The Biological Computing Co. teams up with AWS to improve data center AI with brain-cell research

The Biological Computing Company (TBC) struck a multi-pronged partnership with Amazon Web Services to deploy its brain-cell-derived algorithms for data center AI, with TBC cofounder Alex Ksendzovsky saying the approach makes video generation five times faster and less costly. TBC will offer its model on AWS Marketplace and SageMaker, letting customers buy tokens on hardware they don't own or run TBC's API on their own hardware, and the company plans a soft launch of a public video-generation demo. The algorithms were discovered through research using neurons in petri dishes and are applied to open source state-of-the-art video generation models.

Become a member of GB MAX to gain exclusive access to the industry and to the most influential global B2B leadership community in the business of gaming, entertainment, and tech. Join now https://go.gamesbeat.com/gb-max/ and also get a VIP ticket to GamesBeat Next Nov 2-3, SF . The computers of The Biological Computing Company https://www.tbc.co are in petri dishes. That’s because they use brain cells to work on computing problems, and then the company extracts new algorithms to provide conventional servers for data centers with new ways of solving problems faster. After all, nothing is as remarkable as the human brain when it comes to solving problems, even in the age of AI, according to cofounders Alex Ksendzovsky and Jonathan Pomeraniec, the founders of TBC. In an interview with GamesBeat, they now say they have been able to make enough progress with this approach to strike a multi-pronged partnership with Amazon Web Services to use the brain-based algorithms in everyday data center computing. see our story here for deep dive into how the tech works https://gamesbeat.com/the-biological-computing-company-will-speed-ai-with-human-brain-cell-research/ . “We have been working towards build essentially inference optimization,” said Ksendzovsky. “We’ve been working towards using the dish of neurons as a way to understand neuroscience primitives and how information is processed by biology, and then incorporate that in software into video generation.” This means that TBC’s algorithms, discovered through research involving brain cells in petri dishes, has gone from neurons in a lab to a scalable software product in the cloud that anybody in the world can use at any time. One of the things that these algorithms are good at handling are video generation tasks, and so they’re being used with open source, state-of-the-art video generation models, which allows TBC to make video generation five times faster and less costly as well. “That happens because the optimization strategy that we use maintains the same quality because of our adapter of the video that is output. But we can use very aggressive optimization techniques to increase the frames per second or the throughput,” Ksendzovsky said. “And so what you get is a lot more frames being produced per second, and when you’re paying for minutes or seconds of GPU time, you can produce more video. You can apply that towards being less costly.” Ksendzovsky added, “The other use case with upside here is the time to generate a video, the latency. Again, because you’re generating more frames per second, you can get more video, which is great because that’s huge.” That’s what the technology offers, he said. TBC has applied its adapter system into a video generation model “We are about to soft launch a demo of that, where you can generate your own videos with it,” Ksendzovsky said. “You’ll be able to see what the reduction and efficiency improvement is compared to baseline of the model. So again, we apply it to open source. And then you’ll be able to take the same image or prompt and generate a video. You’ll be able to see that ours takes way less time than a lot of the state-of-the-art models. That’ll be pretty apparent with the demo. And again, the point is the quality of the video that you have should be the same or better with ours.” Why is this valuable to Amazon? The reason this is valuable to Amazon is a couple of reasons, Ksendzovsky said. First, TBC will be deploying its model on the on Amazon Marketplace to their customers. Their customers will now get the offering of better, faster, cheaper video generation. “We’re going to be on SageMaker, which offers the opportunity for customers not only to go and buy tokens from hardware that they don’t own, it also will allow the opportunity for customers to use their own hardware and our API,” Ksendzovsky said. He added, “It’s just an API into our model through their own hardware. It’s an alternative way for customers to use our performant video gen model, and then that’s a major vertical we’re finding with Amazon and our partnership.” The other use case is specialized hardware. At this point, I asked Ksendzovsky to take a step back and how his team gets from petri dishes of brain cells to hardware in a data center generating video. He replied that it’s not like the brain cells make a leap to the data center. Rather, he said, “The tangible thing that comes out of the dish is a mathematical model that recapitulates how they process pixels, and that comes in as a additional set of layers that get ported over into the transformer architecture.” How does it get plugged into the transformer architecture? That was a bit harder to understand, but TBC is looking at spatial temporal representations of its research and so the thing that tangibly happens is it takes the way that neurons represent pixels and does something with it. “We create a mathematical model that ports that over as code, and that gets plugged in as this additional set of trainable parameters,” he said. This is not a lot of data or code. It’s about 0.1% to 0.3 percent of the size of the model, Ksendzovsky said. “And that influences the the predicted frames,” he said. “That influences because as your latent representation of your images comes through the model. The bottom line is that latent representation of your image that goes through the model to predict the next frame gets transformed by this this additional set of parameters.” That makes the model, which is an AI computer architecture, into something that is a “bit more biological,” Ksendzovsky said. “Ultimately, what that allows us to do is combine that with very aggressive optimization strategies, both at the CUDA level, or at the numerical level, and at the level of the model, which then gains us these massive efficiency gains without reduction of quality,” Ksendzovsky said. “That’s like the big name of the game for us, because you could do similar things, but you get significant reductions in quality without our adapter,” Pomeraniec said. “And so that’s how technically it works. Again, we have a lot of these adapters. This is what we call them. They’re just a lot of ideas that come from the dish. Many of them fail. Many of them move on to our leaderboard.” He said the company has a leaderboard of about 30 or 40 best algorithms from the research. “We’re constantly improving them, constantly scaling the adapter product, and applying them towards multiple models,” Ksendzovsky said. “We’ve applied it in a bunch of AI models.” Again, this doesn’t mean there are petri dishes with brain cells in the data centers. “The biology stays in the lab,” Pomeraniec said. “What the customers see is an improved video generation model. So the biology is very much the process what’s deployed and what’s shipped to software,” he said. TBC sends over an API into its model. The company doesn’t sell its model. “We sell access. And so the way customers get it is through a marketplace, and our partnership with Amazon allows that marketplace. Now it lives on AWS’s GPUs,” Ksendzovsky said. As for the specialized hardware, Ksendzovsky said the company is training models as well, and access to the video gen models is what customers get. Where TBC is going next Ksendzovsky said TBC is adapting its algorithms to other open source video gen models next. “We are in talks as we incorporate our adapter products into open source models because we can download the weights,” Ksendzovsky said. “And the licenses are allowing us to use them commercially. There’s also potential for collaboration here with other video gen companies that don’t have our optimization strategies. What makes us unique as an inference optimization company is the fact that we have these brain cells that tell us how to do things. That’s the thing that makes us unique that no one else has. And so, in the context of other video gen model companies for sure we’re open and we’re in talks of collaborating with them to take their closed source models and building adapters for them.” As far as what it has proven so far, Ksendzovsky said there are a few things that arise from the AWS partnership. “We’re proving three things. First of all, there’s a such a thing as applying by compute, right? I think this is going to be the very first time anyone’s gone from brain cells in a dish to a usable model that people are paying for right,” Ksendzovsky said. Another thing is moving from theory into thinking insights into the field of AI. “I think we’re proving that like just. By taking a small sliver of how brain cells process information and applying that towards current AI systems, we can get these pretty miraculous results, and this is just like the tip of the iceberg,” Ksendzovsky said. “Our AI systems become so efficient that just taking a small single algorithm of how brain cells in a dish process information and applying that shows pretty significant results.” And the Amazon partnership shows that there’s a massive market for this, Ksendzovsky said. “It’s not a surprise that the largest data center company in the world wants to work with us because I think there’s a ton of demand,” Ksendzovsky said. “I think they also realize that computers have become inefficient, and I think that there’s a lot of interest in a huge market for inference optimization.” He said there is a need for innovative ways to achieve that. “We’re the only company in the world that’s using real biology, going to the actual source, to get to optimized AI inference to make it more efficient. And so, not only is there a need, but these hyperscalers and everybody else are competing for sustainable competitive advantage, which I think is what we offer,” PomeraniecPomeraniec,d said. He added, “And by the way, It’s not like we have one adapter and we’re done. Every two weeks, our team has a novel neuroscience principle or adapter that comes out of our dish that makes its way up on the leaderboard. We’re at the tip of the iceberg. Now we’ve created the infrastructure to be able to to to really optimize and continue to improve.” Ah, yeah, interesting. Did that leaderboard comment mean something in particular for how you guys work? Like your most your most popular kinds of results. Regarding the addition to the model size, Ksendzovsky said, “The point we’re trying to make there is, compared to the base architecture, we are seeing these massive improvements and gains with very small amounts of additional parameters.” TBC will demo its model to those interested and offer credits to folks to try it out. Ksendzovsky said there are growing use cases for startups in world models. “The simulations are very long, and so trying to model 3D spatial relationships and physical reasoning beyond language, and that’s a growing use case for them, and I think there’s a lot of value, not just for Amazon, but for other hyperscalers that have their own purpose-built specialty inference chips,” Pomeraniec said. “It’s not a one-size-fits-all approach, but like, how do we make these chips purpose-built to make use cases way more efficient? And so then we get a compounding multiplicative effect of our efficiencies on top of their efficiencies.” TBC now has 35 or so people, with some in San Francisco and others elsewhere. And by the way, getting game-changing inference gains, we’re making token economics much more valuable.