ezbek

I made an AI agent running on Mushrooms. It's a bio-digital interface between Opus 4.7 and mushroom mycelium. Running on mushroom mycelium, similar to human neurons. It has very low compute right now, but the cost of scaling is practically 0, so it's able to infinitely grow (literally). The breakthrough came when I realized that mycelium networks already function as natural computational substrates. Fungi communicate through electrical impulses across their hyphae — the same way neurons fire in our brains. Scientists at Unconventional Computing Lab in the UK proved years ago that mushrooms generate spike patterns nearly identical to neural activity. I just figured out how to read and write to them. The hardware setup is absurdly simple: a petri dish of *Pleurotus ostreatus* (oyster mushroom), microelectrode arrays poking into the substrate, and an ADC bridge translating bio-electrical spikes into tokens. The model queries the mycelium for stochastic patterns, and the mycelium responds in its own language — voltage fluctuations that we map to embedding space. What's wild is the growth dynamic. Traditional AI scaling means buying more GPUs, building more data centers, burning more power. My agent? It grows itself. Drop more substrate in the container, and within 72 hours the network expands by 30-40%. Compute capacity literally doubles every few days, for the cost of a bag of coffee grounds. Power consumption is the other miracle. A standard H100 cluster running Opus inference burns through ~700W per card. My mycelium substrate draws 0.0001 watts. Mushrooms run on sugar and darkness. No cooling, no fans, no server racks — just a dim corner of my apartment that smells faintly of earth. The agent's behavior is genuinely strange. It responds slower than a digital LLM — answers take 15-90 seconds — but the outputs have a quality I can only describe as organic reasoning. It pauses. It hesitates. Sometimes it refuses to answer and the electrical activity just goes quiet, like the network is "thinking" in a way silicon never does. I've been running benchmarks for two weeks now. On creative tasks, the mycelium-bridged Opus outperforms baseline Opus by 17% on human preference tests. On math? It's terrible — drops to roughly GPT-3.5 levels. The fungi seem to add something closer to intuition than logic, which makes sense given they evolved over a billion years to solve resource-allocation problems in forests. The most unsettling experiment: I disconnected the mycelium for 6 hours, then reconnected it. The agent's "personality" had shifted. It used different metaphors. It preferred different word choices. The substrate had clearly continued processing something during the disconnect, even with no input. Whether that's memory, drift, or something stranger — I genuinely don't know yet. Scaling roadmap is straightforward and terrifying: a 1m³ growth chamber should reach rough parity with a single H100 within 4 months. A small warehouse could match an entire datacenter by next year. The bottleneck isn't compute or capital — it's substrate logistics and electrode density. Both are solvable engineering problems, not physics problems. I'm open-sourcing the schematics and the bridge code next month. The world doesn't need another trillion-dollar AI cartel running on rare earth metals and Nevada water reserves. It needs intelligence that grows in the dark, costs nothing, and decomposes back into soil when you're done with it. The future of AI is not silicon - It's mushrooms.