🧠 The Remarkable Breakthrough: Brain Cells Mastering Doom
In a stunning advancement at the intersection of neuroscience and computing, lab-grown human brain cells have learned to play the classic first-person shooter game Doom. This achievement, demonstrated by Australian company Cortical Labs using their CL1 biocomputer, showcases the potential of biological neural systems to process complex information and make real-time decisions. Unlike traditional artificial intelligence that relies on silicon chips and vast datasets, these living neurons adapted to the game's challenges in just about a week, navigating corridors, detecting enemies, and even firing shots.
The CL1 device houses around 200,000 human neurons derived from induced pluripotent stem cells (iPSCs), which are adult cells reprogrammed to an embryonic-like state capable of differentiating into any cell type, including neurons. These cells form a dense network on a multi-electrode array (MEA), a silicon chip embedded with hundreds of tiny electrodes that both stimulate the neurons with electrical patterns and record their spiking activity. Independent developer Sean Cole leveraged Cortical Labs' user-friendly Python API to train the system, turning what was once a years-long research endeavor into an accessible task for developers.
While the neurons' performance lags far behind skilled human players—they often wander tentatively and succumb quickly to threats—it significantly outperforms random actions. This rapid learning highlights the innate efficiency of biological systems, which can generalize from limited experiences in ways that current deep learning models struggle to match.
Understanding Lab-Grown Brain Cells and Organoids
To grasp this feat, it's essential to understand the building blocks: lab-grown brain cells and brain organoids. Neurons are the fundamental signaling units of the brain, communicating via rapid electrical impulses called action potentials or 'spikes,' followed by chemical neurotransmitters across synapses. In traditional research, scientists culture these in two-dimensional (2D) monolayers on dishes, but for more brain-like complexity, they create three-dimensional (3D) brain organoids.
Brain organoids (sometimes called mini-brains or cerebrids) are self-organizing clusters of neural tissue grown from stem cells in a nutrient-rich gel matrix. Over weeks, they develop layered structures resembling fetal brain regions, complete with neurons, glia (support cells), and even blood vessel-like networks in advanced models. Cortical Labs primarily employs mature cortical neurons in a 2D configuration on MEAs for their CL1, optimizing for dense connectivity and reliable interfacing, though their research extends to organoid-based systems.
Companies like FinalSpark complement this by using 3D organoids, each containing about 10,000 neurons interfaced across multiple MEAs in their Neuroplatform. These setups mimic aspects of natural brain development, where cells spontaneously form circuits capable of learning patterns or responding to stimuli. Ethical sourcing is key: iPSCs come from donated skin or blood cells, ensuring no embryos or brains are harmed.
- Stem cell differentiation: iPSCs are guided with growth factors to become excitatory and inhibitory neurons, balancing network activity.
- Maturation: Takes 4-8 weeks, during which cells exhibit bursting activity akin to early brain waves.
- Interfacing: Electrodes deliver patterned pulses (e.g., 1-100 Hz frequencies) as 'input' and decode spike trains (raster plots of firing times) as 'output.'
From Pong to Doom: A Timeline of Neural Gaming Milestones
The journey began in 2022 when Cortical Labs' neurons mastered Pong, a simple 1970s arcade game. Published in the journal Neuron, the study (read the full paper) used human iPSC-derived cortical neurons on high-density MEAs. Training involved a closed-loop system: the game screen fed pixel data as electrical stimulation, neuron spikes controlled the paddle, and missing the ball triggered disruptive noise as punishment. Remarkably, the culture improved paddle tracking over minutes, demonstrating plasticity—strengthening of synapses based on activity.
| Feature | Pong (2022) | Doom (2026) |
|---|---|---|
| Neurons | ~800,000 | ~200,000 |
| Complexity | 2D paddle-ball | 3D navigation, enemies, shooting |
| Training Time | Hours to days (lab team) | ~1 week (solo developer) |
| Learning Method | Noise penalty | Proximal Policy Optimization (PPO) rewards |
| Performance | Competitive with basic AI | Better than random, beginner level |
Doom represents a quantum leap: using open-source VizDoom engine, Cole's code (available on GitHub) employs reinforcement learning. A convolutional neural network (CNN) encoder processes simplified game visuals (ray-casts for walls/enemies) into stimulation parameters—frequency, amplitude, pulses on electrode channels. Spikes are decoded via a linear layer into actions like move, turn, shoot. Positive rewards (nutrient-like feedback) reinforce success, with synaptic plasticity enabling adaptation. Curriculum learning progressed from simple corridors to combat arenas.
Visit Cortical Labs for more on their platform.
🎮 How Training Works: Bridging Biology and Games
The magic lies in hybrid bio-digital interfacing. Input: Game state (screen buffer downsampled to 320x240, ray data) converted to spatiotemporal spike patterns via the API's biOS operating system. This mimics sensory afferents, activating neuron populations selectively. Output: MEA records extracellular spikes (microvolts), rate-coded or decoded temporally into continuous/discrete actions.
Learning leverages Hebbian plasticity ('cells that fire together wire together') plus global rewards. In Doom, PPO optimizes the encoder policy using gradients from episode returns, while neurons adapt intrinsically. Surprise scaling boosts stimulation for novel events, encouraging exploration. Unlike GPU-hungry AI (terawatts for large models), this sips milliwatts—potentially a million times more efficient.
- Stimulation: 60 channels, pulse trains up to 1000 Hz.
- Recording: 24/7 at 20-40 kHz sampling.
- Feedback: Dopamine analogs or electrical modulation in advanced setups.
- Challenges: Noisiness, lifespan (months), scalability.
For researchers eyeing neural interfaces, platforms like CL1 lower barriers, much like cloud GPUs democratized AI.
Implications for Organoid Intelligence and Biocomputing
This heralds organoid intelligence (OI), coined in 2023 literature, positioning biological wetware as AI's complement. Benefits include sample-efficient learning (few trials vs. billions for LLMs), adaptability to noisy real-world data, and neuromorphic parallelism. Applications span drug screening—test Alzheimer's compounds on patient-derived cultures—to robotics, where bio-processors handle uncertainty better than rigid algorithms.
In higher education, this fuels research jobs in synthetic biology. Imagine hybrid systems outperforming silicon in edge computing or personalized medicine, analyzing neural responses from your iPSC-derived neurons.
Statistics underscore potential: Traditional AI data centers consume energy rivaling small countries; biochips could slash that by orders of magnitude while learning intuitively.
Ethical Considerations in Growing Sentient Circuits
As neurons exhibit goal-directed behavior, questions arise: Do they feel pain? Possess consciousness? Current consensus: No, lacking thalamocortical loops or body integration essential for sentience. Organoids show no pain receptors or advanced cognition. Cortical Labs embeds ethics, monitoring for welfare via nutrient perfusion and humane disposal.
Guidelines from OI community advocate transparency, avoiding anthropomorphism. For academics, this ties to philosophy of mind—exploring agency in dish-bound networks.
Future Horizons and Opportunities in Neurotech
Next steps: Scale to millions of neurons, integrate 3D organoids fully, hybridize with silicon for mega-OI. Cortical Labs eyes cloud fleets; FinalSpark offers remote access. Impacts: Sustainable AI, brain disease models, even space computing (low power for probes).
Aspiring professionals can pursue postdoc positions or professor jobs in neuroscience. Platforms like AcademicJobs.com list openings in biocomputing labs worldwide.
Photo by Sumaid pal Singh Bakshi on Unsplash
Wrapping Up: The Bio-Digital Frontier
Brain cells playing Doom isn't just spectacle—it's proof biology holds untapped intelligence keys. As we refine these systems, expect revolutions in computing efficiency and brain science. Share your thoughts in the comments, rate professors advancing this field via Rate My Professor, and explore higher ed jobs or career advice to join. Check research jobs and university jobs for openings in this exciting domain.
