Two Neurosurgeons Learn from Brain Cells to Build Faster, Cheaper AI

TLDR

• The Biological Computing Co. is turning brain-cell research into software. The company studies how living neurons process information, translates those patterns into mathematics and incorporates them into conventional AI models.
• Its first commercial model targets AI’s high computing costs. Mischief generates video at least five times faster and at 80% lower cost than the open-weight model on which it is based.
• TBC is applying its neuron-derived technology to robotics and world models while researching a more radical approach in which living neurons could eventually perform some computing directly.

The Biological Computing Co. is taking an unusual route to making artificial intelligence cheaper and faster: studying how living brain cells process information, turning what it learns into software and inserting that software into conventional AI models.

Founded by two neurosurgeons, the San Francisco startup is putting that approach into a commercial product today with TBC Mischief, an AI video-generation model that the company says generates video at least five times faster and at 80% lower cost than the open-weight model it is built on.

“Mischief 1 is the first ever commercialized video generation model that was built or derived from brain cells,” co-founder and CEO Alex Ksendzovsky told The AI Innovator. “This is a turning point for biological computing in general, certainly for our company.”

“It is the first of its kind,” COO and co-founder Jon Pomeraniec added. “It’s the entire process of discovering new and better ways to compute and process information using real biology.”

But there are no living neurons inside Mischief.

“The neurons stay completely in our lab, and that was very intentional,” Pomeraniec said. “So we’re plugging into existing workflows that’s running on silicon that the world already owns.”

TBC’s premise is that studying how biological neurons process information can reveal ways to make artificial neural networks more efficient. The company uses commercially sourced neurons, including rat neurons and neurons derived from human stem cells, Ksendzovsky said.

Built on Wan 2.2, an open-weight video model, Mischief is designed to turn images into video. TBC said its software adds less than 0.1% to the underlying model while making it substantially more efficient. The company is charging 2 cents per second of generated video.

Mischief is available to consumers today on fal.ai and free until Nov. 2. Enterprises can access the model soon on AWS, which is a partner but not an investor in the company.

TBC said it has raised $52 million thus far and is a seed-stage company. It employs 40 people who work in offices in San Francisco and Baltimore.

The science explained

The process starts with actual living neurons.

TBC grows neurons on arrays containing electrodes and converts digital information, such as pixels, into electrical stimulation that can be delivered to the cells. Researchers then record the neurons’ electrical responses and look for recurring patterns in how the biological network processes that information.

“You have to translate from something in this world that we live in, either numbers or images, into electrical stimuli,” Ksendzovsky said.

The neurons respond with electrical activity that the company measures and interprets. The goal is not to copy an entire brain, but to identify useful computational patterns and translate them into mathematics that can be incorporated into artificial neural networks.

The approach is different from merely designing an algorithm based on theories about how brains work. TBC says its optimizations are derived from direct measurements of living neurons. The company describes its commercial process in three steps: grow and observe living neurons, translate their processing patterns into software code and use that code to modify an existing AI model.

Artificial neural networks were themselves inspired by biology, but modern AI systems bear only a loose resemblance to actual brains. TBC is effectively returning to the biological source and asking whether mechanisms observed in living neural networks can help solve problems that artificial networks handle inefficiently.

In its earlier computer-vision experiments, the company encoded images as electrical stimulation and sent them through cultures of living neurons. The resulting neural activity was then analyzed for patterns that could be incorporated into conventional machine-learning systems.

One result of that work is what TBC calls its Neural Dynamics Adapter, a software module derived from patterns observed in living neurons. Rather than replacing an existing AI model, the adapter modifies information moving through it.

Ksendzovsky described it as a modular component that changes the matrix of numbers flowing through a generative model.

“That matrix of numbers gets slightly modified by our additional mathematics, by our additional networks that come from the brain cells,” he said. “And thus, it’s a little bit more biological in terms of how the information is processed and that allows for these quality and efficiency changes.”

The company has also experimented with biological patterns related to spatial and temporal processing, an important problem for video models that must understand not only what appears in an image but how objects change and move over time.

Why video came first

TBC chose video because generating it requires substantial computing power and models must maintain consistency across successive frames. They felt they could make the most impact by building a more efficient video model.

“When we started incorporating our model into just AI systems in general, video at the time, and still does, leaves a lot to be desired, meaning they’re pretty good, they generate reasonable rated videos, but they’re extremely expensive to generate,” Ksendzovsky said. “They still are. As far as tokens go, I think they’re multiple hundreds-fold more expensive to generate than language tokens. And so they’re ripe for disruption in terms of efficiency.”

Secondly, video fits the underlying science as well.

“The way that we are determining the math that we’re building behind how brain cells process information lends itself well towards spatial temporal responses and allows itself to incorporate into these video models,” he said.

Wonder Studios, an AI-native creative studio that tested Mischief, generated 100 five-second clips in about six minutes, according to TBC.

The economics could matter particularly for studios and marketing teams that generate large numbers of clips before selecting the handful they ultimately use.

“If you’re a studio and you’re generating hundreds and hundreds of videos at a time, for example, and waiting two minutes per five-second frame to generate, it’s really stifling to your creativity,” Ksendzovsky said.

Mischief’s lower cost stems from reducing the GPU time required to generate video. Ksendzovsky said the company pays for GPU capacity by time, so faster inference lowers its own computing expense and gives it room to charge customers less.

One of the startup’s more ambitious applications is robotics.

Ksendzovsky said TBC has incorporated its neuron-derived technology into a version of Nvidia’s Cosmos 3 world model and tested it in simulation with a robotic arm. He said the TBC version hallucinated less and produced a more realistic representation of the environment around the robot. The startup is now building a physical robotic-arm experiment and has begun conversations with potential robotics customers.

TBC’s longer-term research is considerably more radical: putting living neurons back into the computing loop.

The company calls that “real-time biological computing,” in which some computational tasks would be offloaded directly to living neural networks.

Ksendzovsky said that work remains in research and development and estimated the company could have a prototype offloading a significant amount of computation within a couple of years.