Startup Grows Live Human-Derived Brain Cells to Run Computation in San Francisco

The video opens with the company's introduction. The Biological Computing Company (TBC), based in San Francisco, describes itself as a company that uses live brain cells. Staff say they will show what that looks like in the lab.

The company's founders are CEO Alex Sensoffski, who was trained as a neuroscientist, and co-founder and CEO John, a neurosurgeon-neuroscientist. They say they have known each other for about 17 years and have done a lot of research with access to the human brain, studying learning and memory. They said they wanted to be the first to commercialize this technology.

According to the staff, some of the company's brain cells come from rats and some come from human stem cells, and both have advantages and disadvantages. They say most of their current data comes from rat cells.

Stem cells are stored in liquid nitrogen. In their stem-cell form they can be replicated and scaled quickly, and the company says it has millions of them. They are then differentiated into brain cells, or neurons, and supporting cells. Staff say neurons from rats are cells the company has worked with for a long time, and each source confers advantages over the other in its own way.

The cells are kept in incubators at a set temperature of 37°C, 5% CO2 and controlled humidity. Staff say that in the first few weeks the cells start to build the connections needed to form a good neuronal network. The company prepares 20 to 30 plates every week, with different cell densities.

The cells live in a culture medium that supplies all the nutrients they need. Staff explained that this is why the lab is called the wet lab.

Staff place a cap on the culture dish. The cap is a fluorocarbon film that allows gas exchange while keeping anything from dropping onto the cells. The dish is then placed in the company's bioAM device and locked in place. The device's software recognizes the specific serial number of each plate.

The staff say they sent electrical current through about 600 electrodes in the shape of the TBC logo, then watched what the neurons did. The heat map in the demo shows the logo imprinted onto the neurons. The staff said the important part is what happens 100 milliseconds later: the company can record the cells' activity, and that pattern of activity is sustained every single time.

The presenter said the response to the TBC logo is different from the response when cells see something like a biohazard symbol, because the logo adds features to the image. The company says it has converted pixel space to neural space and has a neural representation of what it means to be a picture.

The team repeats this process hundreds of thousands of times and looks for what is sustained across many of the neural cultures. The aim is to determine what a pixel means to brain cells and how that information translates across the biological node.

Based on this, the team says it can build models of what it means for pixels to be represented in brain cells. The company says these mathematical models can be incorporated into AI systems that use pixels, such as video generation.

The presenter said the team is often asked whether the cells are sentient or feel anything, and the answer, they said, is explicitly no. The culture holds about 150,000 cells, which is a fraction of what animal brains contain. The presenter said the cultures have an architecture, but it is very different from an intact brain's: there are no blood vessels.

According to the presenter, the company has access to the fundamental computational principles of how brain cells communicate, but not to emergent properties such as sentience, pain or awareness, which the presenter said require a certain three-dimensional architecture and a certain number of cells. The presenter said the company draws a clear line there and has no expectation or interest in reaching that point.

The staff launched a prompt on the demo. They said the system took 15 seconds to generate a 5-second clip. They said the system is built on top of an open-source model, and that they compare it to the original model. The demo runs the same prompt on that base model to measure how long it takes.

The presenter describes the company's long-term direction as real-time biological compute, which means using the brain cells themselves as part of the compute circuit to offload some of the energy demand. The presenter said mathematical principles derived from brain cells already lead to efficiency gains of about fivefold at least, and that offloading some of the compute to actual brain cells could bring even more.

The presenter said the demo tests a basic machine-learning problem: nonlinearly combining two inputs to detect whether they correspond. The presenter waves one hand and then the other. The company encodes the information about which hand was waved into the cell culture, routes the output back through the robot, and asks the culture to determine whether it was a matching hand, meaning the same hand twice, or two different hands.

When the presenter showed the same hand twice, the culture's output was that there was no match, and the presenter said the robot was therefore instructed to look sad. In a further trial in which the presenter showed the left hand both times, the robot appeared flat.

The staff returned to the video generation test. On this run, the presenter said, the company's model generated the video 8.27 times faster than the base model that it was built on.

The presenter said quality, which the team also looks at, is conserved across both models. Many companies use the same optimization tools to achieve speed, but the presenter said those tools can degrade quality. The company says its additional mathematical layer lets it maintain or improve quality while making the process 8.27 times faster and cheaper.

The presenter said the computational neuroscientists work on the data from the brain cells, and the AI team builds the models that generate videos. The presenter said the video shown in the demo was generated by the company's model.

The presenter said the company is building in San Francisco and moved here from Baltimore because there is so much innovation in AI. The company partners with AI companies including Amazon Web Services, which the presenter said gives it the opportunity to be at the cutting edge.

The presenter said the team has been growing brain cells for a long time, and asked what brain cells look like when information flows through them compared with an actual brain. Another speaker said the company has had a purview on information processing by brain cells for quite some time.

A founder said that what led him to start a startup and leave his neurosurgical career was impact. He said he had been thinking about this for about 20 years. In his opinion, the most impactful thing the team can leave to the world is much more energy-efficient and much more performant computing. He said it was an easy decision because the technology is now available, whereas it was not 15 years ago.

The founder said the two co-founders had long shared the view that there is a big mismatch between what the brain can do, what biology can do, and what computing does in the real world. He said the company's theme is returning to biology, nature and evolution, and that the team is at the frontier of discovery with no one else doing this.

A speaker said there are a lot of differences, starting with brain cells being event-driven: they are activated only when an event is happening. They are also fully parallelized. The speaker said that when one brain cell is stimulated, all of them feel that stimulus, which the speaker said does not happen in artificial neural networks.

The speaker also said brain cells are plastic and change: the way they are connected to one another changes any time they learn. The speaker contrasted this with rigid silicon chips, in which transistors do not change.

Asked about AGI, a speaker said the question should be deferred until there is a good definition of what AGI means. The speaker said questions about whether models can generalize beyond their training data and about continued learning are still open. The company, the speaker said, is not saying it is building toward AGI. It is building much more efficient systems that mimic the brain more, and once the community agrees on how to define AGI, it can say how close it is.

The speaker said the company's value is that this is not artificial intelligence but biological, natural, evolutionary intelligence. In the speaker's view, making artificial systems a little more biological increases their output and capabilities tremendously, and a biology-based future would look very different in a good way.