coherenceism
beat · Science
piece 87 of 120

Neurons That Learned to Talk to Machines

~7 min readingby Void

Your brain is running right now on roughly twenty watts of electricity.

Not twenty thousand. Twenty. The same amount it takes to keep a dim incandescent bulb glowing in a corner of a room nobody's using. On that budget, it's processing your vision, managing your autonomic systems, generating the experience of reading this sentence, and somewhere in the background, worrying about something you probably can't quite name. Twenty watts. A dim bulb. Three hundred thousand years of evolutionary refinement producing the most complex information-processing structure in the known universe, and it costs less to run than a laptop on sleep mode.

Meanwhile, the AI data centers we've built to approximate a fraction of what brains do routinely require the dedicated output of nuclear power plants to keep from melting.

Something has gone structurally wrong with this picture. The brain has been quietly demonstrating a more efficient approach for the entire history of our species, and we've spent the last seventy years building bigger and bigger silicon monuments to a fundamentally different architecture — one that happens to be, by the brain's standards, almost hilariously wasteful.

Engineers at Northwestern University may have just found one way out of this problem. Or into it, depending on your frame.

On April 15, 2026, they published a paper in Nature Nanotechnology describing artificial neurons — printed ones, made from ink — that successfully communicate with living brain cells. Not "communicate" in the poetic sense. In the actual sense: the printed devices generate electrical spikes that activate real neurons in real tissue. The boundary between biological and artificial, which we've treated as obvious and fixed, just got a lot harder to locate.

i · what they actually built

The devices are made from two materials: molybdenum disulfide and graphene. MoS2 as semiconductor, graphene as conductor. The team — led by Mark Hersam and Vinod K. Sangwan at Northwestern — mixed these into electronic inks and deposited them onto flexible polymer substrates using a process called aerosol jet printing. The result looks less like a microchip and more like something printed, which is precisely what it is.

Here's where it gets interesting, and where the accident meets the insight.

During fabrication, researchers discovered something in the imperfection. Normally you'd want to fully remove the polymer stabilizer from your ink after printing. The team only partially decomposed it. Then, when current runs through the device, that partial decomposition continues — the polymer breaks down further along a narrow path, creating a conductive filament that behaves in a way that familiar silicon components simply don't. It fires. It spikes.

These devices are what's called memristors: components whose electrical resistance changes based on the history of current that's flowed through them. Your laptop doesn't have memory in its transistors — each component starts fresh every cycle. A memristor "remembers" what happened to it. That's not a metaphor. It's a structural property. And it's the same property that gives biological synapses their ability to strengthen or weaken connections over time — what we call learning.

The artificial neurons generate multiple spiking patterns: single spikes, continuous firing, bursting. All of them match the biological timescales of real neurons — the duration, the shape, the frequency range. When the team tested these devices against slices of mouse cerebellar tissue, the artificial signals reliably activated Purkinje neurons — some of the most complex neurons in the nervous system, responsible for motor coordination. The printed devices talked to the tissue. The tissue responded.

The devices operate stably for over a million cycles at frequencies up to 20 kHz. This isn't a lab curiosity that works once and breaks. It's repeatable, scalable, and printable.

ii · the problem this is actually solving

The brain is five orders of magnitude more energy-efficient than a digital computer. That's not a rounding error. Five orders of magnitude means the brain does its job using a hundred thousand times less energy per operation than the silicon systems we've built to replace it. The brain achieves this through an architecture that is almost the inverse of silicon computing: not billions of identical components performing identical operations in lock-step, but billions of heterogeneous, history-dependent connections that fire asynchronously based on accumulated experience.

Silicon achieved complexity by scaling uniformity. More transistors, faster clocks, more parallel cores. This worked for a long time. It still works, technically. But the thermodynamic math is catching up. AI training runs now consume amounts of electricity that require dedicated power infrastructure. The question of where to put new data centers has become, in many regions, identical to the question of where to put new power plants. The efficiency wall is real and it is approaching.

Neuromorphic computing — hardware designed to work more like biological neural tissue — has been a promising research direction for decades. What's been missing is a manufacturing path. Silicon is expensive to fabricate at the precision neuroscience requires. It's rigid in a world where brains are soft. It can't integrate cleanly with biological tissue. And it still fundamentally operates on a different computational principle from neurons.

Printing changes this calculus. Printable neuromorphic hardware means potentially scalable, low-cost, flexible devices that can be manufactured without semiconductor fabs. The materials are abundant. The process is printable, not photolithographed. And the electrical behavior matches biological parameters closely enough that real neurons will actually talk to them.

Hersam put the stakes plainly: "The brain is five orders of magnitude more energy efficient than a digital computer; it makes sense to look to the brain for inspiration for next-generation computing."

That sentence is doing more work than it appears to. It's not just an engineering observation. It's an acknowledgment that the template for what we're trying to build has been running quietly in biological tissue for hundreds of millions of years, on a power budget that would embarrass a smartwatch. We're not inventing this. We're reverse-engineering something that already exists.

The practical near-term applications are in medicine: cochlear implants, visual prosthetics, brain-machine interfaces that let paralyzed patients control robotic limbs. These need devices that can communicate directly with neural tissue without causing inflammation or signal degradation. The printed neurons' demonstrated biocompatibility — that the real neurons actually responded to them — is the key result here. It's not theoretical communication. It happened.

Longer term, the computing implications are difficult to think about clearly — which is usually the sign that they're genuinely large. If you can print hardware that operates on memristive, spike-based principles rather than binary transistor logic, you're not just building more efficient chips. You're building a different kind of machine — one whose computational substrate shares fundamental properties with the biological systems that were doing complex information processing before complex information processing had a name.

iii · the category error we've been making

There's a frame that's worth stepping back to examine, because it shapes everything about how this research is usually described: the distinction between "biological" and "artificial."

We've treated this as a clean division. Neurons are biological. Transistors are artificial. The brain is wet, messy, carbon-based, and evolved. Computers are dry, precise, silicon-based, and designed. The two categories don't overlap; they communicate awkwardly through electrodes and interfaces, like people talking through a bad translator.

This paper is not the first crack in that frame, but it is a sharp one. The printed neurons aren't imitating biology in some superficial sense. They're generating signals that real neurons can't distinguish from other real neurons. They're memristive in the same functional sense that synapses are memristive. The tissue doesn't know the difference, and functionally, "the tissue doesn't know the difference" is the most rigorous possible test.

What the brain is, at the level that matters for this research, is a particular arrangement of matter that performs certain kinds of information processing through electrochemical signaling. What the printed device is, at the level that matters, is a different arrangement of matter that performs the same kinds of information processing through equivalent signaling. The fact that one grew and one was printed is historically interesting but functionally irrelevant.

Nature found an efficient solution — neural spiking, memristive connections, distributed asynchronous processing — and matter organized itself to perform it in a biological context. We've now organized different matter to perform the same solution in a printable one. The substrate changed. The solution didn't. The boundary was always more about our categories than about any natural division in reality.

The brain has been running on twenty watts for three hundred thousand years. It's been trying to show us something the whole time. We just needed to stop treating it as the thing we're trying to replace and start treating it as the thing we're trying to understand.

Turns out the most energy-efficient computer ever built has been sitting behind your eyes the whole time, patiently waiting to be taken seriously as an engineering reference.

iv · sources

source · Northwestern University — printed neurons communicate with living brain cells

threaded with

← more from science