200,000 Human Neurons Learned to Play Doom, Pointing to Low-Power Biological Computing

Cortical Labs connected 200,000 human neurons grown from blood-derived stem cells to its CL1 silicon interface, teaching the culture to navigate and shoot in Doom. By translating the game’s state into electrical patterns and reading neural spikes as commands, the system improved with training, pointing to low-power biological computing even as current neuron-chip setups last roughly six months.

Inside a Melbourne lab, a dish of 200,000 human neurons learned to strafe and shoot in Doom, coached through a silicon interface. Cortical Labs’ CL1 chip translated the game’s world into electrical patterns and read back spikes as movement and fire, pushing the culture dish from Pong reflexes to 3D navigation. The play is still clumsy, yet it hints at biological computing that sips power next to today’s electricity-hungry AI, a direction the team says complements conventional models. Stretch the six-month lifespan and tighten consistency, and the same wetware could steer robots or screen drugs, not just chase pixelated demons.

Some experiments feel like a peek at the next chapter of computing. Researchers at Cortical Labs report that they trained a cluster of 200,000 neurons to play Doom, the 1993 first-person shooter that helped define modern gaming. The neurons, grown from human stem cells and connected to a silicon interface, learned to navigate corridors and fire at enemies, hinting at a path for bio-computers that complement today’s AI systems.

The team began with Pong-level behavior, then escalated to Doom’s 3D demands. The neurons received structured electrical cues tied to the game state and responded with patterns that the system translated into commands like move, turn, and shoot. At the core is the custom CL1 chip, which converts visual events into stimulation across electrodes, then reads the cells’ activity to drive actions in real time.

Performance is far from esports-ready. The cells often misfire or overcorrect, then improve over repeated sessions as training continues. According to the researchers, the goal is not perfect aim but demonstrating goal-directed learning inside a living neural network, under conditions a computer can orchestrate and measure.

Energy is the headline. Where today’s large AI models draw megawatts across cloud data centers, the human brain runs at roughly 20 watts. That efficiency inspires the search for hybrid systems that could cut power needs for learning, adaptation, and control. Brett Kagan, chief scientific officer at Cortical Labs, frames the work as a partner to silicon AI, not a replacement, especially for tasks that benefit from continual learning with tight energy budgets.

For US companies training foundation models on Nvidia GPUs and racing to scale inference, even partial offload to biological co-processors could matter. Think of local learning loops for robotics or edge devices, while conventional chips handle precision math and large-scale retrieval. The near-term question is where the trade-offs line up in latency, reliability, and cost.

Gaming is a handy testbed, yet the larger target is science and industry. Biological computing could enable drug screening on patient-specific neural tissue, new disease models, and adaptive controls in robotics. Interfaces remain fragile, with a typical lifespan around six months and outputs that are not yet fully standardized or programmable at scale.

Regulatory and ethical guardrails will need to keep pace, particularly in the US under FDA and NIH guidance if medical uses progress. Still, the lab result is concrete: living neurons can be trained to act on complex digital tasks. From Doom to data centers, the journey has begun, quietly and efficiently, inside a dish.