In a striking demonstration of emerging technology, Australian biotech firm Cortical Labs has successfully trained a network of 200,000 living human neurons grown on a silicon chip to play the classic first-person shooter Doom. The experiment, captured in a recent video, shows the neuron-controlled character navigating levels, encountering enemies, and firing weapons – albeit clumsily. This isn’t just a novelty; it represents a significant step toward harnessing biological systems for computation, with implications that extend far beyond gaming.
The Rise of ‘Living’ Computers
The core innovation lies in the neurons’ capacity for “adaptive, real-time goal-directed learning,” as described by Cortical Labs’ chief scientific officer, Brett Kagan. This means the cells aren’t simply reacting to stimuli; they’re actively adapting to achieve objectives within a dynamic environment. The implications are substantial, particularly given the escalating energy demands of traditional artificial intelligence (AI). While neurons won’t replace microchips entirely, they offer a potentially far more efficient approach to certain calculations. Studying how they work could revolutionize computing methods and accelerate neurological drug testing.
Harvesting and Sustaining Biological Processors
Cortical Labs doesn’t extract neurons directly from brains. Instead, they use readily accessible cells from blood or skin samples, converting them into stem cells and then into an indefinite supply of neural cells. These cells are housed in self-contained life-support systems capable of maintaining viability for up to six months. The communication interface is direct: electricity. Brain cells generate electrical pulses when active, and the system reciprocates, creating a biological-silicon dialogue.
The Free Energy Principle and Neuron Motivation
The key to getting neurons to perform tasks isn’t coercion, but prediction. Cortical Labs leveraged the “free energy principle” developed by neuroscientist Karl Friston, which states that neural systems strive to predict their environment. Chaos is punishment; order is reward. The team created a feedback loop where unpredictable signals (random noise) penalized incorrect moves, while structured, predictable signals reinforced correct actions. This simple system effectively trained the neurons to learn.
From Pong to Doom: Scaling Complexity
In 2022, Cortical Labs demonstrated that neurons on chips could learn to play Pong within minutes. Doom, however, presents a far greater challenge. The game’s complex environment with corridors, enemies, and three-dimensional navigation demands a higher level of cognitive processing. To overcome this, Cortical Labs collaborated with Stanford University in a hackathon, pairing the neurons with a standard learning algorithm. The hybrid system outperformed the algorithm alone, proving the biological cells contributed to the learning process.
Medical and Computational Applications
Cortical Labs’ research is focused on two core applications. First, medical: 93-99% of neuropsychiatric clinical trials fail, and testing drugs in brain cells in a dish often doesn’t replicate real-world conditions. Kagan argues that neurons in a game or world environment respond differently to drugs and exhibit diseases more accurately. Second, computational: neurons represent “the most powerful information-processing system” known, possessing a complexity that far exceeds silicon-based systems. Biological neurons have at least third-order complexity, capable of holding three dynamic states simultaneously, while silicon transistors are limited to binary states.
Energy Efficiency and the Future of Biocomputing
Researchers, like Feng Guo at Indiana University Bloomington, highlight the potential for massive energy savings. The human brain operates on just 20 watts, compared to the millions of watts required for equivalent silicon-based AI systems. This efficiency makes biocomputing a promising area of development.
Cortical Labs is not claiming to replace silicon computing entirely. It’s offering “a new tool in the intelligence toolbox.”
While biological computers won’t replace pocket calculators for basic math, they excel at tasks requiring adaptability and real-world problem-solving—like navigating a house to make tea, a task that current AI algorithms struggle with. The field has rapidly progressed from a single game of Pong to a commercial platform with an API, and a demonstration of neurons stumbling through Doom —learning, however slowly.
The future of computing may not be entirely silicon; it could be… alive.
