Beneath the forest floor runs a web of fungal filaments linking hundreds of trees into one connected system, trading carbon, nitrogen, water and even alarm signals. It forces you to revise a default assumption: "a tree" is not an isolated individual but a node in a network. The boundaries of a living thing are far blurrier than the eye suggests.
Mycorrhizal fungi sheath or penetrate plant roots and use their vast filamentous surface to scavenge phosphorus, nitrogen and water from the soil, in exchange for the carbon plants fix through photosynthesis—a barter that has run for 400 million years. When one fungus links its filaments to several trees at once, a "common mycorrhizal network" (CMN) forms. Nutrients and signals can travel tree-to-tree along it, with the fungus as the indispensable middle layer.
Inject a birch with a radioactive carbon isotope and weeks later you can detect it in the roots of a neighboring fir—carbon really does flow between trees via the fungus, and shaded seedlings appear to be "resupplied." But the deeper twist: the fungus is not selfless. Experiments reveal a "biological market"—the same fungus returns more phosphorus to the plant that gives it more carbon, and short-changes cheaters. It behaves less like a charitable courier and more like a calculating market-maker. The underground "cooperation" runs, at bottom, on transactional reward-and-punishment.
In distributed systems this is an overlay network—nodes don't connect directly; a middle routing layer forwards resources and messages, with the fungus as the P2P relay. In economics it is "biological market theory": supply and demand, bargaining, premiums for good partners, penalties for defaulters. In internet architecture it maps to content delivery and intermediary platforms—whoever owns the middle layer owns the pricing power and the "tolls."
When designing distributed systems, don't just watch the "nodes"—watch that unassuming middle layer: the message bus, the service mesh, the shared cache. Like the mycorrhizal web, it brings the whole system to life while quietly becoming the biggest power node and single point of failure. "Who is my fungal layer?" often decides an architecture's fate more than "how many nodes do I have."
In the system you run, what is the "invisible yet universally depended-on" middle layer? Is it forwarding selflessly, or also collecting "tolls" on its own terms, shaping who gets what resources?
Plants have no brain and no nerves, yet they sense, communicate, defend, and arguably "remember." "Plants are passive" is a thorough illusion—they simply live in a language of chemical and electrical signals we barely perceive, doing with their whole body what we do with a brain.
When a plant is eaten it releases volatile organic compounds (VOCs) into the air; neighbors that pick them up pre-emptively switch on defense genes and accumulate compounds that make their leaves harder to chew—a genuine airborne broadcast. Plants also carry electrical signals: a wound triggers a wave of calcium ions across the whole plant, and the glutamate-receptor-like genes (GLR) involved are homologous to the neurotransmitter receptors in animal nerves. The mimosa's fold and the Venus flytrap's snap run on real action potentials. Two evolutionary lineages each stumbled onto the same "excite-and-conduct" logic.
When acacias are browsed by giraffes they release ethylene; downwind kin pile tannins into their leaves within minutes, turning themselves unpalatable in advance—African rangelands have recorded mass "mysterious" poisonings of grazers because of it. More provocative is a "habituation" experiment: drop a mimosa harmlessly again and again, and after a few tries it stops wastefully folding its leaves, as if it has "learned" the drop is harmless; move it elsewhere and that non-response persists for weeks. Not a single neuron, yet a hallmark of learning appears—which is exactly why it remains fiercely debated.
In information theory, a VOC is a noisy broadcast channel—the sender must get the alarm out despite free-riders and eavesdroppers (pests can smell it too), a classic signal-reliability trade-off. In immunology it maps to "damage-associated molecular patterns" (DAMPs): don't identify the specific enemy, just broadcast "there's injury here" so the whole system goes on alert. In distributed computing it is a gossip protocol with no central coordinator—local sensing, near-neighbor broadcast, a coherent response emerging globally.
Plants are the extreme case of a decentralized system: no master, no brain—each part senses and decides locally, and a coordinated whole-body defense emerges from signal spread. When designing multi-agent systems or edge architectures, ask less "how strong should the central controller be" and more "can a local node, like a leaf, jump to the right state on nothing but one broadcast from a neighbor?" Robustness usually hides in the latter design.
Do you coordinate your systems with a central controller, or let nodes self-organize on each other's signals? If you cut out the "brain," would your system seize up—or, like a plant, keep reacting soundly on its own?
Plants have no brain, yet they make "directional" decisions—toward light, against gravity, winding onto supports. They demonstrate a badly underrated kind of intelligence: intelligence without representation. A plant needs no map of the world in its head; its body structure and growth rules are the algorithm.
The engine of phototropism is the differential distribution of auxin: light drives auxin toward the shaded side, whose cells then elongate faster, so the stem naturally bends toward the light. Gravitropism relies on starch-filled bodies (statoliths) that settle like tiny "gravity weights" to sense direction. There is no central unit "deciding which way to bend"—just one local rule (whichever side has more auxin grows faster), from which precise whole-plant orientation emerges. The rule is local; the intelligence is global.
The Darwins, father and son, found a counterintuitive split long ago: the part that senses light is the shoot tip, yet the bending happens below—the signal is "conducted" downward, with perception and action separated in space. More forage-like is the parasitic dodder: lacking chlorophyll, it lives entirely off a host, and it can "smell" host-released odors in the air and grow toward the strongest scent; offered several candidates, it will even favor the more nutritious one. A vine without so much as a nerve, making what looks like a "trade-off" choice.
In cybernetics, tropism is negative feedback—deviation from the target direction produces corrective growth. In robotics it maps to "embodied intelligence without representation": like a Braitenberg vehicle whose two light sensors wire straight to motors, it tracks light and dodges obstacles with no world model at all. In machine learning it is almost the biological version of gradient descent—step locally along some gradient (light, gravity, nutrient concentration) and approach the optimum with no global map. Tropism is gradient descent grown in soil.
Building AI systems, we reflexively think "first build a complete world model, then decide." Tropism points to another path: much intelligent behavior needs no representation, only a good local rule plus a followable gradient. Wherever you can define a clean gradient (latency, loss, reward), a simple feedback loop that "grows toward the gradient like a plant" is often more robust and far cheaper than a heavyweight planner.
Does the problem you're trying to solve really need a "planning brain"? Or would finding the right followable gradient and setting one local rule let the system grow toward the answer on its own, like a stem toward light?
We assume evolution's engine is competition and mutation—survival of the fittest. But several of life's biggest leaps came instead from merger: not one thing outcompeting another, but two independent lives fusing into one. The energy factory in every cell of your body—the mitochondrion—was once a bacterium 2 billion years ago that "wasn't digested." The deepest innovations often come when "cooperation" upgrades into "fusion."
Endosymbiotic theory holds that mitochondria descend from a bacterium engulfed by an archaeon yet surviving, and chloroplasts from an engulfed cyanobacterium. The genomes of host and engulfed then co-evolved, each retaining a division of genes, two lives welded into one. Symbiosis need not end in fusion, either—flowers and pollinators, legumes and nitrogen-fixing bacteria follow long-term mutual shaping: each change in one becomes a new selection pressure on the other. Arms races can build complexity; so can deep cooperation.
Lichen was once taken for a single "species," but it is fungus and alga (or cyanobacteria) fused so intimately they're hard to tell apart—a symbiont that fooled taxonomy itself. Deep-sea vent tubeworms are more extreme still: no mouth, no gut, no digestive tract; the whole body is a living reactor that lives entirely on sugars supplied by chemosynthetic bacteria housed inside it. It has outsourced "eating" completely to the tenants within. The line between individual and symbiont all but dissolves before such lives.
In system architecture this is the eternal "build vs. outsource" choice: internalize an external capability (engulf it like endosymbiosis, couple deeply, max out performance), or keep a loosely coupled symbiosis (call an external service via API—flexible but beholden to others). In corporate strategy it is M&A integration—is the acquired team "digested" into a uniform org, or kept semi-autonomous with its own "genes," like a mitochondrion. In open source it is the endless sway between fork (splitting to compete) and merge (converging to co-build).
Building an AI "super-individual" tech stack is essentially a chain of "endosymbiosis" decisions: which capability should be pulled in like a mitochondrion—built in-house, deeply coupled—and which kept as a loosely coupled API symbiosis like a nitrogen-fixer. The criterion isn't "which is more advanced" but whether the capability has become your core metabolism—if it is, internalize it for control and performance; if you only call it occasionally, keep the symbiosis for flexibility. Evolution has no winner that builds everything, and none that outsources everything.
Take stock of the key external capabilities you depend on (a model API, a platform): which have grown important enough to "endosymbiose"—pull in and control? Which are wiser kept as loose symbiosis? Have you dangerously outsourced some core metabolism to a "tenant" that could raise its price or vanish at any moment?