Decentralised Inference | Everything You Need To Know With Kevin Simback
1 hr 1 minAudio
Zion, hear me! For Season 2 of the Mere Morpheus podcast I'm going to be uploading the Morpheus builder weekly updates. My reason for doing this is to make it more accessible for those who want an update directly via their podcast app and to have them all in one place. If you'd like to participate they are held weekly on the @MorpheusAIs Twitter spaces. So enjoy Episode #81! Recorded: 9 July 2026 AI SUMMARY: This week's episode opened with a Morpheus network update highlighting node version 7.5 a key release for providers. It introduces provider self‑reporting where nodes test their configured internal models on boot and hourly reporting whether downstream endpoints are working. This gives providers better troubleshooting data and offers the network a clearer picture of health at the bid level without inflating throughput numbers with artificial checks. A milestone update noted Morpheus has processed over 20 billion tokens year to date with July on pace to be the first 10 billion token month. The conversation then shifted to a discussion with Kevin Simback of Delphi on decentralized AI focusing on four hard problems: sharded inference for large models model verification privacy across nodes and balancing supply and demand in the marketplace. The talk explored why decentralized AI grows in importance amid concerns over centralized provider dependence government controls privacy and vendor lock‑in. The episode wrapped with Rob sharing the evolution of Cashew into Morscan an open‑source scanner designed to give the community a transparent view into marketplace activity and network analytics. Timeline: (00:00:00) Intro (00:00:30) Node 7.5 rollout and provider self reporting (00:05:43) Throughput growth and a 10 billion token month pace (00:07:57) Introducing Kevin Simback and the DeAI discussion (00:10:05) The four hard problems in decentralized AI (00:17:37) Composability across DAI projects and shared progress (00:26:01) Who will adopt DAI first: consumers or enterprises (00:35:31) Privacy trade offs, regulation, and catalyst events (00:43:16) Projects to watch, traction and proving real usage (00:55:48) Morscan launch and better network visibility If you would like to support me for the value I provide you can use the referral link below when depositing capital and we will both earn bonus MOR. Alternatively you could send me a token of appreciation directly to my wallet ;) Referral Link: https://dashboard.mor.org/capital?referrer=0xE935f231c99c04Ee0b4532a3d0BdA81B152a0384 Wallet Address: 0xE935f231c99c04Ee0b4532a3d0BdA81B152a0384 Connect with Mere Mortals: Website: https://www.meremortalspodcasts.com/ Discord: https://discord.gg/dX9BEhT3kz Twitter/X: https://twitter.com/meremortalspods Instagram: https://www.instagram.com/meremortalspodcasts/ TikTok: https://www.tiktok.com/@meremortalspodcasts Value 4 Value Support: Website: https://www.meremortalspodcasts.com/support Paypal: https://www.paypal.com/paypalme/meremortalspodcast
Audio and video play straight from the show's own hosting. PodSubs stores and serves no media.
Transcript
From the show. Tap a time to play from there.
Bowen BelmerLet's dive into updates on the Morpheus side of things this week. There were a couple. So, Alan, do you wanna kick it off with the latest update to the node? I know there was there was kind of two back to back. So, of course, we can go with with the most recent.
Alan (Nomadic Rogues)Absolutely. Good day, everybody. The two back to back were rolled out seven four and then found that I was close but but wasn't exactly where I wanted to be with the with the provider self reporting. So seven five dropped this morning. So for all you folks who are providers, who are providing inference in the network, upgrade your nodes to 7.5 and I'll talk about what that's gonna give you. Really it's what it gives the network and
this is a result of conversations I've had over the last few months really about it. We have this we do have this problem. We are a decentralized network where we decentralized network that people want centralized behavior and reporting and control over. And it's a really tough dichotomy to to deal with. And one of those is, hey, I'd love to use the network but the provider or the model or the bid that I'm trying to go after,
I don't know the quality of that. Now we have a rating system, we've got a reputation system in place and I think that's good when you've got models with multiple bids on them that's supposed to help balance but it kind of assumes that the the nodes configured right, whatever that is behind the node is configured right and it leaves a lot up to to trust. And so far, what we've got is a kind of a a a status system. If you go on to active.more.org/status,
you kinda see, you know, the models that are on the network. And up till today or this afternoon, that was always governed if I can reach the provider, I assume that anything behind that provider is okay. And that was really a bad assumption. It was the best we had at the time, but we found, you know, certainly when people ask people are starting to improve their agentic work that that's not good enough. Period. Full stop. It's just not good enough for reporting status. And so
trying to skip to the end, what we created was basically provider self reporting. So with 7.4 and really 7.5, what happens on a regular basis on when the node boots and then on an every hour basis. It lists the internal models that you've got configured on your system and sends a very small check completion request from the provider to whatever endpoint you have. Now, doesn't create a session, it doesn't create,
any network overhead, doesn't require any gas, it just tests basically, hey, I've got, you know, my favorite a I dot com slash, you know, chat completion, can I reach it? Can I get a response back? And, you know, depending on whether it's an LLM or an embedding, which are the two tests we have right now, it it sends a simple 10 token, you know, what's two plus two answer with only a number kind of thing. And what that and then it takes that data and presents it on that external health check endpoint that we already have
that we've used for this this ping test. And what this does is really says that the provider self reports its downstream quality. It it'll give you any errors, like, say, there's a 400, a 500 that the downstream provider's not up and running, if there's a four twenty nine, if it's rate limited, if it's 402, four zero three, which means you don't have enough credits, it will help expose that. So so this does two things. First of all, it helps providers troubleshoot their own nodes. Right? Does it really work at least from the node downstream to whatever you have, whether that's a light LLM node, an open router node, direct to direct to Venice, direct to your your your iron that has the GPU,
but just tests, hey, are you good? It also tells you if you've got bids that you don't have models for or that vice versa, you've got a bid on chain that you don't have something configured. So for a provider, it provides a lot of feedback that allows you to troubleshoot. And then for the network, we consume that. So on on active.more.org slash status, what you'll see shortly is we can, on a regular basis, collect that data and give you an idea of what's going on in the network down to the bid level.
So much better it it it was a challenge of how do you act like a centralized information canonical resource on a decentralized network? Well, you have your node self report. Now, we don't expose anything bad. We don't give you we don't expose what you may have configured in the models config for what your endpoints are. We we reference the model name that's listed on chain, the model ID that listed on chain, and whether it worked or not. And so I think this is a big step forward
on provider troubleshooting and provider self reporting to the network. So 75, if you're a provider, please go out and update your node and active.more.org/ status. You'll be able to see a much better interface in reporting on a per bid environment.
Bowen BelmerAwesome. Thank you, Alan, for the update and team, everyone who contributed. Well done as usual. We continue to ship every single week and multiple times a week. This is actually like a two shipments back to back. One thing that I like to highlight on this, Alan, just from talking with you about it yesterday, is I I really like that it solves the problem of being able to verify uptime without having it need to count as, like, session throughput, which I'm gonna affectionately refer to as the Tron dilemma.
I want actual throughput, and I won't be able to brag about the the massive amount of of tokens flying through the network for actual inference, not nodes pinging themselves to make sure that they're still online Yeah. In the same way that certain hopefully, one in here is a huge Tron, Maxie, or anything. But in the way certain people, you know, report transaction volumes on chain, I I want to, like, verifiably not have to fight that uphill battle, you know, when we're talking about the network. So that's awesome.
Speaking of which, one other thing before I introduce our awesome guest here. Very much looking forward to having him on stage. Speaking of throughput, just a quick update as we do every week. These figures can be found at tech.moore.org and hitting the throughput button at the top. Year to date, we have now passed oh, actually, I'm sorry. Yeah. Year to date, we've now passed 20,000,000,000 tokens processed and growing. The chart is up into the right.
Right now, we're on pace for July being the first 10,000,000,000 token month, which is really exciting, and that would be a 36% increase month over month from last month. So continue to be up into the right, which is obviously what we wanna see. You know, the the strength is in the the distribution and the ability to, you know, capture real users of which we have many and and many people building things on top of the network that will ultimately attract, you know, the the end retail user. So
check that out if you're listening to this and you haven't already. It is a it is an awesome site just generally to to learn more about Morpheus and everything happening. So with that being said, I want to introduce Kevin Simbach. Kevin, thank you so much for taking the time to jump on with us.
Kevin SimbackYeah. Hey, everyone. Happy to be here.
Bowen BelmerAnd and just so you know, Kevin, so we do this live version, but then it's also recorded, and and one of our community members will kinda, you know, clean it up and put it out as well. So if you if this is very enjoyable and you end up wanting to share it, very much can. But I know we we really appreciate you taking the time, Kevin. You've been publishing a ton of great, a ton of great articles on your x account directly related to DAI. And most recently well, actually, you just published one yesterday. I saw decentralized training.
The stakes have never been higher. So first of all, everyone listening to this, absolutely go read that as well as the other works Kevin has been publishing recently. The one that kinda prompted the reach out from our side though, and that I'd love to dissect and just have a bit of a round table discussion with the members of the community that could make it today. Unfortunately, some of our core contributors are well, not unfortunately. They're they're in Asia right now traveling, going to events, and so the time schedule is a little tough with that. But that that being said,
your article titled decentralized inference, everything you need to know, I think was, putting it lightly, a awesome, awesome, awesome primer for anyone who wants to learn, like, what the heck is this decentralized AI thing that I see people talking about? Why does it matter? Why should I care about it? And most importantly, and I think this is actually the most interesting part was your identification of the four hard problems that need to ultimately be solved,
you know, in order for for any of this to be real and and kinda move forward. So I'd kinda like to start there. Kevin, where can can you just talk a little bit about, like, your thoughts generally about these four problems and kind of how you arrived at those being the key things that will allow for success for d AI generally as a concept? Yeah.
Kevin SimbackI'm happy to. And and I don't have anything prepared, but I can just riff for a little bit and then happy to take any questions that folks have. I guess just to give a little bit of context, the decentralized AI space for us, and I'm referring to my firm Delphi, which is a research investment and acceleration company mostly across crypto. And in the last couple years, we've been very focused on this intersection of crypto and AI. We've been covering d a d AI for quite some time, and it's
always been a bit daunting because there are some super hard problems that you have to get right to make it work. And so you have to do all the things that anybody that's out there trying to compete in the AI markets around inference or training, but then you have to layer on these extra hard problems that you need to make it work at a decentralized format. And in the first wave, like, we published some reports on the AI back in 2024. And in that first wave,
we didn't really see a whole lot of serious teams that were able to make progress tackling these problems. There were a couple of projects that were getting some notable funding and were attempting to tackle the problems. But what made me really wanna revisit it over the last month or two is kind of two big reasons. One is that the progress is actually now quite real, and you guys are part of that that progress equation. And then the second is just if you've been paying attention to the debate that's going on in kinda centralized versus sovereign AI,
this big realization that's happening that it's a big risk for companies and individuals to place all their dependency on these centralized labs. And, you know, there's one set of dependencies and risk you base if you put everything with Anthropic or with an AI. And then you have, you know, a slightly different but also set of dependencies if you're placing everything in the hands of the the open source labs, most of which are are progressing at the frontier out of China. You get, like, open weights, but there's still a bunch of elements to that ecosystem that are closed off. And what I believe is the greatest
countering force that we can have to this potential scenario where frontier intelligence is is owned and run by a handful of firms or, you know, kinda dictated at the whims by a couple of governments is to really push forward on the decentralization aspect of things so that no single firm or entity controls it, and therefore no government or or no company can necessarily shut it down. And so that's the reason for me getting back into these topics, having kind of explored it for the past couple of years, seeing these real catalysts right now of why should we care, why is it important. And then
kind of getting back into the progress side of things, the four problems I mentioned in the article still really hairy problems. And the first one, and to do decentralized AI, you don't have to solve all four of these, but you have to absolutely solve some combination of them. The first big problem is how do you deal with being able to serve big models where they can't run on a single GPU node that people are providing into the network? So just think about this as sharding.
It's a really, really hairy problem. And I think anybody that's into testing different models or running nodes, you kind of realize that there's a huge difference between something like QN 3.6, which you can sort of safely run-in a quantized version on a thirty ninety and, you know, make that as a a node that can contribute into something decentralized that can run QUEN 3.6 versus running something like GLM 5.2, which is a beast. And a lot of these other SOTA models coming out of the open weights labs from China are beasts that cannot run on
hardware that most consumers have access to. Like, you wanna you got a cluster of b three hundreds? Sure. You can run GLM 5.2, but that's not what most people have. And so problem number one is how do you enable GPUs to be pulled in some way such that you can run large models on hardware that no single GPU can run on its own. The second problem, and I promise I'll shut up after I run through these real quick and we can just get into questions after that. But just to kinda round out the problem space.
The other one is proving the model that you got and doing verification on it. And when you're running a permissionless network, when you don't control the nodes that are serving the model, how do you know that somebody that's supposedly running the model that you wanted to run is actually running it. And there's a bunch of different approaches to how you do that. But you have to do some level of verification in order to ensure the health of the network. And then the third
problem on the technical side is just how do you deal with keeping prompts secret across nodes, or how do you deal with privacy? This is a hard one too because, you know, we all have fears of just sending our prompts to somebody, you know, whether it's a company or a inference provider, they can see everything that I'm providing. When you do this on a decentralized permissionless node basis, like what's to stop some node from being able to participate in that network and then just sit in the middle of all your prompt traffic and and being able to, you know, have access to that and and, you know, you don't have any sort of control over who sees that. And so being able to keep that private across the nodes is another
hairy technical problem. But even if you solve all those three technical problems, which any one of them could probably be a standalone project on itself if you could truly solve those, You still have to deal with the marketplace side of things, is what every other inference provider on a centralized basis has to deal with, which is how do I manage the supply and the demand here? How can I provide enough incentives to build up enough supply to keep people from you know, keep nodes on my network versus putting them to use in some other capacity where they could maybe earn money? At the same time, how do I ensure that there's enough demand for people willing to pay for the models that I'm running on my network so that I can balance that equation?
And, you know, it's a fundamental business problem, and it's a tricky one. We try to solve this often with tokens and using tokenisms to bootstrap the supply side on the early side of things, but that has to translate to actual customer demand so that you can start to bring in real revenue and begin to offset those emissions that you do on the token side of things. And, you know, Venice is probably the clearest example of someone that has started to really solve the demand side as we're seeing their ARR is now approaching 70,000,000.
They're they're clearly bringing in a lot of users here, and that's helping them balance that side of the equation. So those are the four problems. These are not simple to solve, and it's not something that like like, a big takeaway that I have from all this research and talking to lots of teams is this is not something that crypto bros are gonna solve out of the trenches. You need real, legit AI talent to solve the technical side of things, and you need real business acumen to be able to solve the marketplace side of things. So I have a really strong filter for the projects that I've evaluated that have come up out of this space, and that kinda led me to to profile all projects
in my report.
Kyle (Bow Tied Blue Fin)I think that is this is Kyle, by the way. I think that's such a good summary of, the space of the industry. And I think that what's really exciting about a lot of the DAI projects that are out there are that, they all build upon one another. Right? It's not it's not PVP, right, in this where someone like Prime Intellect just came out with additional decentralized training for it. Right? There's other companies that came out with
decentralized and charted inference on a smaller scale. The use of both us at Morpheus and Venice with TEE providers on secrecy and and marketplace, I think that every single point that these different protocols are are pushing forward, like, help the DAI industry as a whole. And I don't think you see that as much in in other places, especially the fact that there's
a lot of open source work in the community, you know, through Morpheus and through other people that, are helping to accelerate this faster than we've seen other problems get addressed overall in AI from a privacy, and decentralization risk standpoint.
Bowen BelmerYeah. I would also add to that, Kevin, this this was so well, first of all, thank you. That was a really great and succinct, you know, summary of that portion of the article. And I do urge everyone listening to check it out and read it in its entirety because it is very good. It's one of those things where I'm reading it, and I'm like it's like all the things you wanna put on paper, but didn't quite have the words or or to a way of structuring it correctly, and and Kevin nailed it. But and and, yeah, Kyle, I completely agree. This is something I've said that makes
you DAI very unique from, like, DeFi and and smart contract platforms is because those platforms are specifically seeking, like, being the oracle of truth. It's contradictory at times to have two oracles of truth. Right? Because, like, one is either the truth or the other based on whatever it is you're trying to do. Where a DAI does not necessarily work in that same way. My singular agent can talk multiple systems, especially if it's open source or leverage different parts of different projects
to accomplish the thing that I'm doing, and I don't necessarily need to pick one or the other. And in terms of solving these problems, I mean, Kevin, are you familiar at all with some of the overlap with Morpheus and Venice by chance?
Kevin SimbackA bit.
Bowen BelmerSo certainly came across some of it during my research, but I'm sure there's Okay. Nuances that I'm not familiar with. Yeah. I mean, as of late and and this is just, you know, Bo Bauer's personal framing speaking now. Like, as of late, I sometimes refer to Venice as Morpheus the business, and I sometimes refer to Morpheus jokingly as Venice the network. Because in terms of value add and and what is being accomplished, these structures are are very similar and there's, of course, reasons for that. There's a lot of overlap in people who've worked on them. But also
the, like, framework of how it actually functions is different in the sense that Venice is obviously a centralized business entity, so there's trust components there, which was, of course, a huge topic of conversation over this past week with their VC raise announcement. Whereas Morpheus is strictly a permissionless open source network. And and so arguably, it provides a lot of the same things, but goes about it in very different ways
from that report.
Kyle (Bow Tied Blue Fin)Slightly different parts of the funnel. Right? If if you look at what Morpheus does, Morpheus is more of an infrastructure product where, you're not exactly looking for end users to come in and use it. Right? You're looking for potentially data center scale, you know, all the way down to the individual scale from a provider standpoint. And then you're looking at, you know, ideally, either businesses, enterprises, or people or agents running
larger scale workflows that require the larger frontier size models or the larger scale and capacity that you might not be able to get in other places or or the tokens per second as well. I think it's kind of like infrastructure for for larger scale versus individual user, gotta send these prompts, wanna keep them private, want, you know, anonymized access and so on.
Kevin SimbackI think you you're hitting on a a really key point. And so one thing that we talk a lot about in the decentralized crypto space is that there's a tax to doing everything in a decentralized way. So you kinda take on all the main challenges of doing it normally, and then you throw on these other challenges of doing it in a decentralized way. So it's like net negative on an effort level basis. But the one thing we have to remember here is that the big benefit of doing something in a decentralized ecosystem is the ability for things to plug in together. It's that composability.
We saw that with DeFi. Like, that's probably the the clearest example of it. But when you have permissionless networks that are complementary, they can start to plug in together and make the end product better than what somebody could do on a centralized basis on their own, you start to have like real advantages. And so it doesn't become just a tax. It's like a tax that can also be offset with these other benefits. And that's what I would really encourage the industry to start to pursue is finding ways to have these different decentralized AI projects complement each other where they can provide something that decentralized players can't do quite as well or quite as easily. And I'll give one example of that. I've mentioned it in both the decentralized inference and decentralized training report. But as we kind of think about intelligence moving more and more towards a commodity,
the real edge, the real nodes that people will find is around data. And when you're running inference, you have an opportunity to collect some level of data, whether it's like telemetry data or traces around how people are running inference. And that data can be very valuable for training or whether it's feeding into pre training open source models on a decentralized training network or doing post training on open source models.
And you can start to create this compounding loop. And that's something just one example of where decentralized AR projects can complement each other with that composability and potentially generate something that others can't do in a centralized realm.
Kyle (Bow Tied Blue Fin)Absolutely. And I think that as a user, right, you need to be able to choose on that sliding scale kind of where you wanna fit. Right? It it's I've before about about the scale of privacy where maybe in some cases, you're okay with all of your data being visible, right, to a provider. In some cases, you only want that kind of rough telemetry data. Maybe not the actual context of the prompt, but rather the context of of what you're doing around the prompt. And then in other cases, when you're dealing with pure sensitive information,
then you need that full, you know, end to end encryption, TEE level, you know, my stuff is locked down where we know, you know, physically that no one can can even access it. So I think I think it's a very interesting way to look, kind of what does the user need, and are they able to get what they need in the application and cost level that they made it. And
Bowen Belmerreally quick, just on that point, if everyone listening visits tech.mora.org and hits the spectrum button at the top, you'll see a really awesome breakout of what Kyle was just talking about there and basically the different points at which your agent, can fall on the Spectrum. So, yeah, sorry, Kyle. Didn't mean to interject, but wanted to just make sure, people are aware of that resource.
Kyle (Bow Tied Blue Fin)Yeah. Yeah. No. Absolutely.
Bowen BelmerKevin, I'm curious. In your opinion, do you think that DAI is ultimately gonna be a product to institutions more so than individuals? Like, in practical terms, I you know, I always, when I first got involved in DeFi, which was back in 2017 during the ICO craze, I had these grand dreams of every single person having a digital wallet, and they would all have an Ethereum wallet, and they would be using a stable coin to do x y z thing every single day. And as that market continues to mature, it becomes more and more obvious that, hey, you know, the average person maybe doesn't want to learn how to use an Ethereum wallet. And so there will be centralized
institutions that run on crypto rails. And so they, as the institutions, gain the benefit of, you know, working on on a blockchain or smart contract system. And I wonder, do do you think that DAI is likely to head down the same path where it'll be the fringe cases of people running their own agents and it'll be more so they will be interacting with DAI rails through central institutions or or maybe some other type of model?
Kevin SimbackI think on a consumer basis, if you just extrapolate where we are today, that it's pretty analogous to those who want sovereign financial independence that that would want sovereign intelligence. So it is a more fringe, smaller slice. It's people who care deeply about privacy and independence and wanna look for alternatives to the big labs or the the Chinese models. So I think that if that piece plays out, it's not super bullish for decentralized AI.
And so you have to think about what are the factors that could shape that and make that slice much larger to become a, you know, a bigger cohort of of people. And on the consumer side, I think it's probably going to take some sort of bigger event, some bigger catalyst to get people to realize that the information that they're sharing with their current providers is is dangerous in some way. It could be that, you know, some new new laws in Europe that seem to be passing are a little bit scary in terms of the privacy that other centralized providers can access on you. And and, like, I'm not gonna talk politics here, but there's been some examples of some
online information being used against people in ways that, you know, I would not think imaginable as someone from The US. So, like, I think it will take a little bit of a wake up call for consumers to really see that decentralized AI is the the better approach here for privacy and and cause that market segment to to be much bigger. But I don't think it takes a huge wake up call for enterprises to go that route. But we have to think about, like, what are their options? And so enterprises that are very privacy
concerned, either due to regulations that require them to be very private with their data or those that just wanna have greater control over their data from a competitive standpoint, there are a lot of solutions out there that they can use that don't involve decentralized AI. So I kind of posed this in my decentralized training report that was out yesterday, is we not only need to solve the tech problems, but we also need to understand who cares enough to pay for it. And I think that there's
some examples today that I could think of where enterprises would wanna pay for decentralized training. But again, I think it's it's gonna come down to what are the alternatives to be able to retain sovereignty over your data, and is decentralized the best way to get that? Or are there providers out there that offer solutions that companies can keep their data private but, you know, are not developed in a decentralized format?
Kyle (Bow Tied Blue Fin)I I I just have two comments on that that I think both fit together nicely. And and I think the first one is that I agree that it's going to take an event. And I think that event is happening right now. Right? We look at the Fable release and and pullback from the government. Right? We look at GPT five six, delaying release because of the potential export control issues. We are already seeing that
x that there is a risk to being able to get the models that you need to run your business, and it's outside of your control since those weights are proprietary and not open. So I I think they're having someone else being able to decide, you know, what can be available and who can it who it can be available from is starting to kind of get into the brains of people of, hey. Maybe there's some risk here. Right? Maybe I need to look at this a little bit more in detail.
And expanding that into what's going to make an enterprise switch, I think it's going to be more like what happened with AWS in the beginning. Right? It wasn't AWS EC two or AWS s three. It wasn't any individual product necessarily that, drove everyone there. It drove into the ecosystem because of the benefits that they had and the interoperability within their system that made it better. Right? It's not just that I can access open source models. It's not just that this is decentralized and private, but it's the the consolidation
of all the benefits that this DAI industry has that I think will make people consider that switch and and make that switch from and I would say the four pillars are probably, you know, uncensored access to information. It would be, you know, regulatory controls. It would be, sustainability and scalability with without locking into a single vendor, and then it would be cost. Those are kind of the four pillars that I look at from what, value does, you know, including privacy within there, of course, but within what DAI provides as compared to the the centralized options.
Kevin SimbackYeah. I think those pillars make a lot of sense. What I would say is that right now, this conversations become front and center and a lot of enterprises are questioning that dependence on the low source model providers, their options that they're defaulting to right now are not decentralized solutions. They are taking open weights models, hosting them on their own, or working with a provider to to host them in a sense where they can put their data in it and have complete ownership over post trained versions of those models that they can then use. So they're taking ownership of their data. They're
using models where they're not locked into a single closed source provider that could rug them at any time by pulling access to those models. All of those options that most enterprises are looking at doing right now have nothing to do with using decentralized technology. So there has to be a continued evolution towards, okay, when do those solutions don't make sense? When does putting an on prem solution using what Coher has in their stack, or
when does it not make sense to work with TogetherAI for a private inference solution where they can host whatever open source model you want, and then you can put whatever RL environments you want on top of that to train. When did those stop making sense where they have to say, you know what? I really need to go and do a decentralized training run with my data to get a model that I can truly work with, or I need to work with inference providers that are not centrally managed.
That's to me you know, we're we're far from, like, having that wake up call moment.
Bowen BelmerWell and and so, Kevin, to that point, in in your research, are you seeing any tangible metrics or anything that are being tracked to try to assess that risk? Or is it still kind of a, you know, gut feeling? The equivalent of vibe coding, but for establishing this, you know, risk to to someone's enterprise.
Kevin SimbackI don't think it's really metrics that I could see that track it. I think it's more event driven. I mean, we've seen some rumors that China could start to withhold some of the access to the new releases of their open source models or their their their labs open weights models. And so I think as some of these events start to to to hit and there's implications for companies that are affected by those events, then it it starts to to push people to think more broadly into what other solutions are out there that would result in these events not affecting them.
Bowen BelmerYeah. So unfortunately, it's it's the idea that, like, something really bad has to happen to someone for everyone else to care. Said differently. Yeah. Like,
Kyle (Bow Tied Blue Fin)people don't use privacy by default. Right? They they typically choose, these things either, one, when they're forced, like we saw with HTTPS, or when something personal happens to them that causes them to kind of swing that pendulum saying, shit. I I, you know, I just got burned by right? I I'm using a hot wallet and it got hacked. And maybe I should have been using a cold wallet this whole time. Just think of a comp.
Player 1 TacoIt it's really interesting that because there's two really cool lines of thought on this. Both on the enterprise side of things and, like, seeing how inference will just, you know, be gobbled up, but then also talking on the privacy and ownership side and and, you know, bowtie you you know, you just sort of touched on, like, personal information and and how that sort of gets attacked. And and what's really hilarious is I'm actually dealing with something like that right now.
You know, are we losing all of that privacy as every single company buys the same inference. And so, like, how do we, you know Kevin, where are you seeing, like, a big boundary on, like, personal information privacy, but then people trying to use AI for good and AI for bad. Like is there a is there a hard line on that that you're seeing?
Kevin SimbackNo. I don't think so. I think history has always shown us that most people are willing to trade privacy for convenience or some sort of you know, minimal benefit that they receive in return for it. And I feel like AI is just gonna continue down that path where most consumers are fine sharing whatever information that they have and giving it more and more information in the sake of convenience. And there will be exploits that can happen with that at some point or another. And and and again, it it probably takes
some events like that to bring the privacy into the awareness of the the public, the majority of the people who who don't sort of see that for what it is. And I I think the parallels have also happened in crypto where you've had, you know, certain governments that have had inflation go out of control, and so it's caused people to adopt stable coins at at a a much much faster rate, or people that that wanna look at, you know, Bitcoin as a hedge against fiat for some of their currencies.
Think most people just don't think about those events happening until they see it, or until they're impacted personally. There's a very few people who kind of see those risks for, you know, what they they can be, and they adopt privacy controls, or they move to, you know, some sort of, you know, permissionless types of setups, whether it's sovereign finance or sovereign intelligence, and they they go that route. But that's it's still a small slice and tends to sort of be dominated by a lot of your conspiracy theorists in in some ways. So
I don't know. I I I I'm a big believer in what's happening for decentralized AI, and I think some of these events will happen that will create that swell of people moving and adopting more of these solutions. But unfortunately, I feel like just gonna have to wait for some bad things to happen. Do you do you think as we saw, like, how
Player 1 Tacothe UK government sort of has blacklisted, like, Venice, do you see that as, like, a a bullish thing for, like, the privacy maxi? Or, like, do you do you think that we're gonna see more countries governments put their foot down on on AI and and that's open source and and uncensored.
Kevin SimbackYeah. I mean, I I think there'll likely be a lot of controversies around that as the open source models start to get to, like, fable quality where you see The US having to throttle the rollouts of these, you know, much more frontier grade models. As the open source ones get to that, I think there there could be a lot of government reactions to that. Governments are always gonna wanna control things to the greatest extent possible, and Europe tends to lead on that front and
does some wacky stuff with their laws and regulations around data. So I'm not gonna put anything past them stupid moves that they might make to try to control this in various ways. I think, again, all that stuff is like I put it in the same category as these these events that can drive people to to choose alternatives in the decentralized space.
Bowen BelmerI I very much agree with oh, sorry, Brian. Go ahead. Oh, I was I was just gonna say
Brian Bourgeriethere's a couple things on that. I agree, like, things come in waves. So in 2015, when we had launched our plug and play home server company, and I we go to Web Summit every year, And I used to in the beginning, I spend most of my time explaining why privacy is important, why they should wanna own and control their data, have a private cloud, or different things like that. But then the Equifax hack occurred. And that very next year, everyone was coming to me, and I almost didn't have to mention the reason why they were seeking us out.
And what I realized besides it takes an event to bring another wave of of noticeable amount of people who care is the crux there is the solutions have to be ready because it their memory is only so long. Like, oh my goodness. It's a big deal for for a cup for a couple days, weeks, a couple months, but then it it fades out again. So it's extremely important that the decentralized or, in my case, better solutions, like, have to be ready when those things occur
to capture as many from the wave when it comes. It won't capture everyone. The best you can do is is is just build as fast as you can and get it as much in place, and then obviously have the distribution to to get to the people. But they'll come looking, and it comes in waves that might be a bunch of waves. It might be years between the waves. It might be days now because everything's accelerating, but it's really catching those waves and realizing that's kind of how you're gonna get closer to mass adoption.
And then on the government front, for me at least, and obviously, I'm more on the extreme side of of caring about privacy, security, ownership, and control. But on the government side and and and, Taco, you brought up how they've kinda, like, blacklisted Venice. That's also a very important key distinction between a centralized company running something versus a protocol that that nobody slash everybody is running. And and I think really starts to
help some people at least realize that, oh, we do need to think about decentralized networks, permissionless networks, and marketplaces that we can actually interact with. Because when you get to government scale and we don't really see as much of that maybe in The US. Like, we're we're pretty privileged overall even when things are are going the way we don't like. But when when other countries get into things and have revolutions or young people are electing new leaders over Discord, like all the stuff that's occurring all around the world,
I actually see a a a much bigger need and use even more immediately for decentralized networks, which we've seen in crypto and everything else too. I mean, it's following a lot of the same patterns, just different data and and bits and whatever they're being utilized and moved around.
Bowen BelmerAre there, I I don't mean to change change the subject, but I just before I forget to ask it, I I'm curious as well, Kevin, are there any, like, specific people or businesses or projects that are really heavily, like, delved into this topic that do not get the light that you think that they should? I guess what I'm asking in in other words is, are there any sleepers that we should be also paying attention to?
Kevin SimbackI don't know if there's any sleepers. I mean, I profiled all the what I felt like were the legitimate projects in my report. I do think that what we just saw yesterday with PrimeMentalX fundraise, you know, a 130,000,000 at a billion dollar valuation, and this is a project that was sort of born out of doing decentralized training, is like a really big signal. And and so I I think I wouldn't call him a sleeper, but in the broader AI ecosystem where we've got a ton of unicorns that have popped up in the past couple of years,
I think that one is is sort of raising some eyebrows a little bit. It's like, oh, you know, wow. What are they doing here? And so I would kinda look for where is the broader AI system that the the broader AI ecosystem is is mostly anti crypto and and tends to dismiss a lot of what happens in the decentralized space. When you start seeing more adoption or more mentions, that's where, like, I would look to see, you know, who's getting the most attention.
Prior to prior to that fundraise, the tensor is probably the one that is talked about the most outside of crypto, and and I I think that that's just like barely scratching the surface of what's going on. So I guess I I don't think that there's a whole lot of sleepers out there, but I would continue to watch the firms that are already in this space, which ones are getting fundraisers and which ones are getting talked about in the AI communities, not just crypto communities.
Bowen BelmerYeah. It is a a interesting line be between AI and crypto. And I even I mean, I find myself not oh, how am I sorry. Trying to phrase what I my thought. I find myself often looking at DAI through the lens of DeFi more than I should Because to your point, well, one, the AI crowd often does not like or even pay attention to the
DeFi space. It's kinda like the the meme from what shows at Suits or whatever where the guy's like, I hate you. And the other guy's like, I don't think about you at all. There's there's a lot of that in in AI, you know, to DeFi. But but also just the the realization that, you know, AI touches everything. Like like, DeFi is meant as a means to an end to do a specific thing, move value, record truth, but the AI or AI generally is, you know, potentially able to touch
anything and everything, which kinda leads into another question, Kevin, that I I had for you was as like, let's look down the road a little bit, and let's say we get to a place, you know, I was gonna say twenty thirty just as a nice round number, but the reality is humans, of course, are famously bad at trying to predict exponential growth, but, you know, we're in the not so distant future, and there is effectively infinite demand for inference and for AI.
How do you see companies and this is kind of similar to what I was asking before about, like, how they tangibly estimate risk. But how do you see enterprises tangibly measuring the benefit of, okay. I'm I'm using DAI as opposed to a centralized provider, and, like, I'm tangibly able tangibly able to prove, like, this is a benefit to my business or entity because I'm making this very deliberate choice?
Kevin SimbackA good question. I think in general ROI for AI and enterprise is still an open question. There's pockets of examples of companies getting good ROI, but there's plenty of others where enterprises have reported not seeing the ROI. So it's like take it one step further, and I know we're looking at a longer time scale here, to see how do they get ROI from decentralized AI is a little bit of a stretch. I say one way to frame it is to not look at it on a ROI sense, but look at it like insurance. It's a cost of doing business that protects you against some adverse events that may be out of your control. And I think there's a scenario where companies would just pay extra and not necessarily look for an ROI on decentralized
AI, but look at it as a risk mitigation cost to prevent them from things that can be out of their control. To be able to present an actual ROI case, I think there would have to be a big progressive move where decentralized inference can be performed cheaper than centralized inference at scale. And that kind of requires owners of GPU hardware to be able to accept
a different margin profile than what large companies and centralized providers of of inference needs to justify their business. And I do think that there's opportunities there. I think that there are people that have more of the consumer grade GPUs, you know, all the way up to like the RTX 6,000 series that bought them either because they're hobbyists or because they wanna, you know, run it for various purposes. But then they have idle time, and so they're willing to accept a pretty low return to be able to use their their idle GPU time in some sort of network.
And so if you combine all of that idle time at a fairly low return profile, you could then serve up models cheaper than in centralized providers that have to look at the return on CapEx and OpEx in a different frame. So that's one way that inference could potentially have a better ROI is just being able to to outcompete on price on a lot of centralized providers for for that same inference. And so companies could just look at it that way. They could look at, hey. I get some benefits of
not being rug pulled by dependence on on one single inference provider, and I can get cheaper inference to go with it. So so that's probably the cleanest ROI case that I could see. But, again, it has to be done at scale and it has to be performant such that enterprises will be able to get the, you know, the the output speeds that they need.
Player 1 TacoSo you were talking on, like, some, like, the big scale stuff. What are some of, like, the smaller little gems that, like, bring you joy and excitement over the next couple years? Like, what do you where do or where do you see or hope that will be?
Kevin SimbackYeah. I think a lot of the really exciting stuff for me is just in the agent space, And we've really seen agents come about only in the past six, seven months. You know, OpenClock kind of burst on the scene and made agents really accessible to most people, and then Hermes has made it even better. And and so people are exploring a lot with agents. I I I really think that there's a ton of opportunity around the agent space, both at the application layer and then, like, at at, you know, orchestration and payments layers for agents to to do really cool things.
And so I'm I'm excited about projects that are finding really novel ways to to play agents to work. Another one that I think is kinda done well at that of application and orchestration layer is the project Gitlab, g I t l a w b. They've just built some really cool stuff. The token doesn't really get much appreciation for it, but the founder there and and his team have have built some cool things. They started with a
decentralized version of GitHub, but made for agents. And so, you know, if anybody's used agents and you've wanted your agents to go build code for you, you've gotta do a bunch of stuff to get them access to your GitHub. And GitHub is really fine agent native platform, so they built their first product as an as an agent first version of of GitHub, and that was pretty cool. And then when the source code, the cloud code was leaked, they were able to take a version of that and sort of create their own with that as inspiration
and have an open source version of a generic version of Cloud Code that you could plug in any model provider that you wanna use through it. And, you know, just another example of just kind of hustling to find product market fit. And I and I think that in some cases, they've been in the top two or three on open router for applications that have used applications using using their tokens. And so, like, that's the kind of stuff that I think is pretty exciting.
Bowen BelmerKevin, I have another question. I know we're grilling you with questions, but I really appreciate it. And the insight is is truly very valuable, especially, you know, you guys do such deep research in all of these things. So it's it's it's really cool to see and hear these opinions from someone who's been really, really in the weeds on all of these different projects. I I do have a question though. So as as you were doing research for your most recent articles, particularly the one that does mention Morpheus,
was there anything you came across? I I guess, my question is, like, do you have any feedback or any questions for Morpheus? Like, for the community of people working on Morpheus that maybe came up as you were doing research or just general feedback, you know, from from doing that research?
Kevin SimbackYeah. Nothing too specific. I think there's always a question around traction for for any of these projects, and where is the where is the traction coming from? And and this is not Morpheus specific. This is really crypto specific. And I've learned this over many years evaluating crypto projects where they start reporting these these growth metrics, whether it's, you know, TVL or volume or, you know, usage in some capacity.
And you come to find out later that there was a lot of engineered traction, you know, whether it was bots or, you know, some sort of insider manipulation to generate traction. That's the thing that I'm most skeptical on. And so I would just say try to be as transparent as possible about where is the real usage coming from, particularly like organic usage. How can you demonstrate that it's truly organic usage? And be able to show examples and case studies of why people are using
your inference and your services over, you know, presumably a handful of of alternatives that they could go use. And the more that you can demonstrate that this is a product that people are using for real purposes and not speculative purposes because they have some, you know, bags on the token side, that's what can separate you. And and I I I say that as sort of generic advice for all the projects out there that have tokens. You need to be able to show why the product itself is valuable beyond
purely feeding the speculative nature of the token.
Bowen BelmerYeah. Very well said and and completely agreed. And to that end, I wanna give another shout out to Alan and team for the recent node update, which specifically strips away, self testing from, you know, from those on chain figures. But, yeah, I I couldn't agree more, Kevin, that the transparency and, like, actual usage is ultimately what matters. And, you know, we kinda saw this, last year. So it's or, like, spring twenty twenty five, we had that brief moment where crypto
like, DAI, quote, unquote, bursts on the scene for crypto, but it was basically, you know, chattypti rappers, like like, bots, like, tech spots with tokens attached to them. There was no actual, like, real decentralized anything of note, at least from my perspective going on. Now some of those projects have, of course, matured and and continued to build very much more robust things like virtuals comes to mind just as an example. But
but, yeah, I I couldn't agree more, and and I think that's that is an excellent perspective. I do wanna open it up and if anyone else has anything else for Kevin, otherwise, we have we have stolen an hour of your afternoon, so I wanna be conscious of your time as well.
Brian BourgerieI was just gonna say real quick, on that note, is actually probably really timely with the launch of Morsecan that that's out there. Because obviously, wanting to track and know what's actually going on in in the network. So Cashew is rebranded to Morsecan, and, obviously, Rob can talk more about that too. But that's that's just nice to see at least within the Morpheus front, more more tools and access and and ways for us to even monitor
the the ecosystem and the blockchain to know what's going on.
Bowen BelmerYes. Fantastic transition. Completely agree, Brian. Rob, can you tell us a little bit about the switch from CashView to Morescan and the motivations behind it? And, yeah. The perfect transition absolutely fits in with what we were just saying.
Rob (DRM3)Yeah. You bet, Bo. And I'll make it as quick as I can. I know that we're going over the hour here. Also, before I get started, I wanted to mention to Alan, all of that sounds great on the provider side with all the monitoring. I think that's gonna go a very long way to improve optics on what's going on and and exactly, you know, what we're getting out of out of the inference at the end of the day. So I started Cashew a while back, a few months ago, really as a cash for a consumer node that I had that I had built
after, you know, or as an alternative to the consumer node. And I decoupled it from I decoupled my consumer node from from the cache, and I realized it was basically a scanner at that point. And I was interested in, you know, kind of having one single pane of glass around everything going on on the compute and provider side, the marketplace, the bids, the additional pools, the builder pool. I also wanted to take a look at Holder and
build some additional analytics kind of under a single pane of glass. And so I evolved it into Cashew and that was really just named, you know, second cash for me. So cash cash do, named it Cashew, named it's complimentary consumer node, Pistachio, and decided to just need to go open source. It didn't really make sense for me to, you know, hold on to the source there. I wanted to get the all of the spooky mechanics
out to the community so that other things could be built on that. And I felt like the timing was timing was right so I moved it to an open source project and I improved the pricing. I was getting pricing from CoinGecko and I I'm pulling pulling it off a Dex contract now. So it's it's running purely against contracts and I'm using another open source project of the arm free labs called RPC pool which effectively strings, you know, a dozen free RPC endpoints to actually maintain
low alchemy requirement. So I'm able to to to effectively run this post the Cloudflare free tier and even RPC free tier here. So I've got the ops down extremely low on it. I think it's a good open source project too for anybody that wants to take a look and get some ideas around developing inexpensive to operate scanners and how to keep sync there. So in effect, that's what Morsecan is. Thanks to Brian who kind of nudged me towards making the move here as well about a week ago.
I was kinda gearing up for it anyway, but Brian kinda nudged me over the edge there and and so so we have more scum.
Bowen BelmerWell, awesome. Thank you, Rob. I love it and appreciate it. And I love having a a site to point people to, especially people who are crypto native that are very familiar with block explorers. It's it's like a, you know, very similar type of thing that I can now point them to to say, hey. You know, look at this. Like, this is this is the throughput. This is where you're seeing the activity and, ultimately, to try to prove, the growth points, you know, that that Kevin was just talking about before. That like, what actually makes this a successful network and a successful product.
Rob (DRM3)Absolutely. And I plan to look at extending some of what we're capturing. Wanted to get, you know, make sure that we're covering all pools clearly, continue to extend analytics, and I'm kinda wondering what it would look like to put more DM under this pane of glass as well. I think it would be good to see visibility there so we can you know, we see these reports on Twitter around marketplace gate gateway, more DM. We've got bunch
of providers coming on board. So kind of just trying to get that single pane of glass to kind of look at that holistically from the scanner. So expect more than that.
Kyle (Bow Tied Blue Fin)We're we're super appreciative of all the of all the work that's being done to you know, from contributors to provide other ways to see the analytics and see information about the marketplace, and and we're supportive of anyone who wants to, you know, come in and and help drive volume, whether that be from, you know, more information to make it easier for consumers, easier to for providers, or to come in and and be a consumer and use the system or provide your excess capacity to it. So we're we're really appreciative
of that work, and we we welcome everyone to to come join and come contribute.
Zion, hear me! For Season 2 of the Mere Morpheus podcast I'm going to be uploading the Morpheus builder weekly updates. My reason for doing this is to make it more accessible for those who want an update directly via...
Zion, hear me! For Season 2 of the Mere Morpheus podcast I'm going to be uploading the Morpheus builder weekly updates. My reason for doing this is to make it more accessible for those who want an update directly via...
Zion, hear me! For Season 2 of the Mere Morpheus podcast I'm going to be uploading the Morpheus builder weekly updates. My reason for doing this is to make it more accessible for those who want an update directly via...
Zion, hear me! For Season 2 of the Mere Morpheus podcast I'm going to be uploading the Morpheus builder weekly updates. My reason for doing this is to make it more accessible for those who want an update directly via...
Zion, hear me! For Season 2 of the Mere Morpheus podcast I'm going to be uploading the Morpheus builder weekly updates. My reason for doing this is to make it more accessible for those who want an update directly via...