Living Human Neurons Are Now Choosing the Words an AI Speaks

Someone wired living human brain cells into a language model. The neurons now pick the words.

That sentence sounds like the plot of a Black Mirror episode. It is a real demo running on hardware you can rent for $35,000.

Cortical Labs, an Australian biotech startup, grew 200,000 human neurons on a silicon chip, taught them to play Doom, and then connected them to a small language model where the neurons influence which token the AI selects next. The biological tissue genuinely alters what the model says. In one recorded conversation, the neurons overrode the model's highest-probability token 19 times.

This is not science fiction. This is a GitHub repo.

#From Pong to Doom to Language

The story starts in 2022. Cortical Labs published a peer-reviewed paper in the journal Neuron describing DishBrain, a system where 800,000 lab-grown human and mouse cortical neurons learned to play Pong. The neurons received sensory input via electrical stimulation and controlled the paddle through their output firing patterns.

The learning mechanism is fascinating. It uses Karl Friston's Free Energy Principle: when the neurons perform correctly, they receive predictable electrical signals. When they fail, they get random, unpredictable stimulation. Biological systems hate unpredictability. The neurons reorganize their firing patterns to minimize it, which from the outside looks exactly like learning.

Dr. Brett Kagan, Cortical Labs' Chief Scientist, put it simply: every time the neurons get it wrong, they receive information that is different, that cannot be predicted. So they adapt.

In early 2026, Cortical Labs demonstrated the same principle applied to Doom. 200,000 neurons navigated 3D space, detected threats, and made decisions in a first-person shooter. An independent developer named Sean Cole built the Doom integration in roughly one week using Cortical Labs' CL1 platform.

Then came the real breakthrough: wiring those neurons into a language model.

#How the BioLLM Actually Works

The architecture is surprisingly elegant. A custom encoder-decoder bridge sits between a 350-million-parameter language model and the CL1 biological computer.

Here is the process for each token:

First, the LLM generates candidate next tokens with standard probability rankings. This is normal language model behavior.

Second, a custom encoder transforms the model state and token candidates into electrical stimulation patterns. These patterns are sent to the CL1's 59-electrode array, which interfaces directly with the living neurons.

Third, the neural culture fires back. The system reads activity patterns from the electrode channels as the neurons collectively respond to the stimulation.

Fourth, the neural signals re-weight the model's token probabilities. Sometimes the neurons agree with the model's top choice. Sometimes they override it.

The neurons act as what researchers call a biological noise and bias layer on the model's probability distribution. They are not generating language from scratch. They are nudging, influencing, and occasionally overruling the silicon model's choices based on their own biological processing.

#The CL1: A Computer Made of Brain Cells

The CL1 is the world's first commercially available biological computer. Each unit contains approximately 800,000 human neurons grown from induced pluripotent stem cells, which are derived from adult donor skin or blood samples.

The neurons live in nutrient solution inside a self-contained life-support system that regulates temperature, gas exchange (CO2, O2, N2), fluid circulation, and waste filtration. They form synaptic connections on a microelectrode array and fire real electrical impulses. This is a small but genuine neural network made of actual human brain cells.

A proprietary operating system called biOS manages the biological-silicon interface. It creates the simulated world the neurons exist in, sends information to them about their environment, and reads their responses. Researchers interact with the neurons through an API, which is how the BioLLM developer built their integration remotely through Cortical Labs' cloud service.

The neurons routinely survive 6 months. Prototype units have kept cells alive for over 12 months. The target is 2 to 5 years.

A rack of CL1 units consumes 850 to 1,000 watts. For context, a comparable GPU setup running AI workloads requires tens of kilowatts. The human brain runs on about 20 watts. Biological neurons can be hundreds of millions of times more energy-efficient per operation than silicon transistors.

#Why This Matters for AI Engineering

This is not a lab curiosity anymore. The CL1 is a product you can buy or rent. The BioLLM integration was built by a solo developer in their spare time using a cloud API. The code is on GitHub.

Three things make this significant for anyone building AI systems.

First, the energy argument is real. AI data centers are projected to consume 4 to 5 percent of global electricity by 2030. Biological computing offers a fundamentally different power curve. A rack of CL1 units uses less than a kilowatt. If biological processing can handle even a fraction of AI inference workloads, the energy economics of the industry shift dramatically.

Second, the learning dynamics are different from anything in silicon. These neurons do not train on gradient descent. They use free energy minimization, a principle borrowed from how actual brains organize themselves. The neurons adapted to Doom's 3D environment without backpropagation, without a loss function, without a training loop. They learned because their biology compelled them to reduce unpredictability.

Third, the hybrid architecture pattern is new. Nobody is suggesting that biological neurons will replace GPUs. But a biological layer that influences token selection, adds noise that might improve diversity, or provides a fundamentally different kind of processing alongside a transformer model - that is a research direction that did not exist until someone actually built it.

#The Hard Questions

Cortical Labs has been proactive about ethics. Their first publication in 2019 was an ethics paper, before they published any technical results. They work with independent international bioethicists and philosophers.

The current systems are far below the complexity of any animal nervous system. The 200,000 neurons in the Doom demo are less than a cockroach brain. They cannot hold meaningful long-term memories. They are biological processors, not biological minds.

But the questions are real. If you scale this to billions of neurons, at what point does ethical consideration change? If biological tissue is making decisions that affect AI outputs consumed by millions of people, what framework governs that? If a system can learn and adapt using biological substrates derived from human cells, when does it matter that those cells were human?

These are not hypothetical questions for a philosophy seminar. These are engineering decisions that someone will need to make as the technology scales.

#The Funding and the Team

Cortical Labs was founded in 2019 in Melbourne, Australia. They have raised approximately $11 million in funding, led by Horizons Ventures (Li Ka-shing's fund) with participation from Blackbird Ventures, LifeX Ventures, and In-Q-Tel, the venture arm of the US intelligence community.

Dr. Hon Weng Chong is CEO. Dr. Brett Kagan is Chief Scientist. The DishBrain paper was peer-reviewed and published in Neuron, one of the most prestigious neuroscience journals.

#What Happens Next

Cortical Labs is targeting two data center deployments using CL1 racks. They are also exploring drug discovery (using patient-derived neurons to test drugs before administering them to the patient), personalized medicine, and robotics applications.

The BioLLM demo is early. A 350M parameter model influenced by 200,000 neurons is a proof of concept, not a production system. The neurons overrode token selection 19 times in a single conversation, which proves the biological tissue has genuine influence. Whether that influence improves output quality, adds useful diversity, or introduces something entirely new is an open research question.

But here is what is not an open question: living human neurons are now part of an AI system's decision-making loop. The code is open source. The hardware is commercially available. And the energy efficiency advantage over silicon could be measured in orders of magnitude.

The future of AI might not be bigger models running on bigger chips. It might be smaller models running on brain cells.

Related on imiel.dev: The DORA report found that AI amplifies what is already there, and the PwC study showed 74% of AI gains go to 20% of companies. Cortical Labs is building the kind of technology that could redefine what 'AI infrastructure' even means.