The Resurgence of Decentralized AI | Roundup

Transcript

It's very similar like AI reminds me is why we got into crypto in the first place back in the 15-20 years ago when it was created. Financial institutions lost trust with the world. And I think right now in AI, I wouldn't say Frontier Labs have lost trust in the world, but I would say that we're at the very very early beginnings of what it looks like when they're so powerful that they can actually impact the world with their decisions. >> Hey everyone, quick disclaimer before we get into today's episode. Nothing said on Bell Curve is a recommendation to buy or sell securities or tokens. This podcast is for informational purposes only and the views expressed by anyone on the show are solely our opinions, not financial advice. Our guest and I may hold positions in the companies, funds, or projects discussed.

>> What's going on everyone? Welcome back to another episode of Bell Curve. You got me, Miles and Zave. How are we doing, guys? >> Good day. Good day. >> Got World Cup fever. >> Yep. >> We got the battle of Bell Curve, USA versus Australia on Friday. >> Yes, I've been following this closely. >> [snorts] >> How are the Scots doing in Boston, Miles? I see they're they're taking over. >> Excellent. They bring such good energy. I think the it's a match made in heaven.

Everybody in Boston loves them and yeah, it's just fun. You look outside and just see a bunch of drunk huge men wearing kilts, you know, bringing good vibes. >> I feel like Scotland and Boston are a match made in heaven in general, but that might We were talking about this last night, too, but the my favorite content right now, and I hadn't seen it until you flagged it to me, but all the Europeans coming over and enjoying like Big Gulps and um, you know, grits in the South, just like huge platefuls of fried chicken and gravy. It's like, this is great.

>> I love it. It's wholesome con- content. Um, you know, it's cultures cultures understanding each other better. Uh and yeah, it's like kind of makes you patriotic. Yeah, America is [ __ ] awesome. Yeah. >> Yeah. I almost FOMO'd into a ticket. It was It's Australia versus US on Friday. And I was in West Coast US. And someone offered me a ticket this Friday. Apparently, the game's in Seattle and almost did it, but yeah, unfortunately, held it off. So, hopefully, I don't have to regret that. But yeah, it was awesome being in US last week where you guys are right now, being a a part of the World Cup fever. I'm very jealous of that you have it for the whole month.

>> Damn, man. That's a hot ticket. Wow. You know, the US right now is I mean, the next one in New York, we got the World Cup. I also I don't even know if I want to say this out loud here, but I I I caught the first like the opener of that UFC on the White House lawn. And I logged on. I was prepared to hate it. I was like, "Ugh, I hate this." But then I watched the opening and I was like, "This is actually kind of metal." >> [laughter] [snorts] >> Yeah.

It was wild fun. Yeah. I don't know. It's I don't know. There's a lot of stuff happening over here recently, but >> [laughter] >> Uh All right. Speaking of stuff that was prepared to hate, but might potentially be metal. Uh we're going to be talking about the intersection of AI and crypto today, which has largely been left for dead. Um And I think an area that you know, most most people when they think about the intersection of AI and crypto, you might think of There was kind of a run of some of these projects a couple of years ago. Uh decentralized training, that type of thing. Even a couple, you know, application layer things that were going to get built. I I obviously, uh BitTensor Tao uh really

took off and one of the teams that we'll be talking about, news, we're still hanging around news or new, um, depending on if you're I don't know what we think. Okay. Um, but um more more recently, I think the vision that people have been excited about is agents paying on stablecoin rails. Um, and so, I I think maybe that has reignited some amount of excitement that there could be an overlap here. But, really, the reason that we're doing this episode is, um, because of some of the announcements that came out of Anthropic over this last week. So, I'm sure people have been following this, but in case you haven't, uh, Mythos was obviously a very hyped model coming out of Anthropic. Um, whether you think it was go-to-market or

political theater or regulatory incompetency or partisan strategy, or if you think Dario's just actually very concerned about the capabilities from a cybersecurity standpoint, um, they rolled out they indicated that they were going to roll out Mythos, worked with a couple of partners to patch any holes that might be in existing systems, and then they launched it as Fable. Um, although there are some there are some pretty significant guardrails on Fable. Um, a team of supposedly Amazon researchers were able to jailbreak it pretty uh pretty quickly.

Um, I guess the US government hoped that they would patch, you know, whatever the the fixes were. For whatever reason, they didn't at Anthropic, and then they announced this export control. So, here here are the the areas where centralization around AI, um, I think are maybe showing some limitations and creating opportunity. So, I'm in Canada at the moment. I can't use Fable. Like, I literally can't use Fable, which is pretty nuts. Um, uh, and in addition to that, they also got into trouble because they were nerfing some some use cases. Um, so, anything that leverages Fable to train other models, um, they got into trouble because they were nerfing the responses but not indicating to users that they were downgrading the

responses. But they really really limited you in terms of anything related to, you know, biology, um a lot of pretty basic science, um cybersecurity, cybersecurity, that type of thing. Um and again, the the most charitable explanation is that there are real concerns, right? You obviously need some amount of guardrails. You don't want people having access to how to create you know, the modern equivalent of biological fertilizer bombs. But on the other hand, um you might also think, well, maybe this is Anthropic one trying to get regulated and put in the seat of incumbency. And then two, maybe they're thinking about ways to monetize outside of just the

$200 a month or the enterprise um subscription for paying for compute. Maybe what they're trying to do is create a foundation for if some form uh you know, pharma or or bio you know, sciences company, uh if they create some new drug, um maybe they want to profit on that by more than just whatever the cost of compute was. Maybe they want some sort of share. So, anyway, the the uh I I think the this has this has been a pretty big development here. I know everyone's eyes are on SpaceX and GPUs in space and all that stuff, but the limits of these large um you know, frontier models um are starting to show some cracks and maybe either open source or even some parts of the decentralized AI stack

don't look quite as silly as they did um a couple months ago. So, Miles, Dave, whoever what you guys are both more informed about the subject than me, but um Miles, why don't you take us away, walk us through. Also, if we could just define what does it mean like the different categories of decentralized AI? Like what does that even mean? Where is it actually interesting versus where is it not? >> Yeah, I mean, I think we're doing this episode because um maybe like a year ago like pretty much all three of us were fading anything that was considered like this decentralized AI uh or you know crypto AI intersection. Um and for this episode I I think we're going to like separate just stablecoin payments made by AI agents over you know

blockchain rails. That's like a different equation. Um but we've hit like it's been seems like it's an interesting inflection point in AI broadly where one like the era of just token maxing the state-of-the-art models is kind of over. Um people are realizing from a cost standpoint that it's just you it doesn't make sense um to use the latest and greatest most expensive model for every single task. And so you know the harnesses have gotten much better at like smart routing, right? So you there's like some you know efficient frontier of like, "Okay, for this task, what is the um minimum viable like cost that I can get the best sort of result with, right?" So one, there's been a resurg a

resurgence of interest in open-source models or cheaper models. Um and a lot of these harnesses are are are routing to those models now and it's I think you know we're we're sort of spreading out like the the usage broadly across more models than even two or three months ago. Um and then the second piece is there's clearly like an indication of platform risk now with uh the government in basically pulling back Fable. And so um I think there's you know there's there's now like very much more legitimate reasons to think about open-source models, uh which is not crypto necessarily but could be.

And um there's uh from a like platform risk and from just a cost perspective, um I think distributed or decentralized training um is looking a lot more interesting than maybe a year or two ago. Um I'm still pretty bearish on decentralized inference just because I think that, you know, will always perform worse than something that is centralized. Um But when you think about like what makes um DePIN as a category interesting, it's okay, there's some product that people want that requires a ton of capex, right? To on the supply side to actually stand up.

Um and I think about like my favorite DePIN project, Geonet, right? They're try it they're putting basically GPS sensors all around the world. And that has led to basically a product that couldn't exist without a massive um company basically investing as much as, you know, more than what what is realistically possible to to get there. And the same thing kind of applies to decentralized training, right? So like these, you know, GPU clusters are extremely expensive. They're distributed kind of all around the world and a lot of them are just sitting there in like university labs and things like that.

And so it's saying, okay, if we don't want to rely on just OpenAI and Anthropic forever, right? We need new models. We can train those models a lot more cheaply if we deep like distribute the cost of training them. Um and so I think we'll see, you know, uh shift of interest back to open source models and then soon, I think, some of those open source models will have been created through like, you know, projects like Prime Intellect, um projects like Pluralis and and and Hermes that that have been, you know, basically training um them for the last 2 years. But yes, they have curious to hear your thoughts.

>> Yeah, I think really well said by both of you guys. In general, I think uh this podcast is very timely. Miles, you mentioned before that some of us were bearish on crypto AI 6 to 12 months ago. I have to admit, I've always been bullish, maybe too bullish at the start as Mike has pointed out several times. For anyone listening to this podcast for a while, and I think Mike was really correct there. Definitely was too early for quite some time. What's interesting, actually, if you look at new research, uh, Pluralis research, etc., they all actually have the word research like in the title. And I think what you're seeing now is research over many years actually now coming to production somewhat in sort of tandem with, you know, other things

happening within open source, which allow the entire technology to catch up to to closed source. I think in the past, like, in in a world where Frontier Labs uh allowed to do anything, and governments allow them to take over the world, probably open-source AI matters less and and decentralized AI. In a world where governments do care about Frontier Labs taking over the world, and we've just seen that last week, I think all of a sudden decentralized AI becomes much more important. For people like us, listeners as well as speakers in this podcast, like, it's pretty clear we're all aligned that we can see AI as a technology that needs to be handled and treated in a way that's actually good for society overall, whereas maybe some

people in centralized AI don't have that ambition, and they really just care about making money. It's very similar, like, AI reminds me is why we all got into crypto in the first place back, you know, 15 10, 20 years ago when it was created, financial institutions lost trust with the world. And I think right now in AI, I wouldn't say Frontier Labs have lost trust in the world, but I would say that we're at the very, very early beginnings of what it looks like when they're so powerful that they can actually impact the world with their decisions. And that's exactly what finance used to be 20 years ago. >> Yeah. You ever play game of Settlers of Catan? I don't know if anyone in this call is a big Settlers >> Just a few times, man.

>> player. Um, but you know how when one player starts to win, everyone kind of teams up on them? Like I think that's that's what generally happens in these industries and one of the unique things about AI, like even if you forget about how cool it is, you're creating intelligence, all of this stuff. By the way, I know Sam Altman is taking a lot of flak. I think Dario is rightly getting a lot of flak for writing these essays about, you know, benevolent overlords, AI ruling over humanity and naming stuff around like, dude, read the freaking room. Like >> [laughter] >> who is supportive of that?

Um but also I think that one of the when an industry or company or entity gets big and powerful enough, it is it is inevitable that the public at large will turn on that and say, "Hey, you have too much power here." and try to limit it. What's unique about these LLMs or these frontier labs is how fast they have gotten to scale. I mean, it's unlike anything in all of even, you know, 5 years ago you would have said Google is the, you know, best money printing machine ever invented. Um and it looks like you know, these companies are going to dwarf even Google in a relatively short period of time. So, I I just think that the scrutiny um there's going to be a lot of scrutiny and it's it's inevitable. They've just

gotten so big and so powerful so quickly um that this was kind of we're kind of speed running, you know, save the point that you made about the banks. >> Yeah, absolutely. And I think, you know, Miles mentioned before a few companies working on this problem which I'm sure we can go into in a minute. But um yeah, in particular companies like Prime Intellect themselves and and I believe Noosphere potentially one or two others. Decentralized training is a concept. I mean, there were very few believers in this the last year, two, three years in crypto. Like also shout out to Jake Brukhman from CoinFund has been talking about this for a long time. Um but he was it's very contrarian actually for a very long time if this is even possible.

For a while it looked like they might catch up and now for the first time it seems like actually the probability of decentralized training and actually people using these models is actually increasing every day which is actually very bullish for all of us in crypto in terms of why we're in this space in the first place and I actually really want to just mention a point which might everyone might not be familiar with. There is actually a difference between decentralized training and decentralized AI and just like open source AI. So I think Miles alluded to it a bit earlier and potentially can go into it in a in-depth soon as well but on on my side how I would explain it is you have kind of companies that are decentralized like News Research, Pluralis, etc. Prime Intellect. Uh, you know, a lot of these

companies are actually training AI models in a way where you can actually contribute, you know, either data or compute, you can share in the upside. Uh, you also can align the models with your own values and biases in certain ways that everyone It's all transparent. It's like what blockchain's there for in the first place. And really most importantly, uh, you can't be censored with these decentralized models. So what's happening right now actually in AI is that we're seeing early versions of closed source AI being censored like, you know, stable. In future, like right now actually in open source AI as some of us know, like a lot of people are using, you know, your Alibaba's, etc.

Quen, you know, some of these kinds of open source models. These are very good in capability, actually better in capability than decentralized models and they're two different things. However, they can still censor you as well. Like right now they're just opening the weights, etc. but it's very easy for them to stop building it or relicense it or censor you in different ways, you know, on the supply side of actually using their open source model. So I think all in all what I'm trying to say is that it's also early on that side and then I think when that side comes fruition decentralized AI will really take off, I suppose.

>> Can I cuz you're something that I sort of wonder if crypto got right versus wrong. I mean, when when focusing on the subject of censorship, I feel like crypto, for better or worse, really focused on individual on the individual level, which to me has always been a little bit more philosophical versus I think where you actually tend to get a lot of traction here is censorship on the business level. Use a different word there, it's called platform risk. But I I I think that that's going to be the driver here for AI and I wonder if at the end of the day, like we're watching the very beginnings of this with privacy, with like Zcash, Morpheus announced a confidential vault with Zama today. Like the business use case, at least in the United States, holds far more water than

you'd think, right? Cuz the Americans are so crazy with individual liberty, give me liberty or give me death. But actually it's the the business use case, which usually ends up getting things across the finish line. I have a feeling that's what is going to be with I have a feeling that that's what it's going to be with uh AI as well. >> The the interest in open source models is Yeah, exactly. Two things. It's cost and there's no censorship resistance, so there's no platform risk, right? Um but the bottleneck of like AI in general is that it's, you know, incredibly hard to create and improve models that are, in you know, some ways comparable to to like the state of the

art frontier labs models, right? The way that these economics works is that they spend about like, you know, let's say 100 100 billion or something like that to train Fable, right? And then they assume that they're going to basically get, you know, 150 billion in revenue or 200 billion in revenue. And by the time they're getting that revenue, they're already, you know, sinking another, let's say, 200 billion in to to train the next model. Um well, what's happening is that the amount of revenue net new revenue that will come in from each, you know, state-of-the-art model is going to go down over time as people just begin smart routing to, you know, cheaper models for like tasks that they don't need it, right? And so, the real demand

is for like cheaper but effective models. And how do you basically, you know, uh break that bottleneck of needing 100 billion dollars to train them? Like, well, you spread that out, right? Over many, many, many people and individuals and you let like, you know, again, like this is the whole thesis for decentralized training. It's I think but to in the way I see it, it it is more of like a a cost hack, a supply-side cost hack, than it is platform, you know, I think like the the platform risk value prop is just inherent to any open-source model. But the question is, okay, if we want those, how do we get more of them and how do we make them as competitive?

You know, more competitive with like the state-of-the-art models. Well, we can spread like the cost of the the upfront, you know, train. >> Let me just ask uh do we need So, I mean, very broadly, right? I mean, a lot of the the crypto and AI overlap has been focused on decentralized training, um which hasn't had a ton of success yet. I will say Jensen has gone on and said we need to eventually crack something like decentralized uh training. He used exactly those words something like a year and a half ago. Um at least when I first heard him say that. So, in theory, that should be interesting. But as one of the one of the other you know, what a an additional bottleneck is just the inference

bottleneck um as well and improving that and having some of these lower cost uh open-source models potentially live on chain as well around like protocols also somewhat interesting, right? There's multiple different And And also, I'm curious Miles to get a sense like we had that whole episode on open source as a strategy, how much of this is an open source versus closed? >> It's like it there's a lot that rhymes here in the past. >> We've had about like you know, okay, all of crypto is open source because, you know, that's inherently in like the nature of the the product, right? But not all open source is obviously crypto. Most of it's not and there's a

uh this totally independent open source community. But then there's also like large corporations that will open source things strategically, right? To commoditize, you know, their their complements, right? And so, just we're talking about Jensen, right? They open source the weights of a model um that, you know, companies like PayPal have taken and now they're training their own foundational, you know, let's say like finance payment specific model based off of, you know, decades of transaction data that they have, which is proprietary, right? Um And so, those are the three things.

There's like, you know, there's the the large corporations that are, you know, open sourcing something for strategic reasons. Um There's the open source sort of community which has overlap with the with the crypto community, but, you know, the real the real turning point was will be like when people want open source models for cost reasons, right? And the only way to get more of these open source models is to distribute the actual training of them. Um and all of a sudden that looks like crypto right? >> Yeah. What I was going to say is like I have been banging this um drum for a little while.

I do think, you know, this has been very clear since Open Claw was released 6 months ago that we are entering a world I mean there's so many layers of the stack that you can think about right now where AI is impacting crypto. We mentioned in this podcast, you have, you know, AI agents and that's kind of like the highest level of the stack of like, okay, they're more like applications using stablecoins. That's happening for sure. There's experimentation there. Uh and and a lot of that is working to the to the sort of all the way to the fact that Tempo is now being released by Stripe and that of course is a settlement layer for this in itself. So, obviously people think highly of that as a potential opportunity. What's more that's consensus, let's say. What has been less consensus is lower levels of the stack

like infrastructure layer. Uh you know, for actual AI use cases, you know, beyond just payments and sort of stablecoin transfers etc. So, Nous Research, I think honestly this even surprised me and I think probably even surprised other people in the space like I would say with some confidence now that Nous Research is pretty much better than Open AI and maybe even the best in this kind of like harness space, which is super cool because it's built by an open-source and decentralized research team, uh which is doing it with crypto ideals and values. And this is a product that's better than actually closed-source AI companies can build.

So, we saw Open AI that came out, ChatGPT acquired it, it got really bad. I actually churned as soon as they acquired it because Open AI got terrible. I'm sure anyone listening using it felt that as well. Never returned back. I'm now back with Harness and that's epic cuz it's like it's a great crypto team. It's a good product. Um and actually they're just getting started. Miles and I were speaking before that they're not even just working on the harness. They have, you know, training their own models etc. But also worth mentioning is that the harness itself is everything around the model. The best way to think about this is like it's the car and then the model is the engine. So, in crypto decentralized AI uh you know, someone like Prime Intellect might be the engine long-term if it can actually compete with capability with Frontier Labs. But

the actual car itself, what we're seeing early on in crypto is that this can actually be decentralized and open and built by crypto company. And I think that is one of the coolest things right now and most underestimated things right now. And actually, a lot of value could accrue here. You know, I I I still like even the way that I'm using Hermes, how cool it looks. Like, they own me as a user. Like, I am the end user and ultimately I can switch out what model I use. I can use, you know, Opus or ChatGPT or Primates, like whatever it is. But, the harness itself and all the context that it has and all my workflows and how I control it, etc., and how cool it looks, I think that's where a lot of value will be. And so, if a crypto company ultimately ends up owning that

layer as well, I think things can get very interesting. Not to plug in one of our own portfolio companies right now, but I will. So, we're working with a company called Index Network. They went live with their product last week at ETH City, and they're actually an agent within the Hermes harness, which is super cool. So, ETH City last week, which I was saying was a great event, I was every morning I'd be, you know, proactively prompted on Telegram saying, "Hey, here's all these calendar invites. Here's what might be relevant for you." Classic. But, what Index does inside the Hermes agent as a kind of school, almost, is it looks for other people at the event and who might be relevant for me in terms of like I'm an investor, I'm trying to meet startups.

And then it tries to connect me with them. And then it asks me a few questions to get to that end state. And so, what I'm saying here is that I am actually envisioning a world where, you know, Index had quite a bit of PMF at ETH City with like a lot of people were connecting, finding valuable opportunities. They're also crypto AI related team. I can see a world where Hermes does really well, harness is actually accruing the value, and a lot of crypto teams actually as like agents and schools within a harness like that do super well. So, I think there's so many layers of the stack. Like, I'm very excited as of this week. It's been a really slow 6 months. I actually just tweeted before the podcast saying that crypto and AI has like been a little bit early the last 6 months, but I'm like

very excited about where it's about to go cuz it feels like last week was the first sort of crucial moment that we needed things to get started? >> Yep. And maybe just a double click on like new strategy here as it's cuz it's really interesting and and I think they're executing it well like the It's very clear today that the harness is like the user-facing wedge. That is what like that is the the component that owns the customer relationship, right? Um and today that you know, it's basically smart routing to a number of models um including you know, the open-source ones and the non-open-source ones um and it's competitive with open claw

basically like the the a largest like non-crypto version of it, right? And at the same time another part of their project is the um is training their own model, right? And they have all these distributed nodes like all around the world and so you see where where this is going like um once their models are you know, uh competitive with with basically what what everything else is out there, yeah, they're going to route to their own model, right? Um and that will allow them to vertically integrate and make even more money. So I think that one's a super interesting one.

>> How do you guys feel about Venice? I feel like Venice is probably the most talked about crypto project at the moment. It has a token um Eric Voorhees you know, very early uh sort of like the beating heart and soul in many ways of like the early crypto ethos is the is the founder there. What do you guys think of that? >> I mean Zev, I don't know if you have like much stronger thoughts. I guess I was skeptical of Venice at first because it felt like it was you know the story was more about like the token and the tokenomics than it was about a product that people want to use um and I still I think it's like feels like usage has um exceeded

expectations, um but you know, I'm still still it like not not exactly sure where where I land on it. Um It seems like it's just kind of like a wrapper around, you know, >> Yeah. Look, I am very interested in Venice. I think it's showing a lot about what people want actually. And so Venice, even though it may be technically it's not the most profound architecture of all time, what it is demonstrating and what I think Eric got very right is like how people actually want to use AI in the first place. So think about how all of us use AI right now. We have ChatGPT, we have Flo and

etc. It knows absolutely everything about me, literally everything. And you know, maybe sometimes I'm asking you about my sleep or something like that. You know, but it actually knows about my sleep and then it's like kind of getting into my life and it's almost like intrusive. Uh I think what Venice is doing which is cool and it was also early in terms of the insight is like allowing people to use AI in a way that is fully private. And to the best of my knowledge, it's because a lot of this is kind of client-side and and private and none of the information goes server-side, which is something I've been following for so many years now. In fact, I felt very early for a long period of time, but again, this is kind of showing that timing finally after all these years seems to be coming now. So

like I yeah, I'm actually pretty pumped in this podcast because like there's been so many years and months where I feel like I personally been waiting for things in decentralized AI and edge computing and stuff like that, privacy, which a lot of us as crypto people have felt like is underrated. And I think, you know, with AI blowing up, you know, you can kind of think, you know, maybe you're wrong. But also sometimes being too early can be the same as being wrong where it's like right now I definitely see this S curve emerging where a lot in AI and crypto is really about to take off. And and Venice on the consumer side, you might make the argument that it's probably the best performing consumer crypto application right now. It only uses a few simple crypto mechanics, but I could see a world where a lot of the best apps

in crypto actually do this anyway, where it's not necessarily the app itself is not like a DeFi. It's just like a regular app with crypto mechanics on the back end that make it better UX. I think Venice is the first of many that will come out in future that look like this. And and and maybe just to finish on this point, I do think that one of the cool things about Venice, similar to what what we were talking about before, is that people who don't want to be censored by models. And so Venice gives you access to open source models, which can still censor you. So it kind of gives you an extra layer of not being censored and using the, you know, open source models. And it also removes your identity so no one can see what you're doing, which is really a privacy thing.

And again shows you in crypto, another thing we talked earlier about last few years is privacy. And so I think all these macro themes are actually coming together right now for the first time in quite a long time. >> Can I ask though, on the censorship discussion, does it Is there a broad like Do you as an individual care about censorship in some major way from the models? >> I mean, I think like there's two things, censorship and then data collection. So also and they're kind of related, right? So the last we with Fable, Anthropic came out and said, you know, there's we're going to like collect your data.

>> Right. And Mike, you mentioned earlier like the this idea of like platform risk and censorship resistance like typically resonates stronger like with businesses than most people, right? Like the more of my personal information I give to AI or I give to even, you know, think about like Apple on my iPhone, the more useful it becomes. And I think from like looking at my, you know, Strava data, like health like with most of us don't really give a [ __ ] right? Like Um but but businesses I think do you a [ __ ] about platform rugging and they really give a [ __ ] about data leakage. And so what was interesting is that they Anthropic and I

think this is a pretty another dumb misstep like they came out and said, you know, oh, all right, sorry guys, like we're collecting your data now. And then it was like if you're a, you know, a health care provider or you're, um, you know, a doctor and they're they're using like, you know, Claude religiously right? They're now breaking the law, um, from like a HIPAA violation standpoint, right? So like the bull case I think for for for privacy is, you know, I I think it starts originally starts to look like this crypto kind of like niche ideological thing like I want to I want full sovereignty and but then like something like this happens and then

it's like, well, actually maybe there's something there. >> You know, you know what you know what this like, sorry, this is just total speculation and I understand how this could get but on the point Anthropic didn't exist as a company. They're they're five-year-old company. Um and Dario is mostly a researcher. I I I saw this I saw some this article where Dario has one direct report which is his chief of staff. Um and he he does he's not very involved in the day-to-day operations of the business. And it reminds me of a structure that you saw in crypto businesses that no one thought too much about until it did become a real problem. Um which is you need more than just the

you know, ding a brain founder, all the mech design stuff. It's like that's ultimately at the end of the day not super useful. It It scale at scaling a company. It might be useful in putting you in the right market and niche. It might be useful attracting the right talent in the early days. But I wonder if I mean, the decisions that Dario and Anthropic are going to need to make are so difficult, right? So impactful at this early stage. You You would expect to see some blobs, right? From from anyone, right? It's not a criticism of It's just Yeah, you would expect to see These are tough decisions for any five-year-old

business to make. >> Yeah, exactly. Which is why now if if you want to get funded, you know, $100 million so you just put the word applied in front of like whatever you're saying is right? >> I mean >> This year just from two years ago, now we're applied compute. >> I think another topic worth talking about, which is also relevant to this conversation, is the concept of like token maxing as well. I feel like there was this narrative that Anthropic had for a period of time of like spend as much money as possible token maxing using board code, which Mike told me personally he really loved this narrative. And so And so there was a narrative where it's like smash the AI button and see what happens. Now we see that like the AI button was smashed and the outcomes

weren't that fantastic and now everyone's going back to being like, "Well, hang on a second. Like we don't want to smash the AI button and let's just see what actually makes sense here and how we're using the AI." And, you know ultimately I guess that's another bull case for like some of these models. They are actually at the end of the day cheaper and it's unclear right now what the capability difference will be between the frontier models and the kind of open source and more decentralized AI models. But what we do know right now is that it's way cheaper and you can potentially even host it yourself. So this is another catalyst and tailwind into like this open source decentralized AI movement, which is actually happening at the same time as all these other macro factors are playing out, including sort of, you know, political statuses from

the US government. So yeah, again here, if you speak, you know, as I was mentioning before, if you speak with hardcore agent engineers, most of them actually aren't using core and Anthropic and closed source for like a lot of the activities. Maybe for like evaluation or checking at the very very end or maybe even planning the specs at the start, but the hardcore execution a lot of agent engineers are actually turning to open source models, which is way cheaper to get more done. Which again here is on the bull factor for this this the second. >> Anthropic owned 5% of like uh I'm sorry, Microsoft owned 5% of Anthropic and they just turned off Claude. Right? Like so like this that that is 100% happening.

Um >> I mean, you know I okay, I had this long uh you know, shout out to our uh our old head of research who just left still at crypto and just talking there was a public we were going back and forth on Twitter and then had a long conversation with him about this, but you know, there's a perspective um and this is also very one-sided cuz I'm only representing my own perspective here. Um and there's obviously a lot of really good counterpoints. Like this would actually be the out of consensus perspective, but you know, uh it's it's the token maxing is the spend is the return on investment. Um And really smart guys who've gotten this cycle totally right like Gavin Baker, you know, have been shouting that the ROIC is there.

But I would it's very difficult to pull apart all the different factors that are going into that calculus. I mean, the you know, that gets the return on invested capital calculation also would look better if you had fewer employees, again, this is now a pretty well understood thing. You know, the all these companies over hired in 2021 and they're getting rid of people. Um but also, you know, one thing that's difficult to really put into words and it's it's a uh like is the top-down pressure that builders and founders get, you know, to incorporate AI. I think there's a there's a really solid mix of you're obviously looking at this technology which is super transformative.

Everything is going to be different. But, you know, the the game theory for founder is to just say, "Yeah, this stuff is transformational. Get on board. Token max." All that stuff. And then like quietly kind of figure out what actually makes sense and what doesn't make sense. Um and I think that we're getting towards that part of the cycle. Um cuz I don't like if you guys I mean, so here's like a great example, right? Like Notion. We a lot of the Blockworks we use Notion pretty religiously. Um bottom. Anyway, uh they've kind of jammed this AI agent in there. And so I don't know. It's like kind of shittier than the granola agent that I use. It's certainly not like the end state of of AI. It It's like a lot of the use cases

like every single software product that I use right now, HubSpot has jammed an AI agent in there. It's like, "Ask me anything about your sales." It's like, "I kind of already know what Like I already have my report set up. I don't need this." Um and also Claude does this for me. So, it's just there's a lot of stuff that's jammed in a very naive way into software products. And I I think the bottleneck here actually isn't compute. It is a creativity thing of what is the right form factor for these different use cases. Like it's it took us a long time to come up with the right form factor for a smartphone. Like Steve Jobs kind of had to look at Xerox and and do all this stuff. It It that's just a creativity.

You got to look at and I think we haven't moved past the naive chatbot UX. And I don't think that's a compute constraint. It is a creativity, a product, what should this look like constraint. And a lot of the AI and I know this isn't the point of this episode, but that's my When people are like, "Well, how is it going to What Where is it being overstated?" I think that's People have jammed very expensive nonsensical AI applications into a bunch of software. That's kind of distribution. That's not real.

>> Which is why they made so much [ __ ] revenue, right? Cuz like all of us are using them directly and then every SAS company in the world is like, "Oh, [ __ ] We have to do something." And they use them directly. So like, when you go to their website, you're like, "Okay, now it's routing through the same thing." Yeah, I think that like, you know, the the of everybody like using the max the those state-of-the-art models for everything, including your SAS. Tools that you don't even want, right? Um is over.

>> Just briefly, I think another interesting point around the privacy for business that we mentioned before, also shout out to me also in this space. We haven't mentioned this episode, but you know, Nym, which was I think a bit under-loved at the time. They worked on Iron Claw, which is kind of like a Hermes that was private. And that was for more personal use. I mean, I still think that like for, you know, this sort of brain of the company that people keep talking about and these agent factories etc. that I guess various businesses are building right now, privacy is still a big thing and like crypto uh cryptography in particular still makes a lot of sense for like these organizational brains that touch sensitive data. And I think this as a thing overall, which hasn't even really

even hit crypto yet. I've spoken to a few startups working on this problem in in more recent times. It can be hard to differentiate and the question is like can the hyperscalers just add privacy and security and then it's all good, but yeah, there there there still feels like a lot of things in the space where Frontier Labs, hyperscalers etc. they could be doing certain things, but there's also a wedge for these startups to come in and earn real customers very fast and get them the privacy and security that they need. This is like beyond just training actually like agentic workflows. So there's so many layers of the stack. So much is happening, so many exciting things. It it is becoming clear to me right now that like I think we've We've about it a bit on this podcast already that, you know, I think quarter four was always

the time I thought things might start coming back. Still think quarter three could be tough even though there's been some some great news, but I do think that it's very obvious now that for this next cycle, this is going to definitely be something in in crypto to get excited about. And I think it was just a matter of time where AI capability continued to get better, but it was actually a good thing for crypto cuz now we know the end state and now we know that we have technology, you know, directly applicable to to serving us as society in a much more fair and transparent way. So, yeah, it it looks looks good ahead for for this sort of category in crypto.

>> Yeah, I agree. I mean, we haven't even talked about like wearables and like those are probably going to come next. And you know, with that, I think the does demand for even more local models will be higher, right? Uh you'll likely want to run like whatever AI is in those wearables doing like day-to-day stuff, right? Doesn't need to be using state-of-the-art models. Uh You know, you know. Solving cancer here on a day-to-day. Um So, yeah, you know, I think there's like a bunch more catalysts that are coming and uh yeah, like all of these things, local source, uh sorry, open source or like local, like they're going to have

crypto, you know, decentralized projects going for them, but they're not not just, you know, >> I think it's it's worth shouting out Miles pre-podcast here. He did a little table on, you know, all the companies that are crypto AI related. The companies that are loved by crypto and loved by AI or hated by either one. And quite interestingly, there are various companies that, you know, the model came back with in terms of what companies both the AI crowd and the crypto crowd likes. And then there are other companies where the crypto crowd like them, not the AI crowd. But I think the point being that you know, it's very clear and real that use cases using crypto or blockchain tech in the AI space and now becoming

more consensus by the AI crowd. And I think that's probably the main thing here if we're really going to get this technology adopted in mainstream, which something like stable coins are today. We really need the crowd of people outside crypto to believe it's actually solving a real problem and at least right now it actually seems like it definitely is. So I think there's been massive wins and then props to the teams, you know, working on this problem for many years being non-consensus around like now for the first time it looks like it's catching up and you know, BitTensor, you know, there's one network template which I think we covered, you know, many podcast episodes ago which is also working on decentralized training and also a great team in the space, but there's been a few teams circle it, but again, yeah, now in different ways of kind of going

about it, but yeah, it now seems and and I guess as a VC more interestingly it's going to be very interesting to see the kind of next wave of applications that go live, you know, as a result of infrastructure being developed using crypto that is in the AI space. >> So news um they're probably at like a billion valuation, right? Like like Pluralis or like they're up there. Um they have crypto investors like, you know, Variant invested in Pluralis. And these if they do launch tokens, like these are not going to be small cap, you know, [ __ ] They're going to be you know, actually like but are probably a big deal. Um and so uh and and and they'll launch in post

PMF, right? Like they are not going to be this um sort of ultra speculative thing that needed to be bootstrapped for a very long time and I think like, you know, some of the early stuff like like Venice, um you know, today is like the only way to get token exposure to something that has like real usage um in this intersection. Um But yeah, you know, keep an eye out for that. Or or maybe they never launch a different. I have no idea. No idea. Um I mean What do you I mean it depends on you know Yeah, I mean this is one of those unanswered things on how far post PMF and what are the you know, there are pretty they're much more well understood

metrics that you want to see before becoming a public company in the US. Um those standards have yet to be set around crypto. Right? Like what are the revenue thresholds? What the profitability? How does that differ for a more recurring revenue model versus you know, transaction based revenue like hyperliquid? Unclear. Um unclear. I mean I I think the yeah, these are open questions. I don't know the answer. >> I'm I'm bullish tokens. I think all these teams at the crypto and AI intersection are going to launch tokens. Pretty clear. Unless you really pivot hard into the pure AI space and like every customer you serve is like a Fortune 500 company. Uh I think if you

have any relationship to things outside of that, you may as well launch a token for all of the reasons we've talked about in this podcast before. So yeah, I think tokens will come back. These crypto AI teams will have great tokens. You're already seeing, you know, as the industry consolidates, people want access and they want to buy these tokens up. And so for the first time ever you're probably going to see fundamentally driven valuations for these products and networks, which is going to be very cool to see. I don't know if you guys saw there was a paper that came out, you know, a day or two ago saying that you can now fundamentally value crypto assets for the first time.

But um yeah, we should put that in the >> A factor based model. Yeah, I saw this. >> Yeah, yeah, something like that. Like I can't even remember >> God, I'm waiting for someone to tell us how to value these So I found >> To be Yeah, like I didn't read it and so I'm just like told you, you know, >> Oh, can I actually You know what? I am curious to get you guys' take on this. So, I Did you see Sam Lesson of Slow Ventures put out a a paper or like a blog post on you know, we are now in a He was talking about SpaceX Let me actually go here so I can get the right quote.

So, the real SpaceX IPO lesson. We live in a capitalist multiverse of value. And the headline of this is DCF is no longer the only globally scalable story of value. The internet lets minority belief systems create trillion-dollar markets. Uh So, >> I heard of that. >> I was like, this is sounding familiar. Uh you know, that that If you've been in crypto for 8 years, that would be a very familiar sentiment to you. And um it's just it's funny, man. I like I have talked about this uh quite a bit like why is the valuation methodology in equity so sophisticated?

And in crypto, a lot of the same people who are involved in buying and selling equities seem to take several steps backwards. But yeah, it's funny. I mean, I and it seems like the the takeaway here is that crypto has once again front-runned. The the rest of the world is becoming more familiar to what we've seen and observed in crypto as opposed to the other way around. And if you were to summarize what happened in crypto over the course of the last um however many years, I think that the the couple of trends that have come together on this are having large a large audience and being able to pump small-cap assets, which is apparently happening at larger scale in more obvious ways than used to happen.

And we don't have an administration that cares a lot about thinking that that's a bad thing. Uh but the other thing is that these like levers that you can pull to sort of manipulate price mostly around the float, it ends up screwing with people's heads a lot because the value of an asset can be can diverge so long from what the DCF multiple would tell you is appropriate that people start to say, "Well, we can't be valuing this on the basis of a DCF. So, what's the rational way to value this stuff?" And you start coming up with ludicrous, you know, ideas about it. Um and I feel like that's kind of happening here. Like when I look at the SpaceX IPO thing, first of all, SpaceX is incredible. Honestly, it's probably a

generational company. I I think it's the coolest story of entrepreneurship in modern history. Uh I'm very much on that wavelength. I'm also on the wavelength of not totally forgetting what's happening in crypto in the last 10 years. And it's a super low-float stock that is that's pumping the value out of here. So, I I think both of those things can be true at the same time. I don't know. I'm kind of going on a rant here, but it is interesting. It's like crypto does lead the way Crypto tends to forecast like social and financial trends. You know, you keep thinking crypto's this crazy wild west thing, but it's like directionally right, even on the stuff that I don't want it to be directionally right on.

But >> I think like one point to make on that is just even crypto AI itself, like this is what society wants, you know? And like, you know, same as meme coin trading, like society wanted that. That's just It's so macro, such a macro technology, crypto. It's like people want to trade cuz they feel like they need a way out. So, hence they find crypto and try and, you know, invest and speculate. Same as AI, it's like AI is coming, but the people were scared. And they're like, "Well, I don't even know what the hell this is doing or what it is, and it's taking my job. Like, is there any way to solve this problem and actually make it in some way or I understand it and can explain it and it actually makes sense in my life?"

So, yeah, I I totally agree. And also, on the flip, like when crypto's mad and everyone's speculating and the valuations don't make sense, everyone's like, "Well, crypto doesn't make sense." Well, maybe it actually does. Like what Michael was just saying around the SpaceX valuation, it is kind of funny looking at the AI crowd right now and like everything is just the exact same as what it once was in crypto and people like this is on pure memetics and the story is so good and this is a trillion dollar and the whole world is going to use it etc. and evaluation goes out the window. Then you know, the bear market hits and they're going to be having the exact same conversations we've had on this podcast the last six or months. So yeah, it's it's full cycle.

>> Yeah. I'll tell you I just dropped this in our chat and put in the show notes but a look from from Standard had a really fun article yesterday where he's like basically this is looking exactly like you know, the how crypto matured where Bitcoin and Ethereum these were like the two immaculate conceptions, right? Like they accrue outsized like crazy amounts of value incredibly quickly and create this new like legible category and like a massive like at least a perceived massive prize to turn this from like a two horse race into a an N horse race, right? And then you get the altcoin basically alt lab like frenzy of investment everybody trying to

again like just and and you know, like how did it turn out on our side? Like oh my gosh, so many chains were funded with so much money um and basically Solana, you know, like managed was the only one that managed to get it like the the seat third seat at the table. And then Hyperliquid comes with like a super apply, you know, version of of a chain, right? And and that's what succeeds. So like I don't know. Reasoning by analogy never always the best idea but um I do encourage you to check out cuz I you know, it always smile when you read it and it's thought provoking.

>> You know what? One thing I also on that same alt labs like yeah, worth worth a read. I mean it makes them reasoning by analogy doesn't need to play out the same way. Although, I do wonder if there is a like there are a lot of similarities in between fees, like block space fees, and how that developed in Ethereum and Solana versus what we're starting to see around ROI uh calculations for the labs, and what's clear is that people are overpaying, I think, for Claude and ChatGPT. The question is just by how much, and how much is are the open How good are the um are the uh

you know, open source models going to be, and how much can it eat into the the fee stream of of uh ChatGPT and and Anthropic? The or OpenAI and Anthropic. The You know, the there's um And people always just say Jevons paradox, Jevons paradox, which I'm sure is going to be right. But, if you look at how that played out in crypto, it's very difficult to you know, when you have a very CAPEX intensive when there's a big gap in between your ability to bring supply online and demand, it's just difficult to balance those things, right? Like, demand is inherently unknowable, it's very difficult to predict. You know, it has the impact on demand is uh is difficult to forecast, and when supply is very expensive, it's just

tough to get the right balance. So, in crypto, what we saw is there was huge demand for Ethereum block space in 2021. That they responded to that by this like modular road map, and Solana and other chains. And then, the uh the sk- we successfully scaled block space to the point where we collapsed fees by I don't know, like two orders of magnitude, um or, you know, fees are couple cents now on Ethereum a couple couple dollars versus you know, fees were at one point like $200 um when there was congestion. But, the problem is the demand didn't materialize in time. Or, because demand was so stymied for a long time, it kind of spread it out and did all So, now we're just reconsolidating to the point

where people don't even believe as we host this podcast. This will be an interesting thing to go back and listen to in a couple years. People don't believe that Ethereum can generate sustainable fees. That is a way out of consensus belief at this moment in time. So, Jevons paradox I'm sure will work, but there is a gap. Right? It it like if the if the if the if there's too much of a deflationary impact on prices before that amount of demand materializes or there's a gap for some reason, it is meaningful. Like that that creates a real stumbling um So, I wonder if we'll see the same thing with the the labs.

>> in like we're in like the fat protocol thesis era of uh of Web 3, right? And then block space got commoditized um and then it turned out that you wanted to be an app. Like, oh yeah, that makes sense. Um so, now we're in the fat app thesis. The The big difference, of course, is that like Open and that Open AI and then Dropbox, unlike Ethereum and Bitcoin, are really good at building apps themselves, right? So like it's Ethereum was This is almost like if Ethereum just kind of like showed up a bunch of of the major apps like just built their own Amazon and things like that and just, you know, crushed the competition, but like didn't, you know, um push people off of the platform, which which maybe they

are, right? Like maybe part of the reason you'd want to get off of Dropbox is because you're afraid they're just going to, you know, make your entire company a feature. Uh uh Which is which is real. Yeah. >> All right, guys. I think that was a good episode. We can end it here. Um Fun one and uh we'll see you guys um See you guys next week.