Computing Evolves: P-bits Cut AI Energy Costs

Welcome to Heliox, where we illuminate tomorrow's technologies with warmth and wonder.

Today we're exploring a fascinating shift in how computers think - moving from the rigid world of ones and zeros to something that looks more like the beautiful uncertainty of nature itself. Just as a butterfly emerges from its chrysalis, traditional computing is transforming into something more organic, more efficient, and perhaps more profound.

Join us as we delve into probabilistic computing - where randomness becomes a feature, not a bug. We'll discover how this elegant approach mirrors patterns we see in everything from falling leaves to firing neurons, and how it might just revolutionize AI while using a fraction of the energy.

Whether you're a tech enthusiast or simply curious about where technology is heading, this episode offers a gentle introduction to a revolutionary idea that's both powerful and poetic in its simplicity.

So settle in, and let's explore how the future of computing might be less about precision and more about possibility.

New Computing Breakthrough achieves
100 Million Times GPU Performance
00:00 - Probabilistic Computing
9:24 - Thermodynamic Computing
https://www.youtube.com/watch?v=hJUHrrihzOQ

Thermodynamic Computing System for AI Applications
https://arxiv.org/html/2312.04836v1

This is Heliox: Where Evidence Meets Empathy

Independent, moderated, timely, deep, gentle, clinical, global, and community conversations about things that matter.  Breathe Easy, we go deep and lightly surface the big ideas.

Thanks for listening today!

Four recurring narratives underlie every episode: boundary dissolution, adaptive complexity, embodied knowledge, and quantum-like uncertainty. These aren’t just philosophical musings but frameworks for understanding our modern world. 

We hope you continue exploring our other podcasts, responding to the content, and checking out our related articles on the Heliox Podcast on Substack. 

Transcript

0:24

Welcome to our deep dive. And this time we're diving into something, well, you could call it a revolution. It's probabilistic computing. And we've got a lot to unpack today. We've got this groundbreaking research paper. It's from UC Santa Barbara. And then we've got some insights from the Anastasi in Tech blog. And get this, there's even a company light paper. Yeah. This company is called Extropic. And they lay out this bold vision for the future of this whole thing. And there's even a company called Normal Computing. They're already building tools around this. So our mission today is to help you not just understand what this tech is, but really why it matters. Yeah, it's fascinating. You know, when you think about it, the timing of this, this emergence of probabilistic computing, it's like we're hitting these limits, the limits of traditional computing. Yeah. Just when we need computing power the most. Oh, yeah. When you look at all these sources, they all point to the same problem, the way we've been computing. With these zeros and ones, it's running into limitations. Absolutely. I mean, Moore's law, that prediction about transistors doubling, it's slowing down. We're reaching the physical limits. You know, how small can we make these transistors? Right. And on top of that, the energy demands of AI. They're just going through the roof. Here's the thing. So many real world problems, they're inherently probabilistic. Right. Like if you think about weather forecasting or financial markets, even understanding how our own brains work, these systems are about probability, not just certainty. In traditional computers, they're deterministic. They're really good at tasks with clear steps. Yeah. And predictable outcomes. Right. But when we try to tackle probabilistic problems, we're kind of forcing these computers to simulate randomness, which is really inefficient. And so that's where I hit this research from UC Santa Barbara. And it was really an aha moment for me. They used these probabilistic bits, or p bits, p bits, to crane a deep generative AI model. And they did it with this incredible efficiency. And the key difference with a p bit is that it's not fixed. At 0 or 1, it fluctuates between those states. Based on probability, you can kind of imagine it like-- Yeah.--like a coin flip. But you can control the odds of heads or tails. That's interesting. Yeah. And the Anastasi InTech blog had a great analogy. They compared it to the shift from analog to digital. But this time, we're embracing the noise instead of trying to get rid of it. Yeah, for decades, we've been trying to get this perfect precision in computing. And now, with probabilistic computing, we're harnessing that randomness. It's wild. So how does this actually work? I'll admit, some of the technical details in these sources kind of went over my head. Can you help us visualize this? Sure. Imagine a box, and it's filled with gas molecules. And they're just constantly bouncing around, interacting. And eventually, they kind of reach this equilibrium. And probabilistic computers, they're very simple level. They operate similarly. So they're not going through these precise calculations. No, it's more like the interactions of many of these people. Interesting. And that's kind of like our brains. Yeah. That's what they were talking about in the Anastasi InTech blog, like our brains. Yeah. With trillions of connections, they're constantly firing. And they solve problems in this incredibly energy-efficient way-- Right.--using probability. Yeah, the brain. It's like the perfect example. Yeah. Of probabilistic computing in action. That's so cool. So let's get into the P-bits themselves. Like, how are they actually created? I know the team at UC Santa Barbara, they're doing some cool things-- Right.--with magnetic memory technology. Yeah. And then you've got Extropic, the company behind that leak paper. They're exploring. They're exploring these things called Josephson junctions. Josephson junctions. Yeah. These special electrical components. And they can make these probabilistic computers super efficient using super conductivity. But before we dive into all the technical stuff, I want to talk about why, why the listeners should be excited, because this isn't just theory anymore. Companies like Normal Computing, they're already building real-world applications using this technology. It's amazing. And the potential benefits, like they highlighted in that ArcSyth paper, they're pretty amazing, potentially. Orders of magnitude faster. Yeah. And more energy-efficient computing for specific AI tasks. The paper used Gaussian sampling as an example. Yeah. It showed that. Traditional GPUs, they're good with low dimensions. But these probabilistic computers, they scale much more efficiently-- Wow.--as the complexity increases. They're built for these really complex problems. And then there's this idea of a thermodynamic advantage-- Yeah.--which is basically a point where probabilistic computing-- Yeah.--becomes way more efficient than traditional methods. Right. That has big implications-- Yeah.--for all kinds of applications. So this is where I get really intrigued with, imagine generative AI, the kind of AI that makes those deep dream images. But now that process is way faster-- Arrow.--and more efficient. What could we create then? Exactly. And it's not just about generating images-- OK.--we could tackle. These incredibly complex optimization problems, like logistics, finance, scientific modeling, any field where finding the optimal solution for millions of possibilities is a challenge. And the arcs of paper also mentioned this example of SNGP, which basically means probabilistic computing could make AI predictions more reliable-- Right.--by quantifying uncertainty. So instead of just giving you an answer, it also tells you how confident it is. That's huge. Yeah. Especially in fields like health care or autonomous driving, where the stakes are high. Yeah, for sure. But of course, it's not all easy. There are some serious hurdles. What are some of the biggest challenges facing probabilistic computing right now? Well, for one, there are these fabrication challenges, creating these b-bits reliably. At scale, that's no easy feat. Right. And then there's the issue of control. How do we precisely manipulate and measure these probabilistic systems? And then there's the software, too, we're basically talking about. Building a whole new software stack-- Yeah.--to work with this new hardware. It's a massive undertaking. It is. But the extropic light paper, they argue that it's worth it. The potential benefits, that way the difficulties. And normal computing. They're already working on this gap-- Yeah.--between probabilistic systems and our existing digital world. They talk about these digital robots-- Yeah.--collaborating with human experts. It's a cool vision of the future. It is. So the question is, how quickly can we overcome these challenges? How soon can we unlock the full potential of probabilistic computing? That's a great question. And we've touched on some of the applications, and there are surely many more just over the horizon. But before we get into those, I think it's worth taking a step back and looking at the bigger picture. I agree. If probabilistic computing is closer to how our brains work, could it lead to AI? That not only performs tasks-- Yeah.--but actually thinks more like we do. That's a question that we'll delve into in the next part of our deep dive. We just scratched the surface of how probabilistic computing works. And now I want to talk about what it can actually do. In the last part, we were talking about how this new approach to computing it might be closer to how our own brains work. That idea really blew my mind. It's a fascinating possibility. One of those intriguing aspects of probabilistic computing is this idea of uncertainty, quantification. Current AI systems often present their conclusions as if they're absolutely certain, but the real world is messy. It's full of unknowns. Right. In that Arcus paper, they had that great example with SNGP, where probabilistic computing can help us understand not just an AI's prediction, but also how confident it is in that prediction. It's about understanding the nuances, the probabilities, not just black and white answers. Exactly. Think about the implications for fields like medicine. A doctor needs to understand not just what an AI diagnoses, but the level of certainty behind that diagnosis. Probabilistic computing could be transformative there. It would make a huge difference in how comfortable people are with using AI in medicine, because it's not about replacing doctors. It's about giving them a powerful new tool to enhance their decision making. Exactly. And think about finance probabilistic models could help us better understand and manage risk. They could lead to more informed investment decisions that acknowledge the inherent uncertainty of markets. So much of the financial world is built on probabilities and predictions. So having AI that can actually quantify risk more effectively could really change the game. And in the realm of autonomous vehicles, imagine AI that can express how certain it is about its perception of its surroundings if the AI detects something, but isn't 100% sure what it is. It could alert the human driver or take a more cautious approach. That ability to express uncertainty could be a critical factor in safety. So it's not just about making AI smarter. It's about making it safer and more reliable. Precisely. It's about building AI that can reason about uncertainty, just like we do. We focused a lot on AI, but the impact of probabilistic computing, it goes way beyond that. What about scientific research? There's a lot of excitement there, many scientific problems, from climate modeling to drug discovery. They involve massive amounts of data and these complex systems that are inherently probabilistic. And that Extropic Lake paper, they really highlighted that potential. Didn't they believe that their thermodynamic computing approach could dramatically accelerate scientific research, enabling scientists to tackle problems that are just way too complex for traditional computers right now? They're envisioning simulations that could delve into the intricate workings of a cell or model complex climate systems with a level of detail and accuracy that we can only dream of right now. Those kinds of breakthroughs could revolutionize our understanding of the world and lead to some incredible discoveries and innovations. But amidst all this exciting potential, we also need to consider the ethical implications. As these AI systems become more sophisticated and more deeply integrated into our lives, we need to make sure they're used responsibly. That's a crucial point. We've talked about how probabilistic computing could make AI seem more human-like, capable of expressing doubt and nuance. But that raises some questions about accountability. If an AI makes a mistake, who is responsible? How do we ensure fairness and transparency in these systems? These are complex questions that society will need to grapple with as we enter this new era of computing. There's a lot of work to be done to make sure that this powerful technology is used ethically and for the benefit of everyone. It's a journey full of exciting possibilities and challenging questions. But one thing's for sure. Probabilistic computing has the potential to profoundly reshape our world, opening up new frontiers in AI scientific research and far beyond. It's a world where randomness and uncertainty, once seen as obstacles, become the keys to unlocking new levels of understanding creativity and innovation. That's a great way to put it. But for now, let's shift our focus to some of the challenges facing this emerging field and what the future might hold for probabilistic computing. We've explored the how and the why of this probabilistic computing now. Let's talk about what this new era of computing could really mean for us, for our world. It's a vast landscape of possibilities. But one of the most immediate impacts, I think, will be felt in AI, artificial intelligence. We've talked about how probabilistic computing could lead to AI that's not only more powerful, but also more trustworthy, capable of understanding and communicating its own limitations. Yeah, the ability to quantify uncertainty could really be a game changer, right? Absolutely. It could lead to much greater trust in these AI-driven systems, especially in critical areas like health care, finance, and autonomous vehicles. Absolutely. Imagine a world where medical diagnoses generated by AI come with a clear indication of the level of certainty doctors could then make much more informed decisions, weighing both the AI's insights and the potential for error. Yeah. That would make such a difference in how comfortable people are with AI and medicine, because it's not about replacing doctors, right? It's about giving them a powerful tool to enhance their decision making. Exactly. And think about finance. Probabilistic models could help us better understand and manage risk. It could lead to more informed investment decisions that acknowledge the inherent uncertainty of markets. Yeah, so much of the financial world is built on probabilities and predictions. So having AI that can actually quantify risk more effectively could really change things. And in the realm of autonomous vehicles, imagine AI that can express how certain it is about its perception of the surroundings. If the AI detects something but isn't 100% sure what it is, it could alert the human driver or take a more cautious approach. That ability to express uncertainty could be a critical factor in safety. So it's not just about making AI smarter. It's about making it safer and more reliable as well. Precisely. It's about building AI that can reason about uncertainty, just like we do. We focused a lot on AI, but the impact of probabilistic computing goes way beyond that. What about the potential for scientific research? Oh, there's a lot of excitement there. Many scientific problems, from climate modeling to drug discovery, they involve massive amounts of data and complex systems that are inherently probabilistic. Yeah, the Extropic Light paper really highlighted that they believe that their thermodynamic computing approach could really dramatically accelerate scientific research, enabling scientists to tackle problems that are currently just too complex for traditional computers. Yeah, they're envisioning simulations that could delve into the intricate workings of a cell or model complex climate systems with a level of detail and accuracy we can only dream of right now. Those kinds of breakthroughs could revolutionize our understanding of the world and lead to some incredible discoveries and innovations. But amidst all this exciting potential, we also need to consider the ethical implications. As these AI systems become more sophisticated and more deeply integrated into our lives, we need to ensure they are used responsibly. That's a crucial point. We've talked about how probabilistic computing could make AI seem more human-like, capable of expressing doubt and nuance. But that raises some questions about accountability. Like, if an AI makes a mistake, who is responsible? How do we ensure fairness and transparency in these systems? These are complex questions that society will need to grapple with as we enter this new era of computing. There's a lot of work to be done to ensure that this powerful technology is used ethically and for the benefit of everyone. It's a journey full of exciting possibilities and challenging questions. But one thing is for sure, probabilistic computing has the potential to profoundly reshape our world, opening up new frontiers in AI scientific research and far beyond. It's a world where randomness and uncertainty, once seen as obstacles, become the keys to unlocking new levels of understanding, creativity, and innovation. It's been incredible exploring this emerging field with you. And to all our listeners out there, keep exploring, keep questioning, and keep imagining the possibilities the future of computing is probabilistic and it's just waiting to be discovered.