TBC: AI adapters trained on live neurons in a dish, $25M seed

TBC startup trains AI adapters on live neurons in a dish and raises $25 million to commercialize the technology

In the standard setup, an AI model is trained on data: text, images, video. The Biological Computing Company (TBC) inserts live neurons between the data and the model. Not a simulation, not a neuromorphic chip, but real biological cells in a dish about the size of a grain of salt.

TBC's device consists of ~100,000 neurons and 4,096 electrodes. Visual data is converted into electrical signals, the neurons process them, and the electrodes record the activity. From those recordings, the company extracts a mathematical model, something like a compressed neural “fingerprint” of the processed data. That fingerprint is then used to train a software “adapter” that plugs into an existing AI model as an additional layer.

“If a photograph is worth a thousand words, the neural representation of an image is worth a million. It captures complex spatial and temporal dynamics” — Alex Ksendzovsky, TBC co-founder

According to the company's заявлению компании, this kind of adapter doubles the length of video an AI model can generate before degradation begins. At the same time, the adapter is less than 1% the size of the base model and does not require retraining it. If those numbers hold up, it could be a cheap way to improve generative models without adding more compute.

TBC has raised a $25 million seed round from Primary Ventures. The money will go toward launching a lab and commercializing the technology. Both co-founders are neurobiologists: Jon Pomeraniec and Alex Ksendzovsky.

“We are communicating with the best computer ever built. It was designed, evolved, and refined by nature” — Jon Pomeraniec

TBC is not alone in this space. Австралийский Cortical Labs showed neurons trained to play Pong back in 2022. Switzerland's FinalSpark is building a platform of biological processors based on neuronal organoids. IBM has long been developing neuromorphic chips, while Groq, the creator of energy-efficient AI chips, was acquired by Nvidia for $20 billion. But TBC is approaching the problem from a different angle: instead of imitating the architecture of the brain in silicon, it uses the brain itself as an intermediate processing layer.

The limitations are serious. There are no peer-reviewed publications, only company claims. 100,000 neurons is a negligible fraction of the 86 billion in the human brain. The Forbes article does not specify what kinds of neurons are being used, how long they survive, or how the system scales. The energy-efficiency question is also still open: the brain consumes ~20 watts, but keeping living cells alive in a lab is a separate engineering challenge. And finally, using live neurons for computation raises ethical questions that regulators have not even begun to discuss.

For readers following brain emulation: if live neurons can be used as a computational substrate for AI, the reverse is true as well. Every such experiment produces data on how biological neurons encode information. And that is another brick in the foundation of mind uploading.