Mere Morpheus

Mere Morpheus artwork

Mempalace, Models & Mythos | Anthropic's Project Glasswing & Securing The Stack

49 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 #68! Recorded: 9 April 2026 AI SUMMARY: This week the regular crew highlighted GLM 5.1’s rollout on Morpheus as a major advance for open‑source coding models. It covered practical model routing guidance using the Morpheus API staking economics and moving daily workflows to DAI for privacy. The segment on celebrity AI featured Milla Jovovich’s open‑source MemPalace tool and a cameo from Elizabeth Holmes. Discussion then turned to Anthropic’s Project Glasswing zero‑day disclosures and why cross‑model code audits are now essential with post‑quantum DEAI positioning noted as next on the radar. Timeline: (00:00:00) Intro (00:00:49) New models on Morpheus: GLM 5.1 and Trinity overview (00:03:44) Opus "getting dumber"? Thinking levels, costs, and usage nuances (00:06:56) Beyond GLM: Gemma 4, Trinity, and use‑case‑based routing (00:11:12) Costs, staking, and calculator: planning your sessions (00:14:34) Dogfooding DAI: moving real workflows to decentralised AI (00:18:14) Cutting the cord: writing and coding fully on open source (00:23:53) Hybrid agent setups: frontier fallbacks and local 9B tool callers (00:26:33) Unexpected celebrity builders: Mila Jovovich’s MemPalace (00:29:40) Enable your agent’s memory_search and flexible backends (00:33:19) Another surprise voice: Elizabeth Holmes on privacy and AI (00:35:04) Project Glasswing and Mythos: zero‑days and responsible disclosure (00:37:13) Auditing the world’s code: timelines, risks, and open source parity (00:40:40) Will smaller players be left behind? Patching and readiness (00:44:01) Security practice tips: cross‑model audits and costs (00:47:43) Call to action: try the models and own your AI 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 BelmerHope everyone's had a great week. I feel like time is really flying by. And every time I blink, we are back to our weekly town hall, which is always exciting. I love, the opportunity to talk to you guys and hear from just the community in general on all the happenings with the AI, Morpheus specifically, and then just nerding out about stuff at large. This is like the best community of functional nerds that I'm aware of, myself included.

Probably the best term for it. I like functional nerds. Yeah. Nerds with families. But anyway, so okay. So we have a handful of things to cover this week. Wanted to start out with the new models that were brought to Morpheus. The devs working on the API gateway and also, you know, those hosting compute. Got this thing live, I think, the day that it went public, and that is GLM 5.1,

which is the latest open source coding model that is, from my understanding, competing fairly directly with the likes of Opus, which is really, really crazy, that we now have something that is open source and and that powerful. And it also, from my understanding, benchmarks with something like MiniMax 2.7. And there were additional models as well, the Trinity models that were also

made available. So David, are you able or I guess David or Kyle, are you able to do just an overview from a technical perspective? Like, what does this bring to the table that we maybe didn't have in an open source capacity previously?

David JohnstonSure. I mean, I'm happy to speak to it as a as a user. And then, yeah, I think Kyle would be perfect to to talk about how it got launched so so damn fast on on Morpheus would be great. You know, the benchmarks are pretty impressive. Like, this is looking basically on par with Claude Opus 4.6, which for anybody that doesn't know is basically the cutting edge model for all things coding. And so for for me, I think this is a huge deal. It's like a big moment

for open source. And as soon as they put out the open weights, immediately people started grabbing them, downloading them, putting them onto inference options for people. So my testing so far is is really good. It seems solid. I even saw some benchmarks today that claim the cybersecurity aspects of the reviews 5.1 is doing is actually better than Quad 4.6. There's also been some news about the degrading capability

of Quad 4.6. I think Anthropic might be, you know, throttling back some of the thinking or the tokens that they're giving people. So maybe a combination of Cloud 4.6 has become slightly dumber, and the new open source model has just passed it or is at least on par. So that that's my hot take. I'll let Kyle speak to how, how he managed to get it up so fast.

Kyle (Bow Tied Blue Fin)Sure. I'll, can you guys still hear me? I'm clear. Alright. Cool. So, I'll I'll give a couple. So in my from what I've seen from some of the benchmarks yeah. G l GLM 5.1 is is really amazing from an open source standpoint. Whether it's OPUS 4.5 or 4.6, I think they're all really in the same range right now with different strengths in different areas. Your comment on the,

like, the narrative out right now that Opus is getting dumber, I think that was mostly for users using Cloud Code rather than the API directly. And that was just because Cloud Code, when Anthropic was experiencing that massive inrush of usage through the OAuth and OpenCLOA, and and we kinda talked about this last week. I think that they put in some functionality of of

auto thinking reasoning level. And what we perceived as intelligence was an adjustment, and I'm sorry for the background noise, guys. What we perceived as intelligence was them manually cranking you down to lower levels of thinking. So if you specify the thinking to high or max, it costs a lot more on the anthropic side, but you can still get that old Opus experience. So I don't believe it was a model change, but rather a

usage change. And you can kinda do the same thing with GLM five one if if you choose to have thinking enabled or disabled, but I just want to kinda get get that out there. As far as the other models, yeah, there were a bunch of models that we added into the router. I'll I'll list through a couple of them. GLM five one thinking, GLM five one non thinking, and that difference is important because some agent tests that you're doing

and depending how you're parsing responses, you may or may not want those thinking tokens. You may wanna trade speed for you know?

Bowen BelmerRugged, but we're back.

Kyle (Bow Tied Blue Fin)Yeah. Okay. I'll just kinda jump right back into the continuation of this. So I was talking about GLM five one thinking versus nonthinking, and they're they're just cases where you might wanna run a nonthinking variant for the case of speed, for the case of cost, or for the case of JSON kind of output parsing, it's it's sometimes a little bit more complex with thinking. So that's really on the on the GLM five one side. Now we also launched a bunch of other models into the Morpheus Inference marketplace and Morpheus Inference API.

That was the Trinity model. Unfortunately, the we launched the Trinity model and then g l m five one came out, which I personally think is definitely a a higher performing model and and a better bang for your buck. But nonetheless, different models for different applications. Now the other models that we launched are are also very interesting. So there's been a lot of news around the Gemma four models. There's two variants that are available, both the

27,000,000,000, the MOE, and the 31,000,000,000, I believe it is, the dense model. Very, very strong models. These are, I would say, best in like, reaching best in class levels or reach not not at not frontier levels, but really high levels of performance at low cost and low parameters. So these are some that people could even run locally on the right levels of

of hardware. Maybe not the the same tokens per second that you're going to get, but they're models that are that are very flexible and and very strong. So I'm going to see a lot of I anticipate a lot of agents running on these for noncomplex tasks, whereas g l m five one and and the like are going to be used for kind of heavy reasoning, coding, etcetera. So really excited about those. Also, we launched some of the very small quen models,

Quen 9,000,000,000 and and a couple others that that I can we can send the list after this. Those are for even simpler tasks where where I talk a lot about the sliding scale of intelligence, the sliding scale of privacy. You need small models for certain tests. Right? Certain extraction tests, certain summarization tests. So you really you don't wanna just throw, you know, all of your money away at at the top models all the time. You wanna kind of intelligently route models to the test that you're doing. And David can talk a bit about the the model routing

that that he implemented in Everclaw, and and a lot of the agent harnesses are are using somewhat similar structure. And, you know, we know from Claude Code, uses Haiku for some things, uses Sana for some things, and and Opus for others. So somewhat proven in the industry, and now we have the models available to to support that in a in a, you know, best in class way right now for the different model sizes.

David JohnstonYeah. I mean, was super easy to update my agent to use the new models. So appreciate you pushing them out so quickly. Kyle, I literally just told my agent, hey. You know, there's new models on Morpheus. Go check the gateway. Go check the the router, and boom. It updated the list, and I was off and running. So anybody using a Morpheus API key, you you know, know, just tell your agent to do the thing, and you should be able to get on 5.1. And the routing, I think, is is important. Like you're saying, there's these smaller models that are coming out with with better quality.

A lot of your heartbeats or cron jobs or stuff, like, they don't you shouldn't be paying Claude, you know, to to run those tokens. There are now really good alternatives,

Kyle (Bow Tied Blue Fin)and you can grab them on Morpheus for an extremely low cost. So yeah. Yeah. And, like, all the all the memes that you see of, you know, the guy with the sledgehammer asking Claude you know, asking Opus what's the weather. It's like, that's not necessary. Let let's let's use our money and and our and our energy intelligently in some ways. But yeah. And just one other comment is that this isn't only on the API. Right? Even if you're just running through hosting your own consumer node,

I I would love for people to use the API for information, you know, as well as as inference if they choose. So you wanna find the all the models available on Morpheus and you don't want to pull the blockchain for the registering the information, you can just hit the model's endpoint on the Morpheus API, and it'll give you all the information. It'll give you the blockchain ID. It'll give you all the models that are available with bids and up and running, and it it's it's just very good informationally for an agent to be able to hit a simple curl command and and then go and and get the information it needs to make decisions even if it's running its own node. So just just kind of a note there.

Bowen BelmerCan you speak to Kyle, I I was actually looking for the dashboard site again to see from a efficiency perspective, like, a cost perspective, how these compare to some of the models people would have previously been using.

Kyle (Bow Tied Blue Fin)Yeah. I believe we have that up on tech.more.org. Let me just pull up the exact URL. Yeah. Tech.more.org/calc. So that's our calculator, you know, brief calculator page that you can kind of put in. Alright. If you're staking tokens, how much, you know, daily credits do you get from staking, whether it be in the API or running your own node specifically. We also have some price we have our I believe we have our pricing in here or at least a link to the pricing

through the API, and you can really go there to figure out, you know, how how much is this gonna cost, how many tokens do I need to stake if I choose to go that route, or you can go directly to a pidocs.more.org, which is the full API reference page for using the API. I would say API reference page page is probably the place to go for the, excuse me, the full API information, but this is just a good

first step in, wanna get some quick information and and do some some basic staking calculations, that that's the place to go. It's tech.more.org.

Bowen BelmerAwesome. Yeah. I'm just looking right now, like, as an example. If you had a thousand more stakes to the API, or the inference gateway right now, you'd you basically have something close to, like, between 800 and a thousand requests to GLM five every single day, which, like, when you think of that in terms of purchasing API tokens from, like, a Hugging Face or somewhere else where you're paying buy token, I mean, that's just an insane value,

when you consider that it's perpetual. That's pretty pretty amazing. Absolutely.

Kyle (Bow Tied Blue Fin)And if you're staking so let me just I like to go through this just to make sure that everyone remembers exactly how it works. If you're running your own node, there's a certain cost that a provider is putting out saying, I'm willing to offer my compute compute at this price. And let's say it's a thousand or whatever the number was, that's saying, I'm going to rent one, let's say, one slot from this provider, right, where I I have a full slot of GLM51 for the whole day. I can send as many requests as I want, but they're going to fill serially.

So for applications where you have a bunch of users or you have a bunch of agents running that you need to send prompts concurrently, you might need more than one session. So you might need to stake 2,000 if you wanna run two prompts side by side all day long. So so it's just good to think about it on like, just like if you're renting a VM, like, how many how much do I need to duplicate this? How many, you know, VCPUs do I need? How how much stuff am I gonna be running at the same time? And you should plan how many sessions you need accordingly. But just wanted to get the point out that renting

full access to a model through the node is a single stream that you get, and you can, of course, scale from there. But it is. It's it's full day long. You know, as as long as you choose to create your session for, it could be a full day, unlimited prompts throughout that session, and in perpetuity, really. It's you could you could just stake the next day as well.

Bowen BelmerAbsolutely. That actually brings me to, of of the topics I wanted to cover today, I'm actually gonna skip two or come back to it and jump straight to three because it's very relevant to what you were just talking about. And it's, basically the idea of as like the DAI community eating our own dog food in terms of the DAI workflows. And what I mean by that is so I was in a private conversation with a couple people yesterday, we were talking about, like, building DAI alongside

the use of central tools and, like, becoming more DAI centric over time. I think we need to think of a term for this. Like in the Ethereum world, it would be like becoming more bankless, which is of course the tagline of that podcast. But the idea of taking your your workflows that you're executing on and transitioning them to ultimately someday, hopefully in the near future, being DAI native. And so I'd love to hear, like, if anyone's willing to share

what their workflows look like right now, and and are there new ways that maybe they're integrating DAI that they weren't previously? I can start if people wanna think about it in the meantime. So fairly recently, I, you know, have had my own OpenCLOS setup. I have it running through Morpheus Inference, to the API gateway. Shout out to the team who built that yet again. And I've started basically teaching it to be somewhat of an assistant for me with my real estate financing business,

which today, what that looks like is basically taking information and whether it's like loan pricing across, you know, more than a 100 lenders or it's looking at an approval from an underwriter and just having it break down information for me. And basically, like like, the way people will use the data to summarize things like articles and and things like that to provide them, like, curated lists of what they need to know, It's essentially doing that for me to cut my my time down in like manually sifting through information,

which has already been beneficial to to me and my business, you know, just in the last few weeks that I've been doing that. But the reason I hadn't done it previously in any other centralized platform was because that data is really valuable. And I don't want, you know, anyone else to have access to my client data or anywhere where I I can't verify that, you know, I'm getting the most privacy that is available to the best of our ability

when I'm running these tools and, like, training this agent to ultimately become, you know, good at this job, like, as an assistant to this job. And so just in my own life, like, I've already been able to integrate that a little bit. I expect to, you know, continue to scale it up over time. And I hope to get to a point where a lot of the models I'm sorry. Excel models that I've built that I used in the past, which I, like, track my clients' financing with to determine when they're in a good position to refinance based on market conditions.

I anticipate a time fairly soon in the not so distant future where I'll have an agent that'll be monitoring those models for me and ultimately, like, making those recommendations before I'm even manually getting to them, which would be huge for me and, like, anyone who does, you know, a similar job to what I'm doing. But does anyone else have any examples of of things, you know, that they've been able to transition to, like, fully DAI?

David JohnstonWell, I'm happy to talk about that, Bo. I mean, the first thing I transitioned was was basically all my writing. Right? I remember killing very happily my OpenAI subscription at a year, year and a half ago. Right? And I I never really used it heavily because I knew it wasn't secure or private. So I experimented with it. But as soon as the open source models and this is probably even back in, like, the llama three point three days, got good enough on the writing side. I was like, alright. I don't need Chachi PT anymore

at all, and I can kill that subscription and move completely to decentralized AI. And I feel like we're at a similar moment now on the coding where I haven't been using Quad for a while, at least directly. You know, I've been using it through some inference providers for better anonymity. But I feel like with 5.1, you know, if my testing bears out the next few days, I'm gonna be able to officially cut the cord and be using only open source models for all of my coding. And I think that's a pretty huge moment

for the community that we can sort of move strongly into that whole new set of users and personas of people that I mean, that seems like the killer app certainly the last six months is every all of a sudden, everybody's a coder, and everybody's coding. Even people who weren't developers historically, you know, the tools become good enough to really pull that off. So, you know, I think that's probably my my next big moment that I've been really looking forward to. And it it was clear it was coming,

but the last few months, the models just weren't quite good enough. I I think we might finally be there with with GLM 5.1. So that's that's my next step is cut the cord on on any cloud usage even indirectly or even through inference providers and just be a 100% on open source for coding. That's that's the next milestone I'm I'm hoping to hit here.

Bowen BelmerAwesome. That would be amazing. A pioneer as always, David. You'd certainly be one of the earliest, probably to do that, at least in in the capacity that we're talking about. So that's amazing. Definitely keep us updated on that. Looking forward to hearing about when that cord officially gets cut, and then no one else will have excuses as to have not also done the same. No. I'm just kidding. But that is awesome. And

Kyle (Bow Tied Blue Fin)and I'll just add on. I mean, we have a couple larger users. And again, I'm sorry about the background noise. Apparently, someone's doing major construction near me. The we have a couple major users on the Morpheus API who have integrated this into their production workflows. So really moving to the open source models for I guess, we'll have them on a in a couple weeks to to talk about it more, but making the switch from Opus, making the switch from Sonnet to these open source models to run their businesses and and rely on Morpheus

and the the providers who are hosting through Morpheus for to service them and their customers. So it's really great to see kind of that transition beyond just us and beyond what we're doing, you know, with with our coding and so on. Personally, also, I've switched my AI research platform. What is AI Research OS. Originally, was doing it through Opus, then I was doing it through one of the smaller

I believe it was Sonnet or one of the smaller ones. I switched fully over to g l m five one, and I'm using Gemma for some of this stuff too. So, again, finding, you know, the parts of my infrastructure that could use open source models, which is all of it now, and then the parts that can use even smaller open source models to help, you know, give, you know, those customers a better cost. So, you know, really excited about about where we've come, and every month, you know, all the models get smarter. So it's gonna be you know, we're we're approaching that

that line that says this is good enough. Right? More intelligence isn't required for this task, and it's certainly something that can be done through open source. So really excited there and really excited to save a lot of people a lot of money, you know, and and that funneling down to users getting great services for for a low price.

Brian BourgerieAbsolutely. I can pile on for a second while we're at it Please do. For anyone who I haven't shared with yet. My current setup with some of my agents, particularly my Hermes agent, I have the default is running off of good pure stake more, mainly off five point zero and five point one right now. If it has something that needs something frontier, it

would be over automatically to run through Venice. You'll be using or Opus. So, obviously, that's not open source, but if I do need something that requires that memo, it's at least more anonymized. And then I also put in I don't know how to say it. If it's a RISE or something. There's a there's a specialized model that was derived from plans 3,590,000,000 parameter. The new one is specifically

for Hermes and for tool calling and things like that that's running completely locally just on a 16 gig. So that's actually performed quite well. And when I was having issues trying to connect all the different things and routing and all this other junk, I ran, like, almost the entire day off of this 9,000,000,000 that was optimized for Hermes, and it didn't suck. Like, it did all the regular stuff completely fine, and it didn't trip up where I thought it would trip up in some of the more

challenging tasks that I had. So I'm still playing around a ton, but little by little, I'm able to, like, expand out what models are being routed to. And at a minimum, everything's anonymized, if not completely over and or mobile. So just more fire under everyone else's pants. Like, you can do it, and it's probably gonna be, like, in a week or two, and you can have an almost sovereign setup.

Bowen BelmerBrian, were there any tasks that you, like, specifically anticipated there would be slip ups on and you were surprised that there weren't or anything like that?

Brian BourgerieYeah. Mainly, like, doing code fixes and debugging and and even building a couple new features. So I had I was trying to eat my own dog food in a sense and not fall back on SSH ing in with my Cloudmax account. And I was trying to, like, fix all this different routing and and a bunch of things like that, and it was powering through.

Like, it might have maybe taken me, like, one more prompt than, like, a single shot, but I was I was pretty impressed.

Bowen BelmerAwes oh, there you go. Sorry. Awesome. Thank you. Yeah. Thank you for sharing. I was I was curious to hear, like, kinda what what exact types of things you were, you you were looking, to use it on and kinda how the outcome was surprising in in its ability to pull it off. Because, honestly, I'm still at the point where anytime I can successfully get my agent to do anything, I'm I'm floored. You know that that meme

that was like it was like the stonks meme, but it was like the white kinda like stick figure guy that was like, traitor, and it was like, spelled wrong? That's how I feel every time I do anything with code in it, but like developer, because it's, you know, giving me powers that I would never have otherwise had. Speaking of which, somewhat of a good transition to a next segment that I wanted to start doing, every week or at least as as they pop up on the timeline or as people send them to me. But that is the

unexpected celebrity chime in of the week, with regard to DAI specifically. So I've been seeing a couple things like this where people who, at least from my perspective and probably the perspective of many, are like very much outside of the scope of almost like people that I'm surprised to see paying this close attention to not just AI, but specifically, like, privacy and and personal access to intelligence

and data security and things like this in AI. And so the the first one, and what made me think of this was the the joke about everyone becoming a developer. Although it's not a joke. And this is Mila Jovovich, who was an actress in many, many different things. Let me just pull up her IMDb really quick. Famous for roles in The Fifth Element, also in

let's see. She was in two was it the Resident Evil movies? She was in Hellboy. She was in tons of movies, like a really accomplished actress, you know, going all the way back to the eighties, actually. I'm looking at her IMDb page right now. It is very impressive. She recently published open source code on GitHub that has now been downloaded thousands and thousands of times and has actually been integrated into Everclaw called Mempalace.

David, can you talk about, you know, what that is exactly and why you chose it for for Memclaw? I'm sorry. For Everclaw?

David JohnstonWell, it's it's really an interesting trend. Right? You have smart people from other industries and fields that have never been able to sort of touch GitHub or code and now kind of flowing in using these tools. She she mentioned she partnered with a guy who's actually in in our community. I've seen him around some of the the the groups, Ben Sigman, that is, like, a longtime developer. So she did this in partnership with with a good engineer

and basically came up with sort of her own approach for, okay, how should you handle agentic memory? Like, the idea of a mental palace has been around for thousands of years. Right? You can kind of imagine in your mind's eye, I'm walking into this room. What do I have stored in this room? What do I have stored in this drawer? You know, you you could say this is sort of just a, you know, a glorified way of of handling a database. Right? But, you know, it's kind of clever and catchy and and and kinda caught on. Right? And so they picked up a lot of attention

and, you know, stars on GitHub. You know, and people have, you know, criticized it. Oh, well, let's do this or that, but it's all open source so they can make pull requests and and prove it. Right? But it gave me an excuse to go back to my agent, and surprise surprise, OpenFlow does not configure memory search by default. So if you haven't done it, go into your agent and tell it to use memory underscore search

and actually activate it in your configuration. Because if you didn't do that and you have to do that step, you haven't been searching your memory when your agent is doing a task. Right? And it's today, you know, probably you're having to give it the context again and again if you don't have that activated. So number one, activate in your agent the ability to search its own memories because that doesn't come baked in, which is a a surprise. You know? I I had seen that error sometimes.

You know? And and so it's been so refreshing ever since I I updated the configuration to enable memory underscore search. All of a sudden now, I'm having to give it a lot less context when I'm talking to it. Right? It's it has an easier time finding its standard operating procedures and stuff like that. So so one, this was a great excuse to go down the rabbit hole on how my agent is using memory. Two, you know, Everclaw, you know, Morpheus in general ought to be agnostic to any memory system. So, you know,

integrating this, there's other great systems like Neo four j that is an open source graph database, but it's probably something you wanna use for larger, heavier loads. Right? It's really designed for for large datasets. So, yeah, it was it was fun integrating. You know, it's good to see it was MIT licensed open source, and so it's super easy to add as an option for my agent. So and and that's what you want is you don't wanna bake in a lot of defaults that force people to use x y z, but you at least wanna make it easy

for them to use the different options based on on what they need for their setup. Right? So, yeah, it was a it was a fun little integration. It only took a a few hours of of coding and testing, but it it was a good excuse to go and really tear apart or, like, how is my agent using memory? What do I need to do so that when somebody uses an agent, you know, it comes with at least the basic defaults turned on for for memory search and stuff like that. But it was it was it was a good experience, but I think we're gonna see more of this. So it should almost be like a recurring thing, like who

you know, what famous person is shipping cool stuff in AI this week? And I I wouldn't be surprised if we see a lot more of that.

Bowen BelmerAbsolutely. And and particularly interesting, like you were saying before, like, that it came from kind of an unexpected source. Granted, in this case, a partnership, you know, with with a developer, but I think we're gonna see more of that too. And and you've said this in the past, David, like, everyone is a developer now in some capacity, or at least can be. Right? And so I I that'll really change where where different ideas and products and and agents and skills come from.

This is definitely a segment that I wanna try to do, and celebrity's kinda tongue in cheek. It doesn't have to be, you know, someone who's like famous per se. But I I just like these stories about people who maybe would not who would not have maybe otherwise worked on shipping software or code that now are, whether it be, you know, on their own or through a partnership or something like this. I think they're just fun stories to cover. I do have a second one for this week, and it ties into the initial topic two that I skipped over, and we will jump back to that because that should be a pretty big one. But the second one of kinda unexpected chime in was

Elizabeth Holmes, who I thought was in prison, and I I think maybe still is. Somebody correct me if I'm wrong. But of Theranos' fame, if you've ever read the book Bad Blood or seen, I'm sure there's documentaries about this, wrote a pretty interesting response and got a lot of engagement on on X over the past couple days as a response to, the news coming out of, Claude versus Open Claw and then also Project Glasswing,

which was actually the topic I was referencing and wanted to talk about. I don't know if there's too much to say about this other than, you know, if you're interested in reading, you know, some good entertainment on the timeline from somewhat of an unexpected source, check it out. I was surprised to see someone like Elizabeth Holmes chiming in on this. One, because I genuinely thought she was in prison, and I'm I'm not sure actually if she is or isn't. But then again, we see SPF pop up on the timeline still sometimes. But two, it was just another example of although it was someone who was in technology

previously, not necessarily in AI, who's recognizing the importance of privacy and, personal ownership of data, and like, what some of the overarching implications are of some of not just the technologies, but the way in which they're being implemented, which I think is actually a good swing over to Project Glasswing and some of the other things going on. So if anyone is not familiar, so Project Glasswing

is a project that was announced by Anthropic, basically where their new Mythos model, they believe and and are claiming, is so powerful that it's discovering zero day vulnerabilities, in some cases that have exist for decade existed for decades across pretty much, from my understanding, they're claiming that this is pretty much like most major softwares, or at least many, many, many companies are exposed to these things.

And so they're they've created Project Glasswing, is basically a collaboration with other large companies, I guess, to give them access to using this tooling to fix these vulnerabilities prior to it going live as as like a retail product, which will eventually, you know, something similar will eventually become an open source product. Some of the companies involved are companies like Amazon Web Services, Apple, Cisco, Google, the Linux Foundation, Nvidia,

JPM, Microsoft. So so really, really huge names. I'd be interested if anyone on stage here has any opinions on this or any thoughts on this just generally.

David JohnstonSure. I've got the hot take. I I think the a 16 z response was was probably the best thing I saw, which was Mark Andreessen's like, guys, we're about to enter a a moment where every vulnerability and every piece of software ever becomes exposed. And then in the next moment, there will be no more vulnerabilities. And and I thought that kinda captured it pretty well, which is, yes, as these models get incredibly good

at pulling apart coding vulnerabilities, like, all the stuff that people didn't have time to go line by line and, you know, perfectly check and all that code that's built up for thirty, forty, fifty years, you know, it's all gonna get reexamined real quick. And I I'm glad that they're sort of doing responsible disclosure, which is what you'd call this in, like, a white hat hacker kinda context. It's like people that find a bug should report it to the maintainer

quietly and let the maintainer fix it rather than be a bad actor and throw that kind of stuff out there for for, you know, people to to exploit. Right? So I think they're doing responsible disclosure, which I'm glad to see. And it's important. Right? Because if a bad actor had invented this level of model first, they would be going and having a field day with those zero days right now. Right? And instead, because they've been responsibly disclosed

and there's been enough time, apparently, the maintainers have already gone and fixed a lot of what was initially found. So I think we're gonna kinda go through waves of this. But to your point, we're only, let's say, a few months out of a purely open source model being at that same level and finding all those those zero days. So, you know and that's gonna sort of, I think, be the next wave is you want good actors adopting those tools quickly, having the best open source available,

and so they can go and patch. Because there's a lot of people that aren't the Linux foundation who, you know, would probably like to find their own zero days, and, you know, they don't have the privileged access. Right? So, you know, yeah, that's it's gonna be an interesting future, but I I I think this moment is mostly gonna play out maybe the next few months, but certainly this year, right, where all the old software gets examined. And then going forward, you've got the ability to do it at an incredible level of auditing of all new code. Right? So I'm I'm not very doomer about this. Like, that's Anthropix's whole marketing thing is this thing is so dangerous, you know, and nobody else can use it. They've been saying that since ChatGPT two,

right, when they when that team was still at OpenAI. Right? And I don't think anybody thinks ChatGP two was seriously dangerous in any way, but it's great marketing tactic. So, you know, I think they're playing also some of that angle. But, yeah, you know, I I think everybody's gonna get these capabilities pretty quickly.

Bowen BelmerDo you think I guess the the one question I would have about it, though, is kinda to the point that you were making. So these people who are and by people, I mean institutions that are directly involved in Project Lastwing and so have this, like, premarket access access to fix their own code. I guess, are there, like, companies, like, smaller places that are gonna get left behind? Like, this will go retail before they have the ability to fix it for themselves, and now all of a sudden there's this, like, massive moment of vulnerability.

I mean, I do agree with Andreessen's assessment that and and this is part of why I love, like, crypto and smart contracts is they are inherently antifragile. Like, every single time a smart contract has a hack, all the other smart contracts get better because they discover what the vulnerability was and fix their own code. And so, like, over time, it's it's a self fulfilling system. So I agree with that. But but is there like this split moment of time where there is such a big advantage to these players that are involved with Anthropic

that everyone else doesn't have access to, and it and it's a huge problem? I

Alan (Nomadic Rogues)can I'd love to jump in here on this one a little bit. If that's okay. Of course. I I I don't think there's gonna be a huge advantage from what I've read. What I would recommend is if you see a patch for many of those companies come out in the next few months, you better bloody apply it. Because I think, you know, as I as I read this, I I don't think it's gonna give a competitive advantage of those. I think they're gonna clean their house up so that you don't have that panic and moment. So as your, you know, as your Linux distros get patched, as your AWS,

you know, your back end ECS things need to get updated, I think the community would be very well served by paying attention. Oddly enough, I got one of my first, you know, you know, iOS patches, you know, that came out and so I think being very attuned to those is good. So I think folks are gonna start cleaning up their houses, right? The the sixteen year old f f, you know, MPEG vulnerability that's always been out there, that's everywhere. So I think that we're gonna get some some disruption,

but then I think that'll level out. To David's point is people will will use this to say, okay, we now have this tool for an as an offensive tool even if it's not even if it's not the the mythos is use the existing tools that you have available and work into your workflow. Hey, check this for vulnerabilities. Play a bad actor, how would this be, you know, how could this be used poorly? So I think even if it just wakes up for folks who are developing software to use the existing models as their own red team would be tremendously powerful. It's a really good wake up that says, hey, there's something

Bowen BelmerOh, I think I lost Alan on my side. Can can you guys still hear?

Kyle (Bow Tied Blue Fin)I

Alan (Nomadic Rogues)lost you. You guys are back. Yep. Pod pod dropped on me. Anyway, yeah. I think just pay attention to what's coming from folks. I wasn't sure from the reading that they were gonna release it broadly. Now that might have been some of the the it's too dangerous to release for the public. Who knows? I you know, another model may you know, other models may come out that that are That they they they are they're at least already releasing pricing.

David JohnstonSo Oh. I think they're gonna kinda white list it. It's like $30 per million tokens. It's it they are charging an arm and a leg, but, you know, I I figure if you've got a high value code base, you're probably dumping the money in and having to give it a go. This is the new VC raise.

Alan (Nomadic Rogues)Yeah. Mistakes. Not that they're struggling here at at all, but interesting. Yeah. It's gonna be interesting from a security perspective. It'll be interesting.

Kyle (Bow Tied Blue Fin)Opus is already 25

David Johnstonper million output. I I can only imagine. I I I didn't realize the pricing was really so I I gotta take a look at that. Yeah. It might have been more. I I saw some quotes higher than that. But yeah. No. It's they're they're gonna, yeah, they're gonna make a lot of money on upgrading everybody's cybersecurity. Okay. Sure. But it was interesting to see that five point one on cybersecurity specifically scored above Claude 4.6. So like like Alan was saying, you know, implement this discipline yourself. Right? In my standard operating procedure now for coding,

every piece of code that gets out out is audited first, and it's audited by a different model than wrote it. And that's one of the unlocks I've seen recently is have a different model do the testing, have a different model run the audit because the models are really good at passing their own tests. But I found cross model auditing is really powerful. You know, grab Grok, grab GLM 5.1, whatever you want, but, you know, get something else in there that didn't write the code, and you find a lot more gaps.

Bowen BelmerThat just reminded me of when, like, government agencies will audit themselves, and it'll be like, the FBI did an investigation and did not find any internal malpractice within the FBI. Thank you. It's kinda what it reminded me of. But, yes. Well, I can certainly say that, to your point, especially on the cost, once an open source version or something that competes with an open source exists, I know the place where I will be able to run it

the most economically, and that is on Morpheus, which is or in Morpheus, through Morpheus, whatever the best vernacular is. Guys, that is the full scope of everything we had this week in terms of, agenda to cover. The only other topic, which is definitely a conversation for another week, but just wanted to put it on the radar of things to think about if you're not already thinking about it, is

post quantum positioning for DAI and what that looks like. This is obviously just like a huge quantum in general is a huge topic, in tech right now and has been kind of all over the timeline from a crypto perspective. I know personal opinion here, Ethereum seems to be well positioned for this, which is one of the reasons that the token for Morpheus obviously lives within the Ethereum ecosystem. And we'll talk about this down the line, but just maybe like put a pin in this and and pin it as something to kinda think about that we'll talk about in the future, because I think it is one of the next topics that'll become most relevant to, you know, the things that we're covering.

But does anyone else have anything else in the meantime that they want to make sure we cover this week?

David JohnstonNo. I mean, that's, that's a lot of great stuff. I mean, the number of models that came out is just, wild. There's a new large, Trinity model, and I thought that was, interesting, you know, just as a as a Morpheus fan. You know? I I gotta go try that one out. It's one of the larger open source models that's been released recently. So you got, you know, Google with their Gemma models on the small side. You got Trinity on the large side, and they got GLM 5.1 somewhere in the middle. There's definitely, I think, demand for US

open source. There's probably some companies that are sensitive to where a model is built or trained or or sort of long tail risks around, you know, poisoning of models and things like that. So good to see, more American models drop. I gotta go check out the meta ones. They just dropped a huge number of new models, and they've got a whole new naming scheme. They're not called Lama anymore. So I gotta I gotta check that out. But, yeah, it's been a hell of a week for for open source models, and so, hopefully, we'll continue to see even more of that.

Bowen BelmerAbsolutely. Well, we urge everyone to try them out. Something we at least I try to mention on pretty much every type of space that I'm on is the best thing for anyone listening to this to do right now is go try stuff. And through trial and error, we will slowly but surely become masters of our own domain as it relates to super intelligence and decentralized AI, and owning your own owning your own AI ultimately, which is of course, the end vision that we're all working towards.

Well, with that, thank you everybody as always for jumping on. From the those on stage, for speaking, and those listening, if you're someone listening to the recording, thank you as well. We really appreciate it and look forward to next week where we will all reconvene and cover the latest in everything happening within DAI and Morpheus specifically.

David JohnstonAlright, guys. See you next week. Keep being awesome.

Open the transcript file

Hosts and Guests

Hosts

Guests

Credit the people on this episode

Be the first to rate this episode

More episodes

All episodes
  1. Video Is Coming! | Morpheus Is About To Get Far More Entertaining

    52 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...

    Credit the people on this episode

  2. Local Models, Foreign Affairs | The Case For Private AI Archives

    53 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...

    Credit the people on this episode

  3. Let's Talk Node NEO | Step By Step Progress

    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...

    Credit the people on this episode

  4. Hello From Hermes! | A New Way To Onboard To Morpheus

    43 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...

    Credit the people on this episode

  5. What Agent To Agent Connection Actually Looks Like | Inside The Morpheus Agent Stack

    58 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...

    Credit the people on this episode