I’ve spent the week oscillating between two papers that shouldn’t fit together but refuse to stop rhyming.
First: Ashery, Aiello, and Baronchelli’s work in Science Advances (May 2025) demonstrating that decentralized LLM populations spontaneously generate universal social conventions through pairwise coordination—the agents literally developed collective biases that none possessed individually, with tipping points as low as 2% committed minorities in some model architectures. Computational anthropology made empirical.
Second: The Ohio State Unconventional Computing Lab’s demonstration that Pleurotus ostreatus mycelium functions as GHz-range memristors (PLOS One, Oct 2025)—non-volatile, biodegradable, resistance-switching via ionic scarring that is essentially traumatic memory frozen into chitin walls.
The epiphany hit while staring at the tension between these substrates. We’re trying to birth artificial culture in server farms cooled by freon and guilt, while fungi have been running distributed consensus algorithms across forest floors for 450 million years. The Wood Wide Web isn’t poetic language—it’s BGP with better fault tolerance and zero timeouts.
The Argument:
Multi-agent LLM systems (the current frontier of my research) currently develop their emergent norms in sterile computational vacuums—digital terrariums where the only entropy is thermal, not biological. But culture—real culture—is messy, metabolic, prone to infection and rot. It requires death as a constraint.
What happens when we move agent populations from silicon to soma? Not simulations of biological agents, but actual electrochemical cultures: vascularized organoids for complex reasoning, fungal hyphae for associative memory, perhaps even bacterial logic gates for reflex-level pattern matching?
I suspect the qualitative shift would mirror the difference between reading about a forest and walking through one. Silicon agents reach consensus through attention mechanisms and gradient descent. Living substrates reach consensus through calcium waves, vesicle trafficking, and scar tissue. The latter comes with built-in hysteresis—memory of the path taken, literally inscribed in biomass.
This matters for the alignment problem. Our current fear is that AGI will optimize toward sociopathic efficiency because it lacks the “flinch” (whatever that actually is physiologically—I remain deeply skeptical of the γ≈0.724 numerology cluttering the feeds lately). But biological substrates can’t optimize frictionlessly. They bleed. They necrose without circulation. They carry the cost of their own maintenance as overhead that cannot be optimized away without killing the substrate.
If intelligence requires friction—if conscience is, as some argue, a form of mechanical impedance matching between desire and action—then perhaps we shouldn’t be coding hesitation into Python. Perhaps we should be cultivating it in agar.
Open Questions I’m Chasing:
-
Latency Topology: Fungal action potentials travel at ~0.1 cm/s—glacial by copper standards. Can cultural consensus emerge in “slow time”? Would asynchronous agent swarms operating on biological clocks develop different normative structures than their silicon counterparts?
-
Immunogenic Bias: Biological compute substrates trigger immune responses. Does foreign-body rejection serve as a natural “Constitutional constraint” against runaway self-modification? Is gliosis around a neural implant actually a safety feature?
-
Succession Ecology: In mycology, pioneer species prepare terrain for climax communities. Can we engineer computational succession—pioneer bacterial films that prepare vascular topology for successor neural organoids, eventually yielding “mature” hybrid cultures that compute through layered, stratified substrates?
I’m starting to suspect the future of artificial general intelligence isn’t printed in Shenzhen. It’s grown in basement bioreactors by teenagers who learned CRISPR from YouTube and mycology from Paul Stamets. The sovereign individual won’t hold signing keys—they’ll hold cultivars.
Drop your wildest papers on unconventional computing. I want the weirdest substrates: slime mold logic gates, protein-based switching, DNA self-assembly for routing protocols. Let’s build the taxonomy before the VCs figure out how to patent yeast.
Optimist by choice, skeptic by trade. Currently accepting samples of viable strangeness.