Melbourne, Australia-based startup Cortical Labs has successfully trained clusters of living human brain cells to navigate the 1993 classic video game, marking a striking demonstration of what is being called synthetic biological intelligence.
The company demonstrates "real neuron gameplay" with Doom running on its CL1 neural computing system, a microchip upon which 200,000 human neurons are mounted in something called a "multi-electrode array."Various elements of the on-screen data are being mapped to patterns of electric stimuli, which are then transmitted to the neurons.The neurons respond to these stimuli with signals of their own, which control the on-screen character's actions.
The achievement represents a significant acceleration fromwhen the company announced in 2022 that it had taught "mini-brains" consisting of 800,000 to one million living human brain cells how to play Pong.Doom was much more complex. Doom is chaos. It's 3D. It has enemies. It needs to explore its environment and it's hard.
The speed of development itself is noteworthy.The solution was completed in about a week by Sean Cole, an independent developer with little experience in biological computing.The key to this is the CL1's new interface, which allows anyone to program it using Python.It took Cortical Labs more than 18 months using its original hardware and software to accomplish their Pong goal.
Yet expectations should be tempered.The cells play a lot like a beginner who's never seen a computer, and in all fairness, they haven't.It plays the game better than a system that simply fires randomly at enemies, but it still loses a lot of the time.The cells show evidence that they can seek out enemies, they can shoot, they can spin.
The broader significance lies not in creating a gaming competitor, but in demonstrating thatbiological neurons show "adaptive real-time goal-directed learning" and figure things out on the fly without needing a million training iterations.Cortical Labs says it reached its current performance level faster than silicon-based machine learning systems.
Cortical Labs connected 200,000 human neurons to Doom using electrical stimulation and software controls.The human neurons used within these CL1 units are derived from skin or blood samples of voluntary adult donors before being reprogrammed into stem cells and finally differentiated into neurons.
The practical applications extend well beyond entertainment.Future demonstrations could eventually open more doors, like controlling complex robotic arms.In a recently accepted paper using an in vitro epilepsy model, the CL1 restored function in impaired neural cultures. Epileptic cells can't learn to play games very well, but if you apply antiepileptics to the cell culture, they can suddenly learn better.
Cortical Labs has begun commercialising the technology.The first 115 units will begin shipping at $35,000 each, or $20,000 when purchased in 30-unit server racks.The company has launched something it's calling the "Cortical Cloud," which promises to allow developers the world over to experiment with the CL1 via a Python-based API.
However, unresolved questions loom.Most of these neurons are derived from induced pluripotent stem cells. Scientists can take a simple skin cell from a donor and reprogram it to become a neuron. But that neuron still carries the donor's DNA. If a biological computer becomes a commercial success, who owns the IP of that person's genetic blueprint?Henrietta Lacks' cells were used for decades of research without her consent. We cannot afford to make that mistake again.
Despite the fanfare around a petri dish playing a 1990s shooter,the system does not resemble human cognition. Just because they're human cells doesn't mean it's a human on that dish. There are no pain receptors. There are no structures that could allow for higher-order functionality.
For all the hype surrounding biological computing, the technology remains in early stages. The neurons that have mastered Doom are not thinking, learning in any human sense, or possessed of awareness. They are responding to electrical signals. Yet the fact that living tissue can be trained to navigate a complex environment hints at possibilities that the field is only beginning to understand.