Don't Trust Me Bro | Morpheus Rolls Out Verifiable TEE Privacy
51 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 #65 (we missed a week as the X space wasn't recorded but hush hush)! Recorded: 19 March 2026 AI SUMMARY: This episode with the Morpheus community unpacks a major privacy milestone the Phase 1 rollout of confidential compute for the Morpheus Inference marketplace. Provider nodes can now run inside a Trusted Execution Environment with verifiable attestation. Consumers can cryptographically confirm a provider runs the exact hardened logging-disabled Morpheus build inside a secure enclave relying on math and open-source code not policy promises. The episode explains how this improves the existing wallet-based non-KYC model and how the consumer node automatically enforces trust with no extra user friction beyond choosing a TEE-tagged model. It looks ahead to Phase 2 extending TEE protection to the GPU inference layer for end-to-end confidential compute and discusses the likely cost and performance trade-offs. The team covers practical UX like TEE tags and technical details including hash verification and a hardened build pipeline. It wraps with ideas for agent-level routing that auto-selects privacy when needed celebrating an open permissionless path to private AI secured by physics and code. Timeline: (00:00:00) Intro (00:01:15) Kyle's overview: marketplace, privacy goals, and TEEs (00:03:31) Current state pre-TEE: wallet-based access and trust limits (00:07:13) Defining phase one: node TEE and attestation basics (00:10:12) Looking ahead to phase two: GPU inference inside the enclave (00:13:07) Alan explains the trust model end-to-end (00:18:06) Hardening the build: signed images, TEEs, and verification flow (00:22:06) Why prove first: open source, physics not promises (00:25:07) Protocol vs gateway: where TEE is implemented now (00:27:00) User experience: choosing TEE-tagged models, no extra steps (00:33:24) Costs of privacy: where premiums may appear (00:38:30) Agent routing idea: auto-select privacy in Everclaw (00:41:43) Surprises from building TEE: follow the math (00:45:20) New use cases and comfort for personal AI (00:48:16) Why open source and permissionless matters (00:50:04) Closing thoughts and sign-off 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 kick off this week. And as kind of an overview of what we're talking about, anyone who's involved in our social channels or in the Discord will have seen that there is a fairly big announcement that we and by we, I mean, certainly not me, but the community of builders who are actually building the code base that is Morpheus have shipped a pretty awesome update over the last day or so that's been being worked on for quite a while,
specifically with regard to privacy and compute. And so as always, joining us today are multiple members of the community, many of which are actually hands on in the actual code. Again, unlike myself, I just talk about it. However, I really wanna dive into what was just shipped, what it does, what the implications are, and kinda cover it from a technical perspective, and then also maybe a more, like, user focused practical perspective. So for someone like me, like, how how does it impact my experience,
using Morpheus or or inference? So I'm not sure exactly who the best person to jump in first is. I I think Kyle maybe doing if you wanna do an overview on, you know, what the heck I'm talking about.
Kyle (Bow Tied Blue Fin)Yeah. Yeah. Sure. So I guess just to to start off the conversation, yeah, a bunch of people in our Discord have been very closely tracking our GitHub commit. So the the cat was a little bit out of the bag that, we are or we have launched, the first phase of a deeper privacy solution, within the Morpheus Inference marketplace. And,
I'm I'm gonna talk more about what privacy means, what TEE means, which you're gonna hear over and over, and and kind of what our road map is for this. So right now, everyone knows that the MoreCase Inference Marketplace is a permissionless two sided marketplace for offering and renting compute, if you will. And the way that that works is a provider will host a Morpheus node
to essentially access the marketplace and will within that node, provide access to an AI model, whether that's running on their local machine, an API pass through to one of their servers, hosting machine, whatever it may be. On the other side is the consumer who is also hosting a consumer Morpheus node, to send prompts, right, that that then get picked up by the provider. And and that's kind of how the system works, the marketplace being in the middle and using the nodes to access that marketplace.
Right now, there is, you know, pre privacy, pre TEE, there is kind of a a deeper layer of trust and privacy than what you get with fully, centralized and and proprietary services like ChatGPT and so on. Right? With Morpheus, you're connecting with the wallet address. Remember, this is the node, not the API. We'll talk about the API at the end. On the node, you're connecting with the wallet address. You're sending prompts to open source models.
There is no GPT in in you know, who has your KYC and everything else. Right? This is your wallet is sending a request and a provider is picking up. Now in that infrastructure, we don't exactly know what the provider is running. Right? Are they running the Morpheus node that they get cloned directly from our repo? Did they make some modifications to it? Do they have logging enabled? Where are they sending their requests?
All of those are, you know, unknowns, which is completely fine because we have a flexible system. And it's fine for certain security applications. I'm gonna go down a couple different sorry about that. A couple different rabbit holes as I go through this to to help you guys understand. Right? We're talking about, about the sliding scale of of privacy in some ways. We said this isn't ChatGPT, and this isn't, you know, Claude directly.
It's in some cases, that's okay. Go you know, send a basic prompt to to chat GPT, you know, that that you're okay attributing to yourself and that you're okay with them training your data on. There are other things that we're sending that that we're not okay with with our name being attributed to and training on our data. And that's when you have to look at these more private solutions. So kind of stage one of that privacy is is breaking the KYC, and that's where Morpheus is, you know, as of yesterday, let's say, where where the basic Morpheus,
know, offering through Morpheus is. Non KYC, yes, someone can see your your prompts, someone can log your prompts potentially as a provider, but it's not directly attributed to you. It's attributed to a wallet address. And
Bowen BelmerCan we pause on that really quick, Kyle? Just just really quick. I just wanna pause on that point. So that was prior to this update how things work, basically. Correct? So meaning that if I provide any type of prompt that might have some type of personal identifiable information, potentially, prior to what we're about to be talking about, the person who's actually running the inference, so the the hard metal, wherever that may be, could potentially see the
those prompts and that inference being created, but doesn't necessarily see, like, who I am, meaning, like, my email or my account or something like this. It's strictly tied to a wallet address. However, if I were to say, in that prompt, I, like, drop my name or I drop something that's, like, very personally identifiable, that theoretically would allow that person to tie my prompt to my address name. In the same way that if I send you a transaction on Ethereum and I have it tied to an ENS name that's publicly disclosed,
you would know that, like, that specific address was me. Correct? Exactly.
Kyle (Bow Tied Blue Fin)That's exactly true. And I'm just going to nitpick one thing that you said in there that that we need to address separately is one side is the node that I mentioned. Right? The the Morpheus node that you're hosting, and the second part of it is the bare metal. Right? The the node is what's your your entry into the marketplace. The bare metal is where you're actually running the inference. That's the GPU.
And it's going to be more clear in a second why why I separate those. So sliding scale privacy, no no privacy, you know, go to to ChatGPT and and send a message. Some privacy, obscuring the, or or removing the KYC and just interacting with the wallet address, you know, risks of prompt logging and and so on. That's kind of what Morpheus is at at the basic layer.
The next layer, which is what we introduced yesterday, is what we're calling and we have to work on naming a little bit, but what we're referring to as, let's say, the node TEE. And our goal and what we're doing here is we're going to get through this sliding scale privacy to maximum privacy where no matter what, no one can see anything. But this is the the first phase in that. And what we're doing with this phase is
we are allowing providers to put their node within a trusted environment. Right? They're they're hosting their entry into the marketplace on a TEE machine. And hosting it on the TEE machine, great. Right? Anyone can do that. You could always have done that. But the valuable part is the attestation and proving that this node is running this version of the software
with these configurations, and I can't get into it. So before I said, yeah. You're a provider. You can host it. Right? You can have logging on. You can have custom code running. You can you can really modify the product. In the latest version of this, we're providing a handshake between the consumer and the provider so that the consumer can say, provider, what are you running and and prove it. Right? And that provider now is going to take and and I'm gonna let Alan jump in and and
explain this on a deeper level, but the the provider is essentially giving their attestation report that says, this is the VM that I'm running on. This is my address. This is the code that I'm running, and here's the hash of it. And that's, programmatically verifiable back to our GitHub releases, and it's programmatically verifiable by the consumer. So when a consumer is requesting a model that that is enabled for this or or a path that's enabled for this, it's saying you're either verified and I'm going to continue or you're not verified and I'm going to stop.
Right? And and that's kind of where we're at now, from what we've released. The next stage of that is not just proving that the provider, you know, entry into the system is within this private enclave, but that the actual GPU inference is in the private enclave as well. Right? And, when I say enclave and and I'm I'm going to spare you all the the deep technical details for our article, but it's essentially an impenetrable box. Now
you can go up to the box with and connect all wires to it and scope the PCIe lanes and try and find the information, but you're just going to get an encrypted output. And the key only lives within that box, within where where the Morpheus node is running. So it doesn't matter. It doesn't matter if someone steals the steals the server rack and and tries to inspect it. They they can't get it with without the key that they can't get. Now
on that final level of privacy, you have your entry into the marketplace within this encrypted box, and you have the bare metal GPU within this encrypted box. Nobody knows what you're prompting or who is prompting it because of the the no KYC that we mentioned. So you're, the provider is just providing essentially a piece of hardware and electricity and an Internet connection, and and they can't see anything else.
And that's where we're looking to get with in in phase two. And and we're really proud of of the release that that we just did now with phase one showing, alright. We can do this. We can have a consumer, fetch information from this provider and prove that it's authentic and prove this attestation. And it's a relatively straightforward, maybe not simple, but straightforward path to saying, okay. Now you need to prove that your GPU is there too. Right? There can be nothing behind that proxy router that you're running downstream where the inference is that is catching your request. So I I know I said a lot, and and I'm going to let Alan
kind of jump in here and and go a little bit more into the into the specifics of of the tech, and then we can have everyone kind of rapid fire your questions that that that that Alan and I can answer. And we'll get into into some more fun stuff on on kind of where phase two is and and how to access this and and all that and all those things. So with that being said, Alan, I guess, my question for you is if you can just explain
the the structure of the of the handshake. Right? And Mhmm. Kind of how do we know that the TEE is authentic? How do we know that the code running is what that person says is running and and how does the consumer, you know, entry into the system actually confirm that and and why can you trust it?
Alan (Nomadic Rogues)Oh, well, because trust me, bro. It's all in there. It's it's good. Thanks, Kyle, for the overview. I think to answer that question, wanna step back a little bit and and I use that, you know, the the the trust me, bro, line a little bit tongue in cheek, it was really how we looked at the communication path. And so we did is kinda sat down and looked at, you know, everything from the consumer node or the or the c node. If I if I start short cutting words, I apologize, but the consumer node
that's running on your laptop with your wallet, where does that go? It, you know, goes out to your ISP, goes over the Internet, finds the provider node at whatever registered address it's on. It might go through a load balancer, it might go through all sorts of network stuff, it might go into AWS or Azure or some other, you know, someone's data center where you, you know, you could have port sniffing and other fun things and then it ends up in the software, right, the provider node software
that's running on someone's server somewhere that that you don't know of. And then sends that prompt either on box or off box again through all sorts of fun network things and load balancers and etcetera etcetera to someone else someone else's server. And you kinda line that whole thing up and you go, where is there a trust me bro kind of handshake, gentleman's agreement? And where do we have provable
encryption, security, privacy, decentralization, etcetera, etcetera. And the good thing is, I think we started, we didn't start from ground zero. When the Morpheus Luma node was first developed between directly the consumer node to the provider node, at the software layer, that communication was developed to be private and encrypted. The c node and the p node, consumer node and the privacy node, the provider node have, they handshake and exchange their own keys at the beginning of the conversation
and that is actually the data that's transferred between the two. So then we're like, well, okay, great, but the provider node, how do I know, you know, kind of to your point that that's running the right software? How do I know that it's not doing things? And in the provider node, are switches. We've provided the ability to do logging and, you know, chat context for other, you know, chat retention for for other activities that people might wanna do. We try to give maximum flexibility.
But when you start talking about TEE and and privacy, those things have to be turned off and you have to know that they're turned off and stay off. And so our first tackle on this was to secure the build pipeline. So when we update the Morpheus, the Lumira node software, we present it in, you know, downloadable form. You can you can, clone the git repo. You can,
grab the container image. You know, we try to make it pretty easy to to kinda get the software and use it. But we've added an extra step, which is really about taking that containerized image, feeding it the right variables to make sure that no logging you know, logging is turned off or detailed logging is turned off, that no chat can be can be saved. We turn off all of those features that can be leveraged for other things, but we said, no. We're gonna lock it down. So we had to harden the image,
and then we had to sign it to make sure that, you know, if someone says, hey, it looks like the same image, it's got v six on it, you know, must be the thing, that's not good enough. That's another trust me exercise. Nope. We need to sign it and say that if the the the hash of that software is different than what is posted on the tested GitHub build, then it's not the same software, so you should not trust it. So that's the first part is kinda harden the pipeline and make sure that the images that we post
can be verified that that came from our source and have all the right things turned off and allows you to then go put this on the hardware. And that's where the second piece came in that that we worked on over the last couple of weeks, which is putting it in hosting provider that has TEE execution environment. And what they allow, and we partnered with Secret Labs in particular for our first article here, it allows us to
bring all the elements of the t build in, create the image, lock it in so you've got that watermark or that cryptographic hash of the image, so that when I take and I make that exact same docker compose file, that yaml file on secret labs and I run that exact thing, I will get the exact same watermark or that exact same hash. So now I know that's running on the TEE, you know, the trusted execution environment is the software that we built with the right settings
for your provider. Like, great. Okay. That's awesome. Now, who checks it? And this is the third piece. And the third piece is when that consumer node knocks on the door and says, hey, mister provider, I wanna send you my LLM prompt. The consumer node says, oh, you're a t model hosting provider. You you know, I can I need to know if I can trust you? It goes out to the endpoint and grabs the attestation quote. The quote on the machine that says,
I am this cryptographic hash. The consumer node then goes out to the GitHub repo, looks up the latest version or the version that it's been told it needs to go check. So we are version aware, so 61 or six O. It says, hey, this provider is running 610. It goes to GitHub, looks at the 610 release, looks at that attestation hash, and then compares it through secrets verification engine that they match. And what that then tells the consumer node is that it's running the exact software, the software has not been modified,
and you can trust sending your prompt to that knowing that there's no extraneous logging going on, your personal privacy is protected and the there's no PII that's flowing back and forth. So those three pieces, the hardening of the end hardening of the build pipeline, getting a a verifiable image, hardened image, building up the the VM to make sure that you know what it's running on. And then the last piece is that double check is where that c node is able to go out, knock on the door, and make sure
that they know who's answering. Now, to Kyle's earlier point, and I'll and I'll pause here after this, that protection perimeter right now stops at the backside, for lack of a better term, of the provider node. So when that provider node sends its prompt off to an LLM server somewhere else, it's going through the network, it's going through you you not you don't necessarily know if anyone's logging on that other side. And that's really where phase two comes from because you wanna make sure that
when it leaves the provider node that it's still safe. And that's why we're gonna have, you know, pretty fast follow as as fast as we can to if I'm hosting an LLM, let's say on on secret non secret labs enclave, that from when you knock on the door of the provider node till that prompt reaches the actual inference engine and returns, you know that it's trusted and safe. So I said a lot of words there, Kyle and Bo. Did they illuminate
or confuse?
Kyle (Bow Tied Blue Fin)No. No. I was a great explanation. I think that understanding for for other people to understand actually how this works and and breaking out the trust me, bro, is is super important. So the last thing that you said, I think, is probably the most critical for anyone listening now or on the recording of, like, what's phase one and what's phase two. And, it it might seem a little backwards for some people that,
it's kind of like, alright. First, give me privacy and then prove it is is maybe what you would think it would be, but we really went the direction of how do we prove it and how do we extend it. Just focusing on trust, being being the most critical piece. Right? If, you know We gotta look at OpenAI's private policy, like, I don't trust that for shit. They might say that they're that it's private, and they might say that they're not collecting any information, but we we've seen that policy change a 100 times already.
We've seen a lot of policies change and a lot of leaks of data in in different places. The the whole point of phase one and and being able to see the cryptograph cryptographic proof is that there is no trust in a policy. There is no trust in anything but, as Alan said, physics and code, and and it's all open source. You can watch it happen in real time. And,
what makes Morpheus special is that you can see it. Right? You can see the open source code, and you can see the substation happening. And and there is no layer of like you might have with other companies who run on TEE that that you do still have to trust it a little bit. You have to trust that they're running the TEE and the encryption's happening where they say it's happening, and,
we just felt that that wasn't enough, and and we needed to to be able to prove it and to
Bowen BelmerCan so okay. I have a couple questions here, you guys. Can we talk about so so everything you guys just said was basically, like, from a architectural perspective. Right? So from the perspective of somebody who can can actually go in and do the verification that you guys are talking about. Right? And so it it it is important, of course, because it eliminates the trust me, bro component, which is obviously huge and is ultimately what differentiates Morpheus
from other projects in the space. Basically, is we're trying to do everything we can to have a, you know, completely open source permissionless decentralized way to access ultimately intelligence, but but AI tooling. From a user perspective, can you talk through what that looks like? Like, if I'm an end user that's plugging the API gateway now actually, let me take even a step back earlier than that. Is this implemented at the protocol level or at the gateway level? Meaning, do I have the same benefit if I'm somebody who's building directly into the Inference marketplace,
or is this something that's specific to somebody who's using the API gateway?
Kyle (Bow Tied Blue Fin)Good. So, let let me let me break this down in into probably three points. So first, I'll I'll hit the API gateway side. Right now, the implementation that that we launched was at the core marketplace level. The API gateway is essentially a service that runs in front of, the marketplace level. So as of right now, we have not
implemented this into the API gateway. That that's kind of on the list. Right? Let's call that phase three. Right? Phase one, is in the is in the actual inference node and in the marketplace. Phase two is similarly hardening and expanding that security within the marketplace. And then phase three is going to be figuring out how we can incorporate this through the, through the API. But that's a a little bit more intensive than something that we're we're a 100% going to do. Now
to to follow-up with something that you said earlier is, if I'll if though, if I'll reframe kind of what you said, why are you talking about technical stuff that, I don't care about, and why does this matter to me? Or how do I actually see that this is working? So
Bowen Belmerfor That was a less delicate way of of framing it, but okay.
Kyle (Bow Tied Blue Fin)For for a typical user, what your process is going to be is you're gonna find a model that's being hosted in the marketplace with, a TEE tag on it. Right? Models in the marketplace have various tags. Some might be model size. Some might be model capabilities. One of the tags will be if it's a TEE model. And if you request that model as a user, the node is going to do the rest of the work. Right? It's going to say,
this user wants privacy. I'm running my scripts to validate that the provider is, you know, providing this attestation and it matches and and everything that Alan spoke about. So there is no u users don't need to do anything other than choose a privacy model and have a version you know, a compatible version of of the node, of course. We are talking about this underlying infrastructure for the sole reason that,
it's more than a policy. You know? It it's you need to be able to prove that that the the validation and the attestation that's happening is actually legit, and it's not just, you know, blowing smoke. But from a user perspective, no additional effort other than choosing the model with the level of privacy that you prefer.
Alan (Nomadic Rogues)I I I think if I can jump in on this, I I like to consider it our journey similar to signal. Right? So everybody considers signal as kind of the gold standard communication platform for for privacy. Right? And why is that? You know, part of it is, you know, you know, nonprofit structure. Part of it is their policy they've posted, part of it's their verifiable trust, their open source and proven track record. And I'd like to see us, you know, kinda get to that where
you go to use signal because other people have beaten the hell out of it and know that it's secure and safe and they've they've passed the test of time, they've passed that hypothesis. So I think we're counting on the community to the the the group of folks that are that are on the phone now who are really excited about this to go out and poke holes in and and look at the source and look at the verification. But then I think, you know, bow to the to the to the normal user, just like I'm a normal user of of of signal,
is you get to the point where there's there's trust over time and you know this the foundation is built, you know that the foundation of privacy is built in and when you go and select that, you know, t branded model, if you say, look, I wanna be in super, you know, the cone of silence mode, that all the right things are happening. And it doesn't provide more friction. It actually the only friction it provides is if you knock on the door of a bad actor and it says, nope, you don't pass attestation and I'm not sending your prompt. That actually should be a really good feeling.
Yeah. That that that you see that that bounces off and, you know, one of the things we're gonna do is I'm gonna, you know, we're gonna put up a a fake bad actor with a t model that is not t attested. And so that you can actually see from a user standpoint, whether using your own consumer node or the API gateway of a failure. Because I think that gets us that that public support that the right right nerds in the audience have gone and looked through the code and tested it and poked and prodded it and and given feedback,
that there is no friction. The goal here besides the, you know, less trust me bro, more physics, is also it can't add friction. The only friction I want is when you're talking to a bad actor or you're not in a good trusted pipeline. That's when the friction that's and that and at that point, it should be a hard wall, not a not friction. It should be stopped dead because we prom we've promised trust. And if we can't deliver it, it it shouldn't the prompt shouldn't go through. Yeah. That that's really, like it's an exciting failure. It's like
Kyle (Bow Tied Blue Fin)it's I I'm expecting privacy here and, you know, my firewall, if you will, has has successfully blocked the the bad actor attempt. So it it's a it's the right it's the right friction in the right place. And Mhmm. Again, we're we're we're looking at this from an infrastructure perspective, and we're looking at this for, you know, setting the the pieces in the right place so that hundreds or thousands of providers can, you know, be offering this
t enabled, you know, inference at at phase. You know, we we see phase one really as temporary, but at phase two of full end to end confidential compute. And, we're we're just really excited that I I don't know of any other platform in the industry that is built on open source in a marketplace format that can achieve this. So it it's a it's it's a a a very deep moat when you say something being completely open source and completely permission. So it's it's something that I'm personally very excited.
Bowen BelmerYeah. I mean, this falls under the category. I put a message in the Discord the other day, but, like, I've seen so many projects in either the crypto space or now the DAI space since it's, like, eight years ago or 2017 approximately when I first discovered the space. And I have never seen a project that ships more actual usable practical utility stuff than you guys are with Morpheus and has, like like, the signal to noise ratio is, like, off the charts. So I wanna give you guys, you know, hats off for that, first of all.
One thing that I I did wanna pull on a little bit further as well is, Alan, you had touched on the idea of, you know, not having friction from a user perspective. So I understand that you are, of course, referring to technical friction in in using the actual product. Is there economic friction, through having private compute this way? Like, does it impact if I'm I guess, to reframe that, if I'm a user who's taking more tokens, and then having my inference, you know, covered perpetually by
by the fact that I'm staking this more tokens. Am I getting more bang for buck not going private and it's more expensive being done in a trusted execution environment? Does it not make a difference?
Kyle (Bow Tied Blue Fin)Yeah. So I'll I'll I'll jump in.
Alan (Nomadic Rogues)There you go.
Kyle (Bow Tied Blue Fin)So, again, right, it's back back to the original sliding scale, right, of of privacy. You're you're gonna pay no additional right. Your baseline is is no privacy and no additional cost because people are making money off your data. As you move up, there is generally a higher cost. Now I'm gonna explain where that cost comes from. For providers running phase one, right,
node attestation and node validation, the cost difference I would expect is near zero. Right? It's it's minimal. The cost of running a basic program on a very small CPU on TEE is more expensive than than, like, on an AWS instance, but it's not tremendous. Where the cost really comes in is when you have the GPUs within the TEE. And the difference there is
you can't run it through the massive data center racks that exist where there's, like, a 100,000, you know, h two hundreds or b two hundreds. Right? You're not accessing the same pool of resources. You're running something on a closed node of either one h 100 or eight h one hundreds that is specifically configured for this one instance and can't do anything else. So
there there is a premium for that both from a cost standpoint and from a an inference time standpoint. There's some overhead associated with how much you can inference and and the adder that that t adds to it. So do I have a cost estimate right now for you? No. No. I don't. And and each provider is going to be different in what they add. I could probably go and pull, like, the the hourly cost for GPUs
for TEE versus non TEE for you guys to get get a basic proxy of that, but, that's something that is just going to have to be seen. And the the great thing about Morpheus is that Morpheus doesn't set any price. Right? It's it's an open permissionless marketplace. So as the demand for this comes in, the competition for providers will will, you know, become significant and and generally reduce to to that
efficient point of the market.
Alan (Nomadic Rogues)Yeah. Go ahead. I I I think yeah. I think you finally got to where I was I was gonna jump in was we don't we don't set the prices. I think what you talked about is certainly theoretically true. Right? Is is as you walk up that security chain, it does get more expensive in general. Now who charges that? It's not something that the protocol layer charges. It's We don't charge anything, right? But the protocol is the protocol. The providers have the option based on their costs to set what they offer those models for.
And each provider has to say, look, I'm either gonna run at a profit or breakeven or at a loss for their own their own reasons, right? If I'm running at a loss because I wanna, you know, attract more people to my provider node and and, you know, build a build a reputation as that as a good provider node. But I think I I I would land on it's, in general, the more security you add into a system, it's probably gonna have pay a premium. You're gonna have to pay a premium. What that premium is is gonna be set really by the providers and as they do their own math in the back.
But we we certainly don't set that at the protocol level.
Kyle (Bow Tied Blue Fin)Yeah. And and it's not that premium is not needed for everything. Right? We've I don't remember which talk it was, but I I was on a talk with, I think, Secret Labs and and and Fala and Eliza and someone else. And we talked about, like, you don't need full confidential compute when you're asking about the weather. And you don't need full confidential compute when you're asking to have a public article, you know, summarized. So it's I I think there's
there's a the option of privacy is is what's valuable, not necessarily blasting privacy every every single, you know, place that that you can get. So it's kind of you use that cost center strategically when when you need it, and and that that's what I think is is really kind of the game changer.
Alan (Nomadic Rogues)Yeah. I agree. It's another it's another arrow in the quiver. It's it's like, look. I can go talk to the car salesman in public and not and and be okay having a chat about it. If I'm gonna go talk to my doctor, I really want that to be private. So I don't need I don't need a a lockbox for everything.
Bowen BelmerThat actually brings me to a follow-up question. So I was basically hoping you guys were gonna stay something to the effect of what you did. The the follow-up question that I have is actually for David if he's able to speak. Is it possible to build into Everclaw, like, into the skill package, a skill where my agent can, basically, I can outline for when I would and wouldn't want to use a like, specifically a private model, and it just naturally knows when to execute that and when not to?
David JohnstonThat's a really great question. I mean, it's almost a logical extension of Everclaw's router. So the router already looks at the prompt and says, oh, well, this is a really simple cron job. I'm just gonna throw this at a small model and save tokens. Whereas if it's a big coding job that's really complex, it's gonna go to GLM five or one of the more powerful models. Right? And so it would be interesting to think about, like, a privacy sliding scale for the same thing. I'm I'm gonna I'm gonna go talk to my agent and add that to the backlog today.
Bowen BelmerYeah. That'd be awesome because this was actually something that dates a little ways back for me. I I did a podcast interview with Lisa Loud over with Secret a handful of months ago, and we were kinda talking through what TEE looks like, like, through a compute interface, and I I was imagining, like, a toggle, basically. Like, I'm turning on private mode and not private mode. But, yeah, via the agent, like, it just could be the toggle, theoretically.
Kyle (Bow Tied Blue Fin)Yeah. You you can but it is a toggle. Right? Because you're going to have public versions and you're going to have private versions of the same model. So wherever that toggle exists, it just has to change the model name, if you will, that that you're calling through the router. But it it's pretty much the same way that that you expected it.
Bowen BelmerWow. This is, to me, it feels like, like, another really big step forward in you know, we always talk about how fast things are moving, but I feel like if we look at it in terms of, like, progress cliff, basically, or, like, you know, a staircase, some steps are, like, much larger than others. And this step towards having actual privacy, not just trust me, bro privacy, is is massive. And it's one of those things that I don't think people are gonna fully appreciate.
Well, actually, the the average end user might never fully appreciate it because by the time they discover it and are using it, it just exists. But I'm really glad that, you know, we're here, anyone involved in the call today, and our communities are here to see, like, what that actually looks like from the earliest stages. Yeah. It's ultimately just so cool, and I I cannot, express my gratitude enough to everybody who's working on being able to actually build it. Can can I ask you guys, like, is there anything that was maybe surprising
about, you know, figuring out how to do it that you weren't anticipating or maybe something that changed your mind and what the best way of doing the execution is or something like this?
Alan (Nomadic Rogues)That's a that's a good question. I don't think there was any real I'll say for me. Don't think there was any
Kyle (Bow Tied Blue Fin)Let me say for me and then Right. Go ahead. You'll be the the depth of what I of under what I have to say. I was I was surprised how straightforward it becomes for the reason that it's math. Right? It's you're you're following the math. Right? I I thought that there would be a lot more handshaking and negotiation and and and testing and figuring out, you know, how everything comes together. But at at the end of the chain
at at the end of the day, the way that Alan and team implemented it is you're just following the hash. You're following the hashes. You're following the math, and and you're just confirming attestation and numbers. And it it becomes simple when you can follow something all the way from the deployed code all the way back to when it gets verified. So it was probably something that I that I shouldn't have been surprised about, but but I was,
you know, I guess at the end of the day. Yeah.
Alan (Nomadic Rogues)I think it's funny. There's also a sneak maturity in the industry a little bit. I mean, if we we talked about this months ago and the how to solve was still unclear, but I think the progression in general of you know, zero knowledge proofs, is just cryptography and checking hashes. Once that, you know, that matured to a certain point that then as we looked at it, you're right. It's it's follow the math. In fact, I you know, being fully transparent, you can sort of see it on the github releases.
We found a we found a little bit of a flaw, which is, you know, we were checking that attestation when we created the session, but not every prompt afterwards. Like, well, wait a minute. That's not gonna be good. How do we fix that? Well, turns out the fix is really easy. In fact, we've got it running on tests right now. We'll be pushing to main later, is you don't have to do the full attestation again. You just grab that that cryptographic hash that you started with, that you attested, and then every time you go to send a prompt,
you check it. You check it that that crypto hash. And so having it in that math base, again, that's why I like the physics physics not promises, is it makes a lot of those checks, especially at the technical level, so much easier. You know, I'm checking I'm checking two shot two fifty six results rather than trying to recompute everything all the time. So that's efficiency, that's speed, that's all sorts of stuff. So I think the ZK maturity in the industry helped
bring this at the right time, right place, right time. David always talks a lot about, well, just happens at the right time. I think this was another one of those, a lot of things had to coalesce, but yeah, physics not promises is really good.
David JohnstonI love it. Now this is this is this is great. Yeah. But I just wanna say, I think this is a huge improvement. This unlocks so many new use cases and privacy, I think, is gonna be just this explosion of going from personal AI being purely local to now people will be a lot more comfortable that their information's really being honored. So, yeah, thanks to the team for awesome stuff.
Bowen BelmerYeah. Couldn't couldn't agree more. I'm trying to think. I I don't know that from a user perspective, I have other questions at this exact moment. I mean, I definitely am looking forward to giving it a try. Although, to your point before, giving it a try from a user perspective should just be a a different tag on the model that I choose, which is a good thing. That's what we want it to be ultimately. Right? So Yep. Yep. I should only know that there's a problem if there's a problem.
Kyle (Bow Tied Blue Fin)The the fun test that that Alan mentioned earlier is that we're gonna give a fake model out there. We're gonna put a model that claims to be private that's not, and you're gonna be able to go and and hit that model, and it's gonna tell you, no. I I don't approve of your actions of of thinking that you're private when you're not. So it'll be the the positive what is it? The not the false
Bowen Belmernegative, but whatever it is. Validating the the bad path is bad. You know, what I need someone to do is take, like, a sizable amount of Ethereum or Bitcoin or something like this and start releasing a seed phrase into their, their session and seeing if the money moves. If the money doesn't move, you know it was private.
Kyle (Bow Tied Blue Fin)Sure. Why why don't you just just send me your key, Bo, and and I'll do that for you. Yeah. Thank you. Somebody
Bowen Belmerwith more liquidity than me, I need you to run that experiment, please. Thank you. No. That's great. Well, that's good stuff, guys. I should we see here? Is there any, anybody else from the community or or in, in the group session here that has any questions or wants to jump up, and get involved? Alright. We have silence, so it might not be the case. That is okay. Yeah. I mean, again, ultimately,
the team continues to ship. The community continues to I actually love, Kyle, you mentioned before that the the cat was a little bit, released from the bag because our community was actively monitoring monitoring the GitHub, which is very cool. That actually makes me really happy to see that there's, you know, there's people following closely enough that they're actually, like, specifically watching for commits to see, you know, what what's going on under the hood in real time. I think that's very cool.
Alan (Nomadic Rogues)Yeah. I agree that. I was surprised to see that yesterday as well. I'm like, how do they know? Like, oh, that's right. The shipping channel. So now I have to make sure that our our release notes are a little more clean. So since people are listening, I'll make sure that, they're a little bit more articulate sometimes.
Bowen BelmerOh, no worries at all. I mean, I I think it's a good stuff. It's it's a again, it's a feature, not a bug of building in the open and and having everything be open source. And, you know, this is a point that I had made actually in in the Discord as well over over the last couple days was just a reminder as to, you know, why Morpheus is what it is and why it's being built in the way that it's being built, being, you know, open source and permissionless and decentralized. And and it's ultimately
because similar to something like Bitcoin or something like Ethereum, if everyone who's working on it at this exact moment today, you know, suddenly disappeared and could no longer work on it, the project itself can and will survive because the project is public open source code. So at any point, somebody or a team of somebody or whoever would have the ability to start working on the code base, and that that just would be what it is in the same way that it you know, as if one person is mining Bitcoin, the network survives and similar,
to Ethereum with Ethereum nodes. So just as a reminder for everybody that, you know, that's that's ultimately why, you know, this exists the way that it does. It's also why it's such a powerful, network and tool, and it's also why, in some cases, things probably don't move quite as fast as they would I mean, definitely as fast as they would in, like, a centralized format. Although, I would argue that you guys are still moving very fast and there's updates, you know, all the time. People are not waiting months and months and months between updates. I I think the the amount that has been shipped from
certainly two years ago, but even one year ago is, you know, astonishing. So here's to the future. I mean, I'm certainly looking forward to that, and we'll we'll keep engaging from the community perspective. Did does anyone have anything else? David, do have any final thoughts kind of as we wrap up this week?
David JohnstonNo. This is just amazing milestone, and, you know, it's great to see it spreading all over the ecosystem. So congrats, to everybody for this big unlock. But this is something we've all been looking forward to for years, but it's so cool to see it in reality.
Bowen BelmerAbsolutely. Well, thank you everyone for joining us then. If you are listening to the recording of this that was remastered by the community, first of all, thank you to the, the community, Mir Mir Morpheus, for doing that remastering, and then also, thank you to you for listening. And with that, David, you have our our final sign off. As always,
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...