Earley AI Podcast - Episode 97: Biological Computing, Brain-Derived Algorithms, and the Future of AI Efficiency with Alex Ksendzovsky

Why Making AI More Biological May Be the Most Consequential Development in the History of Computing

Guest: Alex Ksendzovsky, CEO and Co-Founder at The Biological Computing Company

Host: Seth Earley, CEO at Earley Information Science

Published on: August 14, 2026

 

Podcast Transcript: Biological Computing, Brain-Derived Algorithms, and the Future of AI Efficiency

Transcript introduction

This transcript captures a conversation between Seth Earley and Alex Ksendzovsky about a frontier that most people in enterprise technology have never encountered - biological computing. They cover the historical divergence between AI and neuroscience, how TBC's research platform works at the level of real neurons on electrode arrays, the adapter products derived from those experiments and their measurable performance and efficiency gains, how the catastrophic forgetting problem is being attacked through closed-loop biological learning, the roadmap from software adapters through neuromorphic hardware to real-time biological compute, and the ethical framework that guides work with human-derived brain cells.

Transcript

Seth Earley: Welcome to the Earley AI Podcast. My name is Seth Earley, I'm your host. In each episode, we talk about how artificial intelligence and data are reshaping how we operate, how we do business, and how we interact with customers and employees and colleagues. Today, we're going to be talking about something that really sits at the frontier of AI, neuroscience, and computing - something that most people in the enterprise technology world have not yet encountered.

The brain consumes roughly 20 watts of power - less than a dim light bulb - and yet it outperforms some of the most powerful data centers on Earth in certain ways. That gap is not a coincidence. It's a result of hundreds of millions of years of evolution solving the same problems we are now trying to solve in silicon. My guest today is asking whether we can close that gap, not just by drawing inspiration from biology, but by actually using biology itself.

Alex Ksendzovsky is CEO of The Biological Computing Company, TBC. Alex is a neurosurgeon and neuroscientist who spent nearly two decades studying how the brain processes information, including implanting electrodes into human brains to understand epilepsy and growing neurons in a dish to study them at the molecular level. He co-founded TBC to turn those insights into a new generation of AI algorithms and, ultimately, computing architectures that are orders of magnitude more efficient than what we have today. Alex, welcome to the show.

Alex Ksendzovsky: Awesome, thanks a lot for having me.

Seth Earley: Let's start with common misconceptions. When people hear that AI is inspired by the brain, most think it means neural networks are basically how the brain works. What is wrong with that assumption, and how far has the field actually diverged from biology?

Alex Ksendzovsky: That's a great question. It started in the 1940s when the first AI algorithms - the McCulloch-Pitts neuron - were actually based on how we understood neurons and brain function at the time. Then, moving forward to the 1980s, when backpropagation happened - which is an algorithm that's very non-biological - that's when you had a pretty massive divergence. Moving forward past the 1980s into 2026, you have extremely performant AI systems, but they're being trained with brute force, and they're extremely inefficient.

At the same time, in the 1980s, neuroscience itself advanced significantly. We started to really understand how the brain works through brain cells growing on electrodes, through implanting electrodes into patients for epilepsy - the labs I was part of. Over that time, our understanding of neuroscience increased, but those insights were not really being used for AI. At TBC, we think the reason for the inefficiency is because current AI systems became extremely non-biological.

Seth Earley: What did we gain from backpropagation, and what did we leave behind?

Alex Ksendzovsky: We gained performance. That's what led to the models we have today, and they're extremely performant - language models are used by people worldwide and giving us extremely good outputs. But they're very energy inefficient. We found ourselves as a society in the position where we need these tools because they have improved the lives of so many, but they take so much energy. The idea behind TBC was: now that we understand neuroscience and can actually derive principles from biological networks, can we improve on these AI systems, make them a little bit more biological, and in doing so improve both their efficiency and their performance?

Seth Earley: Talk about the distinction between drawing inspiration from the brain versus empirically deriving principles from it. What is the difference, and what does that look like in practice?

Alex Ksendzovsky: Drawing inspiration from the brain is not new. There are several large companies that look at recordings from human or animal brains and ask how that leads to things like memory, and how it can improve current AI systems. The problem with that is those are not directed experiments.

What we do differently is run directed experiments. We take brain cells - either from rat cortical neurons or created from stem cells - and we grow those brain cells on electrode arrays. The array I can hold up here has 4,096 electrodes. Hundreds of thousands of neurons grow on those electrodes, and that gives us access to both sending information into the biology through electrical current and recording the biology's response.

The way the brain takes in information - vision, sound - is through electrical current. Your eyes convert light, your ears convert sound waves, into electrical current that propagates through the brain. When we don't have those senses in our biological cultures, we send information through electrical patterns. A big unlock for the company early on was figuring out how to take an image, or a token, convert it into an electrical pattern, and send it into the dish of neurons. Once we've encoded that information into the biological network, we record at very high resolution across all 4,096 electrodes over time - a time series of how those neurons process an image, for example. From that, we build mathematical models. We say: a pixel is spatially and temporally represented by the brain cells in a certain way. Can we create a mathematical model from that, and apply it to current AI systems? That is what we call an adapter.

Seth Earley: So you're building a representation, a mathematical representation of what that perception actually is, and then applying those rules to AI systems.

Alex Ksendzovsky: Exactly. Adapters are lightweight, small-parameter models or learning rules that plug into current AI systems - diffusion transformers, other transformer models - to see if they can improve the output and efficiency. As information in the form of a matrix flows through a transformer model in the context of video generation, we augment that flow by building a mathematical model that changes the matrix as it flows through, makes it look a little bit more biological, and predicts the next frame or the next token in a more accurate or more efficient way.

Seth Earley: What are the results? What is the actual difference in performance between a standard model and one with your biologically derived adapter?

Alex Ksendzovsky: We've been benchmarking this constantly. Looking at metrics like VBench and PAI Bench - five to six metrics covering motion quality, background quality, aesthetic quality, image quality - we're showing pretty significant improvements. About 13% improvement in aesthetic quality, for example. To give that a frame of reference: to get on the published leaderboard in this field, it's typically a 1-2% improvement. We're showing 13-20% improvements in some of these metrics, which is very large.

On human preference ratings - which is the gold standard for understanding how well your video generation algorithms perform - when we compared our recent adapter against the base model, 65% of the time human raters chose our model. To get on the ELO leaderboard in the published literature, 53% of the time is considered a significant improvement. So human raters are rating our videos as much better a very significant portion of the time.

On the efficiency side, we're about 4 to 5x more efficient in terms of how many frames we can produce per second. We recently released an interactive demo on our website where you can play a Minecraft game that's being generated in real time in front of you - and that 4 to 5x efficiency improvement is visible in the frames per second as it generates.

An extremely important recent finding: when you combine our adapter with current optimization strategies, they're synergistic. You can get 4 to 5x efficiency from our adapter, 4 to 5x from other optimization strategies, and what we've found is that our adapter, when you use those other strategies together, actually leads to a decrease in quality degradation - it maintains quality. And we're continuing to improve. I think 20-fold within the next year, 100-fold the year after, and it keeps going. This is just in software.

Seth Earley: One of the fundamental limitations of current AI models is catastrophic forgetting - the inability to keep learning without losing what was learned before. How does biology solve that problem, and what have you been able to derive from neurons about continual learning?

Alex Ksendzovsky: That's our fourth blog post, and I'd encourage the audience to look at our blog because a lot of the science is written out there.

Going back to our setup: we have information that goes into the biological network, the neurons represent it, and we mathematically model that. But the neurons can also learn. Learning in a cultured setup means the synapses actually change. In our system, we have what we call closed-loop learning.

Let me set up an experiment conceptually. We want to classify a number - MNIST is a classic dataset. We encode a number into our biological network, take the response, and build a classifier. In our early data, it was more accurate than the baseline. Now we want to use this to actually train the biology. We take additional numbers, encode them, take the response, and when the classifier gets it wrong, within a couple of milliseconds we can send a second pattern of electrical stimulation that tells the neurons the answer was wrong - or we can send one that tells them the answer was right. Over these closed-loop steps, you actually start changing how the biological network is connected, and the representation becomes more accurate.

The reason this closed-loop setup is important in the context of memory is that it gives us insight into how the synapses are actually being changed over time. We can take statistical measurements, model what the synapses look like from a neurophysiological perspective before training and after training, and start building rules from that. In our fourth blog post, we derived a learning rule that more closely mimics how the neurons change their structure in response to closed-loop training, and found that when this learning rule was incorporated into our MNIST continual learning task, we showed significant improvements over backpropagation. So zooming out: we study how neurons change their network over time, mathematically model that, build learning rules from that, and apply them towards continual learning tasks.

Seth Earley: Talk about the longer-term vision beyond the current adapters. Where does this go from here?

Alex Ksendzovsky: The current product is adapters that plug into existing transformer architectures. We're not completely changing the baseline architecture yet, but as we add more adapters and change more of it, it eventually becomes a completely new architecture. That's where the company is heading within software. Then you think about hardware - creating hardware that more closely mimics the structure and function of our biological networks. Neuromorphic computing has been around for a while but hasn't really scaled, and we think we can do better by using the biology directly as a framework.

And then finally, the north star of the company: real-time biological computing. That's using the actual brain cells to perform compute in real time as part of the inference circuit - either at the edge or in data centers. We're heading in that direction, but that's not production-ready now. We have a significant research effort underway, and all of the experiments we do to derive principles from biology are building the foundation to do this. In the next 5 to 10 years, the goal is to actually incorporate brain cells in the compute circuit. The biology, when it's in the actual circuit, gives us a window of opportunity that is massive - the brain is millions of times more efficient than silicon.

Seth Earley: The human brain has 86 billion neurons, each connecting to up to 10,000 others, with over 100 neurotransmitters operating as analog gradients. The complexity is orders of magnitude beyond anything we can build today. And yet, you're saying even capturing a tiny fraction of those principles is producing gains of this magnitude.

Alex Ksendzovsky: Exactly. In our adapter product today, we're not capturing nearly the complexity of how the brain operates. But the fact that just making it a tiny, tiny bit more biological reaches these massive gains - it's a testament to the brain's extraordinary efficiency. And the more of these principles and primitives we can derive and apply to create adapters and novel architectures, the more improvements we'll get in terms of both performance and efficiency.

Seth Earley: The energy consumption of AI is one of the most pressing infrastructure challenges of the next decade. Where does biological computing fit into that picture?

Alex Ksendzovsky: There was a recent Computing Research Association dinner I attended where we discussed a projected 10,000-fold increase in energy demand for AI compute over the next five years. The group identified three buckets of solutions. Novel architectures - which is where we live - are expected to contribute a 50-fold improvement in efficiency. Hardware more closely mimicking biology is another 50-fold. And novel energy sources like solar and wind account for the remaining gains.

Right now, today, we're at 4 to 5x in terms of using our adapters. I think 20-fold within the next year, 100-fold the year after. And when real-time biological computing is in the actual circuit, the ceiling is extremely massive. The brain is millions of times more efficient than silicon - so the potential is enormous.

Seth Earley: Using human brain cells in computing raises questions that go beyond engineering. What ethical and philosophical considerations does your work surface, and how are you thinking about those?

Alex Ksendzovsky: The main framing is around the difference between our biological cultures and the emergent properties of an actual brain. Our cultured neurons are fundamentally different from a human brain - they're hundreds of thousands of neurons compared to the brain's 86 billion, they're connected very differently, they lack the three-dimensional structure, and they're not enough to create emergent properties like sentience or consciousness. They're enough to pull out computational primitives and principles, and to perform specific computational tasks - but not enough to create anything resembling awareness.

As a company, we have no interest in creating anything even close to a three-dimensional structure of a human brain. We are setting our own guardrails, and we've already worked with bioethicists in the field to make sure that, as this develops, it doesn't create a concerning future. As the field starts to define neurophysiological biomarkers for things like sentience or consciousness, that will allow us to build even better-defined guardrails around what we're doing. Right now, the difference is so large that I'm not super worried, but we are actively monitoring it.

Seth Earley: Where do you see this going in 10 to 20 years?

Alex Ksendzovsky: In 10 to 20 years, I think we're going to have real-time biocompute in practical form. Robots operating with neural cultures that make them significantly more performant. Edge devices that have biological components enabling continuous, real-time learning. The challenges are real - we have to figure out how to maintain the cells, feed them, remove waste - which is why this is a 10 to 20 year problem, not a 2 year problem. But the trajectory is clear, and all of the AI environment will look much better when it's more biological.

Seth Earley: Alex, thank you so much for sharing your thoughts and your innovation and your approaches, and really giving us that glimpse into this frontier that most people are not aware of - but that is probably going to be one of the most consequential developments in the history of computing. Thank you so much.

Alex Ksendzovsky: Absolutely, thank you very much for having me on.

Seth Earley: And thank you to our audience for tuning in. Be sure to subscribe to the Earley AI Podcast for more conversations about how AI is shaping the future of business. See you next time.