Video Transcript
probabistic chip. There was three probabistic bits or three pro, you know, thermodynamic neurons, whatever you want to call them. A couple designs and then and then the little snakes you see are like control lines and to control the parameters, the tilt and so on. Um, and uh, you know, that first chip, you know, everything worked out. We have some cool experiments. And really, it's like if you had a different supply chain, you know, you might want to do things in superconductors. Of course, like the problem is you have to to cool it. Ideally, you could use a much smaller fridge, but, you know, it's not it's not as dense as as we like it, right? This chip is basically microscopic. could see the, you know, the details like by the naked eye. And so what we had to do, well, so okay, the reason we gave a uh was that, you know, people are like there's like these crazy like this crazy space-like race to have the most efficient or the biggest transformer uh possible. And our theory was like if we open source it, then maybe the governments will race to build a crazy superconducting supercomputer to have the most energy efficient transformer. And so, you know, we mapped how to break down a trans whole pause chat chat pause. [music] Okay. Did Did he just say that their whole objective here was, hey, maybe if we build this, it'll force governments into making a a an AI supercomputer, a super inefficient AI supercomput. Hold up, chat. Um, pause. Are you just such a good person that you just decided that, oh, we're going to force the governments to make one of these things? I don't know if I like this guy more now or I'm just super suspicious of him now. Uh, I want to let's repeat that. Let's go back like 10 seconds that maybe the governments will race to build a crazy superconducting supercomputer to have the most energy efficient transformer. And so, you know, we met >> I'm all about that chat. Yeah, let's get our governments to build a super crazy superconducting supercomputer. I gotta go ahead and just say I in in the back of my mind I'm getting an itch. It feels like they might already have one. And the other thing I want to point out is that I'm not convinced that what these guys are doing is novel either. I am convinced that they're doing something that's novel to them. Yes, but I'm not convinced that if this is real and this is like a more efficient way of measuring the ether using a computer to measure the ether, then the government's already got it. Loheed Martin's already figured this out. In fact, what I really want to know is I want to have Charles Chase in the room with this guy. I want Charles Chase in the room with this guy and I just want to watch them. I want to be like Steve Irwin Crikey. It's a It's a rogue. We've got a rogue Loheed Martin fellow and he's talking to the extropic engineer about thermodynamics, stochastic thermodynamics. Let's see how this plays out. Let's see who will win this battle. Right? That's how I feel about this. I just want to watch them in the room together and I just want to glean information. What a rare sight. Charles has gone and asked him does he think that the thermal fluctuations are related to zero point energy? mapped how to break down a transformer and its operations into superconducting thermalization physics and then we did some benchmarks. I think those lines are out of order. Um it's like the legends messed up but but you know the problem this is just simulations right is it's you scale to a freaking room size computer of this type. Uh it would be crazy, right? Like the energy efficiency gains are like insane. They're like 10 million x plus right? 100 million x plus. So so that's cool. Um but obviously it's not tractable cuz you got to like super cool it and you wouldn't be able to put in your phone or something. Uh and so uh basically we had to figure out how to port this technology and what we learned to room temperature uh and in silicon uh in order to make sure we could deploy it at scale and mass manufacture it. And uh you know like a year ago put out this plot and and then we were manufacturing our chips and then you know as we was reported and and wired we you know we we we successfully uh created a bunch of new primitives in silicon. There's proistic bits there's other silicon primitives in there. Uh, and basically, uh, I'll talk about it, but we're going to give people access to this chip in a dev kit, uh, uh, very soon. Uh, I'll get back to that in a second. Um, I don't know if there's a movie, but you know, you can see the the proistic bit. >> I think that dev kit already happened. That's that weird box we were looking at >> signal there. It's basically we control how much time that the signal spends in zero or one. And we have like pretty good control there and like sort of reliable manufacturing. That's kind of the that's kind of the hard part. And if you get your components right, then you can scale. Of course, like none of this matters if we can't scale it, right? And so not only do we have to scale the hardware, but we got to scale the models, right? And so how are we going to scale the hardware? So this is like uh yeah, I mean this is our game plan. So yeah, so thousandx year over year uh number of degrees of freedom. So we start off with three proistic neurons or proistic bits. We have 300 components uh proistic degrees of freedom on the on the new chip and then uh we're designing a chip that's coming out next year with uh in the order of millions, right? And so yeah. Yeah, we're scaling this Uh yeah, thank you. Um so yeah uh no no no tweeting this yet. Uh yeah, eyes only please don't no leaking. Uh [snorts] and uh you got to you got to also scale the algorithms, right? So you know there's a there's a paper from Google called denoising recovery likelihood. It's like chaining EBMs to uh to replace the fusion models and using like 100 times less steps, right? And so uh basically we like supercharged that algorithm for our hardware. And um so the next slide you're going to see 1,000th of this chip. This next chip we simulated it running the algorithm and um you could basically do what we call denoising thermodynamic models uh in a way that's much more efficient. DDPM is like a diffusion model. Uh the circles are running on GPU and the axes are running on our computer. Uh again this is in simulation. Um but essentially it's uh you know you know if you run a cheap algorithm it's a thousandx better. If you run an expensive algorithm like 100 millionx more energy efficient and this is sort of the number of uh jewels you need you know how much energy per sample and so you could have you know more generation more throughput in parallel or so basically what he's saying is that because the old system of doing generative AI the way the old system of generative AI works is it's almost like uh what did they say transformation or something like that where it has to go through these chains one at a time right and this is what makes it so inefficient makes it take so long. It makes it take a huge amount of energy. Their system can skip all that. They can skip all that. So that huge chain you saw, they skip all that. And so when you look at these numbers here on the screen, they're talking about efficiency levels a million times more efficient in the energy consumption. So we look at the energy consumption down here and we see like under 10 uh jewels per sample and it goes all the way down to 10 the minus 7. So every minus every minus you see increase adds another zero before or after the decimal point. So 0.000000. So minus 7 has minus has seven zeros. 0.7 zeros and then one. So the amount of energy these things are using is significantly less than current architecture. This is the kind of thing where I look at this and I go, "Wow, if that's real, if that's true, then that is a that is such a huge amount that there's definitely value there. There's definitely value there, we still are going to need fusion. And even if those numbers aren't accurate, even if it's only half correct, it's still huge. My bigger question would be, is there some kind of larger reason why this won't work? like is the microchip going to break down after it's been used for 15 minutes or something like that? You know, is there some kind of like what's the catch? What's the catch? Is there a red flag here? Because otherwise I look at this and I'm going this seems like a no-brainer or overclocking time using more energy. But um but yeah, this this was I think for CFAR. Um, and so, uh, yeah, I mean, this is pretty exciting. Like, we're designing this and we're going to, you know, have it next year, uh, from the manufacturer and you could run some pretty pretty interesting algorithms on it. And so, you know, I'm I'm really excited to, um, you know, finally be able to give people early access to the software and some of the hardware. Um, so you know, people ask us like when when the dev kit will be ready and when they'll be able to like experiment with the dynamic computers and essentially it's this summer. So summer 2025 >> and uh got the chip right here. So we're packaging it into devkit and essentially you're going to be able to have it on your desk and just plug it in your computer with Ethernet and start hacking called a Bolson distribution which is this exponential distribution there that is normalized. Right? So you what we're doing is programmable bowls that have parameters of how you shape the bowl and we're letting electrons be our bouncy balls hop around this highdimensional landscape and then we get samples from the computer 100 times less steps. So there's a paper uh diffus deninoising recovery likelihood with EBMS uh it's a paper by Google couple years ago uh but you could do basically diffusion models with a chain of deninoising EBMs right so this machine can do basically everything that everything like diffusion models can do but better so that's pretty great that's a lot of applications the next slide is very fresh just specially cooked for today it's very preliminary so Yeah, you know, wait for the paper. >> Thermo superconducting thermo Aai hardware. Our first prototype, our most macroscopic thermodynamic computer. So, this is our superconducting lab. This is as macroscopic as you can make a thermamic computer. Anybody claiming they can do a breadboard in their garage that's a thermic computer injected synthetic noise and it's a LAR. So, we had to go to great lengths to make macros as microscopic of a prototype as we could. This chip is about as big as your thumbnail, but the features are visible. And we have a whole cryogenic lab to fabricate and test these devices. These use superconducting Joseph and junctions to create again general energy functions over the continuum. >> Chat, he said the words. That's all I wanted to hear. Chat superconducting Joseph and junctions yatsi. So this is why when the man's talking about it's not a it's not a quantum computer. I'm like what? But quantum, we just found out that quantum computer is just two Joseph injunctions built into a squid connected to a capacitor. That's all a cubit is. So when you're telling me, oh, my microchip's totally not a quantum computer. It just it uses all the exact same functionality, the exact same microchip design. That's where I'm like, uh, okay. I mean, technically, you're doing something different, but is it really not a quantum system? Okay. >> Tip is about as big as your thumbnail, but the features are visible. And we have a whole cryogenetic lab to fabricate and test these devices. These use superconducting Josephus and junctions to create again general energy functions over the continuum. And this is there's a whole movie online. You can check it out. We showed our fab and lab and how we've manufactured these devices. >> But essentially where you to scale a massive superconducting supercomput, you can have the most energy efficient transformer you can build. And so that's probably for us this was a stepping stone. This was like our breadboard. >> So a superconducting computer is the most efficient system you can possibly build. Why? Because superc conductivity has no resistance. No resistance means there's no energy loss due to heat. That's as efficient as it gets, chat. Well, not quite, but as close as you can get from a standard thermodynamic standpoint. Prototypes of this fab is in Sherbrook. Uh on the left is our chip. So you know we have 25 patents in counting on this on this paradigm. So come talk to us if you're interested in scaling this. So that's our first thermodynamic computer. This is a programmable energy based model device. We're actually I'm announcing today that we're going to be open sourcing the blueprint for the software and hardware. So how we did how we can do all sorts of algorithms and how we can map algorithms onto super. Sweet. Okay. One last thing we're going to So there's a brand new video they played. So that was what I wanted. There's a little bit more about the wobbly physics and utilizing brownie in motion like 3 minutes into the video. Uh but that's fine. I think I kind of explained it. So this is I think their lead engineer. I don't think this is the founder, but their lead engineer, the other the one the other guy basically. And he kind of explains some of the science behind what they're doing here. Like what is Extropic doing? What this is what we're trying to understand. What what is this actually doing? >> Leverage that hardware efficiently to do something similar to chat GBT or Midjourney, right? So the neural networks we have today like transformers were really built to leverage specific things about GPUs. Um and if you built so if you build a new type of hardware, you also have to build a new type of algorithm to go along with it. And so where are we now? Well, over the last two years, uh we've kind of built a version zero of everything I just told you about. We've built chips that let us test these new probabilistic circuits in the lab, right? So it's not just theory. We've actually built them and validated that they work how we think they should, which was a lot harder than you'd think because, you know, all circuits have noise, but it's a whole another matter to actually use that noise to do something useful reliably. We've integrated those chips into simple systems. So, you know, we package our chips, built boards around them, um, built interfaces from the talk FPGAs and kind of do hybrid algorithms. And then we've also done a bunch of work on the machine learning side to figure out how to leverage this new type of processor to do something useful in generative AI. So, all that's been done over the last two years. And what that gets us is this plot here which shows that simulations of a chip we're building now could be around 10,000 times more efficient than a VE running on a GPU on some simple generative AI benchmark. Now great okay so basically we did a bunch of research. Where are we going? Well now it's all about scale. We've built these simple systems. So now we have to scale up both the hardware and the algorithms to raise the capabilities of our systems to be more comparable to what you can do with LLM today. Right now everything we've done so far has been extremely scale small scale. The company's only around 15 people and those people are split between integrated circuit design, statistical physics research and machine learning. Now let's get a little bit more specific about what each of those new things is. Right? So what do I mean by a new type of integrated circuit processor? Well again an integrated circuit is just a chip made out of a bunch of transistors that you can get from TSMC. So there's nothing exotic like quantum computing or photonics. The way it is different from what exists today is mostly architecturally. So a GPU is essentially a giant array of floatingoint multiple accumulate units that get orchestrated. So he's saying that instead of it being a quantum chip using quantum mechanics, it's more like a standard GPU standard processor, but it uses a different mechanism. Here we go. >> Created by some central controller to implement things like matrix multiplication for deep learning. That's great for doing neural networks. Our new type of processor is called a thermodynamic sampling unit. A thermodynamic sampling unit is extremely different from a GPU in that instead of being a large array of floatingoint multiple accumulate units, it's actually a giant array of sampling cells. Right? So a thermodynamic sampling unit looks like this picture on the right where you have a bunch of cells and within each cell you have some kind of specialized sampling circuitry that implements ultraefficient random number generation. You have some circuitry that computes the parameters of that random number generation process such as the bias of flipping a coin. And you have a register that stores the state of the program. And so when you roll out a program on a TSU, what basically happens is all of these cells talk to each other and orchestrate some kind of sampling procedure to uh sample from some computationally useful probability distribution. What do I mean by probabistic circuit? Well, traditional circuits compute functions, right? So, if you have an okay, I think we'll probably call it there. It's already getting a little bit too deep for me, but [clears throat] what he's saying is that instead of actively inputting our data, instead of having our input be as he describes in that uh screenshot matrix and a vector, he's saying instead we just have these sensor units. These sensor units that are sampling units, they're just sampling the data that's already out there. So it does seem like they're quite literally just measuring the ether. Now in this case it's not the zero point energy ether. They're measuring the thermal fluctuations that are occurring and then they're using that as their resource instead of creating their own binary ones and zeros through an input. So if you take the input away and you just sample the parameters and the data that's already there, then of course that becomes a lot more efficient. So I think that's probably the most important takeaway if we if we really go high level on this and say what is the big takeaway from this entropic entrop entropic right extropic chip is that it is sampling energy that's already out there instead of inserting energy to create a result right sampling the energy it's already out there instead of inputting energy to create a specific result my last thought of the night on this is if I was developing a sentient AI. An AI that meets the three criteria that Salvatore Pais mentioned the what do he called the the triarchy of sentience. What did he say? He said number one number one is you need to have energy. You need energy to be able to pull this off. You need the processing capability to pull this to this stuff off. I would say okay, we've got that met. Number two, we need the ability to have good data, good vector analysis, I think they call it. We need to be able to pull the best sources. Okay, that's also pretty easy. The third one is the hardest one. Creativity. Ultimately, when we talk about creativity, I equate that to randomness. Can a computer produce true randomness? When a computer can produce true randomness instead of the fake random Monte Carlo thing they do right now, that could be seen as a form of creativity, unique inspiration, new content that's original. I [clears throat] think that sentient AI will only come from microchips that sample the ether. If our consciousness exists in the ether, if we are just receivers and we're tuning in to our consciousness, then these kinds of microchips would be exactly what's needed for a sentient computer to also tap into that consciousness in my opinion.