Biology Meets Silicon: Neural Networks Built on Unconventional Computing

For a long time, artificial intelligence copied the brain only in its mathematics: layer after layer, matrix after matrix, chains of linear algebra. Now researchers are starting to copy it literally, using biological mechanisms in neural networks and building hardware that computes like a living cell. Unconventional computing, whether molecular, optical or neuromorphic, is becoming the leading contender for the architecture of strong AI.

When silicon hits a wall

Modern GPUs are built on transistors that keep getting smaller, but not indefinitely. Moore's law is slowing down. Training large models takes a staggering amount of energy: a single Perplexity search consumes as much electricity as a car uses to drive 5 km. Training GPT-4 cost hundreds of millions of dollars and released as much carbon into the atmosphere as a small city. AGI would require planetary-scale power if it were built on conventional silicon. Engineers are running into physics.

Unconventional computing promises a way out. Instead of pushing electrons around a chip, you can use photons in optical systems (parallelism at the speed of light), DNA molecules (an incredible density of information storage), or electrochemical processes in neuromorphic chips (energy efficiency comparable to the biological brain).

Neuromorphic chips: brain architecture in electronics

Intel Loihi, BrainScaleS and SpiNNaker are not GPUs imitating neurons in code. They are chips physically built like a brain: spiking neurons, synapses with adjustable weights, and asynchronous, event-driven processing. Instead of multiplying all weights by a batch, only active neurons send impulses, and only when needed.

The result is striking: power consumption a thousand times lower than a GPU of the same scale. The human brain does on 20 watts what current neural networks need 20 kilowatts for. A neuromorphic chip reproduces that result.

• Real-time learning (without waiting for a full batch of data)
• Processing streams instead of matrices (ideal for sensors and cameras)
• Adapting to new tasks without re-optimizing an entire model
• Resilience to noisy inputs thanks to the biological logic of impulses

Companies such as Intel and academic labs have already used neuromorphic chips in robotics and sensory data processing. Scaling is harder, but the potential is huge.

The molecular machine: computation in a test tube

DNA is not just a storage medium for genes. It is a parallel computer. A single molecule can encode gigabits of information, and computation can run simultaneously across all the molecules in a solution. Researchers at MIT, Caltech and other labs already use DNA computing for machine learning tasks such as classification and matrix multiplication.

Scaling is harder than with chips: it requires methods to control billions of molecules in sync and to extract results without errors. But if it works, computing density would exceed any silicon equivalent, and energy consumption would fall several-fold.

Researchers are also exploring optical computers (photons instead of electrons) and quantum systems. The picture is clear: the monoculture of silicon GPUs is on its way out. AGI will be a hybrid.

What this means

The boundary between biology and silicon is blurring. Unconventional computing is not a laboratory curiosity but an engineering workshop for strong AI. Within a year or two we will see the first hybrid systems: neuromorphic accelerators in data centers next to GPUs, molecular computers for specialized tasks, and optical processors for linear algebra. When AGI arrives, if it arrives at all, 90% of its hardware will consist of what living nature proved effective billions of years ago.