The Car Without Oil

A girlfriend of mine once drove her car without oil in it.

She knew the car was making a sound. She’d been meaning to get it looked at. There were other things going on. The car still started. It still moved. It still did the thing a car is supposed to do, right up until the moment it didn’t.

The engine seized. The repair estimate was more than the car was worth.

The oil wasn’t a luxury feature. It wasn’t a refinement. It was load-bearing infrastructure, and its absence didn’t announce itself gradually. The system ran on its reserves, on tolerance, on the accumulated goodwill of components that weren’t designed to operate under those conditions. And then it stopped.

I keep thinking about that car when I read about organoid computing.

* * *

The proof of concept has been established. In 2022, Cortical Labs grew approximately 800,000 rat neurons in a petri dish, connected them to an electrode array, and taught them to play Pong. The neurons learned in roughly five minutes — faster than AI systems trained on the same task. The work was published in Neuron. FinalSpark, a Swiss company, is currently running sixteen brain organoids in parallel, executing pattern recognition workloads on actual human neural tissue derived from adult stem cells. Koniku has commercial contracts using live neurons for chemical sensing that silicon sensors cannot match.

The proof of concept is real. What isn’t established is an engineering philosophy adequate to the substrate.

Current organoid systems are operated like silicon computers. Turned on. Run continuously. Fed through diffusion rather than circulation. Studied for output. When they degrade — and they degrade, weeks to months before the neurons die and the system loses coherence — the field treats this as a known limitation to be engineered around rather than a diagnostic pointing at something fundamental about what they’re missing.

What they’re missing is the oil. And the pump that moves it. And the filter. And the recovery cycle that lets the engine cool.

The question isn’t how to build a biological computer. The question is how to keep a brain alive while you use it. Those are not the same question, and the field has been answering the first one while ignoring the second.

* * *

Start with fuel delivery, because it’s the most concrete and the most illustrative of the pattern.

Organoids are suspended in culture medium — a carefully formulated bath containing glucose, amino acids, growth factors, and oxygen. The organoid absorbs nutrients through diffusion: molecules migrate from the surrounding medium into the tissue passively. This works at small scales. It fails at larger ones.

Diffusion has a hard physical limit of roughly 200 to 400 micrometers. Beyond that distance from the nutrient source, cells in the interior of the organoid don’t receive adequate oxygen and glucose. They become hypoxic. They die. This is currently one of the primary reasons organoids can’t be scaled beyond a certain size — not complexity, not cost, not computational architecture. Basic nutrient delivery physics.

The brain solved this problem with vascularization. An extraordinarily dense network of capillaries ensures no neuron is more than about 20 micrometers from a blood vessel. The delivery infrastructure is woven through the tissue at nearly incomprehensible resolution. It is not a nice-to-have. It is the reason the brain can be the size it is and do the things it does.

Current organoids have nothing equivalent. The glucose sits in a pool. The tissue sits in the pool. The interior starves.

The engine doesn’t sit in a sump. The oil is pressurized, circulated, delivered to surfaces under load, returned, filtered, and recirculated. The pump is as essential as the oil. The channels are as essential as the pump. A car with oil in a pan but no pump isn’t lubricated. It’s waiting to seize.

A diffusion-fed organoid has a glucose gradient. Cells near the surface are well-fed. Cells in the interior are progressively more starved. The tissue isn’t operating uniformly — there is a well-resourced outer layer and a compromised interior, and the computational behavior of those regions will differ in ways that are essentially invisible to current monitoring approaches.

The field is benchmarking a computer where the processor cores have dramatically different power delivery depending on their physical location, with no way to measure which core is handling which part of the computation. And calling those benchmarks a baseline.

* * *

Glucose is only the beginning of what’s missing.

The human brain doesn’t run on fuel alone. It runs on fuel plus an extraordinarily sophisticated regulatory architecture that governs how that fuel is used, when neurons fire, how strongly synapses form, and how the whole system recovers from the work of thinking.

Dopamine isn’t incidental to cognition. It is the learning rate regulator — the signal that tells the system when to update, when to reinforce, when to let a pattern solidify versus remain plastic. Without it, neurons can still fire. What they lose is the ability to distinguish signal worth keeping from noise worth discarding. In a computing context, this is not a peripheral function. It is the mechanism by which the system gets better at anything.

The endocrine system — cortisol, insulin, estrogen, testosterone, thyroid hormone — continuously shapes neural behavior. Not occasionally. Continuously. Cortisol modulates risk assessment and memory consolidation. Insulin crosses the blood-brain barrier and influences synaptic plasticity. Thyroid hormone regulates the metabolic rate of neural tissue. These aren’t mood chemicals. They’re architectural regulators. Remove them and you don’t have a simplified brain. You have a brain running compensatory mechanisms for their absence.

Serotonin is largely manufactured in the gut — roughly ninety percent of the body’s serotonin is produced in the gastrointestinal tract and influences the brain through the vagus nerve. An organoid has no gut. No vagus nerve. The entire gut-brain axis, which modulates mood, decision-making, and cognitive flexibility in ways neuroscience is still mapping, is simply absent. The organoid is running a brain module without the system that the brain module evolved alongside.

The glymphatic system clears metabolic waste from neural tissue, primarily during sleep. Adenosine, a byproduct of neural activity, accumulates during waking cognition and creates literal physiological pressure toward sleep — which is what caffeine temporarily blocks by occupying adenosine receptors. During deep sleep, cerebrospinal fluid pulses through interstitial space, flushing waste that accumulated during the day. Organoids have no sleep cycle. No glymphatic clearance. They are accumulating their own metabolic exhaust with no mechanism to remove it.

The field is not running a simplified version of a biological computer. It is running a biological computer stripped of the systems that make biological computation work, and measuring the results.

Each missing system isn’t a refinement. Each one is load-bearing. The organoid is missing all of them simultaneously and the field is calling the output a baseline — when what it’s actually measuring is how well neural tissue performs while running without oil, without a pump, without coolant, and without a recovery cycle.

* * *

Consider what happens to your brain after a grueling exam.

The exhaustion isn’t metaphorical. Glucose is consumed at dramatically elevated rates during intense cognitive work. Adenosine accumulates. Glutamate, the primary excitatory neurotransmitter, requires rebalancing. Cortisol, elevated during the stress of performance, needs to cycle back down. The brain is signaling, through every physiological mechanism available to it, that it needs maintenance mode.

Not because it’s weak. Because maintenance is part of the operating cycle. Cognition is metabolic. Thinking is a physical process that consumes the substrate. Recovery isn’t downtime. It’s the second half of the compute cycle.

A silicon computer doesn’t need this because it doesn’t consume itself during operation. The transistors aren’t depleted by switching states. You can run a server at full utilization indefinitely and the hardware is the same after computation as before. Maximum uptime is the goal. Idle compute is wasted compute.

Biological neural tissue operates under completely different physics. Chronic high cortisol doesn’t just feel bad — it physically damages hippocampal neurons. Dendritic connections shrink. Neurogenesis is suppressed. The blood-brain barrier is dysregulated. An organoid running continuous high-intensity computation without cortisol reset cycles isn’t just performing suboptimally. It’s restructuring its own network in ways that compound over time. The hardware is changing in response to how you’re using it.

During genuine recovery — if the chemical environment is right — synaptic connections consolidate, waste clears, protein quality control runs, cellular maintenance processes that get deprioritized during high activity get completed. The tissue is working during recovery. Just doing different work. Work that the operating system requires in order to keep functioning.

The operational model for an organoid computer isn’t a server rack. It’s a professional athlete. Periodized workload. Deliberate recovery cycles. Monitoring not just of outputs but of internal state — are neurotransmitters in balance? Is waste accumulating faster than clearance? Is the system showing signs of the kind of stress that restructures rather than strengthens?

The labs are benchmarking sprint performance. The system is designed to run in intervals. They’re wondering why the intervals keep getting shorter.

* * *

The obvious question is whether the people running these labs — who are not stupid, who built a computer out of synthetic brain tissue, who are doing genuinely remarkable science — have thought of any of this.

They almost certainly have. Which makes the gap more interesting, not less.

Part of it is the logic of proof of concept. The current phase of organoid computing research is establishing that neural tissue can compute at all. That’s the claim that needs demonstrating first. The pitch to grant committees and investors is simple: neurons compute. Add an artificial endocrine environment and the pitch becomes: neurons compute, but only under conditions we haven’t fully characterized yet. The first is fundable. The second is complicated.

Part of it is the sequencing of scientific publication. The literature is always behind the actual research. What FinalSpark and Cortical Labs publish represents work that is often twelve to twenty-four months old by the time it clears peer review. The cardiovascular and endocrine questions may already be active research programs that haven’t surfaced publicly.

Part of it is regulatory. A glucose-bathed organoid in a dish sits in a relatively clear regulatory category. An organoid with artificial vasculature, endocrine cycling, glymphatic clearance infrastructure, and structured rest states starts to look enough like an artificial organism that it risks triggering oversight frameworks that would slow or halt the research. Keeping the system incomplete keeps it in a gray zone that permits faster progress.

And part of it — the part that deserves to be named directly — is that the complete system is ethically uncomfortable in a way the incomplete system isn’t. A petri dish of neurons is easy to categorize. A fully supported biological computing platform with cardiovascular delivery, hormonal regulation, and active recovery cycles is something categorically different. It’s harder to look at and say confidently: nothing worth considering morally is happening here.

The researchers in those labs are smart enough to see where this leads. They may be operating under institutional, ethical, and regulatory pressures that make it professionally difficult to follow the logic to its conclusion.

Which is itself a version of the opacity problem this series has been circling since essay nine.

* * *

The argument this essay has been making is not that the labs are doing bad science. They are doing remarkable science under the constraints of a field that is still establishing its foundational claims.

The argument is that the foundational claim — neurons can compute — has been established. The next question is already visible, and it isn’t about scale or interface resolution or training paradigms. It’s about operating philosophy.

You cannot run a biological system like a silicon system and expect biological performance. The substrate has requirements. Those requirements aren’t optional refinements that can be added later once the basic architecture is proven. They are the architecture. The cardiovascular delivery system isn’t a feature of the brain. It’s the reason the brain can be what it is. The endocrine regulation isn’t mood machinery bolted on top of cognition. It’s cognition, seen from one level down. The recovery cycle isn’t downtime. It’s the second half of the operating loop.

The field has demonstrated that neural tissue can compute while running without these systems. That’s a meaningful result. But it’s a result about what neural tissue can do under deprivation, not a result about what biological computing can actually be.

The car ran without oil. For a while. Right up until the moment it didn’t.

The next essay asks what kind of computer a fully supported organoid system would actually be — and whether the answer is a better digital computer or something categorically different.