The Tokyo Metro Experiment
In 2010, Atsushi Tero of Hokkaido University did something that sounded a bit silly. He placed a slime mold on a map of Tokyo, with oat flakes (the slime mold's favorite food) at the locations of 36 major train stations, turned off the lights, and left.
When he returned 28 hours later, the slime mold had spread from the central station outward, connecting all 36 oat flakes with its network of veins. The stunning part: the network was almost identical to the Tokyo subway system—the same loops, the same radial lines, the same connection logic, the same fault-tolerant redundancy. A transportation network that engineers spent decades optimizing was recomputed in 28 hours by a single cell with no brain.
This was not mysticism. Tero published the results in *Science*, with comparison figures showing the striking overlap between the slime mold network and the Tokyo metro.
What Is This Organism?
*Physarum polycephalum*, the many-headed slime mold, is not a plant, animal, or fungus. It is a giant single cell: one cell membrane enclosing many nuclei. A mature specimen can spread across a square meter with tens of thousands of nuclei—yet it remains one cell.
It has no brain, no neurons, no synapses, no hippocampus, no amygdala—none of the structures found in neuroscience textbooks. It only has cytoplasm flowing inside its membrane.
Solving the Maze
In 2000, Toshiyuki Nakagaki of Nagoya University placed Physarum in a maze with two exits, one holding oat flakes. The organism spread its veins in all directions, exploring every path. After a few hours, it found the food. Then:
It retracted all veins that were not part of the shortest path, leaving a single optimal connection between the entrance and the food.
The paper appeared in *Nature* under the title *Maze-solving by an amoeboid organism*. This was not a reflex or simple chemotaxis—it was selecting the optimal solution among multiple options.
Externalized Memory: Escaping the U-Shaped Trap
In 2012, Chris Reid of the University of Sydney tested Physarum on the classic "U-shaped trap" problem from robotics: the goal sits behind a U-shaped barrier, and gradient-following agents get stuck at the bottom of the U. To escape, an agent must move away from the goal and go around the barrier.
Physarum escaped. Reid found why: as it moves, the slime mold leaves a layer of extracellular slime—a transparent, non-living glycoprotein secretion. Crucially, Physarum avoids its own slime. Attracted by food odors, it first rushes into the bottom of the U, but detecting its own slime tells it "I've been here, take another route." After a few attempts, it navigates around the barrier to the food.
Reid put it in the first sentence of his abstract: "a brainless slime mold constructs a form of spatial memory."
The memory is not in a brain. It is on the ground. Physarum externalizes "where I have been" into the environment and reads it back later—the same logic as ant pheromone trails or humans marking trees. Reid proposed a bolder idea: externalized spatial memory may be a functional precursor to internal memory. In evolution, the ability to write memory into the environment may have preceded the ability to store it in neural networks. The primitive form of memory is not the synapse—it is the slime.
Learning Without Neurons: Habituation
In 2016, Romain Boisseau of the University of Toulouse asked whether Physarum could learn. The simplest form of learning is habituation: repeated exposure to a stimulus weakens the response—like eventually sleeping soundly next to an airport.
Boisseau placed Physarum in a petri dish with a bridge to a dish of oats. Some bridges were harmless agar; others contained quinine (a bitter substance Physarum dislikes).
On day one, slime molds facing the quinine bridge showed clear aversion: taking 2.5 hours to touch the bridge and 4 hours to cross it, with narrowed veins. The agar-bridge control crossed within an hour. Over the following days, the quinine-group got faster and faster; by day six, they crossed the quinine bridge almost as quickly as the agar bridge.
Then Boisseau swapped the stimulus: quinine was replaced with caffeine. Physarum showed strong aversion again—it had habituated only to quinine, not to caffeine. If the cause were sensory or motor fatigue, the response should be blunted for any stimulus. Instead, sensitivity returned immediately—the habituation was stimulus-specific.
Withholding quinine for two days restored the slow response—spontaneous recovery.
Declining response plus spontaneous recovery, with stimulus specificity ruling out fatigue, are the gold-standard criteria for habituation. Boisseau wrote: this is the first clear demonstration of learning in a non-neural organism. A single cell without neurons learned to ignore a bitter substance—and forgot it after two days.
What This Means for Learning
Humans have about 86 billion neurons with thousands of synapses each. Our learning is synaptic plasticity. Physarum has not a single synapse—yet it learns.
This means learning requires no neurons, no synapses, no specific hardware. Learning is something more fundamental: a biological system's decaying response to repeated stimuli, parameter adjustment in feedback loops, statistical adaptation of any stateful system facing its environment. Physarum's "hardware" is a network of chemical oscillators in its cytoplasm: local regions oscillate at frequencies modulated by local conditions (faster near food, slower near repellents), coupled by cytoplasmic flow—a distributed oscillatory computation, not neurons, but computation nonetheless.
The molecular mechanism of this learning remains unknown. But one thing is clear: learning is not the nervous system's invention; it is a fundamental capacity of life. Nervous systems merely optimized it to the extreme.
The AI Connection
Large language models have two kinds of memory: one in the weights (written during training), one in the context window (prompts, retrieved documents, conversation history). The former resembles "synapses"; the latter resembles "slime."
RAG (retrieval-augmented generation) pulls external documents into the context so the model can answer based on them. The model itself is unchanged, but its "working memory" is filled with external information—the same architecture as Physarum writing memory into slime and reading it back later.
This similarity is structural, not metaphorical. Physarum's slime and a RAG vector database follow the same design principle: externalizing memory saves internal capacity, at the cost of re-reading every time. Ant pheromones, human notes, LLM contexts—different implementations of the same strategy.
Reid said externalized memory may precede internal memory. In AI, the order was reversed: we built internal memory (weights) first, then discovered we needed external memory (RAG). Evolution went outside-in; engineering goes inside-out. But the destination is the same: an effective intelligent system needs both internal and external memory.
Emergence Without a Designer
When Physarum solved the Tokyo metro problem, it did not know it was "solving" anything. It only spread from food, connected food sources, retracted redundant paths, and kept the shortest ones. The rules are local: cytoplasm oscillators respond only to local chemical signals, with no central controller or global planner. Yet after 28 hours, a globally optimal solution emerged.
There is a deep resemblance to Transformer training. Each attention head does one local, simple thing: compute weights from a query and key, then weight the values. No head has a global view. But after trillions of training tokens, global capabilities emerged—translation, reasoning, code, jokes.
The emergence of intelligence requires no global designer. It requires enough local units, the right feedback mechanism, and enough time. Physarum evolved its oscillatory computation over a billion years; Transformers emerged attention over trillions of tokens. In both cases, no one designed "intelligence"—it emerged from the repeated iteration of simple rules.
Nakagaki's 2000 *Nature* paper concluded: "even single-celled organisms may possess a primitive intelligence for solving complex network problems." Twenty-four years later, we built systems with no cells, no organisms, nothing but matrix multiplication—and they, too, exhibit a kind of primitive intelligence.
Perhaps "intelligence" was never meant to be bound to "brains." It is something more fundamental—any stateful system, given enough feedback loops and iterations, will grow some form of intelligence. Brains are one implementation. Slime molds are another. Transformers are a third.
Next time you see a slimy blob on the roadside, don't rush past it. It might be solving a problem you can't.
---
References:
- Nakagaki T. et al. (2000) "Maze-solving by an amoeboid organism." *Nature* 407: 470.
- Tero A. et al. (2010) "Rules for biologically inspired adaptive network design." *Science* 327: 439-442.
- Reid C.R. et al. (2012) "Slime mold uses an externalized spatial memory to navigate in complex environments." *PNAS* 109(43): 17490-17494.
- Boisseau R.P. et al. (2016) "Habituation in non-neural organisms: evidence from slime moulds." *Proc. R. Soc. B* 283: 20160446.