Projects / Read and costed
Computing with living neurons
Horizon An idea with sources behind it and nothing built. Published so it can be argued with.
Living neurons in a dish, wired into silicon, change how they fire when given structured feedback. The product is real and on sale. Most public claims about it fail against the sources, and the survey lists which.
The area is as much about reading claims as about the biology.
Drawn from Jo's survey, Part I.
What is established
These are the survey's own gradings, carried over unchanged.
-
Established
Cultured neurons and organoids can be put in closed loops with silicon and will change their firing statistics in response to feedback. Known since 2001 to 2008.
-
Established
Neural tissue works as a physical reservoir: nonlinear response, fading memory, enough dimensionality for a linear readout to classify inputs.
-
Established
The CL1 is a real shipping product. 115 units in 2025, $35k, 59 electrodes, sub-millisecond latency, six month viability, Python API.
-
Established
The interface problem is being solved, which the survey calls the most important development in the field: eighteen months and an internal team to run Pong, then one week and an outside developer to run Doom.
What is not
Every item below is commonly stated and does not survive the sources.
-
Commonly claimed, not established
"A million times less power." It compares brain metabolism to whole silicon system power, ignores the incubator, and rests on an energy figure that appears inflated around eightfold. Measured per-neuron power of the shipping product is about ten million times worse than a brain.
-
Commonly claimed, not established
"The neurons played Doom." The builder's own notes say the decoder tends to start becoming a policy head, meaning the software may be routing around the biology entirely. No published ablation establishes that the neurons contributed.
-
Commonly claimed, not established
"Code deployable biological computer." The code deploys to the interface, not to the neurons. There is no programming model for the tissue.
-
Commonly claimed, not established
Neuron count as a capability metric. Doom used around 200,000 neurons and Pong used around 800,000. Capability went up as the count went down, because the interface improved.
The one sentence
From the survey, unedited: biocomputing in 2026 is a maturing instrumentation field with a working API, a real product, genuine scientific interest in sample efficiency, and an energy story that is currently the exact inverse of how it is marketed. Nobody has demonstrated, with an ablation control, that the living tissue contributes anything a random nonlinear reservoir would not.
What is not known here
The open items for this project. Work on it starts here.
- Sample efficiency, whether living tissue learns from fewer examples than silicon, is a legitimate open research question with a paper behind it and no settled answer.
- Organoid intelligence as a near-term field does not hold up: the roadmap targets ten million cell organoids and deployed ones have ten thousand. That is a thousandfold gap on a decade-scale programme.