Anthropic Outages Vs Morpheus Marketplace | Everclaw No More

It's been a hell of a week for open source, for announcements, for developments in the space. You guys haven't been paying attention. It's basically something new every single day. So Bo, where do you want to start, man? There's all these new models. We can talk about that. There's all the drama with Anthropic. Yeah, I was going to say, I think the best place to start is down goes Claude again. So yesterday, obviously, there was a really large outage with Claude affecting tons and tons of people. I know I was testing it for a while to see how long it took to get back up. And so it prompted me to take a look at the number of reported outages. So in the last 90 days alone, according to isdown.app, which is obviously the authority of if something is down or not, they have had 129 reported incidents with Claude over the last 90 days, with 41 of them being considered major outages in entire regions of the world and a median resolution time of one hour and six minutes. Which, you know, for all the talk we do of data privacy and access to intelligence, which are, of course, key tenants of what Morpheus and decentralized AI provides us with, something we don't always focus as much on and probably should, and this kind of brings it to the forefront, is the ability to have a network with no downtime because you're not reliant on a central server anywhere. Yeah, I mean, we should also say they were not only down yesterday, they were down Monday, and they were down Sunday, and they had downages on Friday and Saturday. I think I looked back at their status bar, and in the last 30 days, they've had a major disruption of service, 18 out of 30 days. So, you know, it's a bit ironic where, you know, they're sort of going out there with their new model and say, look, our model is so amazing. It found all these zero days in, you know, these other pieces of software. It's like, dude, you can't even keep your own system online. It's like, what are you talking about? And so if you're going to build anything serious in AI, like 18 disruptions in 30 days is completely unacceptable for anybody building a serious product. I mean, if you go to most industries, they want three nines, meaning 99.9% uptime. And Thropic has 1.9 right now. They're 98% uptime, right? And so, like, that's not even close to good enough to build serious products. And so to your point, Beau, it's not just cheaper. It's not just private. Morpheus can give you a system that won't go offline all the time because you have different providers. You can switch models. You can, you know, have transparency into how they're working. You don't have this black box that just randomly turns off or becomes dumber. That's the other thing we should talk about. Like, the benchmarks are pretty crazy. People have recorded 4.6 has gotten something like 40% dumber the last two weeks. For whatever reason, they've been cutting back on the thinking. And, you know, they're not only making it more expensive, but, like, they're degrading the performance. Like, how do you build on that? Absolutely. I want to linger. I do want to talk benchmarks because that actually ties nicely into another topic I want to make sure we cover today. But before we get to that, I do want to linger on the downtime piece because, to your point, I mean, it's, again, like, we obviously make the Bitcoin analogy often or something like Ethereum. If you're relying on your agent to execute important tasks for you, which is, of course, the goal of this, right? It's like you want it to actually be able to do useful things that are ultimately important, whether the you in question is literally you as an individual or, like, an entity that you represent or even maybe a nation state. You don't just need a system that won't go down because they're good. You need a system that theoretically really can't because it has redundancy and no one can stop you from accessing it. So, for example, I mean, if one, technically speaking, if one laptop somewhere on Earth is running Bitcoin, Bitcoin survives. All of it. Whatever. A full node, right? Granted, that would be maybe hard on the laptop today. But the point persists that that is the type of system that we ultimately want to build towards. And even if so, if we look at just the private enterprise component, why would you want to even take the risk? Like, let's say, OK, let's say Anthropic did have, you know, three nines. Let's say they got to 99.9% uptime under regular conditions. They have the ability to give you specifically not that. And the other parties who they want to have that can have that, which also poses a problem. What if you're doing some type of, you know, unfortunately, my brain like exclusively works in finance terms. But the first thing I think of is if you're, you know, a fund trading against another fund and I'm using AI as a tool, but the other fund can pay them more money to have better access to data than I can. I'm at a perpetual disadvantage. Or if they can even shut me down for 10 seconds, that's a problem, you know, and like now apply that to health care, apply that to any plethora of things. Like, it's a huge issue. And the last point that I like to make, and I used to bring this up all the time, I feel like I haven't talked about it in a while, is I personally believe that Morpheus is a national security tool. I think that the government and the government agencies should really take a look at things like what we're building and consider incorporating it when they're building their AI stack. Because it's not, you don't have a central point of failure. Like if someone's trying to attack your network, well, where are they going to attack? There's no physical specific location to even go after. Yeah. Well, I mean, it sort of harkens back to the early days of the internet, right? ARPANET, you know, exists because, you know, they wanted to create a network that would be robust in case of a big war. Or, you know, any one particular city, you know, was bombed or fell, it wouldn't bring down the whole network, right? And that ended up, you know, starting to become the backbone of the modern internet. And eventually it was everywhere. And so, yeah, like there's no CEO of the internet. There's also, you know, not a lot of centralized risk, right? I mean, you could name out some of the providers like Cloudflare, AWS, but they're running one piece, right? One set of data centers, one set of nodes. But because you can switch to anybody else, it's really anti-fragile. So, yeah, it's worth making the point that AI should be similarly anti-fragile. It shouldn't rely on one or two big players and they have a hiccup or a bug or a leak and all of a sudden everybody's stuff goes down. So, yeah, no, it's like Anthropic has just been gifting us users every day. Yeah, I know. And again, right? Well, this is where I knew that one advantage that decentralized AI has over something like DeFi and something like Bitcoin did is the systems are being built in parallel with one another. Whereas like DeFi and Bitcoin was kind of looking to jump into and replace pre-existing systems. Now we're getting to see what it looks like when it's built in parallel. So every time something like this happens, more and more people awaken to the fact that, oh, this is bad. There's a solution for this. And, you know, ultimately, of course, we believe that that's Morpheus. Yeah, 100%. So what should we get into first? There's talk of 4.7. But, you know, my hot take is Opus 4.7. Honestly, it just kind of feels like they degraded the performance of Opus 4.6 so much that then they just gave back the existing model with, you know, slight improvements and a new number. Right. And so somebody was doing some analysis in it and it looks like they did something similar when they did the launch of 4.5. I don't know if they're like trying to make the model look better or like give people this experience like the old model doesn't work. And so they better shift to the new model because apparently they've enabled more thinking. They want you to burn more tokens. And it's sort of like if your experience the last week was, oh, this thing isn't working. Oh, thank goodness they released a new model and this one works. It kind of feels like this bait and switch that they're doing for marketing purposes. I mean, that or they're just incompetent and can't keep it online at the same level of performance. But, you know, I tend to feel like it might be marketing. I'd love to hear Kyle and Tomas's stances on this because I think I'm too normie of a user to be as impacted by some of this stuff as some of you guys are that are like more deep in code. Yeah, definitely. So I'll come out first and say that I was wrong last week. Last week I came out and said that we think Opus 4.6 is dumber because they changed just our thinking settings. But after going in and manually adjusting my thinking settings, it was still dumb. So, David, I think you were spot on that Anthropic must have been doing some quantization decreases or some sort of limiting over the last two weeks because I think fundamentally the performance has dropped. They also have had a lot of issues on keeping it live. So I assume that's something with splitting their GPUs to try and deploy 4.7 while also running 4.6 and having problems. So I think it's probably a mix of those things. I have not tried 4.7 yet, but from some of the early feedback I've seen, yeah, it looks like some incremental improvements. It probably looks like cranking that quantization back up to where 4.6 was originally. But I think the biggest takeaway from the new Cloud launch is their tokenization and that the same amount of output is now resulting in higher token counts and higher charging. So it's pretty much I sent a meme to someone earlier. It's like the trade offer meme of you get slight improvement. We get 35 to 50 percent more revenue. And I think that every single time this happens, people are going to get more and more pissed off about their monthly token costs on Anthropic and consider the open source alternatives, which really are getting closer and closer and closer to that frontier model intelligence benchmark. So I just started using GLM 5.1 like you have, David, and I've had really, really great results coding so far. Unexpectedly great results to be honest. Yeah, no, I've been impressed. And it's helpful, Beau, to understand something that, you know, people see these benchmarks. A lot of them are what's called like one shot. You know, basically you asked it to do something and did it get it right the first time. That increasingly doesn't matter. Right. How these models work these days is the model does an output. Okay. It tried to make the website. Let's say it got 60 percent of the way there. Okay. Well, it's going to look at its output the next 10 seconds and say, oh, well, I didn't get this part done. I'm going to go ahead and fix that. Okay. Now it's 75 percent done. Well, it's going to do another turn. It's going to do another cycle in the next 10 seconds and say, oh, okay, let me fix this, this, and this. And it fixes the last three things. Maybe it's 95 percent of the way there. Great. It did it once more. Now it's on its fifth or sixth shot and it's 100 percent perfect. Right. I don't need it. And maybe that whole process took a couple of minutes. Let's say that even took five or 10 minutes. Like I didn't need it to work perfectly in five seconds. Right. It took me longer than five seconds to come up with the idea for the website I wanted to call it to build. Right. So we're crossing this chasm, like Kyle is saying, where once the open source, and I think 5.1 is it. But it's good enough that it pretty much works. And if it has an error, it's going to go back and do another cycle and fix the error. Right. And I don't think people are that sensitive to whether it took, you know, two minutes or four minutes to get a perfect result. Like I'm going to go grab a coffee and I'll come back and it'll be done. Right. And why would I pay 10x as much, right, to get it done, you know, slightly faster. Right. So we're kind of crossing this chasm where if it's good enough, okay, it'll take a couple extra cycles, but it'll get the same thing done at basically the same quality. That's been my experience with 5.1. So I'm happy to say I can't officially say now I've cut the cord. I haven't used Claude, I think, in probably almost a week, almost a week since we last talked. I don't think I've used it since Friday or Saturday last week. Well, to be fair, there was only about 15 minutes of uptime that you could have been using it. Well, and that was the forcing function. A bunch of people flooded into my DMs and they're like, oh, my claw doesn't work. I can't, you know, do all of my, you know, normal work tasks. How about that Morpheus thing? Like, how do I get that? And I couldn't, I can't tell you how many people I gave app.more.org, right? And they're like, okay, how do I get more? How do I stake it? Oh, you know, go here or download this skill. You know, and 20 minutes later, they're like, okay, great. I'm all set up. I'm switched over. I'm set. And they've got all their agents stake. So it's really cool, you know, the hard work that Kyle and Alan and Thomas and everybody else did to get the API and the whole process smooth is like just starting to pay off. And so that's really rewarding. It is. Well, and the other thing, too, I was just thinking as you were talking, the example you gave was you're specifically, you're specifically prompting and then waiting for a response. But if it's like a recurring task anyway, you won't even know necessarily that it took slightly longer than it maybe otherwise would have. Yeah. If it's going to run a cron job at, you know, 2 p.m. on a Tuesday, you know, okay, it took an extra 30 seconds in the back end. You know? Yeah. You won't even necessarily see it. Yeah, exactly. Sorry, Thomas, did you have something? No, I was just going to say, I think what matters most isn't so much the one shot as it is the number of interactions that you have to have as the whatever you want built gets built. Right. So if I describe what I want and then it takes another two minutes or five minutes or whatever, right, because there's more tools being called, which could lead to a better output. As long as it doesn't come back and say, hey, can I do this? Hey, can I do that? Which Claude is increasingly doing now. And I don't know if that's one way for them to sort of better organize the cues of users that are using this system. But if you can just let it run and it doesn't interact with you and to the point, you know, David made where it works its own iterations, then you have something that functions a lot closer to the way the human brain works, right? Like you don't write a novel one shot. You start it and you read it. You think of things you could improve and you go and iterate on it. And that's how you get that output that you like and you're happy with. Yeah, that's a good point. It's about like, does it get done what you want to get done without, you know, having to come back and, you know, sort of redo the work and, you know, have you explain it and do a bunch of interactions. Yeah. Yeah. And, you know, just to go back on the points that were made earlier on Anthropic, I wonder how much of what we're seeing right now has to do with them prepping for the IPO. Because, I mean, it's clear that the usage, whether you're on a pro plan or a max plan, is significantly subsidized. You burn way more than 200 bucks worth of tokens when you're on the max plan. How much effort are they putting in right now to tighten that up so that the pitch deck doesn't show a big gaping hole on the bottom line? And they start showing progress, right, in terms of them reducing that incentive that they're putting in front of everyone with the pro plan and the max plan. Just that trend is going to have to continue, right? At some point, they're going to have to start generating revenue, which I don't, I can't, I cannot believe they would at, you know, 20 bucks a month on a pro plan or 200 bucks a month on a max plan. Yeah. Yeah. Well, that's the other thing is, you know, if they're going to get to go public, all of a sudden the market's going to know, you know, are you making a profit or, you know, are you just lighting a lot of money on fire to get market share? Right. And, you know, that works for a while and then eventually the market wants you to be profitable. Right. And that's when the end certification begins, right? So now you start getting ads and all of this. So I think making the switch now to open source models is the right move to make because you're approaching that parity and you're not being essentially sucked into that trend of we're going to make a profit. And because of that, we're going to start, we're going to start using more data that you're sending to us than we already are or pulling an open AI and pushing ads into the output. Yeah. A hundred percent. Hey. I see Ken. Yes, Ken. I see you have your hand up. We'd love to hear from you. Hey, everybody. Hey. Hey. Yeah. So I was just, you know, catching the news this morning as well. And hearing you guys talk like a couple of comments. It feels like I saw. Remember in like 2023 and 2024, we were talking about how they were never going to give us the best model, the centralized options. Like they were never going to do that. It was always going to turn into the iPhone where, well, that's what I initially thought where it's like, they'll give you a little bit better camera, a little bit better battery life. And they'll just iPhone this so that they have, they can just continue to squeeze profit out for the next 15 generations. David, what I wasn't anticipating is then pull up a page from like the job numbers, the job number data where they push out a big number and then they start like retracting, you know, later. Like they start dialing back the capability. Like, oh, that's it. That's it. I didn't I didn't even think that they were going to do that. So it's like good on them for getting that creative. Yeah. Well, it was what was it? Apple got caught basically engineering obsolescence where at three years exactly the phone battery would die on purpose. They got caught, you know, red handed and lost a big lawsuit over that. But yeah, no, I mean, that's what they're always going to do. Right. Their goal is to extract as much value as possible. And that's kind of where they're at, because I don't think they have a defendable moat. Right. The open source models have caught up. Right. And so they're going to try to, you know, they're back with the scare tactics. Right. Like, oh, our model is so powerful. We can't possibly release it. It would do all these cybersecurity things. And, you know, we've only let a privileged few, you know, handle it right now and, you know, access it. But, you know, I mean, I think a lot of that, again, is is marketing. Yeah, I would think so. And if they are going to IPO, too, I mean, it's a race to IPO between them and SpaceX, who now wholly owns XAI. So, yeah, they need to, like, extract every single cent of value that they can and either beat them to market or try to spoop some type of moat that may or may not be real. Yeah, exactly. But, you know, I think the timing, like you said, is great. You know, Ken, that stuff we were talking about in 2023, 2024, like, it's reality now. Agents are mainstream. People are using this in their daily workflows. They depend on these tools. And between introducing government ID, now required for Anthropic subscriptions, apparently, using and hoovering up all your data, the introduction of ads, racking up the cost, banning users that use a tool like OpenClaw that they don't like. I mean, like, this has become really sort of a hostile relationship with their user base, right? And, yeah, it's good that we've got Morpheus and, you know, it as an alternative. Because I can't imagine if three years ago Morpheus hadn't come along and built all this out, we'd be in a really bad spot. Like, what's your real alternative if you didn't have open source inference and open source agents, right? You'd just be, you know, going between one corporate stack or another. So, you know, it feels like what we built is definitely sorely needed and has come and matured just at the right time for people to have an option. Yeah. I don't know if Bo or Thomas said it because I had my phone down when I was listening. But when you're talking about the subsidization of, like, the pro versus the normal subscriptions, that was my aha moment, like, a couple weeks ago. And I've been, like, stacking Morpheus because, like, it's already expensive, you know, to use their product in their subsidizing. Like, wait till they don't subsidize anymore. Like, what the cost is going to be. Yeah. A hundred percent. A hundred percent. I mean, if you look at, you know, just the cost of renting GPUs and how many GPUs you need to run one of those models, the math just don't work. They work at 25 bucks a million token. But then, okay, well, let's say that's the price. That means you're burning, what is it, two, sorry, eight million tokens on a max plan? Like, what? No. You do that in half an hour. So they have to subsidize you. Well, I see we've got Misha on the call as well. Maybe we can get him promoted. I know he's one of the folks that Anthropic was clamping down on and made the switch recently. So I'd love to get him up here if you want to talk about that experience. You know, it's definitely something we're seeing all over the place. So, you know, the market's really trying to, starting to respond. Yeah. While we get him up there, just as far as how much are people going to spend on Anthropic when that switches over, I think we're probably a couple weeks away from really finding out after they shut down the OAuth connection within OpenClaw. And if you wanted to continue using it, you have to use the, you know, the metered usage essentially or pay as you go. So I think when those bills become due from everyone who has been running it in that manner, we will start to see exactly how much that costs and get some of that outreach. But excited to hear from Mika or Micah. Yeah, I'm going to add him right now. Send invite. Cool. Yeah, he said he's working on getting up there. So, yeah, I mean, that's, I think, the big story. So we saw, and just to go back to the government ID stuff, you know, I can't say how onerous this is, right? I mean, introducing KYC to use an AI is crazy. You know, this should be an open informational tool. Imagine you needed, you know, your driver's license to get an email account, right? Or your driver's license to, you know, use the internet or go to a website or go to Google. You know, it's very draconian. And what it does is it means they're going to block large parts of the world. Like, oh, we don't want users, you know, from Africa. You know? Oh, maybe they burn too many tokens and they, you know, can't pay enough for the subscription or they've got some policy that's blocking a part of the world. You know, Latin America or Africa or something. They could just turn off access. Oh, I'm sorry. You don't have a DMV, you know, license from New York or Texas. Get out of here. So, you know, I think the introduction of that stuff is just going to push more and more people to the open source option. But Misha, I see you're up here, man. Good to see you, bro. Yeah, hey guys. I'm Micah Winklespect and been in the blockchain crypto, primarily Bitcoin movement since 2013. So, I ran a company called Gem and sold it to a company called Blockdaemon. And now I've been spending all my time going back to my roots and writing code. And I've been a heavy user of OpenClaw as soon as it started, kind of changed my world. And have been using, you know, primarily Clawed, Opus, and Sonnet as my engine for writing code. And I'm a heavy, heavy user. And I was on the, you know, the max plan $200 a month. And when they shut me off on that just recently, I was in the middle of a project and I was so productive, I just couldn't make a switch. So, I just paid the toll. And just in the last, like, couple weeks, I'm up to, you know, $900 in API spend on top of my $200 plan. And I'm just spending, like, you know, an insane amount of dollars a day. So, it becomes, like, very expensive. And I don't know how somebody, I mean, you know, I'm blessed to be able to afford it. But, like, I don't know how, like, an average person can spend, you know, thousands of dollars a month on this plan. So, I think you guys are building something that's really interesting at the right time. I've heard David, you know, showing Morpheus in Signal for a long time. And now I'm finally checking it out. And I've got GLM 5 running as my primary agent. I'm still going to probably send my heavy coding work to Opus 4.7. But I'm curious to hear what you guys think about, like, really deep coding tasks with these open source agents. That's great. Open source LLMs. Awesome. Yeah, glad to have you here, Micah. Yeah, no, I've been talking about Morpheus for a while. But I think it's really hitting now because people can see the practical use. My feedback would be GLM 5. I find good for documentation and a lot of the lighter stuff, especially writing. But I would really encourage you to check out GLM 5.1 for agentic engineering. So, for most of the heavier code stuff, I'm using GLM 5.1. And that just came out last week. And so, it's very, very recent. And it's right up there, I would say, between a Claude 4.6 and something like that. Well, pre-Claude 4.6 is getting a lot dumber. So, it's probably actually benchmarking above Claude 4.6. And I find I can get the same things done if I'm articulate and descriptive on what I want done and how I want it done. And I would say I'm getting the same results. But maybe, you know, something that took me 20 or 30 minutes is taking me 30 or 40 minutes. So, slightly longer set of cycles. But I'm getting the same quality. And I would encourage you, if you want to save a lot of money, what I'm finding is I can code perfectly well with 5.1. And then I'll just have a model like Grok do an audit. And so, I'll use 4.2, which is the best cybersecurity auditing model in the world right now, even better than Claude, for just auditing the code. And so, I've included that in my process. So, I'll have, you know, the cheaper agent, you know, GLM 5.1, do most of the coding. And then it hands off the coder, copy and paste it over to Grok. I ended up just giving it the API key to Grok and saying, you know, on the coding step, go and get it audited by Grok. And it'll go back two or three times. Grok will find errors that it didn't see. And that's a big unlock. If you do cross-model auditing, you get much better quality code. Because the models are good at passing their own tests. But if you take it over to another model out of context, it'll find a lot more bugs. So, I basically get it to go back for three or four times and it ends up with really high quality, but at a fraction of the cost. So, yeah. We're trying to use it using adversarial reviews. Often I'll have my, you know, whatever I'm working on. I'll even use the same agent and do an adversarial review. I agree that they're probably better at passing the test, but I still find it really helpful if you just tell the agent to attack its own work from the other side. Yep. I think you made a really interesting point, David, on kind of giving it the right requirements, right? I think that Opus, one of the things that Opus, and not even any version of it specifically, but overall as a model, is very good at working in isolation, right? But you don't necessarily need that trait of the model when you have a coding agent running, or at least a skill that facilitates agent-like experience with it, where set the requirements, create an implementation plan, create tests, run through the implementation, run through the test, validate it, right? When you go through that process, you're able to use something like GLM-5.1 more easily because the structure is there, right? It doesn't have to be inherent in the model weights. It's able to go through and run that process and get the results and iterate on it, rather than the one-shot result that you're looking for, or that Opus was maybe trained on a little bit better. So I think it's a really strong comment on maybe how we're doing things is going to change a little bit, and maybe we're going to act more like software engineers, right, rather than vibe coders. Maybe we're going to get a little bit more there, rather than, hey, build this thing to, hey, these are the requirements, build it, and the coding agent can kind of help facilitate that, which will allow us to use these open source models more and more and more until the market kind of converges there. Yeah, well, and maybe a good word for that is a harness, right? Right, I've ended up, you know, building a standard operating procedure for coding, right? And it has 10 steps, right? Before it starts coding anything, it does the research, right? And sort of grounds itself in reality, like what's available, what are the tools, what are the best practices? Okay, here's the research. I'm going to build an architecture based on this research. Then when the architecture and the specification is clear, okay, go to code. And what I found is if I spend five or 10 minutes on the first couple of steps there, I'll save myself an hour on the coding back and forth because it's not doing that research as it codes and realizes XYZ isn't available or whatever assumptions it had, you know, get blown out of the water. And so, you know, I really recommend to anybody that's using this, build your own process. I'll go ahead and open source my SOP. I'll post it on X. I think I've put it out there a few times, but now it's got all the stuff, right? You're saying, Kyle, we need to get to the point of basically agentic engineering. I think that's what we're going to get to, right? My process now includes not only testing, not only auditing the code, but I'm looking for regression, right? Did I lose functionality? I'm looking for dependencies. Did I introduce something new that the software needs to now know to download or include in the install? So you end up with all these best practices, document what you built at the end of the process, push it out, you know, get GitHub to build the Docker, all those steps. Once you've got that process down, you know, it's become easier and easier. I've gone from basically having to handle the whole process to giving instructions at the beginning and just approving outcomes at the end. It's getting really smooth. Absolutely. And the hour learning curve has reduced to zero because this can be as simple as injecting a skill file and slash start coding, right? And then it kind of just walks you through that process, right? That's where we're at now where, you know, one, our willingness to try new things and our willingness to, all right, we're going to go about this a different way. That's going to save me 90% per month. Like, all right, hell yeah, I can now, you know, develop three times as much, you know, output here, four times as much output considering my time that I have to harness it. But I just think we're in an amazing time in this and seeing this open source trajectory and the tools to support that advance so quickly is, you know, just really excited about it. And I think Morpheus is positioned incredibly well to kind of facilitate, you know, the infrastructure for that market. Hey, David, as the chief normie correspondent here, the people are asking, instead of an SOP, have you simply asked your agent to make no mistakes? Thank you. Oh, yeah, that definitely helps. But, you know, what I found is I'll tell it to follow the SOP and then kind of remind it. One of the big unlocks there recently has been memory. So once I got the memory in my OpenClaw properly configured, because remember, it does not come turned on. If you downloaded OpenClaw and you never turned on memory search, it's just sitting there and not getting used. So as soon as I got that properly configured and sort of really got the memory better flowing, now it's learning more and more of, you know, I don't have to correct it as much. Oh, I know to check for this. I know to go do the audit. I know to come back for the approval. Right. And so I'm finding myself having to give it less reminders. You know, so that's that's one of the things, you know, we talked a little bit about memory last week. I know we're going to kind of touch on it this week. There's been a couple of big things as far as Mem Palace and Neo4j had some announcements. And I love to get into that because I think that's that's one of the other critical pieces of this is not just the inference, but it's using the other open source tools that are out there to improve your context. Yeah. And the information you're giving your agents. Speaking of Mem Palace, so it's a great transition. Last week, our unexpected celebrity chime in of the week was the existence of Mem Palace. Although we don't have a specific unexpected celebrity chime in this week, there was follow up news or maybe even a little bit of controversy surrounding Mem Palace because of the the realities of what it is doing and its output relative to the benchmarks that it was publishing. David, you want to speak on that a little bit? I think that's a great transition to what you were just talking about. Sure. And there's a bit of a nuance here, right? Like, again, a lot of these benchmarks are, you know, strictly does it work if you run it once as opposed to run it a few times. And so that's called zero shot or one shot benchmarks. And initially they had sort of touted a benchmark that, you know, is sort of fair if you run it a number of times. Right. And it checks its own work kind of thing. And then you can get really, really high numbers. A lot of people, I think, to be honest, their competitors did not like that. And we're like, well, technically, if it's only a one shot, it would be, you know, way less than 100 percent, 90 percent or 80 percent or 40 percent. So it's hard to tell what of that what percentage of that is like their competitors don't like the tool because they open sourced it. And there was a bunch of stuff that you could use before, but it's all paid. So I think some of that might be competitive, you know, positioning. But I think some of it's fair. Like the team, Ben and his co-founder both came out and said, yeah, that's fair. You know, here's what we think are better metrics. And we ran it, you know, against these other benchmarks. And here's where it comes out. And, you know, it's an open source project. It's MIT, right? License. So a bunch of people offered PRs and improvements. And I know there's been a lot of iteration. I'd love to get the inside scoop. We have some of those folks in our community. And they're, from what I know, they're Morpheus users. So, you know, maybe we'll get the inside story at some point. But, you know, I think it's a cool experiment. And it caught a lot of people's attention. And for me, the big benefit is it forced me to go and look at how my memory was working, which it wasn't, right? And fix that in my agent. And then, you know, I rolled a skill around MemPalice. But, you know, I think these tools ought to be agnostic. You know, if you're using a big data set, you should probably be using Neo4j or something stronger like that. It's also worth kind of unpacking the technical side. What is MemPalice? It's ChromaDB, right? That's what they're using under the covers, right? Beyond sort of the schema they came up with and, you know, using the MemoryPalice sort of reference points. But it's ChromaDB, right? So what you're really saying is what's the performance of ChromaDB versus Neo4j or graph-divated bases? And there, I think it really depends on your use case. So I don't know if Thomas or Kyle have a hot take on this. But I'll defend the MemPalice people a little bit and say, you know, I think it's a good open source experiment. It's been useful for me. It's working for me, right? So, you know, quantitative stuff aside, qualitatively, like, it massively improved my agent. So, you know, I know what you guys think. To have mixed experiences with memory, right? So I started working on that maybe a couple of months ago. And it got better. But then eventually it got worse. So I think there's a scalability limit that we hit pretty quickly when we go with tools like a ChromaDB or even the QMD that Toby built, which is the searching through your markdown files. And I think the answer is going to be really looking at a combination of systems, at least for now. I really like the approach that the team had taken for MemPalice. But then, you know, would it really improve the results? After seeing all the controversy, I sort of paused the rollout of MemPalice within my own system to see, like, wait until it stabilizes a bit. Because I think it's easy to jump from one system to another to another. I tried one, stayed with one for a couple of months. It was amazing at the beginning. And now it just doesn't work as well. And so I want to, I think I want to wait until there's a little bit more sort of market feedback before I experiment with one. You know, the other side to that is how your agent actually utilizes the memory that they have access to, which I think is another problem and one that personally I've run into, particularly with OpenClaw. And I'll just read, you know, like a response that I got this morning from OpenClaw when I said, why do you keep forgetting? Is that information, the issue that we ran into? Not in your memory.md. And the answer was, the problem isn't missing documentation, is that I'm not checking it before acting. And so that's, I think, the first problem that I want to spend some time on, because you can have the most amazing memory system behind the scenes. If your agent just forgets to check it, then it doesn't really add any value. Yeah, no, that's good feedback, Thomas. I ran into something pretty similar. You know, you got to train it to follow that process of checking and searching its memory before it goes and gives a response. And so some of that ended up, you know, prompting me to improve, you know, the sole and the user files and stuff like that. So it does those checks more regularly. Yeah, they're in there. It just shows to ignore it. And, you know, I know that's a different discussion, but that's my number one reason now why I've switched from OpenClaw to Hermes. And I am using Hermes 95% of the time and OpenClaw is down to 5% to the point where I'm about to uninstall OpenClaw and really focus on Hermes. Because the other experience I had was asking Hermes a question and Hermes coming back saying, we actually talked about this not so long ago. Let me start from where we left the conversation off. And I thought, okay, that's a change. And that's a change that I like, right? It's not like everything got reset overnight. And now you have a new, you have a brand new intern that just started this morning that you have to train all over again. That's really interesting. I've seen Hermes is doing extremely well on adoption. If you track the number of stars on GitHub, it's now, you know, a top project. I think we're going to see more of that. And it's worth pointing out to folks, you know, this is one of the reasons this week I rebranded EverClaw to the Morpheus skill. I think EverClaw maybe sounded more like it wasn't an agent as opposed to this skill that gave you access to inference and decentralized AI. And so, you know, it's worth emphasizing the Morpheus skill works for Hermes, for OpenClaw, for IronClaw, for whatever agent you want to power. However, Morpheus works across all of them, right? And, you know, the Morpheus skill specifically, I put up a placeholder website, MorpheusSkill.com, and you can check out all the content. You know, but it's basically a series of MD files. So any agent you give it to is going to digest those MD files, make it part of their system, add those skills, and then be able to, you know, be powered by this decentralized inference. And so I think that's an important note is, you know, the Hermes folks have been involved in the Morpheus ecosystem almost since the beginning. In December of 23, Noose Research, you know, was fine-tuning models and contributing to Morpheus code base. And so big credit to those guys for continuing to iterate and moving beyond the fine-tuning into agent building, which is really, really cool to see. So it's funny how many of the projects in the ecosystem, you know, were there during the early days of Morpheus and have been involved. So, you know, it's really cool. You beat me to it, David. That was the next thing I was going to bring up was the rebranding of Everclaw to Morpheus skill and why that's important. Yeah. Well, I mean, I think it'll be more clear. I've gotten good feedback from people that that's a lot easier for them to get. And so, yeah, hopefully I'll be pushing out a new version today. I'm just working on a new version of the skill guard that's more intelligent. So, you know, Morpheus, you know, in order to use it, you sort of need this pack of skills, right? You need good security. You need a wallet. You need to be able to hold the tokens. You need to be able to store your passwords. So there's sort of all these complementary skills that go in to, you know, adding the Morpheus skill to your agents. So I'm excited about maintaining that. And, you know, it's been a lot of fun. Like every time there's a new version of OpenClaw or Hermes or whatever, you know, I'll roll an update and push it out to the community. But just gotten a ton of great response. And that's got the API gateway built in. So it makes it easy for people to, you know, never have to even sign up for Claude in the first place. They can just go straight into using Morpheus. And then as a bootstrap, key included. So if you set up your Morpheus skill, you set up your Hermes agent, you know, it'll automatically start using the Morpheus API and get you basically set up with your agent enough that you can go and get your own key. But there's no sort of cold start problem, which is one of the big things that I want to fix. Absolutely. That we ran through our whole list. We had a ton to cover this week, which is awesome. The only other thing I had on our formal list of things I wanted to make sure that we touched is just our eating our own dog food section. Or if anyone has like a specific thing maybe that they use DAI for or AI generally over the past week that they want to talk about. I like leaving some open space for that. I guess I can start just with a really quick one. Since I have now discovered how good my agent is at taking Excel spreadsheets and turning them into HTML dashboards that I can store offline and just access from anywhere. I've been obsessively taking like every single thing I use for my real estate business and turning it into some form of dashboard to mess around with. So not exactly a super crazy application, but it's been fun. It's been enjoying it. It's not a super crazy application, but it's a super valuable one. It saved me time. Yeah, like quite a bit actually. Everything doesn't have to be, you know, we don't have to change the world with every application. I think it's more of integrating AI and integrating these things into what you're doing on a daily basis is going to be like some of the low hanging fruit that expose us to. How do we, how do we, how do we augment ourselves with these tools rather than kind of one-off use them for, for super important things? Yeah, fair enough. That is certainly done for sure. It probably shaves an hour off of like a couple of different tasks that I've been using it for. Did I lose you? No, you're good. Okay. Sorry. I wasn't sure my screen flickered. Yeah, that was, that was all I had this week. I mean, is there anything else anyone wants to make sure that we cover while we're all here? Can I ask you a question? Yeah, go for it, man. Of course. I'm totally new to this world here with you guys. So can you, can you explain to me a bit about the backend of Morpheus and you know, the, the actual compute network? How much of that is, you know, bootstrapped by the founders? How much of that is, is people joining the network and just adding compute? Are you using like a cosh on the backend? Like what's, what's the backend look like? That's a great question. I'm happy to take that. So think of Morpheus, um, as an inference marketplace. So, you know, uh, there's no Morpheus company, right? It was totally fair launch. There's no founder pre-mine or anything like that. So everybody involved is an open source code contributor. Um, you know, folks like Kyle and others have put up infrastructure like the API gateway, which is all open source. A lot of the inference comes today from people, um, running Venice. So maybe they own DM or something like that. And they want to offer models on the Morpheus marketplace. They can list those models. So they just download a node, right? There's a peer-to-peer node system. And in Morpheus, anybody can download a node and say, hey, I'm a compute provider. And they could be running their own GPU, their own data center. They're using some third party. Um, but certainly, yeah, folks have used, uh, a cosh in the past. Um, though less, less now, um, more people have been shifting towards systems that offer the inference, not necessarily just, uh, the GPUs. Um, and so, Kyle, I don't know what color you want to add to that, but, you know, it's this open marketplace and people can, are mostly offering the open source models right now. There's nothing stopping somebody from offering a closed model. Um, but right now it's today. I mean, they're just so much cheaper. Uh, the open source models are, are most of the usage. If I look at the stats, I think the most popular model is GLM. And after that it's Mistral and some other ones, but, you know, Kyle, you could probably get into the details. Sure. I'll just give some quick insight into, I guess, how the system works from the, the actual infrastructure standpoint. So pretty much, yeah, completely open two-sided marketplace. Uh, we're talking more specifically about the provider side, right? So when you want to provide to the marketplace, you're going to host a provider node, right? That connects you locally to the blockchain to allow to, um, for other, you know, consumers to go and run inference against you, right? So within your node configurations, it's as simple as where does this inference come from, right? You're either going to set a URL, which could be localhost. It could be, you know, like David said, open AI, right? If you wanted to offer open AI for some reason that I have no idea of, or if you want a proxy to something like Venice or wherever else. So it allows you to go, you know, localhost connection, a connection to Akash, a connection to RunPod. And then, you know, if there's an associated API key, you can include that. Um, I personally have, uh, done a lot of bootstrapping in the system. Um, I've bootstrapped through Venice. I've bootstrapped through localhost on an, like an AI rig that, that I run. Um, and, um, I've found that although I love Akash, like I'm, I'm a huge supporter. I found that, um, Akash is good for some applications with some GPU types where other products like Vast AI or RunPod, or, or there's a million of them, um, are good for other types of GPUs. So I was hosting some text speech through Vast AI and RunPod. It, it's really non-specific, right? If there's a difference between capability and then what's happening right now, um, I would say, and sorry, this is a long response, but I would say right now it's primarily bootstrapped. And we expect that as more demand continues to flow into the system, providers will see that opportunity to go and capture some incremental revenue, right? Whether they have excess capacity from something they're hosting themselves, or they just find a nice arbitrage where, Hey, I can go host this model and, and charge less than someone in the system or, or they're overloaded. And I can pick up some of that extra demand, latent demand that's there. So, um, bootstrap now, and we, you know, not fully bootstrap, but mostly bootstrap now. And we expect, um, that to, to expand, you know, kind of starting immediately as this demand is really, you know, beginning to pick up. Um, is there, is there any arbitrage now, you know, if I wanted to help contribute inference to the network, is there like, uh, you know, cloud provided GPU services where, where there's enough incentive on top to run inference on this network or, or maybe assuming a little bit of growth in the network or, or do you really need to run your own hardware in order to make it viable? Uh, I would say right now, um, so we could talk separately on some of this, but right now there's probably not the largest opportunity. Um, because through some of my bootstrapping, the pricing is extremely low to incentivize demand. I would say this is going to be a little bit of a, um, of a, uh, cat and mouse type game, right? Where, when, um, you know, there are some limits on my side to, to what I can bootstrap, then the, the next layer of pricing will be much higher. And I will similarly be able to, to increase the pricing accordingly to, to be able to provide more. So I would say today let's, you know, let's get you in there and we can, we can figure out how to, um, configure the pricing properly and, and figure out if there's any incentives to be, to be made there. But, um, I would say current step right now is we're looking to drive as much demand as possible. And, um, from the provider standpoint, that will, that should immediately follow. Well, and Micah, to answer it in a different way, the arbitrage is the $25 per million tokens that Claude is charging versus the $5.50 per million tokens that, you know, um, it costs on, on Morpheus. Right. And so anybody that can come along and convert people that are using Claude today over to, you know, GLM 5.1 or similar models, there's like an 80% arbitrage there. Um, Kyle's talking about the arbitrage, like between, you know, uh, the cost of running the machines and the, and the various providers on the network. But the much bigger arbitrage is, is converting over those people that are paying through the nose for Claude today. And there's a whole universe of resellers, affiliates, people that are, you know, ought to be building products on top using that, that cheaper inference on Morpheus. And as far as their users care, you know, it's 50% less, it's 60% less. And whoever's building that interface, you know, is taking some of that margin. I think that's going to be where a lot of the, uh, demand comes from is people building stuff on top. Yeah. I think my question is more rooted, not, not so much in my personal game, but rather trying to understand whether the model is sustainable at 80% discounts. Like, um, you know, aside from, aside from the founders bootstrapping and subsidizing. Right. So, um, I think that's. Well, again, that's sort of what's unique about Morpheus. There are no founders, right? There was no pre-mine, you know, um, the original white paper authors were anonymous, Morpheus, Trinity, and Neo. Nobody's heard from them since 2023. Right. And so in the absence of a treasury, it is sustainable, right? There's nobody subsidizing that 80%. That's the natural 80% less that it costs to run the hardware and the electricity for these open weight models versus what you have to pay, you know, uh, an Anthropic or a chat GPT. Right. Because there's no license. It's like saying, you know, Linux isn't more compelling because it's subsidized. It's more compelling because there's no Microsoft sitting on top charging $100 a seat for a license. Right. So you're just ripping out that, that license and copyright structure. And you've got an open weight model that nobody's charging a license fee on. Yeah, um, I, I agree. I think that, um, I think on both sides, on the consumer side, there's a clear, um, you know, arbitrage is probably the wrong word, but, uh, uh, discount for the open source compared to closed source. I would say on the arbitrage opportunity as a provider, um, there will, the, the arbitrage will come mostly from posting local, you know, or self-hosted options as compared to someone who just comes in and, and resells open router, but it's going to be a margin game. It's going to be a, uh, what are you offering? And is anyone else offering it? It's, um, you know, no different than the dynamics of any, um, kind of open market. Yeah. Community is basically doing price discovery, um, not bootstrapping or, uh, we're, we're past the, the, the early bootstrapping where, you know, um, it was, uh, experimental and, you know, uh, is come, come build this. Now it's come, come use it. Absolutely. Well, I'm giving it a shot this week, so, uh, I'll let you know my feedback on the signal. Cool. Awesome, man. Good stuff. All right. Well, uh, Bo, thanks for MCing as always, man. This has been a great session. Again, it's so cool to see, you know, people getting real utility out of this. And I think this is just the beginning, right? This is the worst the models will ever be like we're in, in a week or two, we're going to be talking about, you know, deep seek four or whatever the latest, greatest model is. And, you know, it's going to continue to sort of erode, I think, uh, the lead that, you know, Claude and others had. So, uh, it doesn't matter if you're using it for price or privacy or reliability. You know, um, it's really cool that, you know, so we're at this moment right now where we've crossed that chasm and people are getting real daily use out of it. Yeah, absolutely. I think, well, and, uh, thank you for thanking me. This is a huge highlight every single week. I appreciate it. And everyone who joins and everyone who listens. Um, the, the thing that has just rung so true week over week, month over month, quarter over quarter since I discovered Morpheus, um, geez, over two years ago now, which is really crazy to think about. Um, is that all of the principles like the anti-fragile nature, the, on top of the fact that like the tech, the tech is improving to your point, this is the worst the tech is ever going to be. It's only going to get better, but then on the other side, the non-tech pieces like the freedom principles, the access, uh, to intelligence that's unobstructed by central parties, the ability to have a network that can be accessed from anywhere permissionlessly. All of the other aspects of the world, like geopolitically have also just continued to trend in that direction. And so like those stances that the community has from my perspective, just continue to get validated over time. And eventually those two things, in my opinion, will just completely converge. The tech will be at the perfect point, the values and like the recognition from the public will be at the perfect point. And, you know, then we're on a rocket ship for everyone to ride together. So anyway, it's been awesome to see and look forward to, uh, you know, continuing, continuing on. So thank you as always to everyone for joining us, whether you're a panelist speaker, um, or listening live or listening to the recording shout out to the mirror Morpheus guys for doing the, um, kind of remastered, uh, recorded versions. And with that, I look forward to seeing everybody next week. See you guys. Keep being awesome. Take care.