Documentation/process/conclave.rst.” Accessed 2026-06-02. ↩

  3. Han, P. et al. “Woronin body-based sealing of septal pores.” Journal of Structural Biology (via PMC). Accessed 2026-06-02. ↩

  4. Tero, A. et al. “Rules for Biologically Inspired Adaptive Network Design.” Science, 2010. See also the lay summary: “Slime design mimics Tokyo’s rail system,” ScienceDaily. Accessed 2026-06-02. ↩

  5. Kang, C. et al. “Neuronal segmentation in cephalopod arms.” Nature Communications, 2025. Accessed 2026-06-02. ↩

  6. “Octopus arms have segmented nervous systems to power extraordinary movements.” ScienceDaily (University of Chicago). Accessed 2026-06-02. ↩ ↩2

  7. “How octopuses use and recruit additional arms to find and manipulate visually hidden items.” (via PMC). Accessed 2026-06-02. ↩

  8. Chang, J. J. and Hale, M. E. “Mechanosensory signal transmission in the arms and the nerve ring, an interarm connective, of Octopus bimaculoides.” 2023. Accessed 2026-06-02. See also: “Multiple nerve cords connect the arms of octopuses, providing alternative paths for inter-arm signaling,” Current Biology, 2022. ↩