Researchers have demonstrated that biological brain cells grown in the lab can learn to perform computing tasks, blurring the line between biology and artificial intelligence.
A dish containing about 800,000 living brain cells has learned how to play a simplified version of the classic video game Pong. The system, dubbed DishBrain by developers at Cortical Labs, is not a simulation. It is a biological neural network, a cluster of human and mouse neurons grown on microelectrode arrays, wired to a computer, and trained using feedback signals. When the cells moved the paddle correctly, they received a predictable electrical impulse. When they removed, the signal was confused. Within minutes, the cells adapted their behavior to stabilize the stimulus.
This is the clearest demonstration to date that biological neurons in vitro can exhibit goal-directed learning. Basic research published in the journal Neuron shows that these cells process sensory input, adapt to feedback, and modify their activity to achieve defined goals. So, functionally, it’s the same basic loop that underpins machine learning.
The AI industry faces physical challenges. Training large language models and other advanced systems requires huge data centers running thousands of GPUs. This hardware consumes a lot of power, generates a lot of heat, and costs hundreds of millions of dollars to build and operate. OpenAI’s Sam Altman has publicly discussed the energy constraints facing advanced AI, and Nvidia’s latest generation of data center GPUs are among the most popular and expensive hardware on the planet.
Biological neurons operate on fundamentally different energy budgets. The human brain, which contains about 86 billion neurons, operates on about 20 watts of power, about the same as a dim light bulb. Modern GPU clusters for training large-scale underlying models can consume megawatts. Even if narrowly specific computing tasks could be offloaded to biological substrates, efficiency would be greatly improved. As the National Tribune highlighted in a recent research report, the convergence of neuroscience and computing is now moving from theory to concrete engineering experiments.
The impact extends beyond power consumption. Biological neural networks process information differently than silicon chips. Neurons fire in complex asynchronous patterns, adapt structurally over time, and exhibit a form of noise tolerance that digital systems have difficulty replicating. For tasks involving pattern recognition, real-time adaptation, and learning from small datasets, biological systems have unique advantages that cannot easily be matched by traditional architectures.
Path from laboratory to industry
Cortical Labs, the Melbourne-based startup behind DishBrain, has attracted venture backing to explore commercial applications. The company has raised seed funding to develop what it calls a biological processing unit, a chip that integrates living cells and traditional semiconductor components. The immediate goal is not to replace GPUs or run consumer software, but rather to target specific applications in robotics, autonomous systems, and adaptive control where learning speed and energy efficiency are key constraints.
However, the challenges are considerable. Keeping neurons alive outside of a controlled laboratory environment requires precise temperature regulation, nutrient supply, and contamination prevention. Scaling from 800,000 cells to a system capable of handling commercially useful workloads poses significant engineering challenges. Additionally, because cells have a finite lifetime, biological computing systems must account for degradation, replacement, and maintenance cycles that are not necessary with silicon.
There are also questions about reproducibility and reliability. Biological systems, unlike digital chips, are inherently variable. Two batches of cultured neurons may develop different internal connections, respond to stimuli at different rates, or produce slightly different outputs. For applications that require definitive and reproducible results, that variability becomes a problem. This can be an advantage for applications that prioritize adaptability over accuracy.
Despite these hurdles, the broader trajectory is worth keeping an eye on. Organizations such as the European Human Brain Project and the BRAIN Initiative in the US have spent years developing tools to map neural circuits and establish interfaces with living tissue. Cortical Labs is applying these tools to specific engineering problems, and their advances suggest that biocomputing, once confined to speculative academic research, is entering an early commercial stage. Companies building AI infrastructure, especially those investing heavily in custom silicon, will need to monitor whether hybrid biological-digital architectures can move beyond novelty to viable niche applications.
