What Comes After GPUs? Great Sky’s Bet on Brain-Like AI

25:18Jeff Shainline:But like, what's your sense of the landscape and what works the best on this architecture? Yeah, I think that's a really important question. So let me just rant a little bit about the transformer architecture for a second, because I think it provides some important context. So without going into all the details, the most important thing about the transformer is that it basically allows any input to modulate the attention given to any other input. That's the most general utility you could possibly ask out of some neural network type function. And that's what makes it so powerful. It ends up being able to train really well because it has this sort of like all-to-all ability for anything to influence anything else.

25:59But it's sort of like the strength is the weakness because that's also such overkill for a lot of problems that it's what leads to the dramatic energy consumption and really unsustainable training time and power consumption for training as well as inference. And the reason why we work with the transformer architecture today, the reason why the community does, is because it maps pretty well to existing hardware and it takes advantage of all those general hardware capabilities that GPUs give us. However, that ability to take any input and let it influence the attention given to any other input, there's a lot of great work from sort of the neuroscience or the neuro AI community showing that networks in the brain, dynamical circuits that are ubiquitous throughout the brain accomplish something very similar, but they don't have the same painful scaling, this quadratic scaling where if the input sequence gets twice as long, you need four times as much compute and four times as much memory.

27:00So these are typically recurrent neural networks. So recurrent neural networks were kind of where it all started. You can go back to Jeff Hinton's work in the 80s and 90s and parallel distributed processing. And all of that was sort of like intuitively, that's where AI should come from, right? It should be these recurrent neural networks. It turns out that those are less efficient to simulate on conventional hardware. And so even though they have the fundamental principles that we look for in artificial intelligence, brain-inspired computing, and they have a lot of functionality. They can work as well, if not better than transformers in a lot of cases.

27:39They don't map as well to the chips that we've had available for the last 20 plus years. And so they haven't gotten as much traction. That's exactly where we come in. And so there's a resurgence already happening. Leave Great Sky out of this. State space models are very popular. These are linear recurrent neural networks, which allows them to, in a way, map as well as possible onto a GPU so they can still be trained quite quickly. And so they've gotten a lot of attention, no pun intended there. But in our case, what we're doing is we're physically building circuits that implement recurrent neural networks.

28:17So you pay no temporal or energy penalty for connections that don't just go forward, but also go side to side and also go backwards. So you can have these very complicated networks, works much more like the architectures of the brain, where attention comes from the dynamics. It comes from this layer projecting its inputs to the next layer, and then the next layer feeding back and saying, based on what I know, these neurons should be more active, these ones should be less active. And then you get linear scaling where the input sequence gets twice as long, it's only twice as much compute. And it can be a much more efficient way to have a lot of different kinds of data sets, especially when the data is temporally varying.

28:57So your question of which model architectures, state space models map very well onto our hardware. We're spending a lot of attention on that right now. When I was talking about knowledge distillation a second ago, that's where we're focusing right now, taking trained state space models that have been trained using GPUs, but then mapping them to our hardware. So it's just blazingly fast inference. You also mentioned energy models. We've shown that all of the circuit equations, all the dynamics of our system can be derived from an energy model. So you can do this relaxation and things like equilibrium propagation are very natural on our hardware and more brain-inspired architectures where recurrence is just a part of it.

29:38You also have elaborate dendritic trees, dendrites in the brain, where it's not just a neuron and synapses, but these intermediate processing stages. We can implement those very straightforwardly with our circuits. So So the sky's the limit now. We think that the AI community has really been, okay, it's been enabled by existing chips. And of course, we wouldn't even have AI without all the tremendous progress from the semiconductor industry. But now you invent this new chip architecture, this new kind of hardware. We think there's a new renaissance coming where all these new architectures finally have their day in the sun because they have the platform on which they can be properly hosted.

30:19Wow.

30:19Jeff Shainline:Love it. So let's talk about your first generation chips a little. I understand they've shown major gains for workloads like video analysis, even saw a number of like 60 million frames per second somewhere that I was really curious about. And I'm wondering kind of what's being measured there? Because as a guy who watches the slow-mo guys on YouTube and their cameras, I'm like, that is amazing. Yeah, OK, so let's talk about our chips then. So we've been since the last 15 months, we've taped out four wafer generations, starting with just little prototypes of showing that we can get the basic circuits going.

30:55And now we're at networks of over a thousand parameters. OK, a thousand parameters is still not a billion parameter large language model. But, you know, that took us 15 months and less than four million dollars. That's pretty steady progress. And what those chips are doing is high speed inference. They're programmable chips, so you can put in whatever synaptic weights you want and run different kinds of models. With these generations, we've done things that are in the domain of sort of what I'd call speech processing. You could also consider it signal processing as well as video image processing.

31:31Let me zoom out for a second and talk briefly about where we really want to go, and then I'll come back and point to where we are on that roadmap. So with this hardware, one of the things that we think is most compelling is the ability to process multimodal data in a manner that's much more how a human would process it. So you see the world, you hear the world, and you can produce things like universal assistants that interact with humans in our native modalities. And you just talk to them and they provide you actionable insight, basically with no discernible latency. So that looks a lot like video processing, where you have a video stream and an audio stream, which brains are really good at.

32:13You have the visual cortex, the auditory cortex, they merge, and you have this seamless combination of those different data streams. So that's kind of our North Star. That's where we're building towards. So as we're at our initial product generations right now, we want to assess how quickly can we process video, how quickly can we process audio. Because when we just do calculations, sort of estimates of how well our hardware performs at that, say, trillion-plus parameter model where it's handling video, it's the number you said. It's 60 million video frames per second. And that's, I mean, the most simple way to think about that, you can go through all the math, but the simple way to think about it is that we build neurons and synapses like the ones in the brain, connected like the ones in the brain, but everything in our hardware happens about 2 million times faster than biology.

33:05So we humans watch videos at 30 frames per second, and our hardware watches it at 60 million frames per second. Everything gets accelerated by that factor of 2 million. So what we're showing with our early generations of chips is that we can take audio, spoken words, for example, accelerate all those audio frequencies by something like a factor of a million, process it through our chips, and the chip can tell you which digit was spoken. This is a common benchmark task, the Google speech commands. Similarly, we can show small digits, little numbers in a small image and classify them. What we've shown so far is 10 million frames per second.

33:48That's what's measured. We know we can get to 60 million frames per second. That extra factor of six is bottlenecked by the room temperature electronics on the rack, not the chip itself. So unfortunately, and we just need to, you know, fork out a little more dough. Exactly.

34:06Jeff Shainline:That's exactly right. Yep. Fair. Fair. So that's where we're at right now. We're handling these workloads that lead us toward that North Star. I'll also mention one other application with this thousand parameter accelerator that we're really excited about. We're working with the Department of Energy and Brookhaven National Lab. They care a lot about taking sensor data from a fusion reactor, a tokamak fusion reactor. You're basically looking at the plasma, streaming that sensor data into an AI model and predicting what the plasma is going to do. Is it going to become unstable? How can you steer it to maximize energy production?

34:43So we're partnering with Shinje Yu and his group at Brookhaven to build models that are able to stream that sensor data. It's coming in at a million data points a second, megahertz frequencies, and build a control system. So we're working on that demo right now. And we think in the slightly longer term, not that far in the future, we can build systems that take data from all the sensors around a fusion reactor, have this ability to perform really fast inference, and control that reactor so it never becomes unstable and never causes a disruption that leads to damage or a break in the power supply.

35:19That's one of the applications. And it turns out that with only a thousand parameters on the chip, if it's blazingly fast like this, it can start to handle that workload. And we think once we get to around a million parameters, this will be a very state of the art system for that fusion control application.

35:35Jeff Shainline:So what's what's between you and the million parameter version? Is it just money? Like what do you need to scale this? Well, you got to turn money into chips and that's where a foundry comes in. So we've been building a lot of our prototypes using the – so our company spun out of NIST. We were at the National Institute of Standards and Technology for about 12 years, and we developed this superconducting electronics process to make these circuits when we were there. We still use that process in that clean room to build some of our early prototypes. But now we're partnering with IMEC. They're a foundry in Belgium that specializes in sort of taking technologies that are right at ready to pop and bringing them to the final stages of maturity before they get transferred to another foundry, TSMC, Global Foundries, for large-scale, high-volume manufacturing.

36:26So IMEC is our manufacturing partner that brings tremendous lithographic capabilities beyond what we can do in our sort of research-grade cleanroom. And immediately by using their lithographic tools, we shrink our parameters by more than a factor of 100, which allows us to dramatically increase that parameter density. So that's one thing is just moving to a better foundry. I'll also briefly mention that with this technology, based on superconducting electronics, manufacturing stays relatively cheap because you don't need extreme ultraviolet lithography. To build the next generation of semiconductor chips, NVIDIA's next chips, you need EUV lithography that only ASML can provide and really only TSMC can produce at volumes, maybe also Samsung.

37:14With our hardware, superconducting circuits don't benefit from being shrunk beyond a certain point. So you use instead what's called deep UV lithography, which is state-of-the-art in like 2005 or something. And it's just much cheaper. And there's a lot of foundries in the U.S. that can do it. So from a manufacturing perspective.

37:33Jeff Shainline:That's good for scaling. Yes. It brings costs way down. Throughput can go way up. And a lot of different foundry partners can contribute. So as we look to scaling up, another thing that's really important to producing larger systems is going beyond what fits on a single chip and going to multi-chip modules interconnected with either superconducting interconnects or also obviously optics, fibers over longer distances. So we're building multi-chip modules that allow us to scale far beyond what can fit on a single die. So with just those two innovations, not even innovations, just steps forward for the company, manufacturing with iMac multi-chip modules, there's a clear path to state-of-the-art, very large models that can handle a tremendous variety of workloads.

38:19Jeff Shainline:So what does this look like in practice at the limit? Are you building this for data centers? Will there ever be like a consumer version where we have to have cooled in our house, like in our basements or something? Could this even be used in robotics? Like what's the use case here? Like when will we see this or how will we see this in its final form, let's say? So we're kind of focused on basically three sort of commercialization pathways. One, as I mentioned, is around fusion, but a little bit more generally. It's about deployments to high-value workloads where extreme high-speed inference really matters a lot.

39:01So the DOE is a really good example there. They have a few use cases where it really matters a lot that you can perform inference on like a sub-microsecond timescale. So you're streaming data from fusion sensors around a fusion reactor. Another example is particle colliders. There's the electron-ion collider being built at Brookhaven. Other examples like the Large Hadron Collider at CERN, where they're colliding these particles and they can't store the data from every single particle collision. It's just too much. It's terabytes per second. And so they have to take data from all these detectors around the sensors that they're using to detect the particles that have been produced in these collisions.

39:42and with really fast AI determine, does this look anomalous? Does this look interesting, like I should save it, or is it just another whatever collision that I've seen a trillion of them and I don't care? And that's another sort of like high-speed workload that our hardware is absolutely perfect for it. And then a third one that I'll mention is materials development. So imagine that you have a source of x-rays that's firing really frequently and it's interacting with the material and producing a diffraction pattern. and you need to reconstruct that diffraction pattern really rapidly. Same thing.

40:17You have a massive amount of data. You can't save it all. And if you could reconstruct in real time what that diffraction pattern tells you about the material, then you have much more uptime on your X-ray source. So those are some of the applications where we think large-scale, high-value deployments make a lot of sense. Okay, contrast that with another use case where it's not so much about going after a few high-value workloads, but it's really about exploring the breadth of applications. In this case, I'm talking about hosting cloud systems. So we can build a handful of systems that we host here at our facilities, and any user in the world can just remote in, run their job on our chip.

40:59And so this is like a cloud model. It's NeoCloud, whatever, you know, CoreWeave does similar stuff, Crusoe. These are companies that they take chips and they let anybody run an inference job on it or whatever. It could be a training workload, whatever it is. So that's something that we're working on as well. The good thing is that because our chips can handle inference so much faster than conventional chips, you don't need to build that many systems before you can host a lot of users. And the idea here is that, first of all, it's great for us as a company if we can get some revenue from these users.

41:31And second, we want to partner with them. We want to understand, hey, all you brilliant model developers out there who now have access to this new sandbox, what are you going to do? What do you think is exciting? Where is the creativity from the community going to lead this? And so cloud is the second thing that we're working on. And then the third thing is sort of a bucket, but I'll call it hyperscalers. And there's a number of workloads there. So, for example, in the near term, I mentioned that we're interested in this video comprehension workload in the long term. The very simplest, earliest stage of video comprehension where you have traction at a smaller model scale is content moderation.

42:12And so in this case, picture your YouTube. You're getting so many uploads every second that it's a computational challenge just to assess, is this video clean? Does it violate our policy and need to be blocked? And so right now, those video analysis models, they basically sample every third or every fifth frame. They don't really watch the video. They just pick out frames and look for objects or symbols. They separate out the audio and they turn it into a spectrogram and just look for signatures of harmful content. And we think we can do that better. We think we can handle all of those videos significantly faster while watching the whole thing and understanding it like a human would.

42:52And that's a workload that's worth billions of dollars to those companies. You've seen Meta recently lose in court because they allowed harmful content to be seen by children. And that's a really hard thing to avoid at the scale of Meta. And so they have to contract all these humans to watch the worst videos on Earth, and it's a damaging thing. And so we think we can do that.

43:13Jeff Shainline:It creates issues for the people who have to watch it because they have to see all this horrible stuff. So it would be actually – that would be a great job for AI to automate. Exactly. We want to get rid of jobs that people don't want anyway. As long as it doesn't train on it, I think that's important. Yeah. So you have to be careful about how the models are trained. So that's – and then the final thing I'll say, it's similar customers but advertising. Between Meta and Google, there's$500 billion a year in advertising revenue. And that's all moved into this more dynamical sequence-based kind of processing where they're looking at a long sequence of user behaviors, the ads they've seen, what they've clicked on, what they haven't.

43:52And it turns out that if you can increase the click-through rate by something like 1%, then that's worth like$5 billion a year to Meta. And I think we can do it in a way where it actually not only does it produce more user engagement, but it does it in a way where it's actually offering you the ads that are useful to you because we have a faster model that can take more context and feed it into the model without adding latency, actually still performing with less latency. it's going to be a better product. It's going to be something that, oh, wait a second. That's actually something I really need.

44:27I'm glad they showed me that. And now it's better for our customer, which is meta, and it's better for their customer, which is you. So those are some of the things that we're looking at where smaller models that we can achieve in the near term have tremendous value across society. That's really awesome.

44:46Jeff Shainline:Makes sense. So I saw that Semaphore reported that Great Sky plans to begin delivering chips sometime later this year. Is that correct? Well, right. So our first contract is with the Department of Energy. We already have the chip made, and we're testing it in our facility right now. And the contract states that we will deploy it at Brookhaven National Lab for this fusion analysis workload by the end of the calendar year. So we think we're on track to make that delivery, yes. Very cool. Very cool. What would be your like pie in the great sky dream for how quickly you get up and running with all the, you know, all the other things you want to do?

45:25Jeff Shainline:Like, let's say, you know, you make this first delivery this year, you want to 10x that, you know, in two years or three years, like, what's your goal, long term vision there? Okay, well, we want to get to 100 million parameter multi-chip modules in two to three years. Like through the Series A, at the end of that, we want to be building systems that we can deploy at a customer's facility in their data center. That's why this DOE initial engagement is really promising because we've already investigated what it takes to install our system in their data center. And now we know we don't need to change anything about it.

46:00We can take the power right from the ceiling, the cooling water right from the wall. We don't even need to weld a single pipe to get our system in their data center. So by the end of the Series A, two to three years from now, we want to deliver systems like that to a number of different customers. We want to host multiple models here so people can do workloads. As I was describing, things like content moderation, speech processing, also things in drug discovery, protein folding, molecular dynamics, simulations. And that's when we think we're up and running, when we're hosting our own models, letting customers access them, deploying on site.

46:37And then through the Series B, we want to get beyond billion parameter models. We think we can get to 10 billion parameter models in that time frame and really have a lot of commercial traction. Because once you realize that this costs less and performs better, it's going to catch on. It's not a scary thing. And people hear superconductors and that sounds new, but once they understand, don't worry about it. It's a box. You put your data in, you get your answer out way faster. That's what we're looking for is two to three years for our first real scale deployments and then getting beyond 10 billion parameter models two years after that.

47:13Wow. How much of this can be like manufactured using existing U.S. foundry capabilities versus is there anything in this that requires new processes to be able to manufacture? Okay, so let me sort of split a hair there. So in terms of the equipment and the manufacturing capabilities on sort of the foundry side, there's very little that needs to be changed. We use, like I was saying before, GDV lithography and all the same kinds of sputtering, deposition, patterning, etching, planarization. The whole workflow of a modern fab, we use the exact same manufacturing concepts. That doesn't mean it's exactly the same in every detail because we use different materials.

48:01So, for example, where the wiring layers that are common in a CMOS process are largely copper, we use niobium because niobium is the most ubiquitous superconductor. I think at an abstract level, it doesn't really matter that much because niobium, you sputter it just the same way. It's very easy to process. It's not an expensive material, so it doesn't add hardly anything to the cost of your wafers. But if you ask TSMC, hey, can you go run this process tomorrow? No, not tomorrow, because they have to dedicate their sputtering tools to a new material. You can't take one sputtering tool and do copper and niobium at production scale.

48:44And you take their etch tools, you need to slightly change the etch chemistry. All the processes are there. Our partner IMEC is running the process right now, but not at volume. so they can get to a few thousand wafers a year but we need to get to a few hundred thousand wafers a year and that means basically it's a commercial challenge it means you need to convince tsmc not that they can do the process they know they can do the process you need to convince them that there's a demand for the volume and that's really where the challenge is yeah because there's a there's a big fight right now trying to keep trying to get their capacity Absolutely.

49:19And yeah, it's really interesting. How has AI been used to help in the development of this technology or maybe the testing phase? Has it been a beneficial tool or was this moving before? I mean, yes, it was. So that's a great question. So we started this project in 2014. So for some patent reasons, I had to go back through my old notebooks and figure out when did I draw the first circuits about these superconducting neurons? And it was 2014. So AI was a thing back then, but it wasn't the thing it is now. And so I was more fascinated by ultimate limits of intelligence than like, hey, what's the commercial workload we're going to run?

50:03But so then it's evolved over time. And then, you know, you had the watershed moment with ChatGPT3 in like November of 22, was it? Yeah. Yeah. So that kind of obviously breathed a bunch of life into the whole field. And now I guess what I would say is our company, like I think most companies, most people now are using AI in a lot of different aspects of our life and our work. So most obviously, our algorithm developers that are writing a lot of code are using software coding tools. I won't say which ones, but you can guess. We're using AI to write a lot of our code. We try to be pretty careful because we want our stack to be totally bug-free from top to bottom, so we have to be really careful.

50:52to prompt it, say, write a software stack for this. We use it a little bit more judiciously. It's a pretty vital stage to ensure everything is exactly dialed in the way you want it.

51:05Jeff Shainline:I assume also it's very novel what you're writing software for. I don't imagine that's in the training data, right? Well, I mean, I guess honestly, I'm not the best one to answer that question, but I don't think it's all that novel. All of our code is written, all of the software that we're using to simulate our systems, to pipe data into it. It's all done in, well, it's largely done in PyTorch for the simulations, Jax, back end. So it's like pretty off the shelf stuff. We're using a lot of FPGAs to provide programming signals. And so there's a lot of known C code for how do you interface with FPGAs.

51:43So a lot of what we're doing is pretty bread and butter. The novelty is really on the chip that you're talking to. But you're providing just square pulses for changing synaptic weights and waveforms for putting data into the system. So there's a fair amount of knowledge base around that. I guess I would say beyond using AI in our coding, I personally use AI as sort of just like a thought partner and thinking through maybe different commercialization strategies and things like that where market analysis and doing research around different use cases. But yeah, I think we're using AI in pretty conventional ways at this point.

52:20Also, a lot of like, hey, we're building a neuroscience inspired chip. Help us formulate the analogies between AI, neuroscience in a way that the public can understand communications, that kind of thing.

52:34Jeff Shainline:Awesome. Very cool. We'll have two more questions for you. So the first one is the memory side of this, right? where a lot of ink was spelled earlier this year on the memory shortage and how that's become a big bottleneck now for deploying AI systems, especially with agents. And I was wondering if you could talk just a little bit more before we wrap here about how your system is helping with the memory side of things. And then one more after that. Okay, sure. So we approach memory in a completely different way. I mean, okay, in computing, there's lots of different kinds of memory. There's memory hierarchy.

53:09And so it sort of depends on exactly which aspect you're talking about. But I think the most important one from the perspective of performance is the memory that's closest to the processing. In conventional chips, that's SRAM and it's sitting on the same chip. We take a completely different approach. There's no such thing as a bank of RAM in our hardware. We physically build a circuit right next to the synaptic processing circuit that stores the synaptic weight. We call it the memory cell, and it basically just stores a superconducting current. The value of that current can be ratcheted up, ratcheted down.

53:44We've shown that we can do beyond 8-bit. We think sort of 4-bit is a sweet spot, but it doesn't really matter. We can do whatever value the application requires. And those memory cells are made in the exact same processing layers using the same niobium wiring and the same Joseph's injunctions as all the computational circuitry. So there's not this like tension between a process required for the compute and a process required for the memory. They're one in the same here. And so that means that the primary data movement bottleneck of AI, where you're shuffling information between the processor and the SRAM, it's not that we did better at that.

54:27We just got rid of it completely. It's not present because we store synaptic weights where they're processed. So that's the most important thing. And then in terms of the rest of the memory hierarchy, yes, imagine your YouTube and you need to store, I don't know, a million hours of video before it gets piped into our system. That's going to be in RAM and DRAM, not SRAM. And so, of course, we still need some digital storage and things like that. But But the most important thing is the elimination of that data movement bottleneck between processor and memory. And that's just completely gone. And it turns out that moving to superconductors is the key enabling step there.

55:05It's really hard. I mean, people have been wanting to do this since the beginning, integrate processing and memory. And silicon really forces you to choose. And the digital architecture, this von Neumann architecture, really deeply embeds this principle that they're separate. But when you go to a brain-inspired architecture, which is much more matched to superconducting electronics, and the materials and processing requirements of superconducting electronics just naturally give you co-location and memory with processing, and that leads to application-level things like continuous learning, the ability to not have a training phase followed by an inference phase, but continual learning as inference occurs, which makes things much more user-friendly and able to understand you deeply as an individual eventually as well.

55:51I think is the result of that.

55:53Jeff Shainline:And then what would be the limit, I guess, is just like scaling to just more chips, basically. Right. You mean the limit on like the memory and like how much, because like, I'm so used to thinking of the limit on what type of model I can run in terms of like the amount of GPU I have, right? So like, I'm like, I know that if I have X amount, you know, like 90 gigabytes of GPU capacity, I know I can run XYZ model. Otherwise, I'm using it over the cloud. You know what I'm talking about? Yeah, exactly. So, I mean, it's the same here. You're limited by, we think in terms of the number of parameters, basically.

56:30There's sort of a mapping between the two. And in our case, there's the fact that we have eliminated the data movement bottleneck to RAM and the interconnectivity challenge. Solving that with photonic interconnects means that as you continue to scale up, You add more chips to your system, ultimately entire wafers stacked. You can use optics to communicate between wafers over fiber optics over longer distances. You don't run into pinch points because you have this high fan out optical communication. Every neuron can communicate to many thousands of others at the speed of light, regardless of how far away they are.

57:06That means that the ultimate limits are eventually set by how long it takes light to reach a certain destination. And so it's very large. At every stage, we're not limited. We're far from the physical limits of this hardware. It's going to be basically manufacturing limited and how fast can we build. We want to build systems that have 20 wafers interconnected. That's still small compared to a data center filled with GB200s. But when we get there, we have beyond trillion parameter models, which are profoundly useful, especially when they're going a couple million times faster than a human brain.

57:44And we don't stop there. We go to much larger systems. As I said at the beginning, the cryogenic requirements are already met. There's already off-the-shelf systems you can build. But it's really that. We think there's another path to exponential scaling. Moore's Law was all about lithographic improvements that enabled you to shrink transistors. In our case, exponential scaling is enabled by the ability to double the amount of compute in every product generation without running into painful power consumption bottlenecks or data movement bottlenecks. And so you can keep adding more compute to the system, keep doubling, keep doubling, have exponential growth in a completely different way.

58:25But it's enabled by superconducting circuits for computation and optical communication at the single photon level. Amazing.

58:32Jeff Shainline:Awesome. Wow. I have so many more questions, but Corey, I'll let you wrap it. Jeff, thank you so much for joining us today. This has been really interesting and it's exciting what you're doing. We always love talking to someone who's kind of stepping outside and not afraid to break the paradigm. Yeah. All right. Thanks for having me. I really appreciate it. Absolutely. Where can people go to learn more about what you're doing and keep up, follow Great Sky? Well, you can follow us on LinkedIn, X, Blue Sky, our website, greatsky.ai. Should be able to find us on a Google search. Excellent. Very cool.

59:10If you're watching and you haven't yet, please take just a moment to like and subscribe above or below, actually, I think it is. Either way, we really appreciate it. It helps us continue to bring you more fascinating guests like Jeff here. Also, if you haven't yet, please pop by the Neuron.ai and sign up for our daily newsletter. It's read every morning by 700 ,000-ish people. So we'd love to count you as one of those and have you as part of our community. But I think that's it for today. Farewell for now, humans. Farewell for now, humans.