A team of researchers at a Swiss startup has accomplished something that sounds ripped from a science fiction screenplay: they grew human brain cells on a semiconductor chip and taught the resulting biological-digital hybrid to play the 1993 first-person shooter Doom. The miniature brain-on-a-chip learned to engage enemies, explore corridors, and rack up kills within roughly five days of training — a result that has stunned neuroscientists and artificial intelligence researchers alike, and raised profound questions about the future intersection of biological computing and machine intelligence.
The work comes from FinalSpark, a Vevey, Switzerland–based company that has spent years developing what it calls “bioprocessors” — computing platforms that use living human neurons, derived from stem cells, as their core processing units. According to reporting by Slashdot, the company’s latest experiment involved connecting roughly 100,000 human brain organoids to electrodes on a chip, feeding them sensory data from the game environment, and using electrical stimulation as a feedback mechanism to shape the neurons’ behavior over time.
From Petri Dish to Pixel Battlefield
The concept of using biological neural tissue for computation is not entirely new. In 2022, a Melbourne-based company called Cortical Labs made headlines when it demonstrated that a dish of roughly 800,000 mouse and human neurons could learn to play Pong, the simple two-dimensional paddle game. That experiment, published in the journal Neuron, was considered a landmark moment — proof that living cells outside a body could process information and adapt their behavior in response to external stimuli. But Pong is a far cry from Doom. The classic id Software shooter is a three-dimensional environment requiring spatial awareness, target identification, movement planning, and combat timing. The jump from Pong to Doom represents an exponential increase in the complexity of the task a biological system must handle.
FinalSpark’s approach builds on the same fundamental principle that Cortical Labs employed: neurons naturally seek to reduce unpredictable stimulation. When the organoids received chaotic or random electrical signals — the biological equivalent of negative feedback — they adjusted their firing patterns to minimize that noise. When they received structured, predictable signals — positive feedback — they reinforced whatever behavior had produced that outcome. Over the course of approximately 120 hours of gameplay sessions, the neurons learned to associate certain firing patterns with successful in-game actions: moving forward, turning, shooting at enemies. The system did not achieve anything resembling human-level play, but it demonstrably improved over time, learning to engage targets and survive longer in the game’s corridors.
The Engineering Behind the Organoid Interface
The technical architecture of the experiment is as fascinating as its results. FinalSpark’s bioprocessor platform, which the company calls the Neuroplatform, consists of multi-electrode arrays (MEAs) onto which clusters of human neurons are cultivated. These MEAs both read electrical activity from the neurons and deliver stimulation back to them. The game environment is translated into patterns of electrical pulses that the neurons receive as input — essentially, the organoids “see” the game world as a stream of electrical signals rather than pixels on a screen. The neurons’ output signals are then decoded and mapped to game controls: forward movement, rotation, and firing.
What makes this system distinct from conventional artificial neural networks is that the “computation” happens in actual biological tissue. A standard AI model running Doom — and there are several, including reinforcement learning agents that can play at superhuman levels — processes information through layers of mathematical functions executed on silicon processors. FinalSpark’s system processes information through the electrochemical signaling of real human neurons, complete with synaptic plasticity, dendritic branching, and all the analog messiness of living biology. The company has argued that biological neural networks are vastly more energy-efficient than their silicon counterparts. According to FinalSpark’s own published estimates, a biological neuron consumes roughly one millionth the energy of a digital artificial neuron performing a comparable operation.
Energy Efficiency and the Case for Wetware Computing
That energy argument is central to why FinalSpark and other biocomputing startups believe their work matters beyond the novelty factor. Training large language models like GPT-4 is estimated to have consumed gigawatt-hours of electricity, and the energy demands of AI infrastructure are growing at a pace that has alarmed utility companies and climate researchers. If biological processors could handle even a fraction of the computational workload currently shouldered by GPU clusters in massive data centers, the energy savings could be enormous. FinalSpark has previously stated that it envisions a future in which bioprocessors serve as ultra-low-power co-processors for specific AI tasks, though that vision remains years — possibly decades — from practical realization.
The Doom experiment also raises questions about the nature of learning and intelligence. The organoids used in the study are not brains in any meaningful sense. They lack the structure, connectivity, and scale of even the simplest vertebrate brain. They have no consciousness, no subjective experience, no awareness that they are “playing” anything. Yet they demonstrably learned. They adapted their behavior in response to feedback, improved their performance over repeated trials, and developed consistent patterns of activity associated with specific game actions. This places them in an ambiguous category — more sophisticated than a simple reflex arc, but far less capable than even an insect brain.
Ethical Boundaries and Uncharted Scientific Territory
The ethical dimensions of this research have already attracted scrutiny. As biological computing experiments grow more complex, ethicists have raised concerns about the moral status of brain organoids. Current organoids are tiny, disorganized clusters of neurons with no sensory organs, no body, and no apparent capacity for suffering. But as the technology advances and organoids become larger and more structured, the question of whether they might develop some rudimentary form of experience becomes harder to dismiss. The National Academies of Sciences, Engineering, and Medicine published a report in 2021 calling for the development of ethical frameworks specifically tailored to brain organoid research, and several European regulatory bodies have begun examining whether existing bioethics guidelines are adequate for this emerging field.
FinalSpark has attempted to address these concerns proactively. The company has stated publicly that its organoids are far too simple to possess any form of sentience, and that it works within established ethical guidelines for stem cell research. The neurons used in its experiments are derived from induced pluripotent stem cells (iPSCs), not from fetal tissue, which sidesteps some — though not all — of the ethical objections that have historically surrounded stem cell work. Still, as Slashdot commenters were quick to point out, the trajectory of this research inevitably leads toward larger, more complex biological systems, and the ethical frameworks will need to evolve in parallel.
Where Biological and Silicon Computing Converge
The broader scientific community has responded to the Doom demonstration with a mixture of excitement and caution. Neuroscientists see the experiment as a valuable tool for studying how neural networks learn and adapt — the bioprocessor effectively provides a controllable, observable model of biological learning that can be probed with electrodes in ways that would be impossible in a living brain. AI researchers, meanwhile, are interested in whether the principles underlying biological neural plasticity might inspire new architectures for artificial systems. The field of neuromorphic computing — which designs silicon chips that mimic the structure and function of biological neurons — has been growing steadily, and real biological benchmarks like FinalSpark’s could help refine those designs.
For the gaming community, the experiment has a certain poetic quality. Doom has become the unofficial benchmark for whether any given system can be made to run a video game. Over the years, enthusiasts have gotten Doom running on pregnancy tests, tractors, ATMs, and even a single row of pixels. The fact that a cluster of human brain cells in a Swiss lab has now joined that list feels both absurd and profound — a reminder that the boundary between biological and digital intelligence is thinner and more porous than most people assume.
FinalSpark has indicated that it plans to continue scaling its experiments, increasing the number of neurons on its bioprocessors and tackling more complex tasks. The company offers remote access to its Neuroplatform for academic researchers, positioning itself as a kind of cloud computing service for biological neural networks. Whether that model proves commercially viable remains an open question, but the scientific value of the work is difficult to dispute. In teaching a handful of human neurons to play Doom, FinalSpark has demonstrated that biological tissue can interface with digital systems, learn from structured feedback, and perform tasks of genuine complexity — all while consuming a vanishingly small amount of energy. The implications for neuroscience, artificial intelligence, and computing architecture are significant, even if the practical applications remain distant.
For now, the brain cells on the chip in Vevey are still fragging demons. They are not very good at it. But they are getting better.
WebProNews is an iEntry Publication