Most current neural biocomputers are hybrid machines. Conventional electronics encode an input, living neural tissue transforms it, electrodes record the response, and software decides what to do next.
1. Encode the input
The biological tissue does not receive a text prompt or an image file directly. Software converts task information into something the culture can receive, usually patterns of electrical stimulation delivered through a multielectrode array. Chemical and optical inputs are also possible.
2. Let the biological network transform it
Neurons respond according to their current connectivity, membrane state, recent activity and synaptic plasticity. In reservoir-computing experiments, the complex dynamics of the network provide a high-dimensional transformation that conventional software can read out.
3. Record the response
Electrodes measure extracellular electrical activity. Depending on the platform, software may work with detected spikes, spike counts, field potentials or richer time-series data.
4. Close the loop
In a closed-loop task, the recorded response changes the next software state or the next stimulation pattern. This makes adaptation possible. DishBrain, for example, coupled neuronal activity to a simplified Pong environment, while the 2026 cart-pole work delivered task-dependent training signals to cortical organoids.
2D cultures and 3D organoids
Two-dimensional cultures make electrical access comparatively straightforward because cells grow near the electrodes. Three-dimensional organoids can contain richer tissue organisation but create harder problems in recording depth, stimulation, nutrient delivery and standardisation.
Where the conventional computer remains
The biological component currently handles only part of the computation. Encoding, decoding, task logic, data storage, control software and much of the training machinery remain electronic. Any claim of energy or performance advantage therefore needs measurements across the complete system, including life support, electronics and conventional compute.