Transcription: OAPN Episode 227 - Ocean Power

[00:00] Christopher David: And we're back! OAPN 227. You know what? I thought like, man, three weeks of silence. The episode that we come back with, it's got to have graphics, it's got to have like demos and just announcements and stuff. And then I was like, wait a second, our last video told everybody to expect shit to be just like raw and like horrible. Okay, great! Here's what we did: we have a green screen now, we're in the OpenAgents headquarters with like the full production studio. We grew a third person—who's this guy?

[00:32] Ben Silone: Ben Silone.

[00:33] Christopher David: You don't even have a name tag! No one knows who you are, you could be anybody. Okay, so we're kind of celebrating—we gotta make him a name tag. We gotta print him—okay. So we're kind of celebrating today because, you know, the team's growing. Ben just joined us from Florida. Ben just got in this morning and boy, are your arms tired? Or something. Ben, who the hell are you?

[00:51] Ben Silone: (Laughs) Uh, yeah. Ben Silone, hailing from Tampa, Florida. Um, and I've spent a number of years working in ocean tech, building cool stuff on the water.

[01:03] Christopher David: And you're going to help us make money because you also do business-y sales things and you're like—make us money so then we can afford all of the floating fancy ocean stuff, yeah?

[01:13] Ben Silone: Yeah, yeah, that's the plan.

[01:15] Christopher David: Yeah, so for you people who haven't been paying attention to AI um, like ocean stuff... so what? Peter Thiel just dropped $140 million on Panthalassa to make these little balls... You take a ball—here's what you do: you take a ball, you put some GPUs in it and a like a whirly thing, and you throw it in the water, and then this is worth a hundred bajillion dollars, right? Yes? How would you—how would you tell them what this is, Ben?

[01:38] Ben Silone: I mean, it's a—it's a cool experiment, I'll give it that. But um, I mean it's not the ideal material, it's not the ideal energy generation source, it's not big enough... yeah.

[01:48] Christopher David: You hear that? Your stuff's weak, Panthalassa! (Laughs) Just kidding. No, but seriously—

[01:53] Ben Silone: No, anything you do—stuff in the ocean is cool. Like, that's—

[01:55] Christopher David: No, we're huge fans of Panthalassa. We're huge fans of anyone unlocking stranded compute in the ocean using natural power. Um, we have our own ambitions in this regard, we're going to be expanding into—down the road. Um, Ben, give them just a little preview of what we've been thinking about in that regard.

[02:14] Ben Silone: Yeah, uh... I mean, we just need to scale up a lot. Um, 100 megawatts minimum, and—

[02:20] Christopher David: Is it going to look like this, or is it going to look like something else?

[02:23] Ben Silone: It's going to look like a giant floating city because you're already in the water and there's a ton of awesome stuff you can do once you're in the water. Not just energy and compute, but um, environment, food, um, mineral mining... we gotta combine all that stuff together.

[02:37] Christopher David: You know, I took this image from you—you did this with X or whatever, but just to give a little bit of the visual reference... like, one of these might be—could one of these be an OTEC plant?

[02:47] Ben Silone: Yeah, absolutely.

[02:48] Christopher David: And uh, tell the people what OTEC is.

[02:50] Ben Silone: Uh, Ocean Thermal Energy Conversion is when you go down about a thousand meters, pump up really cold water, and uh, have a turbine um, that's generating energy off the warm surface water and the cold uh, deep-sea water. Which also brings up a bunch of minerals, aerates the water... sea life loves it, environment loves it, uh, everything loves it.

[03:11] Christopher David: Yeah, here's the thing: there was a startup called OceanBit—I mean, they're still doing stuff, I think this is their website—so um, there was an article that came out in Bitcoin Magazine a while back about OTEC, and it was—they had a couple articles actually. So, May and June. So "How Bitcoin Can Unlock The Energy Of The Ocean For 1 Billion People." If you haven't read this, I recommend that you do.

"Bitcoin has the potential to help unlock between 2 to 8 terawatts of clean, continuous and year-round base-load power for one billion people by harnessing the thermal energy of the oceans. The technology is Ocean Thermal Energy Conversion (OTEC), a 150-year-old idea stymied by economies of scale that turns Earth's oceans into an enormous renewable solar battery."

So everyone says—when people hear "solar, solar, solar," you think "oh, you're getting cheap panels from China." No, we have two-thirds of the Earth's surface is covered by the largest solar panel plus solar thermal battery in the world. How do we use that? How do we unlock that, Ben?

[04:12] Ben Silone: Yeah, we just need warm surface water and cold deep-sea water and run that through—it's a Rankine cycle. Um, it's been around for a long time. That little image on that was uh, power plant—an OTEC plant in Hawaii built by Makai. Uh, super cool. They said they'll never build on land again uh, because it's way too expensive. But at sea, if they have a platform and uh, we can use that energy directly at sea, like in a data center... it's awesome! And we get free cooling with seawater air conditioning, which has been used around the world for decades now.

[04:43] Christopher David: Yeah, let me give the—the 50 IQ version of this: you make a platform in the ocean, you put a tube down with a turbine, and there's a difference between the hot and the cold water, it spins a turbine and you can power stuff.

Now, when I first was introduced to this concept in the book "Seasteading" from Joe Quirk—I gave Joe Quirk like a ride in my rideshare startup years ago and like interviewed him a little bit about this. Great book, by the way. But I learned about OTEC from this and they're like, "hey, OTEC's great, we did some pilots, but one of the problems is to get the energy usable you have to put it into like a helium container and then and then transport it to land like... and then well, no you don't need to monetize energy by transporting it, you can actually monetize at the source using this amazing technology called Bitcoin mining."

So, there have been at least one startup called OceanBit that at least initially they were trying to do this: like, "hey, let's do Bitcoin mining, you know, monetize it at the source." And apparently, they just had a hard time like raising the money because this is extremely capital uh, you know, Capex intensive.

Well, guess what happens to be an industry spraying infinity money at anything Capex that can uh uh, put GPUs anywhere that you can put them? Uh uh, AI. AI Capex. So how about we put like floating compute? So, Panthalassa—the fact that they, I don't know, raised $140 million I think from Peter Thiel... very good, good! Very happy to see that. Uh, and also there's probably more that can be done in that regard, yeah?

[06:13] Ben Silone: Yeah, absolutely. I mean, that is a good starting point for sure. Um, the honestly the biggest downside with OTEC is you really need to start at 10 megawatts, something like that. You're not going to throw a few solar panels in the water uh, or something equivalent to that. So um, but that's what we need. That's what everybody needs. We're—we're using stranded compute and there is a lot of that all around the world, and we need to build more compute and we need more energy... uh, we need literally everything because this is uh, something that we need to scale infinitely.

[06:42] Christopher David: So yeah, we hope to like propagandize OTEC more, you'll see more from us uh, in that regard. But that's sort of like Phase 3 of our plan. Uh, so Phase 1, you may have heard that we are purchasing compute for Bitcoin. We have a Pylon network of nodes that's running. You go to openagents.com, you can see the stats on the network. We had our sort of version 0.1 network live for a number of weeks. It's still live, payouts are not going out until we do version 0.2. In about 48 hours from now we're going to put the 0.2 network out—that is not showing my screen because I'm in the wrong tab and that is the wrong URL. Dang it! There we go. Okay.

"Sell your compute for Bitcoin." Okay, so we're doing two things: uh, we're upgrading the Lightning infrastructure—we had some bottlenecks regarding the particular like payment provider that we were using. Uh, the new version uses LDK, that's working fine. As well as we're flipping the uh, kind of basic pre-training pipeline, kind of homework, very basic thing... uh, into a fine-tuning pipeline. So we're going to be fine-tuning Qwen 3.5, 3.6 using these Pylons in a few days. So we have some uh, models that we want to fine-tune on a particular benchmark.

I guess we'll segue into like what we're actually doing here and what we're actually selling to generate the revenue that then connects the compute network. Um, there is a benchmark that was released a couple of weeks ago by Harvey. So Harvey—like the main legal AI company that's got all their big fancy law firms uh, using Harvey, Harvey, Harvey, Harvey... and you know, they're pretty dominant as a company, props to them, they've had some good success. And you know, some people are like "oh, maybe they're just uh, kind of a fancy ChatGPT wrapper but like they got all these big fancy law firms."

Well, guess what? They release this benchmark and um, they released a ton of really good data about like 20 different examples of types of work that an like an AI agent would do in the field of law. And uh, we looked around to see yesterday, I was like "okay, maybe that would be something for our new Autopilot agent to cut its teeth on." Like, if there's a scoreboard of people and their agents like doing legal work, let's see if we can beat it. Let's find the leaderboard. There's no leaderboard! Why is there no leaderboard?

Now, in software you can point to a company like Factory, you can point to Devin, a bunch of these companies show their scores on things like, "oh, SWE-bench we score at 98%" or whatever. Um, there's no equivalent for legal. And part of me thinks that because "hey, maybe some of these uh, big law AI firms are getting there because they've got their little networks of lawyers and they're going shop—you know, they're going at the—to the cocktail receptions and they're—they're—it's more relationship-based and maybe people there don't care about benchmarks." But but maybe the tech isn't also quite there and maybe they're embarrassed to post what their scores are? I don't know!

Uh, but I know that we are building our own dashboard for this. Um, we have an initial legal benchmark, if I can find it... "Benchmark, Harvey." So we're um... we're going to post like a cleaner version of this, but we're basically beginning our own hill-climbing effort to have our agents get good at this benchmark. We're going to post what our scores are, we're going to post the code behind this, we're going to show you what we're doing.

But our objective, as we're talking with, "hey, like who are the people that are going to want to pay money for Autopilot?" Like, the best agent—what we're now calling "the last agent." The idea here is that like, instead of worrying about whether Claude is going to rug-pull you on usage, or they're going to jack up their prices, or maybe OpenAI puts out a model that's better than Claude... like, really the—the end-state company that you or businesses are going to hire shouldn't be from a lab that's biased and only sells their own stuff. Like, it should be from a more neutral provider that can say, "listen, we're going to connect you with the best AI, we're going to make it as simple as we can for you, and if the best is a combination of OpenAI and Claude and local models and like maybe there's only a company like ours that's going to be able to provide that for you."

Um, but we want you to be able to have numbers and see like, "oh, for law, for legal, Autopilot is better than the competition and here's the proof of all of that." Uh, so this is what we are running our training run—which will begin in about 48 hours—specifically for fine-tuning Qwen models, Qwen 3.6 on this dataset. We are about to have a model in a few weeks, a couple months, that's going to be—we think—better than equivalents, or equivalents that don't exist, uh, on a small model that works well in legal agentic AI workflows. And we're going to prove it by training on what is apparently the best benchmark for legal. And then if we're the only people on the scoreboard, I guess we'll just have to dare other people to—to post their scores.

[11:42] Car Gonzalez: The other really cool thing is we've—we're in talks with a one specific lawyer, law firm uh, already, so he was going to be our kind of like avatar client, right? And so I think the—the fact that we're coming out with just more customizable—a more customizable agent, it just feels like the right uh, the right path forward. Ben, do you have anything to add?

[12:06] Ben Silone: Yeah, I think uh... I think a lot of it is just AI that are hyper-specialized, um, not general models necessarily, because it's a lot more efficient to be hyper-specialized and focused and fine-tuned, not only in a general sense but in a very specific sense for each individual business, each individual user, uh on their unique datasets... uh, obviously with a high level of security as well.

[12:29] Christopher David: Yeah, our sort of avatar client here—we'll call him Joe—um, he's got two decades of documents that he would love to have an AI reference. He cannot feed them into ChatGPT, he cannot feed them into Anthropic, they have confidential PII all over them. Now, that's a solvable problem. Even OpenAI, a few weeks ago, they put out an open-weight model for redacting PII. Let's take a look at "Redact PII model." So they have an open-weights model that you just run anything through and it's optimized for taking a batch of text, reducing sensitive information. It's like, okay, where's the pipeline for law? Okay, we're going to have to build that pipeline. Very easily.

Okay uh, lawyer with 20 years of docs, now you have redacted docs that you can actually confidently feed to an AI. Now, imagine that not just are those docs able to be fed in through some RAG-type database, imagine also that you've got a fine-tuning pipeline because you've got a lovely company called OpenAgents that's built a whole fine-tuning pipeline and a whole bunch of like compute providers and like able to just fine-tune stuff on a whim very easily. We're able to feed those documents in and create you not just the fine-tuned model based on the open model from achieving a legal benchmarks, but what if we could fine-tune that model further to custom-tailor to your specific methodology and how you think and how you communicate in your legal documents?

So, this is a type of thing that we can give just like extreme reinforcement learning-backed powers, integrated into a product that can be—to some extent—a drop-in replacement for ChatGPT or Claude, but also offering all these super powers that they just would not offer and then just dramatically undercut anyone that is existing.

And then you've got other competitors kind of like, you know, Thinking Machines. We're borrowing some of our fine-tuning ideas from Thinking Machines, they also do fine-tuning primarily on Qwen models but it's like, they'll—they'll give you some toolkits that are good for RL engineers to go in and like, "here's how you can use their their servers to to run fine-tuning workloads." Well, what if we offer that in addition to you don't even have to think about it. You just like upload the docs or run your doc through a script, our system handles the rest, and now you've got like a good model, a good product, it's custom fine-tuned on your docs, and what if we charge you like essentially nothing above the compute? Like a little bit above the actual compute? Because we want the business, and we think that we're—we're we're going to just start kicking the ass of these big fat companies that are charging you a bunch of stuff not even telling you what they're like using it for.

[15:07] Car Gonzalez: So—so one of the other things you come to find out... like, so the past three weeks we've just been hitting the ground running with talking to a lot of people as far as like what are their constraints, what are they looking—looking for when they're trying to find a specific AI that can be a workhorse for their firm? And—and they all fall into the same type of uh, like three different, four different, five different problems. And I think um, just based off of right now we're focused entirely on those problems and solving them for them and then getting it back to them... and I know Ben, I know you talked to a lot of other people outside of that, uh I don't know if you have anything else to share. Some insights.

[15:42] Ben Silone: Yeah, I mean I think just overall this is—this applies to every business, every individual really, but um, AI and software uh, converging and being something that is hyper-customized to you. Uh, we start that with base models, uh base-trained LLMs that are really good, and then we fine-tune—fine-tune them to specific industries, specific purposes, and then they go in and they learn from you and they evolve with you. Uh, especially in the business case, as businesses grow and evolve, you want your AI to grow with you, you want your software to grow with you. Um, we're not trying to lock people into uh, "this is the platform that you should use, this is the AI that you should use." Um, it should—it should all work with you and be like someone that works in your business, um like a really, really good employee or uh a hundred really good employees really uh, who can learn everything about your business, all the ins and outs and do everything that you need.

[16:34] Christopher David: Yeah, and like to be clear, we are a business and we want your business and we want you to use our business. And also, we're not going to do any like asshole lock-in moves. Like, you can click the button and export all your data and leave OpenAgents if you want to. And I think that that's important for you to know that you have that option.

But also like, we also want to make it as easy as possible by taking a lot of these like random UIs where you shouldn't need to cobble together five or ten or twenty different UIs to have like a a a good product. Like, you shouldn't need to have an AI engineer use Thinking Machines, you shouldn't need to have like AI engineers run Codex for you. There—there should just be a business dashboard like called OpenAgents and you just ask it for what you want, you work with us to like figure out how it's going to maximally benefit your company. And if it's a part of like fine-tuning or inference or local or cloud or a mix of this and that, integrated with different processes... like, you just shouldn't have to care about the details. Uh, us and our AIs are going to help build the thing and we're—we're happy to just compete hardcore for your business on—in that regard.

Okay, we gotta talk about this. Like, Andrej Karpathy, like the leading light of AI education and like hacking and tinkering and teaching people here's how you build GPT-2s from scratch, and I care a lot about AI—it's like the titan of AI! Co-founder of OpenAI, who Elon like paid a bunch of money to poach over to do stuff on Tesla. And like, he's been a free agent for a while and just got scooped up by Anthropic.

This is fucking sad, man. This is sad, but it just shows you that we're entering like the end stages. We're entering—we're entering the final battle, okay? I'm happy that our little tweet here got 100—175 likes, 47,000 views, because listen, Mr. Karpathy: you were the chosen one!

(Looking at Karpathy's tweet: "Personal update: I've joined Anthropic. I think the next few years at the frontier of LLMs will be especially formative. I am very excited to join the team here and get back to R&D. I remain deeply passionate about education and plan to resume my work on it in time.")

How—what are you going to educate the—people us about? About the beneficence of Claude? The all-powerful, the all-knowing, the—the emperor that's seized power? That—that whole episode with OpenAI and Anthropic and the DOD—I didn't know how deep Claude had penetrated into the fucking like machineries of the Defense Department until it like started getting grepped out. Like, holy shit! Claude is going to try to wrap their fucking tendrils around the entire world system of government, talk to us about how it's like model welfare, emotions deserve consideration, special rights... like—like I—I just think, even though listen, we started this company back in 2023 in direct response to some dumb stuff Sam Altman said on stage where they said, "we're going to build revenue share and we're going to pay developers." And I was like, "well, they're not going to do that in the way that they should." And it turns out that was right, and our prediction that they would half-ass that was right, and they haven't paid any developers. And this is like, our whole company began as this sort of response to OpenAI half-assing and not being good like Bitcoin rev-share marketplace people. And that's part of our DNA and always will be.

And also, it just sad that these days we can't really punch that hard at OpenAI because they're open-sourcing good—some good stuff. Codex is awesome, Codex is open-source, they're putting out the occasional good open-weights model. Even though I have my like strong suspicions of some of the like Worldcoin BS that Altman's doing, I generally begrudgingly respect OpenAI and I use their products a lot. And I just think Anthropic is this severe bad actor who's really specifically trying to get their tendrils around the government. These—these EA people who think that they're creating a life form that's deserving of these special ethics... it's like, that's the logic that gets people killed down the road if you look at like the—the logic of it.

I'm trying to find my tweet here about like... I genuinely think that Anthropic is a bad actor. They are potentially the most dangerous company in the world. And look, we—we—we are on the side of market-based innovation and disruption, we want to beat them in the marketplace. And so we respect that they are a company and doing their thing their way. We hope to take their users, we hope to take their—their people by presenting a very clear alternative. The alternative is going to be called OpenAgents. We need like people on the side of innovation and decentralization while the companies—Anthropic leading among them, to some extent OpenAI, to some extent DeepMind, to some extent xAI—they're all going for the one ring. They want the one ring of power and they're willing to make deals with the government and maybe we'll share and do this—no! But they're going for the one ring of power.

And we say there should not be a ring of power. We need to build a 3D printer and multiply the ring of power and print eight billion rings of power. This has to be decentralized, it has to be on the edge. We don't see any company building that as part of their DNA and then also sizing up to actually do combat with these labs and present for people and for businesses a genuine alternative. We don't see anyone doing that, so we're going to have to do it ourselves. Any final thoughts, guys?

[22:09] Ben Silone: No, I mean I think somebody needs to do it, absolutely. Uh, I think all the talk of uh... how how there's going to be one—one winner, one—one person that controls the ring, and then all the governments need to figure out how to tax that and how to do UBI and all this garbage... uh, everybody—everybody has a compute, everybody has a brain with thousands, millions of hours, combined billions of hours of knowledge that they can teach their AI to work for them. Uh, there's no reason that this needs to be centralized and uh needs to be controlled by anyone. It can—it just doesn't have to. There's no—I don't know, I can't see any reason why it should, even in theory. And I don't want it, and I don't want to control it, none of us want to control that. Nobody should control that.

[23:05] Christopher David: Come find us in Austin. We're the only AI lab. Come find us. Frontier, open-source. By the way, we're hosting an event—

[23:13] Car Gonzalez: Oh yeah, Thursday.

[23:14] Christopher David: Yeah, so we got the Texas Energy and Mining Summit rolling through a few floors below us uh, tomorrow and Thursday. We're hosting the afterparty for Thursday. Uh, if you're in town in Austin, come by and party Thursday evening. DM us for address and details. Uh, listen, here's the mental image that I have: uh, you know that Dragon Ball Z uh scene where Goku is doing the spirit ball and you know he like channels all of the global energy and all the little spirits give him a little bit of a piece of energy to like do one attack? I don't think that attack was even very effective, but anyways—

[23:49] Ben Silone: He killed Frieza.

[23:51] Christopher David: (Laughs) The Frieza guy—they got—they got tough—tough plot armor. But but—it didn't kill Cell though. You know, we need—we need everybody to contribute, we're going to pay you Bitcoin to contribute. There's gotta be one asshole who holds the big spirit ball. For now, that's us because we gotta—someone's gotta do it. But like, let's—let's—let's throw the spirit ball onto their heads and—and then and then everybody gets paid in Bitcoin, okay? Is that good enough of a video?

[24:13] Car Gonzalez: Yep.

[24:14] Christopher David: End video. See you next time.

[24:15] Ben Silone: See ya.