What Cortical Labs built and why it matters
In 2023, Cortical Labs demonstrated that living human cortical neurons grown on a multi-electrode array could learn to play the game Pong. The system, which they call DishBrain, could adapt its responses based on feedback in a way consistent with Hebbian learning. In 2024, they released the CL1, a commercial platform that pairs living neurons with digital computing infrastructure.
This is not a simulation. The neurons are biological cells, fed nutrients, maintained at physiological temperature, and connected to the silicon substrate via electrodes. The computation is genuinely performed by living tissue.
Why does this matter? Because it raises a concrete question: if biological neurons can perform machine learning tasks, how do they compare to software-simulated spiking neural networks (SNNs) running on conventional hardware? And more importantly, do they offer any advantage that justifies the substantial operational complexity of keeping neurons alive?
How spiking neural networks compare in simulation
Brian2 is the primary Python library for simulating biologically realistic spiking neural networks. It implements the Leaky Integrate-and-Fire (LIF) neuron model, STDP (spike-timing-dependent plasticity) learning rules, and various network architectures including liquid state machines.
import brian2 as b2
# Leaky Integrate-and-Fire neuron
eqs = """
dv/dt = (v_rest - v + R * I) / tau : volt
I : amp
"""
neurons = b2.NeuronGroup(
N = 1000,
model = eqs,
threshold = 'v > v_thresh',
reset = 'v = v_reset',
refractory = 5*b2.ms,
namespace = {
'v_rest': -70*b2.mV,
'v_thresh': -55*b2.mV,
'v_reset': -80*b2.mV,
'R': 10*b2.Mohm,
'tau': 10*b2.ms
}
)
# STDP synapse
stdp = b2.Synapses(neurons, neurons,
model='''
dApre/dt = -Apre/tau_pre : 1 (event-driven)
dapost/dt = -apost/tau_post : 1 (event-driven)
w : 1
''',
on_pre = 'Apre += 0.01; w = clip(w + apost, 0, 1)',
on_post = 'apost += -0.01; w = clip(w + Apre, 0, 1)'
)The MNIST benchmark: results and caveats
On MNIST digit classification using a liquid state machine architecture in Brian2 (1000 LIF neurons, rate-coded input, readout layer trained with logistic regression), accuracy reaches approximately 95-96% with appropriate hyperparameter tuning.
CL1-based systems have demonstrated pattern classification capabilities, with reported accuracies in similar ranges for simple classification tasks. The comparison is complicated by several factors: CL1 systems are not standardized in the way that software benchmarks are, the number of "effective neurons" contributing to computation varies with culture quality, and the task formulation differs between publications.
The short answer: on standard classification benchmarks, biological and simulated SNNs achieve comparable accuracy. Accuracy is not where the interesting differences lie.
What biological neurons do that silicon cannot (yet)
The genuinely interesting properties of biological computing are not accuracy on benchmarks: they are architectural and dynamic properties that silicon implementations approximate only crudely:
- Continuous-time processing: biological neurons process information continuously, not in discrete time steps. The timing relationships between spikes carry information. Brian2 can simulate this but only discretized to a simulation timestep.
- Structural plasticity: biological networks physically rewire themselves: new synapses form, old ones are pruned, over timescales from hours to weeks. This is distinct from weight plasticity (changing synapse strength) and is not captured by STDP models.
- Metabolic constraints as regularization: neurons that fire too frequently use more energy and are naturally constrained. This biological "cost function" may act as an implicit regularizer that prevents overfitting in ways that are not yet fully understood.
- Heterogeneity: real neurons are not identical units. The diversity of cell types, ion channel expression, and connectivity is partly responsible for the robustness of biological neural computation.
The energy efficiency argument
The most frequently cited advantage of neuromorphic computing (biological or silicon SNN) is energy efficiency. A human brain consumes approximately 20 watts while performing computations that require thousands of watts in silicon implementations.
This argument is real but often misapplied. The brain is energy-efficient at the tasks it was optimized for over millions of years of evolution: continuous, multimodal, embodied perception and action. It is not particularly energy-efficient at matrix multiplication, the core operation of modern deep learning.
For specific tasks: sparse, event-driven sensing, continuous-time temporal pattern recognition, neuromorphic approaches can achieve significant energy savings over GPU-based deep learning. For the tasks that dominate current AI applications (large language models, image classification at scale), the comparison is not favorable to neuromorphic approaches.
What this is not
CL1 is not a path to artificial general intelligence. It is not a replacement for deep learning. The press coverage has sometimes framed biological computing as a step toward "thinking machines" in a way that conflates substrate with capability. A living neuron on a chip is not more intelligent than a silicon implementation: it is differently implemented.
The more honest framing: biological computing is an interesting experimental platform for studying neural computation principles in vivo, and may have specific applications in neural interfacing, sensory processing, and low-power edge computing. It is not, in its current form, a practical alternative to GPU-based deep learning for any major AI application.
Where the field is actually going
The near-term research directions that are likely to produce useful results are not building bigger biological neural networks: they are using biological systems to study learning rules and information coding that can then be implemented more efficiently in neuromorphic hardware.
Intel's Loihi 2 and IBM's NorthPole are silicon neuromorphic chips that implement SNN-inspired architectures in hardware. They offer real energy efficiency advantages for certain inference tasks. They are informed by neuroscience but do not require keeping neurons alive.
The CL1 platform is most valuable as a research tool for studying how biological learning works at the circuit level, an in vitro model system, not a product roadmap.
Key takeaway
Biological computing and SNN simulation produce similar accuracy on classification benchmarks. The real differences are in dynamics, plasticity, and energy characteristics that matter for specific applications. CL1 is a fascinating research platform. It is not a practical alternative to silicon computing for any current AI application, and that is probably not what it is trying to be.