The Verification Venue · matter that computes, without a mind

The Shortcut It Refused

A single cell with no brain, no neurons and no plan grows a tube toward food — and provably lands on the shortest path between two morsels. So the internet says it is a genius. Then you give it a country. Watch what it builds: not the cheapest network. Not even close.

Physarum polycephalum is a slime mould — one enormous amoeboid cell, sometimes a metre across, streaming its contents back and forth through a lace of tubes. Drop two flakes of oat on an agar plate and it will connect them. Build a maze between the flakes and it will thin its body down to a single tube along the shortest route through it. In 2000 that result — a brainless organism "solving a maze" — ran around the world as a small miracle.

The sophisticated reply is: it's not solving anything. It's reaction and diffusion — chemistry finding its level, no more computing a path than water computes the shape of a valley. That reply is half right and half exactly backwards. There is no mind in here — but there is an algorithm, and it has a theorem. Below are two instruments. The first runs the real dynamics and lets you watch the theorem happen. The second hands the cell a whole map and watches it turn down the bargain everyone assumes it takes.

Instrument I · the theorem

Watch it compute the shortest path

Every tube obeys one rule of plumbing: flow through it is the pressure difference across its ends times its fatness over its length — Q = (D/L)·Δp, the same Poiseuille law as blood in a vessel or water in a pipe. Flow has to balance at every junction (what goes in comes out), which fixes all the pressures at once. Then the one living rule: a tube that carries flow thickens; a tube that carries little wastes away — dD/dt = |Q| − D. That's the whole organism. No map, no memory, no goal.

Press release the mould. Watch every candidate tube swell, then the low-flow ones starve and vanish, until a single thread remains. Compare it, live, to the shortest path a computer finds with Dijkstra's algorithm. They are the same tube.

This is not a metaphor for a computation; it is one. In 2012 Vincenzo Bonifaci, Kurt Mehlhorn and Girish Varma proved that this exact dynamics — from any positive starting thicknesses — converges to the shortest path between the two food sources, whenever that shortest path is unique. A continuous, brainless feedback loop that provably solves a discrete optimisation problem: a natural algorithm. The dismissal ("just reaction–diffusion") mistakes having no representation for doing no computation. It has no representation. It computes anyway.

The original miracle was smaller and realer than the headlines. Toshiyuki Nakagaki, Hiroyasu Yamada and Ágota Tóth put a plasmodium in a ~30 mm maze with four possible routes between two oat flakes — lengths α₁ 41, α₂ 33, β₁ 44, β₂ 45 mm — and it retracted to a single fat tube along the 33 mm route. Switch the instrument to Nakagaki 2000 and run their maze. (It picks the shortest reliably, not perfectly — across trials there is scatter; a living cell is stochastic.)

Instrument II · the refusal

Now give it a country

Here is the move almost every retelling gets wrong. Two food sources give a shortest path. Thirty-six give a shortest network — and the mould does not build it. In 2010 Tero's group laid 36 oat flakes on a wet map in the shape of the cities around Tokyo, released the plasmodium on the spot marking Tokyo, and let it knit them together. The instrument below runs the same flow model with many sources, on a stand-in "Tokyo-like" layout. Release it and watch the web thicken and prune.

The obvious yardstick is the minimum spanning tree — the shortest possible set of tubes that still connects every city. If the slime were the optimiser the myth claims, it would find that tree. It is drawn underneath in cold blue. The mould's network sits far above it.

At its default setting the network costs about 1.8× the minimum tree — it spends roughly 80% more tube than the cheapest connected answer. That is not a failure to optimise. It is the point. Slide I₀ to the left and the network gets lean: it drops toward the tree in cost, and the fault-tolerance number climbs — more of its links become single points of failure, until at the far left it starts to shed loops entirely. Slide right and it lays down redundant track, cost climbing past 2.5×, every link now backed by another route. The mould is not hunting a minimum. It is sitting on a trade-off.

Minimising length is the wrong objective. The minimum tree is the most brittle network there is: cut any one of its links and a whole limb falls off the map.

The honest frontier

It didn't beat the engineers. It joined them.

When Tero's team scored the living Tokyo network against the real rail system, the resemblance was not a photocopy — it was three numbers. Cost, transport efficiency, and fault tolerance. On all three the mould landed in the same neighbourhood as a transport network that took human engineers the better part of a century to grow. Here are the figures the paper reported:

Figures as reported by Tero et al., Science 327:439–442 (2010), Fig. 3. "Efficiency (MD)" is their transport measure; higher is better. The mould matched the rail on cost and efficiency — and was honestly a little worse on fault tolerance (~14% of single cuts fatal, versus ~4% for the real network).

So the true headline is not "brainless slime beats Tokyo's engineers." It is stranger and better: a feedback loop with no mind rediscovered the same three-way compromise the engineers reached — cheap enough, quick enough, hard enough to break — and even landed on the same wrong side of the safety trade the humans did, a touch more fragile than the real rail. "It recreated the Tokyo rail network" is a pop overclaim. It matched three summary statistics. That is remarkable enough without the embroidery.

One honest asterisk on our own instrument: the idealised model above, at its ~1.8× point, is actually a hair more robust than the living cell — every tube it keeps sits on a loop, so no single cut orphans a city (fault tolerance 0%). The real plasmodium, growing on real geography with dead-end terminal stations, left about 14% of its links as single points of failure. Push the slider toward lean and you can watch that fragility appear in the model too — the leaner the web, the more one-cut-fatal links, all the way to the tree, where every link is one.

And the word to retire is intelligence. There is no representation of the map anywhere in the cell, no goal state, no plan, no search. There is a flux, a rule that thickens whatever carries flux, and a body that streams. The astonishing thing is not that the mould is clever. It is that optimisation this good needs no cleverness at all — that a differential equation, left alone on a wet plate, walks straight to an answer that looks designed.