A-Shroom

Two pictures. Pick one.

You eat the output layer

This is the part that really got me. Oyster mushrooms yield 10–15 kg per square meter per year. A typical datacenter footprint, converted to mushroom cultivation, would produce over a thousand tonnes of high-protein food annually.

And here's the wild part: you can harvest the mushrooms (the fruit bodies) without killing the mycelium network underneath. The computing substrate persists while feeding people. Dual-use in the most literal sense.

No datacenter has ever fed anyone.

Jobs, carbon, soil, water

A datacenter employs a handful of engineers. A mushroom operation needs farmers, foresters, mycologists, probe technicians, soil health people. These are local jobs. You can't offshore a forest.

The ecosystem angle is almost unfair: mycelium sequesters carbon (glomalin), improves water retention by up to 5×, builds topsoil, supports biodiversity. You could deploy fungal computing on contaminated land and clean it while computing. Mycoremediation actually works.

In 1976, a researcher named Slayman stuck a microelectrode into Neurospora crassa — a common bread mold — and watched the oscilloscope trace something impossible. Spontaneous action potentials. Depolarization, repolarization, refractory period. The whole thing. The same electrical signature we use to define neural activity, happening in a fungus.

That paper sat around for decades, basically a curiosity. Then in the mid-1990s, Olsson and Hansson showed that Armillaria bulbosa generated electrical signals when it sensed food sources just a centimeter or two away. The fungus wasn't just alive. It was paying attention.

The 50-word vocabulary

Fast forward to 2022. Andrew Adamatzky at UWE Bristol publishes a paper in Royal Society Open Science that breaks my brain a little. He recorded electrical spiking activity across four fungal species — ghost fungi, enoki, split gill, caterpillar fungus — and found the spikes clustered into trains. Not random noise. Structured patterns, resembling vocabularies of up to 50 distinct "words." The core lexicon was about 15–20 patterns, used most frequently. And here's the kicker: the distribution of these spike-words follows Zipf's law. The same statistical signature as human languages.

"There is also another option — they are saying nothing. The spikes may just be random fluctuation. I was just curious to compare." — Andrew Adamatzky

I love that quote. Honest uncertainty. But the follow-up work keeps stacking up: psilocybin fungi spiking patterns, electrical frequency discrimination by Pleurotus ostreatus (2023), a definitive state-of-field review in FEMS Microbiology Reviews (January 2025). Whatever these signals are, they're not nothing.

From signals to logic gates

Adamatzky didn't stop at listening. By 2018 he'd published "Towards Fungal Computer," proposing Basidiomycetes as literal computing devices. By 2021 his lab had built Boolean logic gates — AND, OR, NOT — implemented in living mycelium. By 2022, "Mining Logical Circuits in Fungi" appeared in Nature Scientific Reports. In 2023 he published the book Fungal Machines through Springer. The man has been methodically turning mushrooms into computers for the better part of a decade.

The key insight: mycelium-bound composites implement functions from all classes of cellular automata complexity, including computationally universal ones. That's not "kind of like computing." That's Turing-complete. On mushrooms.

Shiitake RAM and mushroom chips

Then Ohio State showed up with something that made me sit up straight. October 2025: they grew memristors from shiitake mushrooms. Actual memory devices, switching at 5,850 Hz. They demonstrated pattern recognition with 90±1% classification accuracy on test tasks. And the weird bonus: radiation resistant. The fungal tissue handles radiation levels that would scramble conventional electronics. Someone at NASA must have noticed.

Around the same time, a bioRxiv preprint (August 2025) described PEDOT:PSS-infused mycelium chips — you soak the hyphae in a conductive polymer and they become electrodes. The fungus literally grows your circuit board. And separate work on reservoir computing showed mycelial networks classifying MNIST digits in simulation. Handwritten number recognition. On a fungal substrate.

The wood wide web

While Adamatzky was building logic gates in Bristol, Suzanne Simard at UBC was uncovering something equally wild underground. Her 1997 Nature paper demonstrated that Douglas fir and paper birch were exchanging carbon through shared mycorrhizal fungi. Not competing. Sharing. "Mother trees" transferring up to 40% of their photosynthates to shaded seedlings — and they can recognize their own kin.

Then came Babikova's aphid warning study in 2013, one of the most cited papers in mycorrhizal research. Bean plants connected by fungal networks. When one was attacked by aphids, its neighbors activated their chemical defenses before being attacked themselves. The warning signal traveled more than 20 cm through the common mycorrhizal network. Unconnected plants showed zero pre-emptive response. The network isn't just a nutrient pipeline. It's an alarm system. And it's compatible with about 90% of plant species on Earth.

The architecture is eerily familiar

The network topology reads like a neural architecture spec sheet. Branching points, tips, and fusion points form nodes; hyphae form edges. The resulting graph has small-world properties: high clustering plus short path lengths, the same topology that makes both brains and the internet efficient. High-traffic pathways get reinforced while underused ones get pruned. That's synaptic plasticity. It's scale-free, fractal, genuinely 3D. Silicon wafers can't do that.

Memory exists too. Fungi remember growth directions toward resources. They make decisions at network forks. They habituate — decreased response to repeated stimuli, same as neurons. Nicholas Money's 2021 "fungal mind" hypothesis isn't as crazy as it sounds.

I stared at these numbers for a while. The speed difference is brutal. Orders of magnitude slower. But look at the energy and resilience rows. What if speed isn't the only axis that matters?

The speed treadmill

Remember dial-up? 56k modem screaming into the void, waiting 45 seconds for a JPEG to load from the top down like a slow curtain reveal. Then broadband hit. Then fiber. Then 4G, 5G. And we're still impatient. I tap my foot waiting for a page to load in 1.2 seconds. My grandparents waited three weeks for a letter.

Every speed improvement recalibrates our expectations. The infrastructure gets faster. Our patience gets shorter. We've been on this treadmill for decades and nobody's asked whether we should get off.

The Slow Movement

In 1986, McDonald's tried to open a restaurant near the Spanish Steps in Rome. Carlo Petrini didn't write a petition. He handed out bowls of penne. That launched Slow Food — snail logo, now active in 160+ countries. Then came Slow Cities (Cittaslow), Slow Media, Slow Fashion. The Slow Media Manifesto of 2010 argued "not about fast consumption but about choosing the ingredients mindfully." There's even a Slow Internet site at slowinternet.info.

These aren't Luddite movements. They're asking the same question I keep coming back to: faster for whom? And faster toward what?

Walking and thinking

Darwin walked his "thinking path" at Down House every single day. Nietzsche wrote that "all truly great thoughts are conceived while walking." Einstein talked about "combinatory play," ideas bumping into each other during daydreams and long walks. Rebecca Solnit put it perfectly: "The mind, like the feet, works at about three miles per hour."

Your best ideas come in the shower, on walks, in the gap between tasks. Moments of deliberate un-productivity. The forest doesn't buffer. It doesn't optimize for engagement.

What actually needs speed?

Financial trading, I guess (though flash crashes suggest even that's debatable). Loading a webpage, sure. Real-time video, yes.

But environmental monitoring? Soil health changes over months. Climate patterns emerge over decades. Agricultural decisions operate on seasonal timescales. The soil doesn't care about your sprint velocity. Some problems require decades of patience, not nanoseconds of compute.

Not every system needs to think in nanoseconds. Some of the most important questions on Earth unfold over years.

Turing showed computation is substrate-independent. Adamatzky showed mycelium computes. So... why not mushroom intelligence?

Honestly? A lot stands in the way. But the gap is closing, and the parallels are harder to dismiss than you'd expect.

Mycelium already does multi-agent AI

Every hyphal tip in a mycelial network makes its own decisions about growth direction. Millions of autonomous agents, no central controller, optimizing for resource acquisition through a combination of electrical spikes and chemical signals. They use external tools (secreted enzymes), they communicate, they coordinate, they self-heal, and the network literally restructures its own topology based on what it learns. If you wrote that as an AI architecture spec, a venture capitalist would fund it.

Honest assessment

Fungi have been computing for 800 million years. Neurons for 500 million. Silicon for 80. We're the newcomers here.

Everything above is other people's work. This part is ours.

The Idea

So how do you actually talk to a mushroom? You need something you physically stick into the substrate that can listen to the network and talk back.

The probe has two jobs: record electrical activity across multiple channels, and deliver micro-doses of nutrients to specific locations as a reward signal. A brain-computer interface, except the brain is a fungus. And instead of dopamine, you use glucose.

The mycelium strengthens pathways that lead to food. So you reward the patterns you want, and the network literally grows toward your desired behavior. Operant conditioning for fungi. Biological gradient descent.

Recording — How to Listen

Reward Delivery — How to Teach

As far as I can tell, nobody is building this part yet. Plenty of labs record from fungi. Nobody is training them with food. That's what we're working on.

• Microfluidic channels — precision, independently addressable
• Slow-release nutrient capsules
• Chemical diffusion gradients
• Electroosmotic nutrient pumping — dual-purpose electrodes

Training loop: Record → Interpret → Reward → Adapt → Repeat.
Learning rate in hours, not milliseconds. But the network literally grows its own architecture.

Probe Design Concept

Software Stack

I've been selling you on mushroom computing for six sections. Time for some honesty about what could go wrong.

1. It could just... not work

The speed gap is real. Orders of magnitude slower than silicon. Biology is messy. Every colony grows differently. You can't git revert a mutation. Reproducibility is the unsexy reason most bio-computing stays in the lab. Papers show proof-of-concept. Nobody has shown proof-of-product. The gap between "a logic gate in mycelium" and "a useful computer" is roughly the same as between the Wright brothers and a 787 Dreamliner.

2. Zombie ants (yes, really)

Let me tell you about Ophiocordyceps unilateralis. It's a fungus that infects carpenter ants, and it's been doing this for at least 48 million years — we have fossilized death-grip marks from the Eocene to prove it.

Here's how it works. A spore lands on an ant. The fungus uses enzymes and mechanical pressure to pierce the exoskeleton. Once inside, it spreads through the body, positioning cells around the brain — not inside it. This isn't root access. It's a man-in-the-middle attack.

The infected ant abandons its colony. It climbs to exactly 25 cm height on a nearby plant. It bites down on a leaf vein with a "death grip" — hypercontracted jaw muscles, fungal cells wrapped around individual muscle fibers. The ant dies. A fungal stalk erupts from its head. Spores rain down on the colony below.

ant.climb(height=25cm)
ant.bite(substrate=leaf_vein, force=MAX)
ant.die()
self.fruit()

Is that a tool call?

3. The substrate might eat your lab

Mushrooms evolved to decompose things. Wood, leaves, dead animals, industrial waste — they're not picky. Saprotrophic fungi don't distinguish between "computing substrate" and "everything else that's made of carbon." What happens when your carefully cultivated fungal computer decides your lab bench is lunch? Or the insulation on your wiring? Or the structural timber in the building that houses it?

"Self-assembling infrastructure" sounds great until it starts self-assembling in directions you didn't plan.

4. Evolution doesn't have a rollback button

You can't pin a dependency version on biology. The chytrid fungus Batrachochytrium dendrobatidis (Bd) — a single species — is responsible for declines in over 500 amphibian species, with 90+ presumed extinct. It's found on every continent except Antarctica. A 2019 paper in Science called it "the most destructive pathogen ever recorded."

Then there's wheat rust (Puccinia graminis), which has caused famines for millennia and evolves resistance to every countermeasure we develop. A computing substrate that evolves is either the most exciting or most terrifying thing you can build. Possibly both simultaneously.

5. The control problem (sound familiar?)

We're already having the AI alignment conversation about silicon. Now imagine having it about a substrate that literally has its own metabolism, its own survival instincts, and 800 million years of evolutionary experience. Fungi survived five mass extinctions. They'll survive us. The question is whether they'll cooperate.