Solving the AI Energy Crisis  ## Metadata - Author: [[The Changelog: Software Development, Open Source]] - Full Title: Solving the AI Energy Crisis - Category: #podcasts - URL: https://share.snipd.com/episode/93e280ef-d185-4697-bb79-5640ef0fff4b ## Document Notes ## Highlights - **AI's Energy Demand Crisis** - AI's energy consumption is projected to reach 12% of US energy by 2028, up from 4.4% currently. - The US grid growth at 1.2% over 20 years cannot keep up with this demand, causing a potential energy crisis. Transcript: Greg Osuri Yeah, so, I mean, AI is taking an incredible, I guess, place in all our lives. And, you know, while there's a amount of progress that's being made from, let's say, solving cancer to securing the homeland, all that progress comes with an enormous cost. And that's in energy, right? So the amount of energy we need is reaching a point of what we call crisis. And President Trump declared energy as a national emergency recently. So I was invited to testify before Congress on this particular subject as to how much energy is being consumed by AI and what are possible scenarios in the future if we don't do something About it and what are some of the solutions we think we can try out nationally as well as internationally to solve this crisis. So yeah, it's a great opportunity. And I think like my covered pretty wide breadth. I went into the report from Department of Energy, the official sanction report from by U.S. Congress, which was conducted by Lawrence Berkeley Lab. And happy to get into it if you want. Jerod Santo Well, I'm interested in just how much energy we're going to need or how short we're going to come up. I did read a 40% figure, I think, in your testimony. Greg Osuri So Department of Energy, you know, it's a government organization. It's known for its conservative estimates, not real actual estimates, we call them. The conservative estimates amount to about 500 plus, I think, 508 10-awatt hours by 2028. For context, that's about 12% of U.S. Energy consumption for AI. The problem is, and that's a huge jump from 2026, I believe, and we consumed about 4.4% of energy. So it's an incredible jump in terms of percentage consumption. And number two is, the problem is the U.S. Grid grew about 1.2% in the last 20 years. So the challenge is the grid is not growing fast enough to service AI workloads. Gartner, which is an analyst, predicts by 2026, that's next year, we will consume about 40% of AI data centers will be out of power. Because we will be, you know, forced to a position where we have to make a decision as to who gets the energy. Is it the homes or is it going to be data centers? I mean, obviously, we're going to prioritize homes over data centers. The challenge is, in the West, our grid is extremely hard to upgrade. And it's also extremely hard to produce energy. Considering we exhausted all the natural resources for the most part, we don't have enough hydro, we don't have enough dams, we don't have enough naturally generating energy, so we Have to either build, you know, nuclear plants or we have to look for alternatives, right? Nuclear plants give about a gigawatt of energy per plant. The only problem is it took about 14 years to build the last one. So we can't build them fast enough to address the immediate need. And even if you do build nuclear plants, we can't transmit them. Because the transmission infrastructure in the States is extremely hard to upgrade. We had energy, electricity, one of the first countries, in fact, the first country to have electricity. So when we designed the grid, it was designed to be, they didn't quite have the scale in mind when they designed the grid. So if you want to upgrade the lines, there are two ways to do it. Number one is, well, you know, put an additional line, you know, power line. And no one wants additional power line running through the backyards because that's real estate values. It's going to plummet. Or you got to do interruptive. That means stop energy, upgrade the lines. These are literally copper wires, right? Like you have thick copper wires, replace them by interruption, and nobody wants their power to be interrupted. So you can't put new lines. You can't upgrade existing lines so how do you upgrade the grid uh china is a communist country you can just basically say you have no option america and the west in general or western Democracies is very very hard to upgrade uh you know lines because it has to do with property rights. So the third option is to underground, right? And that seems to be a feasible option, but the cost is so high that you have to literally justify how you're going to dig a line under the ground and, you know, upgrade infrastructure There. So either way, it's extremely challenging to upgrade infrastructure in the US. So you can say, oh, maybe we can do renewables. Well, renewables, we think about it. Solar is terrible at utility scale. And if you do utility solar, you end up with the challenge of transmitting energy. So you can't have large swaps of solar panels, and the expense of storage, which is batteries, is super expensive. And you need more energy to cool the batteries so the efficiency goes down. So solar is not really an option right now at utility scale. Solar is good at residential scale. ([Time 0:03:41](https://share.snipd.com/snip/bd468e12-ae79-4a34-a71d-86c1883eae84)) - **Doubling AI Training Energy Needs** - Training state-of-the-art AI models requires energy capacity doubling every two years. - By 2030, AI model training could require energy equivalent to a nuclear power plant. Transcript: Greg Osuri Expense of storage, which is batteries, is super expensive. And you need more energy to cool the batteries so the efficiency goes down. So solar is not really an option right now at utility scale. Solar is good at residential scale. And you really don't have, you know, wind is not practical because it's very location-specific. And if you do have wind, you you're not able to transmit. So now we're in a limbo as to what to do with the energy problem. So there's no easy solutions for that way. And it's not stopping. So right now, the only solution is burning fossil, right? So if you look at the XAI data center that trained GROC for, has about 150 megawatts of capacity in Memphis, Tennessee. They consume about 7 megawatts from the grid. The rest, they're burning gas, LNG, to power data centers. Well, the effects of that is increasing respiratory conditions in the southern Memphis area. There's a new lawsuit. That's, I guess, I'm not sure what the status of it is. I think the Southern Poverty Law Center is sued XAI about health issues that are coming out of these data centers. We just heard yesterday, Sam Altman tweeted out saying that they have acquired 4.5 gigawatts of capacity in Adeline, Texas, in partnership, I believe, with Oracle. For context, 4.5 gigawatts is about five nuclear facilities, equivalent. There's no way in hell they're getting that from nuclear because, you know, we don't have free nuclear plants in America. I mean, we have about 96 reactors that are consumed 93% of the time, you know, fully consumed 93% of the time. There's no nuclear reactors energy. So where do you think they're getting the energy from? But in fossil, that much amount of LNG will produce about 2.5 to 2.7 million tons of carbon, net new carbon into the environment. And that's equivalent to about two thirds of all the CO2 emissions from Vermont. So we can shut down Vermont. Two thirds? Two thirds. Jerod Santo That's more than half, right? It's more than half. Just check. That's more than half? Greg Osuri Over, yeah. So over 66% of CO2. Jerod Santo So you're going to pick Vermont or OpenAI, right? Greg Osuri I'll choose chat. So, I mean, we're talking about Vermont being a very agricultural, heavy state with a lot of methane. Just to give you a context, this is one data center we're talking, one company, right? So put that with the heavy competition between, what's his name? Mark Zuckerberg tweeted out saying, or Facebooked out saying that they're going to build a data center bigger than Manhattan, right? I mean, I don't know if you saw that tweet. I did not see that. Everybody is competing. Yesterday, Mark Zuckerberg, I mean, Elon Musk said they're going to, in fact, they're going to get about 5 million GPUs by 2028, I believe he said, or next five years, 2030 five million GPUs he said five million each each 100 equal and says five five times one kilowatt right 1.2 kilowatt per GPU just to give you a brief idea as to how much computer how much energy we're Jerod Santo Looking at oh sorry 50 million I've sorry that taken them 50 million five five zero five zero 50 million so that's what you originally said yes apologies i mean these numbers are so large Greg Osuri 50 million gpus for context nvidia made about 2.5 million h100s in 2024. So uh not financial advice but NPD, I believe, is going to be extremely beneficial. You think the stock is high now. Wait till you see ([Time 0:09:32](https://share.snipd.com/snip/a021a70b-1e76-4dbc-aec5-d095ce28251b)) - **Uniting Around AI Energy Issues** - Public fatigue on climate activism complicates uniting around AI energy challenges. - Clean air concerns can unite people more than abstract carbon emission issues. Transcript: Adam Stacoviak There's some hope though. What I think is interesting though, is that one of the things you mentioned in your testimony, I just glossed through some of the details from that was that, you know, while you said all Those things and while you mentioned all those human nature feats, we must somehow navigate because there's no putting this genie back in that bottle. It's just not going to go back there, right? You're arguing that this isn't just another tool. This is becoming foundational like electricity or internet. This is a foundational thing. So we can't not solve this problem. And I know we have politics and countries and political powers and stuff like that. And I just wonder, like, how do we unite humanity around this movement? Is it possible? Is the geopolitical powers just too strong and too unable to share the information and think versus China versus US or Europe versus US or whatever the, you know, the comparisons may Be. How do we get to a place where it's like human, humanity needs to solve this problem because everyone across the entire globe uses electricity. No one argues about that. We figured it out, right? We're still figuring, we figured out every day or every year iterating towards how we do better electrical grids. As you mentioned, the challenge to even upgrade. How do we unite around this as a human race versus a country or geopolitical? Greg Osuri People are fatigued with the climate activist agenda. So anything that's related to climate or carbon emissions, like I'm, you know, I'm a pragmatist. See my testimony. I was like, well, this is a situation, you know, we got to, we can't sacrifice progress for other things. But I don't know. It doesn't matter what side of the political angle you are. We don't like dirty air because it's literally dirty air. Like if you're burning coal or if you're burning fossil, the air gets polluted and nobody likes dirty air. That's my concern, right? Others, you know, carbon emissions, you know, questionable and whatnot. So that's a challenge. How do you rally politically a crowd that's fatigued with this agenda, the climate agenda over the last decade or so? That's one, right? And second is we have right now a very challenging national security posture when it comes to AI, right? China is outbuilding, and they're doing a phenomenal job with renewables and whatnot. I mean, don't quote me on these numbers, but somewhere I saw a stat where they're building more solar every year than the entire U.S. Grid, or they will be in a few years. So they're building incredibly well, and they had this vision about 10 years ago to build this infrastructure, and U.S. Fumbled because of the climate agenda. And we have this national security hawks to convince, right? So U.S. Needs to have a leg up. The only way to do that is through getting more energy in whatever ways we can get more energy. So how do we unite people? Spread the message. I mean, I'm doing my part. Yeah. But also like show hope, right? So, I mean, it's not all dark days ahead. I mean, we ([Time 0:16:40](https://share.snipd.com/snip/4bb9ee8a-003a-4e42-9795-78c656fcf254)) - **Advances in Distributed AI Training** - Recent advances enable distributed AI training reducing the need for centralized energy. - Companies have made progress in asynchronous training and fault tolerance with 45-50 billion parameter models. Transcript: Greg Osuri So you can't do distributed AI. You can't train half the AI in one data center, another half in another data center. That was about a year ago. Now, over the last year, we saw incredible progress in distributed training. It's not quite as good as the non-distributed co-located training, but the results are very, very promising. The three main areas that we're seeing progress in, the first is low communication algorithms, we call it. What that means is by reducing the amount of communication required by different layers of AI when you train, you can reduce the need for connectivity. So you can have multiple GPUs in different data islands, we call them, and somehow train them. So this method was proposed by Google DeepMind, which is the most advanced AI lab in the world, arguably. They created transformers, which is a GPT thing. DeepMind proposed a framework called DialaCo, which another company called Prime Intellect implemented that and actually was able to train a 32 billion parameter model. For context, LAMA, which is the first open source model, was about 15 billion parameters. So it's a little bigger than the first open-source model. Still not state-of Another context is the GPT-3 was 150 billion parameter model. So it's like far behind GPT-3. GPT-3 was the first model we used for ChatGPT, which was in a head turner, right? So we are far behind, but 32 billion parameter distributed is a great, great outcome. Another company called News Research, which also has another algorithm called Distro, similar, they were able to train a 50 billion parameter model. Now they're training a 45 billion parameter model, fully distributed. Another promising company called Pluralis, this is, it's very, very excited about this company. They figured out how to do the second level of optimization, which is fault tolerance and asynchronous training. I don't want to get too deep in the weeds as to why this is important, but, or how it works rather. Why it's important is using what they figured out, which has been a very challenging problem to do asynchronous machine learning training. Asynchronous means it lets no... So machine learning training or training right now is heavily homogenous. That means if you're using H100s you have to use H100s across the board. If you use any less powerful GPUs that one GPU will hold up the entire training run. It's called straggler nodes. These guys figured out how to do asynchronous training where you can have slower GPUs mixed with faster GPUs distributed across the world using something called a swarm training mechanism. Can all somehow come together and train a model with 1% or less loss in accuracy, which is very, very good in terms of results. So this is very promising. Now, that's not in code yet. And another company called Jensen, which is also, you know, they've raised quite a bit, they figured out how to do another fault tolerance mechanism similar to Pluralist. And there's other companies that are doing like zero-knowledge proofs for gradient synchronization. So what that means, you don't have to verify all the gradients, so the gradients can come from, and you can verify all the gradients without using zero-knowledge proofs and whatnot. So the point being, there is a lot of work that indicates distributed training will become a reality by end of the year. By end of the year, we will produce a model as good as GPT-3. By end of next year, or maybe going into 2027, there's a good chance ([Time 0:21:13](https://share.snipd.com/snip/10a1f447-df49-492e-aec9-1f45373f7921)) - **Use Energy-aware Scheduling** - Move AI workloads to locations with abundant renewables like solar or wind. - Develop energy-aware schedulers to pick cheapest energy sources, benefiting environment and costs. Transcript: Greg Osuri In the US than we consume uh during the day i know for sure in solar solar wind a whole lot of things we have a lot of sources in the u.s i mean there's been enormous investment that went into Renewables over the last few years over the last 20 years um that is a phenomenal solution that's what we are going towards, right? So why burn energy, but instead use that energy to train AI? When it's windy in Kansas, move the AI workload there. When it's sunny in California, move the AI workload. The AI workload is going to be elastic and has to be asynchronous. It has to be these things. So that's why I'm particularly excited about the progress with distributed training. Now we can have elastic workloads that can move where there's energy-aware schedulers. So Akash, we're developing one of those. I'm sorry, I forgot to tell you what we're doing here at Akash. We are developing an energy-aware schedulers that will pick the lowest energy. So we'll start advertising the cost per energy now. There are lots of data centers in the U.S. That can give you one cent. I mean, economics-wise, right? Jerod Santo Yeah, give me the cheapest thing I can get. Greg Osuri Right, it's the cheapest thing. It's actually beneficial not only for the environment, but for your pocket. So that's where capital is coming to play. Wynn will give you one cent per kilowatt hour, which is the cheapest possible energy you can get. Solar, depending on the setup, right? Depending on the battery cost and a whole lot of things. Solar, you can get very, very cheap. And if you can arbitrage, you can get for like almost zero. LNG, the gas, natural gas will give you 10 cents a kilowatt hour, 10 times more than what you would pay for wind. Diesel will give you 32 cents a kilowatt hour. ([Time 0:50:10](https://share.snipd.com/snip/2eb42c0b-53ca-4a8e-87fc-43e062cd5a01)) - **Solar Energy Vastly Outpaces Grid** - Earth receives vastly more solar energy than current grid capacity. - Humanity is currently below Kardashev Type 1, needing to capture all planetary energy. Transcript: Greg Osuri So we are like literally over hundreds of times more energy that we produce from the sun. I think it could be more. I could be wrong. The point is there is this like civilization scale, type 1, type 2, type 3 by Kardashev levels, we call them. Aware of this, but there's a Soviet astrophysicist from the 60s who define what civilizations in terms of advancements they are compared to like extraterrestrials and whatnot. So Kardashav level one civilization or type one civilization is a civilization that figured out how to capture all sources of energy on the home planet, like from planets. So anything that hits the earth, should be able to capture we're currently at 0.7 a level so we're not one yet so to reach one that means we are an advanced species we need to capture all The energy that we got and and figure out how to use it because you know if we capture now we don't know how to use it but with the ai we can know how to use energy level two kardashav level two Is is when we get dyson spheres right dyson sphere is like when the civilization expands from its home planet and captures all the energy from its home solar system right there's always A sun uh so the best way to do that is to build a sphere around the sun and capture the energy. So that's why I don't know how much deep into, into extraterrestrial. Only through science fiction, really. Dyson sphere. Dennis C. Taylor, he's been on this podcast. Adam Stacoviak He talked about Dyson, Dyson sphere in his we are Legion. We are Bob, the Bobaverse series covers the whole entire, there's a book four, I believe. I don't know. Anyways, they talk about it. They're pretty plausibly when it comes to the science, like he's ground, he's a software developer. He's grounded in science. I think while it's science fiction, it may give you some, you know, insight to a Dyson sphere. Greg Osuri A light years away from Earth that looked like there was strobes around the star. So there was some indication this could be a type two civilization. No way. Right. So the star was pulsating. It was not a dying star. So only time a star pulsates is when you have light being distracted. Yeah. So they say it's a Dyson sphere. Anyway, so like type two is the entire solar system. Type three is a civilization that they can leverage or harness all the energy for its home galaxy. ([Time 0:55:52](https://share.snipd.com/snip/f23f92ad-1554-4b62-bc7f-ffa73df2115c)) - **Nuclear Energy's Potential and Challenges** - Nuclear energy is clean, safe, and favored by Greg Osuri. - High regulation and public perception issues slow nuclear adoption and increase costs. Transcript: Adam Stacoviak Oh, yeah. Greg Osuri So if you look at if you look at the pushback to nuclear, it's on safety. And if you look at all the safety, all the incidents that happened over the last since the beginning of nuclear, almost not 100 percent of those issues were due to negligence by the staff. Staff forgot to check something or put something because they, you know, whatever, right? It's all negligence. And there are very few incidents that were caused due to natural disasters. And Fukushima was the only one that we can point to that was, you know, that that melted because of a tsunami, right? Historically, there hasn't been any natural disasters leading to nuclear meltdowns. And Fukushima, nobody died, zero deaths, zero fatalities. And you can actually drink water now from Fukushima, it's pretty safe, So from a safety standpoint with AI now, and we have a lot of other automated systems that are phenomenally good, You can make a case saying that if you remove the human out of the equation to maintain nuclear facility, safety levels are significantly higher. Right now, nuclear, for example, nuclear like Gen 3, we currently have Gen 2, right? So Gen 3s are the new ones that are coming out. Their cost per kilowatt is significantly higher than Gen 2 because there's so much of investment or so much of capital that needs to go to our safety protocols. The recovery of capital is significantly longer for Gen 3. So if you put Gen 3 reactors, you're not going to see cheaper energy for a while on these reactors, right? And I mean, I remember doing the math, I think it was at least 50% higher than Gen 2s. Because most of Gen 2s are like actually now paid off, right? So, you know, nuclear is like very, very good, but you had to make a case on like, is one case and economics is a second case. Like why would someone choose nuclear over, let's say wind, unless there's no option or nuclear is more expensive than LNG right now. Right. Like from the Gen 3 at least. So you had to make a case like why would someone choose nuclear over drill baby drill, right? ([Time 0:58:59](https://share.snipd.com/snip/daa58581-ba5a-4861-9c84-ad5a36ccabbb)) - **Incentivizing Home GPU Sharing** - Offer free energy and free internet bandwidth to homeowners to incentivize hosting GPUs. - Limit GPU quantity to fit home electrical capacity and maintain feasible economics. Transcript: Jerod Santo I mean, that's what you're saying, right? Like your goal is to put GPUs in people's homes through their telecom relationship like that's how you're going to actually reach that's one way to do that yeah we have several ways but Telecoms are the best way because you're sure that's just one way you're going about it and you're going to connect those up to your network yeah right a cache network and they're gonna Become resources available so now consider me as your potential homeowner and you're pitching me on the idea of putting a GPU in my house. Well, why? Greg Osuri You get the package that we're going after is you get free bandwidth. You get free energy. Jerod Santo I get free bandwidth. So you're providing my internet connection. Yeah. And my bandwidth is now free and I get free energy. Yeah. Greg Osuri We'll put solar in your home. We'll put a battery in your home. You don't know how to pay a thing. We will charge a battery during the day. We'll sell it back to the grid. When we make the most money, we'll use excess energy to produce enough energy to power the GPU. So you get free energy, you get free bandwidth, maybe some money on top of it. Yeah. Okay. Jerod Santo And how many gpus can you put ([Time 1:06:05](https://share.snipd.com/snip/0244521a-7ca0-4f8a-8548-e7b50fe42ba6)) - **Pursue Distributed AI Expertise** - Engineers should dive deep into distributed systems and AI research. - Use peer-reviewed AI research and follow top investors to find merit and opportunity. Transcript: Adam Stacoviak What can they do to capitalize on this change, this massive money-making possibility, and also just changing humanity in the positive way fundamentally in the next couple of years? Where can they apply their talents to capture some of these dollars and also innovate? If you are an engineer that has experience in distributed systems, go into AI. Greg Osuri Have you seen how much Meta is paying to poach AI researchers? I've seen offers for up to a billion dollars for an engineer, one engineer for four years, and they're still rejected. But I've seen offers all the way from, I know, the space is so, I'm like, what am I doing? Not enough. That's not enough. So they're offering engineers 300 million, 500 million offers for four years. If you're an engineer that is good with ai that's that's that's the ceiling right now you can make a million if you're good enough um i think if you're a decent researcher now in the space Is about a million dollar cost on average per year so as an engineer you have a lot more to offer so i would like just drop everything do you need to do really deep into differential equations And you need to brush up on your math skills. But if you break it down, I mean, it's just multiplication. I bet you if you spend enough time, you won't be that hard to regain confidence. There are lots of opportunities in AI, like lots of opportunities if you're an engineer, right? You can solve a distributed training, for example, right? If you can, I bet, if you can figure out how to distribute gpus even centralized you know across two data centers really well right you have a winner i mean centralized companies when They there's a good chance they may go distribute because anthropic ceo co-founder talked about distributed training in their newsletter. I wrote an op-ed recently. By the way, I wrote an op-ed in Coindesk last week. I mentioned about all these stuff that you can actually go. It's a good starting point if you're looking into getting understanding. So there are frontier labs that are actually looking at distributed training as a solution. Not decentralized yet. I will get there because decentralization is about adding incentives on top of distributed training, right? That's what decentralization is. But distributed training, I think, is a great area. We need a lot of people to focus on this stuff, right? As an engineer. Now, if you're an investor, savvy investor, if you're looking for that edge, the niche energy, like look at the space. Like now, I mean, there's so many people you can talk to ([Time 1:33:56](https://share.snipd.com/snip/12e1171e-2f38-47ac-be18-d2c1591f588b)) - **Augment Work with AI Skills** - Use AI as augmentation, not full replacement of humans. - Learn math and differential equations to leverage AI effectively and boost productivity. Transcript: Greg Osuri Free bandwidth. You're not literally getting free, but there's always an exchange of value. You know what I mean? That's fine. No one will put something in your home unless there's something to get out of it. About when it comes to UBI and like future of AI taking jobs, there are two things to look at. Number one is so far the society has rewarded intelligence. Human time has value, right? So whatever value that may be, because you're providing some time and you're giving intelligence during the time, you have some value that's called value of labor, right? Now, value of labor, you get some price back. Question you should be asking is, will the value of labor be diminished to a point that's no longer valuable? Number one. Number two, what happens when a world shifts to that mode where the only mode we knew so far in existence of all civilization is the value of time? So question number one, will AI replace human beings fully autonomously? Now, I am a practitioner of AI. I use AI much more than most people do. And I'm also an engineer. I guarantee you the state at which AI right now is you cannot replace a human being not even close one because of one main reason it that happens to be learning capability AI once trained Cannot learn continuously. Whereas a human, if you do something, if you work for me, if you're doing a bad job, if you do something that I don't want you to do, I'll tell you what to do. Next time, you're going to not repeat that, right? You can't tell that to AI. AI is not self-learning. You can talk about reinforcement learning and all these techniques, but it's not self-learning yet. So you cannot replace a human being in any way. But a human being will be replaced by someone that knows how to use AI, for sure. Now, is it a replacement or is it an augmentation? Now, there are two ways to look at it. As a CEO of a company, I was not coding as much as I used to code back in the day. But now with AI, I'm coding a lot more than I used to code, even back then. Why? Because the speed at which I can deliver stuff has gotten significantly lower compared to pre-AI world. So I'm able to do a lot than I couldn't do. Now, is there, as a CEO, do I want to do more things? Absolutely. There's no limit as to what kinds of things I want to do. Like I told you, I wide-coded a Siri replacement because Siri sucks. Now, would I do that without AI? Probably not, because I don't know how to do Swift. I don't know how to do Mac apps. I don't know any of this stuff. I mean, I could learn, but is it valuable for me to do that? Probably not. But I was able to replace this whole thing in, I don't know, afternoon, because it was fun. I reconfigured my entire network on 4th of July, because I wanted to protect against Chinese IP device, IoT device, right? Will I do that without AI? If I have the motivation, but probably not as quick as what I could do. So I could do a lot more as a human being that I would rather need a network engineer, probably I need to hire to do this. Or I need a developer I need to hire to do this, right? Yeah. Am I hiring more developers? I am. And we're equipping them with AI to do a lot ([Time 1:41:36](https://share.snipd.com/snip/45b5b722-4330-4975-b7a0-749bd0d8bf64)) - **Efficiency Drives More Usage** - Increased efficiency typically leads to higher overall usage, not less. - AI's acceleration of productivity will fuel even greater demand and creation. Transcript: Greg Osuri Because of this paradox and not, not the opposite. Right. So yes, what you're saying is dramatic inefficiency in producing code at scale. Adam Stacoviak And now we have a way to produce code at scale. And so the inefficiency has been unlocked and it, it doesn't change our need to create more and produce more. Greg Osuri It only exacerbates it. It fuels it. Yeah. It doesn't matter you just i think it's going to get we're going to accelerate beyond our that's a great argument or at least a you know a great take on it because i think there's a lot of Adam Stacoviak Doom and gloom out there to some degree and i think we need to look at i ([Time 1:47:10](https://share.snipd.com/snip/f1af6ffc-e850-43e6-96ce-5928ce565aaa)) - **AI Glasses Aid Public Speaking** - Greg Osuri used AI-powered glasses on stage as a teleprompter without audience noticing. - This tool helps him access key info instantly while speaking publicly. Transcript: Greg Osuri But you may be able to see yeah put them on let's see yeah what uh put your new mask on become clark kent this is a it's called even realities um is it on right now these are very early We're Jerod Santo Just like a crazy person at first. Greg Osuri He's like, he's staring off into the ceiling. This is the early version of AI glasses. I'm experimenting with a bunch of these glasses, right? And a lot of companies working on different variations and not everything is ready for prime time. But like, okay, that's me augmenting myself. Like it has a teleprompter when I'm on stage and I'm giving like phenomenal, like key notes, teleprompter. Jerod Santo Oh, you actually wore that on stage? Did you wear that at your conference? People couldn't tell. They wouldn't even notice. They have no idea. They thought I was just a good speaker. This guy has great recall. Adam Stacoviak He's got it all together. Jerod Santo Were you just reading your teleprompter on your glasses? Wow. Secrets out. Secrets out now. Greg Osuri I mean, I got the idea from, what's his name? The guy who created Oculus Palmer, Palmer ([Time 1:51:40](https://share.snipd.com/snip/0886aaa8-1be8-4cb6-94fb-2c74ab48e315)) - **Building a Smart AI-powered Home** - Greg is building a smart, energy-efficient home with sensors and local AI. - AI will monitor health, detect emergencies, and ensure privacy with no outside data leaks. Transcript: Greg Osuri Incorporate that in your real life. Like, I have my new home building in Texas. It has, it's a large home and multiple stories, soundproofed. It has cameras where possible around the campus, about two acres. And inside and outside the house, it has sensors, infrared, as well as ultrasound sensors for places where you can't have cameras, like bathrooms and bedrooms and whatnot. The goal is to detect any living form in the facility, be it within the walls or outside the walls, and particularly to detect mishaps, like my mother-in she's moving in with us. She's got her own very private unit inside the house. If she falls down, for example, I want to know. Or the AI should know. So all the information, everything is recorded and analyzed and fine-tuned locally using a massive GPU cluster. Nothing leaves the boundary of my home, the network of my home. It's fully locked down, locked down network because a lot of this stuff is private, right? Every conversation. So this conversation would be recorded, be transcribed. And so that I come back and talk to you next time, I'll remember what I talked about. Simple things of that nature, right? You read the name off your glasses hello security right like for security purposes if someone is there on the campus that's not supposed to be there there's going to be three drones that Get dispatched uh you know triangulate the subject or the target rather and if they still don't comply they'll be like the robot dogs i'll go in with this is now this is your house now and Adam Stacoviak What i'm doing that i'm actually on it. Is that the Westlake house or this is a different place? The drones aren't currently set up? Greg Osuri I'm experimenting with a bunch of drones. The problem is a lot of these drones are Chinese drones and I just don't like the software coming from China. So I'm trying to keep it most American. Jerod Santo What model are you using locally? DJI is the only, DJI is the best drone right now, unfortunately. Yeah, DJ the drones are awesome. Are you using Lama or what are you using locally? Greg Osuri DeepSeek. DeepSeek. Okay. R1 or R5B. I mean, Chinese have way better. I mean, nothing goes to China right obviously. It's local. But DeepSeek is really good for agents. Lama is pretty good too. Lama 70B is pretty good for agents too um but i'm experimenting we don't know like i gotta load test this stuff right really like put through like chaos testing whatnot you documenting This somehow somewhere you got a blog you got a tiktok i'm debating because it's a security posture you can't expose the security aspects of the system. You can pick and choose what you're going to expose, though. Probably. Probably. I haven't thought about how all my friends that I talk to, they all want the system at their home. So I don't know. Is it product opportunity there? I don't know yet. Business. This is definitely a new business, right? I mean, imagine a house where you can talk to the house, like, you know, that's local, running locally. It's responding to things. I mean, particularly about health, right? Like I want to see what I'm eating all the time. Like, you know, what kind of food? Jerod Santo I imagine you can see it going in your house. Greg Osuri Analyze, I meant. I want to know like continuously what I'm putting in my body and, and what I could do better, like things of that nature. Right. Adam Stacoviak I was thinking about that recently in particular, like this whole idea of like, what comes into my home from the grocery store? There's a way to communicate that to intelligence, to my, to my, to my home, to my homework. Greg Osuri You know, all the attempts of like taking pictures of what I'm eating, that's not gonna work because I'm never going to remember. I'm going to be like picking something. I'm talking to somebody. It has to be unob be like out of your way that's the basic thing if it's in your way you're not going to do it doesn't matter how much effort i mean i know how much can we highly motivated so So simple you can use ai to automate your everyday life so much that with the current tech with the current tech you don't need much, right? It's not going to be that expensive too. I mean, maybe, sorry, I take it back. Like, okay, it's going to be expensive. The GPU is going to be expensive. Drones are the cheaper aspects of it. Drones are gaza here. That'll be cheap. The AI, the chips to run a real timetime AI that can respond with low latency to agents, that is expensive. The chips are expensive, energy expensive. Obviously some. So, you know, that's why I think like, I'm so pro like local AI of the future and leasing the unused compute back to the grid, like an AI grid with a cache, right? I'm very, I don't know, I'm very excited about this new future of, like, shared economy. Like, you know, because everybody wants GPUs. If you want AI locally, right, in your homes, you know, say we have, like, pass through the whole, like, you know, limits for whatever your grid imposes on you. I mean, I'm particularly excited about, like, what you can do with AI. And also, the new home is very sovereign right has rainwater collection about 70 000 gallons gallon rainwater capture and it has a greenhouse uh on site that produces food local energy Production with uh you need water for for food 70 000 gallons you said yeah two major massive rain capture tanks underneath the pool that they're hidden. And it has gray water filtering. It has, you know, we filter through the plants. It produces food locally. The more you learn about the grid, the less. This is your home in Westlake you're building? I know. The permitting is taking a little bit. I want to visit the site with you. Adam Stacoviak I want to come and visit the site with you. Greg Osuri Yeah, happy to show you. I mean, the permitting has been, they're like, what do you want to build? I'm like wow they were like blown away with my plans and they were a little challenging with permitting but we're getting through that stuff yeah um nice i even propose a bunker but the Bunker is going to be challenging oh ([Time 1:53:35](https://share.snipd.com/snip/ca1c0577-4601-4cc8-a255-ab95ff0475e5)) - **Decentralized AI Poised to Explode** - Decentralized AI is quietly advancing with major players working on distributed compute. - The market will explode soon as momentum builds behind these innovations. Transcript: Jerod Santo Let's save it for the next time we have you on, Greg. We got to leave something unplumbed. Let's throw the plus plus. Adam Stacoviak Let's end this and throw a little zombies to the plus plus folks. Greg Osuri Alright, thanks Greg. That's good stuff, man. Thanks a lot. Appreciate it. Adam Stacoviak So we do have a bonus on today's episode. I can't wait for you to hear it because it's kind of like unexpected. If you're a plus plus subscriber, well, stick around. And if you're not a plus plus subscriber, it's better. Check it out. Changelog.com/plus plus 10 bucks a month, 100 bucks a year. Drop the ads and get that sweet bonus content. Again, learn more at changelog.com/plus plus. ([Time 2:00:44](https://share.snipd.com/snip/d10e46fe-e47b-4197-9b6f-7d8514238a47))