The Intersection of AI and Crypto: What Worked, What Didn’t, and What’s Next | Roundup

Transcript

It just takes one or two breakouts like maybe it's some vibe coding thing, maybe it's some like prompt thing like who actually knows what it is where a bit of luck and a bit of skill and then the whole AI space is talking about it as well. So it be a breakthrough not just in crypto but also in AI. So there's always that small chance and that's why I think it keeps people interested in the ecosystem. >> 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 forformational purposes only, and the views expressed by anyone on the show are solely our opinions, not financial advice. Our guests and I may hold positions in the companies, funds, or projects discussed.

All right, everyone. Welcome back to another episode of El Curve. You got me, Miles, and Zave. Fellas, how's everyone doing? >> Doing great. >> One short week off. Um, >> missed you guys less. We got a we had a full team offsite in Costa Rica, so I have a good excuse. But uh didn't get any Tanner, unfortunately. But yeah, I have a good excuse. >> What you guys What happened over this last week with the two of you? I'm in Australia right now, so calling in nice and early for this podcast, which is nice. So appreciate you guys in the afternoon over there on the east side of US. But yeah, here for the summer. So, getting into routine this week and yeah,

it's been been nice to be back back home. >> I've been freezing in Boston. It's been absolutely miserable, but uh yeah, no, ears year's off to a good start. No more no more uh I don't know, dog bite incidents or anything like that. We're getting over it. So, >> no more. >> Sorry, buddy. Um, well, I think we can start the zero off with a bang on the bell curve side of things, uh, by tackling, you know, we hinted at the end of our commercial open source software episode that we're going to get into this intersection of AI and crypto. And, you know, I think you can look at crypto and then eventually AI is the next big iteration of software as an industry.

Um, you know, I think crypto took and ran with open-source and, you know, definitely created more opportunities for monetization. And we talked about L1's breaking some really interesting trade-offs and but AI is redefining the game of software at the same time and there are implications from everything like just SAS as an industry and uh like pricing doing kind of you know fixed recurring revenue versus outcomes based pricing um there is you know people are looking at how like how product um works within an organization product and engineering and you know really you it was a very flashy part of the industry uh going into 2025, this intersection of AI and crypto. And I think the thought was, hey, there is this AI thing that's

happening and this crypto thing that's happening. And man, wouldn't it be even more exciting if you smash these things together? And I'm curious what the retrospective was for the two of you. First of all, maybe we could just start with why were people excited about this intersection outside of potentially token go up, right? Maybe there are some bad reasons why people were excited about this, but what was, you know, what's the retrospective on why this might not have taken off as much as people thought and what was the initial real kernel of excitement for where this made sense in the first place? >> Yeah, I'm happy to kick it off. Um, I think in general, like you said, crypto and AI is a massive intersection, and I think speculation maybe uh took everything away from where

it actually was. you had a lot of AI meme coins launch uh virtuals launched and there was a lot of hype around kind of agents but back then agents didn't do a whole lot since then actually virtuals has enabled agents to do more things and it's quite interesting to follow but outside of that there was only a few successful kind of breakout projects another one was bit tensor which we'll touch in this podcast but um not a whole lot and there was a lot of venture funding that went into this and particularly on the infrastructure side which I think looking back on maybe in the last 2 years, it's hard to find many successful crypto AI infrastructure and maybe people jump the gun a bit in terms of like where the value might actually be in the crypto AI space. Like I'm I'm

quite bullish on kind of like the agentic layer and the application layer and I think the infra layer in in crypto AI is actually commoditizing quite fast apart from kind of building the guard rails that make us as humans trust what the agents are doing for us in crypto itself. So we had a bunch of startups working on layer ones kind of like agendic frameworks etc. But I think one of the main reasons where it failed uh was really just in terms of the complexity of what we can do with these agents on chain. I think right now like most of us probably don't use agents in our day-to-day life and that's a problem because in web 2 many of us are using agents in different ways sometimes without even knowing it. So I think crypto has a long way to go in terms of

like understanding the demand and then actually being able to serve that demand with frameworks and infrared applications that people actually want to use and pay for. So I think it's just a market misunderstanding exactly where the demand is coming from and speculating on where it might be and actually realizing in hindsight that it wasn't really there in the first place. >> Yeah. Like I think there's just so much noise in that like first generation of of the crypto AI intersection. And my like I don't know read from from a distance to be honest with you was like um okay we see this AI is finally kind of here. Um it's being led by huge centralized uh research labs. Um, and I

think it was like a a fleeting moment where it was like, okay, can you really trust these infrolabs? Like, do we should we have a decentralized version of this, right? That's, you know, accessible to everybody, open source, like, you know, running on open source like models. Um, and you get like some, you know, like sort of trust minimized guarantees by what you get back, right? Um and the problem was like the um like rate of acceleration and progress from the centralized orgs was so much you know um larger than people kind of expected. And so if you looked at it from afar and you're like okay you know I'm going to ask like an LLM you know a

question. Um, obviously the one that's a centralized or is going to give you like the fastest answer, right? Rather than having a bunch of like folks all run that same question and give you back something like that is they all agree on, right? Um, and they're going to improve the models faster, right? And crypto will always lag. And so like the just like decentralized version of let's say like LLMs or LLM training just didn't really play out. But I think like we've kind of turned a corner here. um where there's I see it as like three categories that are interesting um you know one is more like using crypto like incentives and coordination mech mechanisms to bootstrap a supply side um

that can you know of of like AI inputs right that you can then build a product with um that's like the bit tensor side of things the other is just adding like additional trust assumptions or like trust guarantees to AI um and then the Third one is uh just AI like agents that can move money on chain, right? Um so we'll cover all of this stuff, but um yeah, I think there's like it's becoming clear where this stuff where it's actually valuable. Um and this is not like if you talk about crypto and AI, it's not just one thing. It's many many things. So >> yeah, I think just add my piece here too. I think this is a little bit of a tale as old as time when it comes to I think we're over there was a lot of

infrastructure that needed to get built out and you know we we'll talk about how effective something that we say often is crypto is amazing at coordinating incentives and that's true in some not all instances and I think if we look at the last couple of years the levels that needed to get built out first to make AI usable just in this chatbot that early instantiation that we currently have were really like like close to the metal like you know um FPN heavy that type of stuff. So like if you were to break down the AI value chain and everything that needed to work, there's the compute kind of like the atoms and silicon layer.

There's the infrastructure layer, so like data centers, you know, getting these massive clusters of GPUs and making them network faster. Um there's like the model development layer um which is actually tinkering with them like these individ like these foundation models. There's the inference and then real- time kind of intelligence layer and then finally like tooling agents and middleware and those needed to kind of proceed in chronological order. And I think we're going to, you know, unpack maybe the inference layer, but I think what's clear over the last couple years is that the first four value layers there were just not ever meant to be coordinated by crypto. And I don't think there's a lot of great evidence yet like we'll talk about this with DPIN that

crypto is capable of coordinating that in a decentralized incentive based way better than a highly function highly functional centralized organization. And I think you know so there's just what is the right coordination mechanism for all of these resources and when it's super fixed asset heavy I just don't think crypto is well suited to do that and then also in order for these things that where crypto might be a really interesting use case and fit like agents we needed to build out all this other stuff first and also I don't think we're giving enough credit to like crypto being s I don't know if you guys remember there was a whole narrative around AI is private or open AI is private but you can trade X, you know, this little AI shitcoin.

That was, you know, I think that was the prevailing reason why people were actually excited about it more so than the actual tech. I would pause it to all of you guys. >> Yeah. And I think might just to briefly add on to what you were saying, if you look at the kind of AI stack, you know, starting from the bottom, you have kind of compute networks and you have kind of training and inference on top of that, more like agent platforms, etc. on top of that making use of the infrastructure itself and then you know apps on top of those agents and I would say maybe for us in in crypto like for me it makes a lot of sense that agents will use crypto rails crypto rails are 247 they're permissionless they're much faster than what we have they're open they're not siloed there's a lot of properties that

crypto gives that in my opinion agents will need long term but in terms of where crypto can compete with AI I think the other layers like you were saying Mike just then the comput the training, the inference, etc. It's very, very hard to compete with centralized incumbents that have billions of dollars they're throwing at this. And then crypto kind of copies centralized incumbents and they try and bring it to crypto, but it just doesn't work the same as what we have outside of crypto. And so then really as in terms of opportunities and where crypto can push AI forward and where the intersection actually is I would probably make the argument that I do think it's actually at the application layer at that agentic layer using crypto rails and giving users a better UX and they can get even in the centralized landscape and I think that

is really where Emperor could have been built around and is still being built right now but that is kind of a more emerging area in the crypto and AI space that we'll touch on in this podcast I'm sure that I I personally think is the most interesting where crypto can actually compete. I have uh one thing that I rarely ever hear people talking about and there's a little it's related to what we're doing at Blockworks in very early stages, but I actually I I it's been very difficult to put these words into a category, so I haven't I haven't heard anyone talk about this, but uh I mean Miles, I think this is going to come up when you talk about kind of guard rails and provenence for data um for crypto. But one, Ribbit put out a piece and I'll link it in the show

notes here, um, where they connected the idea of tokenization as we understand it in crypto with tokenizing. Tokenization as it exists within literally tokenizing words and phrases for LLMs to understand. And the connection in their mind, which I thought was beautifully succinct and really powerful, was the idea of making the real world machine readable. Um, and that's a very simple concept, but it's a super powerful concept in my opinion. Um, because when you make essentially everything, and in the case of crypto, that's mostly financial data machine readable and very easy to interact with. Um, you kind of completely redefine what software means and workflows that you build. So I mean

literally if you just think about the the process of what does it mean to create like a piece of fintech or financial software today you would kind of go with this very historical product approach and you look and you'd create a user you know problem and then you'd create this big PRD you'd build software and you kind of think of engineering on the back end versus the way that this might work in the future is like one beautiful thing about crypto data at least in different virtual like VMs is that it's standardized And if if there are agents that are uh actually building software in real time, you might actually not start with or worry too much about that because the user might end up building a lot of custom things.

It might be all about how you structure this standardized data. It's a it's a super unique thing that only exists in crypto that your core critical company data everything about your costs your revenues there is offchain data so there's asterisks to this but huge component of that the whole history of your business could be machine readable and you could easily ask an LLM to build workflows on top of it I mean it is the it sounds esoteric but the idea would just be like if you think about this from the user perspective is instead of having one piece of software ware with a ton of seats. Um, you pay to have your data standardized, versioned, authenticated, structured in such a way that an LLM could easily query it, come

up with customized workflows, and answer any question that you'd have in the whole history of the org. It's like it's a totally uh different way of um I think the idea there is personalization. you're you're going to get much higher degree of personalization and time like quick decision-m once agents are taking these actions on the raw data that's been structured in a certain way. So I think the way that we build products and especially SAS is going to get completely reinvented and it's going to happen on crypto rails first because this very unique property of data but also because crypto is mostly startups and you don't have this incredible tech debt of massive you know multinational orgs which are you know have like their Oracle onrem software and like all these different licenses and they're going to

have to untangle that and it's going to take forever. So yeah, I think crypto foreshadows the future in a number of ways. And I think just by virtue of it's a young industry and has this really unique property around what the data looks like that you're going to see what the future of organizations look like here first before anywhere else. >> I think Mike, I actually might throw a question at you here which the audience might be interested to hear about. I mean outside of crypto they talk about uh AI and you know how agents are going to sort of completely overtake SAS and you know to some extent block works is a SAS company. Are you thinking, you know, as a crypto company, are you thinking about using AI to replace some of your more software as a service or like how do you see it disrupting Blockworks

right now? Like the entire AI category if you have to throw it back on on that? >> Um, I think where the space is generally going to move over a long period of time is like outcomes based pricing. Um and it's like in this rivet piece I thought they did a really good exa like it's a high level framework but I think it's a useful one is thinking of businesses as token factories. Um you have a certain number of like you know you essentially are requesting a specific um output and there's a certain amount of computation that happens before it outputs something. So, I think that it's going to be so the cost is going to end up being so variable in this new model that you won't be able to do standard pricing the same way you were doing within SAS.

Um, and this like per seat model where you kind of lock everything in and try to jam people into like too much structure I think is going to break um pretty soon. So yeah, I kind of agree with all these trends that you hear about that um I think it'll take longer like we're still have SAS products with um subscription revenue and I think that that's going to it'll take a couple years for that to really change. Um the other thing too is I think buyers are very used to paying in a subscription form and like you kind of need big companies to innovate on this first. it doesn't make sense to move on things like SAS pricing before like a big incumbent like a Google or an Apple or

something like that moves first just as a strategy. So >> and to double click on that though like what would that mean? That would mean like okay now the majority of people are not coming to Blockworks directly. Maybe they're asking an agent to go do write like a pull data for a research report or pull data that informs like actions on their behalf. And now you have to like change your product to be like, you know, have like agentbased interfaces, right? Um, and I think one of the big pain points right now is everybody having to give the API keys to their agents that they pay for, right? Uh, and that's where X42 will like I think come in pretty interestingly like down the road is just moving over to like usage based models. So it's either usage or outcome based like I totally agree.

>> Yeah, outcome is really tough to me. I think everyone's going to want to >> Yeah. go towards outcomebased. The problem is quantifying outcome based is really difficult and it that actually is based on a company taking the time to define what an outcome looks like. Right? So this is actually one of the big geniuses of Google is they were the one not only did they create the best search engine that ever uh you know like advertising engine that's ever existed but they also taught marketers how to do attribution. So, not only did they create the best thing, but they took the time to put together um you know a framework for how to do attribution for marketers and then they educated everyone. So, they wrote the rubric, right? They built the machine, then they wrote the rubric and they don't get

quite enough credit for writing the rubric. And it's it's difficult to measure something like an outcome, right? Like let's say you are building um I don't know some sort of marketing module or like marketing piece of software where you're trying to acquire customers. Attribution is super difficult there. How do you measure an outcome? you you need you know it software providers will eventually define a benchmark or create a methodology for but it takes time to educate your customer there >> I think it's you know the way I could see it developing is like it's really what you pay for like the outcome maybe that you want you're willing to pay more for it and maybe it's better defined and then you pay a certain amount for that and then if it's not the outcome that you define maybe you pay a bit less for that because the agent can't execute on

what you want >> I think Mike you also touched on a few interesting points around like relevancy and so if We move into this world where agents are the ones that are actually paying for maybe something like Blockworks, how can a company like Blockworks make sure that when an agent is scanning the internet that it is pulling from Blockworks data over maybe some other peers in the space and like do you have to pay for that to get in front of the agents? Is there some sponsoring ads going on there? Like it's unclear as well in the AI space right now what this even looks like and how things will become relevant. We're kind of going through that Google moment all over again, but in this time for AI. And so I do think in terms of crypto and AI and opportunities like this is

potentially one that even centralized crypto companies, incumbents like maybe someone like Blockworks can get very right and really actually become like a behemoth in the space by understanding like how to get in front of the agents at the right time to get their data in the hands of these agents because agents are hungry for data at the end of the day. So like I do think that the crypto AI landscape is still new. Mike, you said before it could take many years before maybe you switch over to a more agentic kind of software. And I think that's really where we are in crypto and AI overall where there's been a lot of experimentation, but realistically, I think we're probably still two or three years away from understanding exactly where it's going to hit the hardest.

>> Yeah. Why don't we break down some of the major categories that people have been excited in or excited about in the past? And why don't we start with this idea of coordinating compute? Um because this was like we talked about this a little bit at the top of the episode, but I think the very first um the very first application that people were interested in was render um aos uh Jensen these sorts of protocols where you're it's kind of like um filecoin but for you know GPUs and the idea here being you know instead of centralized training um you could maybe do the um painting a broad brush like some of

these categories moved up into inference and but just for painting with a broad brush purposes decentralized training and physical infrastructure was a huge one. Um and then you had some you had some other uh you had some other protocols like move one layer above the stack there um and get to inference. So this was um this is a little bit where I uh is currently playing kind of not 100% but kind of um you also had like ritual if you remember that uh that that protocol and um yeah there are some other there are some other pretty interesting one like some one that flies a little bit under the radar is like hyperbolic um is an example of this but uh and then bit tensor was this kind of you know incentivized marketplace for

bringing um you know inference inference models um to bear. So can we start with the first category here of just decentralized compute I'm just to not bury the lead this one doesn't seem like it worked super well. So what what went wrong here do you think in this category or how how much how how much do we like this idea of you know bring like leveraging incentives in a decentralized way to uh like to improve AI outcomes. I mean I think uh yeah there's it's like some very very important nuance like we were just talking before the show started you know you're not like the the product experience of asking you know

openai or claude you know a question and getting it like a response back is always going to be you know you asking like you know and a validator set of 150 like machines to do like the same task check each other and then give you back some result, right? And so you're not going to like pick the shittier product like product for ideological reasons. Um, and I think where like I think Bit Tensor has found its niche is um, it's the like amount of money and time uh, like and resources that would go into standing up a new AI based service that is not built on one of the big incumbent

um, labs, right? Is like in getting more and more and more expensive, right? Um, in the same way that like you know it's the classic deep end like sort of thesis like if you were to try to stand up a new cell network around the world that would be a herculean task that would require you know maybe definitely billions maybe tens of billions of dollars. Um, and so where Bit Tentress kind of found its niche is like, well, okay, you don't want to be totally dependent and get on on like one of these big models that could just rug you at any point in time or lock you. Um, and you don't want to go out and have to raise like a billion dollars to build

like a defendable, you know, business with a with a like the sufficient amount of resources that you would need to actually like stand it out, right? And so we've just got a big pool of operators, you know, people running machines kind of waiting around for jobs. They get paid, you know, some financial incentive of like the Bit Tensor token, right, to go do a job for a particular service. And that isn't just like doing, you know, running models. That's also like, you know, collecting, you know, data, right? Or like labeling data. Um, it's doing like testing. It's it's a bunch of different stuff and so like you know Z you've dug into it more than I have but it does

feel like you know Bidens kind of stood up like this general purpose you know uh coordination layer um that if you want to build a like a new AI company today and you don't want to go raise ton and of like an enormous amount of money um this is kind of like an accelerator right for your supply side um but that does not mean that the product that you're operating has to be decentralized itself, right? The product you can build a product in totally centralized way.

Um, and you're just kind of, you know, uh, paying for these resources um, to this open supply side and that's seems like it could be viable. Um, but yeah, what's working, what's not working, Dave? Like is how different from the ones that fail. >> Maybe just to touch on two. So like a bit tensor versus like on a cache here. Um I think a cashache you know massive props to that team known them for many years and have a lot of respect for them. The hard part with comput networks is that it's kind of a commodity compute itself and so like what's stopping the next person creating the next decentralized compute network and how you going to compete on costs and like what is the infra that you actually own underlying that is it tokenized like who's the actual customer etc. Uh and

also is it cheaper for like longstanding workflows which are needed continuously instead of like having to set it up per workflow. On the other side you have like bit tensor which I think you know Mike was asking earlier in this podcast like is coordination a thing to do incentives work? I think to the most part it doesn't actually work. I would make the argument actually a bit tenser. It does work and it's actually something that I think should be studied still and it's something that I am studying right now in terms of like I didn't spend enough time thinking about like why it exists or why it's still working and why there are so many builders still going there. Realistically, I think what Bit Tensor does well and maybe better than just standard compute networks that are maybe more commoditizing is it incentivizes miners to provide very very

cheap compute using their token and they kind of play these token crypto games to get what they want and then as a result people who actually want to pay for this who have the model for example that need to train they're willing to go to Bit Tensor because it's probably a lot cheaper. I haven't done the maths here, but I expect it to be a lot cheaper than other comput networks because it's subsidized by the Bit Tensor token and then they kind of speculate on then the appreciation of that token and it's a classic crypto circle. The question is if the token doesn't appreciate is Bitensor still as interesting and does a coordination work in itself? Uh that's like TBD. However, just to finish quickly, I do think there are some subnets in the Bit Tense ecosystem that are showing very early signs of some

level of PMF. And I think their target market is more like indie AI developers and there's nothing stopping indie AI developers now in this day and age with the amount of tools they have and it just takes one or two ideas with enough compute and and training etc to create a product. The next question is can you productize it and make it feel like it's not a crypto bitensor AI app. that super difficult and maybe that's actually where the bottleneck is right now and the constraint in that ecosystem but that's how I see it. >> I could I maybe take the other side of why I actually don't think this is I I I am starting to maybe as a as a hot take.

Um I do incentives work at all? Um, I'm actually not really sure like maybe in a very very select number of cases, but I think I'm starting to get on the train of incentives don't really work. And the reason why is the best thing that you could there's so much reflexivity in crypto. The very best thing that you could do for adoption of your product is to make your token go up. And as long as you're dumping out tons of incentives, even if you actually get more usage, your token will go down and therefore people will not use your product. So you it's you're I think this is why there's why there's so much confusion behind the like my my this is one part of the reason there's confusion around this but

um you know my my metrics are actually improving but my token price is going down. It's like yeah because you're paying a ton of money for that for that activity right you're you're just paying for that activity. So you know I I think it does come down to the maths here. Um, and yeah, I think there's we people should just start looking at this stuff as though it's actually cash compensation because it is. Um, so I don't know. >> The Bitcoin people would push back like that is that's worked out pretty well and it's because there's like a cult sort of, you know, belief around this token. um and like what it represents and I think you know Bit Tensor folks are like I mean they're they're into Bit

Tensor like there are entire funds just dedicated to stacking TOAO and you know I think it's it's super interesting and I think like a lot of us faded it because it was mixed in with all these like VC backed you know typical like turned out to be kind of scammy or like at least low you know short-term plays that the original crypto AI like narrative. Um, and you know, Bit Tensor is not VC backed. It's kind of got like a little bit of hyperlquid, you know, sort of lore to it along with like um, you know, obviously trying to play off the the Bitcoin um, history.

>> And yeah, I mean that you have to like I agree with you. Um, unless there are people that are actually just not selling the reward tokens, you know, immediately or like, you know, they're selling maybe a little bit of it, but yeah. By the way, I don't know what like Bit Tensor's chart looks like. It's probably not amazing. Um, but >> it's actually a lot better than most other crypto projects, I would say, in the last 12 months. So like they've held pretty steady and that's why I think it's interesting because it hasn't been negatively impacted by the current cycle. And I do agree Mike like you know I think we're both aligned that demand matters again more than anything and you shouldn't just incentivize with tokens

on the supply side and get people doing certain things without anyone actually wanting to use it in the first place. We've seen that fail many times. The only time I think it works in crypto is actually if you subsidize enough and you have a goal to achieve and that goal is PMF and actually getting people paying for this stuff and like actually earning you know whatever revenues, fees, etc. Bitensor has this subnet ecosystem that you know I think it's like 128 subnets that are allowed to go live there at least for now. I think they're working on trying to increase that, but if they manage to get 128 subnets, it's kind of like throwing darts on a dart board where like you just need one or two to stick and then probably it gets talked about in other AI circles. And I spoke about this with my intern yesterday.

Shout out to Spencer. Um he's been doing some research for me and there is one I can't remember the name like right now, but um there's one that Subnet that did better than Claude Code for in terms of like performance and what you can do and that's like very cheap to use and was very cheap to build. And there it just takes one or two breakouts like maybe it's some vibe coding thing, maybe it's some like prompt thing like who actually knows what it is where a bit of luck and a bit of skill and then the whole AI space is talking about it as well. So it be a breakthrough not just in crypto but also in AI. So there's always that small chance and that's why I think it keeps people interested in the ecosystem.

>> But is so that that is kind of like the the thesis here, right? that they can offer because their um their supply side like their capex is being subsidized by bit tensor, right? they should be able to build products much much cheaper which means I could sell it for much cheaper and so like that is the thesis in my like what it seems like is like you get these you know potentially more success in like the long tale of like services right but they are going to be cheaper than you know a company that had to dole out like $2 billion of training costs before they could even release the product right um and the question is like is that cheaper going to come at the cost of like quality latency like

reliability because it is, you know, an open supply side without SLAs's and all of that good stuff. Um, and it sounds like, you know, there are a couple projects out there that are like giving uh building really performant, you know, as you just mentioned, save like products. Um, and they were able to sell it much cheaper than their, you know, centralizing like counterparts or compet competitors. Um, so that that to me is a thesis and like uh maybe I'm missing something, but it's interesting and it's worth paying attention to. Um, if they can win demand, right, because it's cheaper.

I wonder if again um you know this reminds me of a little bit is when crypton natives were trying to sell Ethereum they used all these words like credible neutrality and you know you had a gajillion ways of saying we have the most validators at the lowest cost. I mean like all these buzzwords and all this stuff and now that you actually have financial institutions building it's unclear that they actually care about that. But what they do care about is one property which is that it's 24/7. Like that seems to be the the killer app, right? And it's unruggable and all this other stuff. I don't want to trivialize any of that, but I think that's you know when you hear people from the like wow this stuff just

happens 24/7 and that that seems to be a huge component of it. And uh you know I wonder with some of this you know with the bit tensor stuff I wonder actually if the killer feature is that it's just close to being on chain and the proximity of it these models existing on chain already. And um you know I think a lot of crypto businesses will be the first ones to leverage and utilize AI in like when we start to get to aic commerce like that kind of way. And there's value to like proximity of an inference model living on chain so you can take action quicker. Like I I think

it actually might be more of the proximity thing than the cost thing. I just I haven't seen any evidence that I I I've calling me like extremely skeptical that any of these um marketplaces will be capable of producing a better product. But the proximity and it existing on chain first I see as being super valuable. Um, >> yeah. So, I have a few small things here. So, I actually wonder if it's going to be more of a compliance thing long term. So, like, you know, at some point maybe regulations come in and you need to make this AI is explainable, transparent, auditable, etc. And how the hell are you going to do that right now with current systems? like you're not some of the other projects in the past that I've worked with that also others people know of for example like Ien

cloud you know they're working on inference and computing kind of offchain and then verifying the compute was done in some way offchain onchain like you know etc like the classic kind of layer 2 framework that people are familiar with listening to this podcast probably the hard part is like how do you do inference onchain and like get people actually executing things like fully on the blockchain fully transparent fully auditable like no other trust assumptions that is quite hard but I do see a future where that comes to blockchain and people are working on this right now. And I do think there will be a breakout team that understands the compliance requirements and then probably the value will crew to the team that makes it the easiest for companies already using AI that need and must as per regulation to use compliant AI to like make it very easy for them to

integrate the same models onto their stack on chain. Yeah, >> but let me just like to be clear like you cannot run inference on chain at least in a smart like as smart contracts that we know today right >> they are determined >> software like as an input it runs exact same way exact like every single time right validators run this code and the output is the exact same and the validators check each other um that's also the reason that smart contracts cannot do things like call APIs natively right because It's not deterministic. If a bunch, first of all, validators are not going to like set up, you know, APIs for every single smart contract that needs one, right? Um, even if they

could, they would get back different results and therefore they couldn't agree on the results of the next block. And so what we're talking about here I think is like maybe slightly different things like we are trying to get sort of uh deterministic or like smart contract like guarantees from inference that is of course run offchain. You can do that by like having a bunch of operators run this in the same box, right? Um and you know and you can add like um basically give the consumer assurance that the model has not been changed under their feet, right? Um and then you know there's also like agents that we're talking about that do stuff in crypto

like like make onchain like transactions and that can be done also without any sort of verifiability or determinism for like you know the model that's being run. But the agents are controlling money and moving onchain. Um, so I think these are like very different things. And then the Bit Tensor thing, like there is nothing that's even close to onchain there in terms of what what we typically think about it. Um, like tokens move around Bit Tensor and they found a way for like the validators and the miners to check each other, but they're not doing anything that's like DeFi related like to be clear. It's more just um you know it's like a supply side for like regular regular products. Um but yeah, that's why this is so

complicated because it is like you know there's a lot of things that sound similar but I think they're like pretty different and have different considerations. >> I think you're right. Um I think you're right, Miles. And uh yeah, I think that's actually one of the the biggest challenges about um you know, LLMs are fundamentally non-deterministic. Uh smart contracts are super rigid in that sense and getting those two things to marry up. Um maybe there like that feels like there's a big unlock there. Um but you know, if the models don't live on chain currently and that's not possible.

Um you know one thing you know getting back to maybe Z your note about compliance I mean you could imagine uh just imagine the compliance over one thing that is absolutely coming to market soon when clarity happens is if you have a publicly traded instrument which is a token you're going to need like essentially a lighter version of what public companies require which is a basic set of disclosures risks um people need some transparency into the operations of the business and you can imagine And again, if you are generating all of this data in a way that an LLM can easily query and the LLM you can it's provably non-temperable like that should be a dream for a regulator, right? Or any set of investors. Like

imagine you can just ask almost any question of a business's database and get like a pretty good honest authentic answer out of it that like can you imagine what that would do for trans I mean you can still do all sorts of shady things I'm sure like fraud and whatever but the more that's onchain um and you can understand like an LLM can query it and you can make sure that it's operate it's not hasn't been tampered with and that's an authentic answer wow that'd be great what an interesting world right >> yeah but it also like the more if it's lit literally on chain then that is is in itself an attack surface right like that actually is a challenge um >> you know explain that what do you mean it's how is it an attack >> well it's visible right if you like have

something like the guard rails or the policies of like um you know that's going to happen offchain is visible onchain now that's a potential attack vector for somebody to try to you know game this or or something along those lines right um and so like this is I think pretty challenge ing. Um, and yeah, it just depends on like how many use cases are there where like uh either a user or a company is going to conscript some AI model to do a job and they really care about what AI model ran, right? Um, that's like kind of the IGEN thesis is that there's going to be a lot of cases where they really care about that. Maybe that is for compliance reasons sake like I we'll we'll see,

right? Maybe it's for risk reasons. Um maybe they just want the cheapest or the easiest, you know, service to run and maybe that is going to be a bit tensor. I don't know. Um or maybe they just care about the outcome and the outcome satisfying the constraints that they've, you know, basically defined and like the guardrails that they defined. Um that's the the last one is what I'm spending a lot of time on these days. But um yeah, I I'll pause there. Maybe we can move on to uh deepen here as a category which has always been somewhat adjacent to AI, right? Like the idea here is you're kind of I mean again if you kind of look at this I mean

literally I was on the the website for a caution. It's kind of the like Airbnb for compute right? You're kind of coordinating this you know these offchain pieces of infrastructure to produce something you know uh you know to create this marketplace. I mean, deepen is really been very similar to that as well. Um, and like Miles, maybe I can kick it to you here, but I'm curious, uh, how you view deepen within the context of the intersection of AI and crypto and like what if any, you know, you know, fruitful material there is there.

>> Yeah. I mean, I think it just goes back to like um there uh you don't want to be maybe in like the middle ground either. Like, you know, uh releasing products from like one of the big incumbents is going to probably go well and you'll probably win every market that they really go after. Um and then I think on the other end of the spectrum, if you can bring a product to market as cheaply as possible that's still equally performant as the big incumbents, then like you have a a chance at at getting a wedge here. Um, you know, I don't like I I think outside of Bit Tensor, most things have kind of died. Um, and Bit Tensor subnets like they there's a couple that are doing well. Um, and it's not clear if people are like, you know,

using those services because they do love like TA or if like they're actually winning. Um, we were talking, you know, comparing this to like Geodet, you know, before we got on. Um and Geonet clearly has a product that like you know regular people that know nothing about crypto really like because it's better than what had you know gives more accurate like GPS data right for something like precision agriculture right they're going to selling this to like farmers like big big you know middle America like farms and that's great because it's a cheaper better product um so if you can get cheaper better products from deepen

great. If you can't, then like this will just not work, right? Because we've seen deepen like work sometimes. We've seen it, you know, kind of fizzle out. Like I don't think any like Helium, I don't know how well they're doing these days. Um, but yeah, that's that's generally my my take on it. I know there are people that have spent way more time on this than I have. Um, but I think there's there's there's a chance. Um, I'm not writing it off yet. >> Yeah, I can maybe add a little bit of color here. Um so there is one category in the deep in space uh which I think is still potentially interesting. Maybe timing is still a little bit too early, painfully early for me because I've been working with a team working on this problem for four years, but it seems

like now they're getting first interest from enterprise AI customers, which is really interesting. And so like you know many listeners probably have heard of a team called Grass, you know, Grass is like a kind of tab compute network uh you know provides certain compute to AI providers etc. Uh there's another team which is very similar which we work with very closely at practice called bless and we invested in this team many years ago. It's a very similar concept to grass. It's an edge compute network but they've found a new go to market which is actually using the edge compute to you know they will ask users for their permission and then the user essentially gives permission and how they're using certain platforms for example like chatbt and then bless will anonymize this data privatize the data and then

basically collect the data on like 100k of their you know crypto deepin users and then sell that data on to like AI companies wanting to get more insight into how people are actually using LLM for example right now which is actually quite hard to get this data funnily enough. And so yeah, again here it's quite a sort of like out there concept in terms of some deep in things are so early where like we're still only just seeing demand actually on like how we can use this in the first place. And I actually am still like optimistic on the data side. you know, I think Vanna, Grass, Bless, there's, you know, there's probably like five to 10 of these I could mention off the top of my head where I'm sure if you research any of these, they're making some uh, you know, some progress into like actually selling

on to people outside of crypto, which is where I think it gets very exciting. I will say though that like I've experienced this quite frequently recently where a lot of teams realize now postregulation there is a market outside of crypto, but if you have a token, it's like how do you sell on to these customers? cuz then you have the token and they're looking at Bless Network and they're like, "What the hell is this Bless Network token on Coin Gecko and like is this a scam?" Blah blah blah. So, it's difficult with Deepin because you have the token and then probably you're selling on to people outside of crypto and I don't think we're there yet in crypto where there's like a lot of comfort, but it might come very soon. Who knows, >> right?

>> Yeah. >> Well, I was going to say also I think some of this depends on there are different categories of DPIN as well. like you could you could even lump like ionet or a kosh or render into like deep like right that's physical infrastructure that's being coordinated by token incentives right so it's kind of like a compute layer but then there is this data layer as well um and actually the examples that I was think that I was thinking of was hivemapper um so the hivemapper business model right is basically like hey give out token incentives to um you know each one of these things is like hey here is some percentage of the token supply to get you to perform an action which gives in some cases a very specific set of data, right? So like geodet is another example

of this where they can actually pinpoint you know within some crazy small distance right and that like the question is just there's a you know within all of these different data um you know deepin plays I think there's a real difference on how commoditized the data is um so if if the data is something which is commoditized that the LLM scraper gods are eventually going to get around to scraping you probably don't have a great business there right and you're probably equivalent to, you know, a mainstream media outlet pre Google, right? Where Google just crawls all of your stuff and they end up extracting all the value. In the instance of something like Hivemap, so Hivemapper's business model to remind

people is you have a dash cam, you're driving around and you're mapping this stuff and you actually might try to sell that data back um to, you know, large purchasers of geographic data. Um, so the LLMs might actually be a really good customer for you because there might not be a super easy open web solution way to get that and it's consistent. You have to keep buying it. So that actually I think the devil is in the details here in terms of what sort of deepin category you sit in and then also what sort of data vendor and reseller are you and I think the way the only way that you can build potential like ultimately at the end of the day I think that even if you have a very proprietary source of data you never want your whole business to be

selling into an aggregator right like eventually that aggregator will squeeze you um so the question is you know can you build um you know can you find proprietary ways of getting access to your customer that doesn't entirely depend on LLMs as well. Um but I think I think that it's like if you are the way that I would put this is imagine if you're a data reseller to an LLM and the LLM independent of you is what's making that useful like turning that into an insight or an action I don't think you have a very good business.

are the one that is making that useful via LLM in some way. That's a better business. >> Yeah. And that that is I think like become very clear as the winning formula for deep end is like you you know aggregate like the some scarce resource right that that powers your product. Um and you can do that much more cheaply than paying out cash. You can do this like programmatically with the token that you mint out of thin air. And then in between that side and the customer, you have a company that's just like a normal company, right? And so like Zav, you were saying like they get nervous when they're like, "Okay, well, once they actually double click into what this product is, like what's happening

in the background?" It's almost like you want that to be a completely separate website, completely separate everything. And then you just have this buttoned up company that packages whatever resource it's like aggregating, right, into a product that people can easily consume and feels like a normal product. Um, and you know, if you get enough revenue on that side, you can start doing it to like, I don't know, burn the tokens or buy back the tokens or do something. >> Yeah, exactly. make something give like some reason for people to actually want to hold this token that you continue to like shell it you know like mint you know uh inflationary like incentives out of um >> we keep teasing this buybacks episode but we have to do the buybacks episode

>> it'll take five minutes >> maybe as well just to double click on something briefly that you said before Mike I I do think that there is a business like there is data out there which like if you have enough you know compute maybe using crypto incentives to get that comput in the first place that people will pay for in like droves. And the question is like what is that data? And you said Mike before that, you know, the data feeding into LM's like a lot of value there. Like I agree there's definitely value there. I also think there's actually value in understanding how the hell these LLMs work in the first place and how businesses can leverage LLMs and get in front of the right people and like make sure it's a relevancy problem again like I was saying earlier in the episode. So I

think maybe relevancy is actually a topic in crypto and at the intersection of crypto and AI where like if you can help a business become more relevant in this age of AI using either data or agents or whatever it is like I do think there is actually a massive opportunity there in deepin across deepin agents etc to be for crypto businesses to become more relevant in this age of AI and make sure they're always at the top of being discovered. So I'm quite bullish on that as like a category as well say in the deep in space. So rounding up here, how do we feel about this whole bucket of kind of the of one one bucket of AI and crypto as an intersection which is kind of compute marketplaces and you know um and also leveraging incentives to

coordinate uh you know computer data in some kind of way. Maybe we can lump deepin in there and also some of the more um you know decentralized training and infrastructure plays like the IONETs, the renders, the accash and also like the kind of bit tensor and ritual. I don't even know if ritual is doing anything anymore but the inference layer. Um how do we feel about this as a whole category? like where if you had to slice and dice it and to create some rules and takeaways like what makes you know which parts of this are you bullish on versus not like how would you summarize everything we've said so far?

>> Yeah, I think uh on my end after sort of speaking with many crypto teams the last two years I would say at the infra layer I'm a bit more bearish. I'd say Bit Tenser is probably the one team that I think is doing a good job there and they're doing a good job because they're coordinating well and using crypto incentives and they've also found a market for people who actually want to use this stuff and this market it helps them build better products and so I would make the argument that Bit Tensor is one of the very few crypto infrastructure uh projects at the intersection of AI that is doing a great job. I think it's really really early on the demand side for products and like apps and I think that is where a lot of value is going to acrue in my opinion

and I would say in general like we talked about a bit on the deep end side you want to think very hard about demand and like what people are going to pay for here and like why use crypto in the first place and then when you thought very hard about that and you actually build the product you then want to think very hard about how you distribute the product to the customers who are actually paying for that for example if you have the token and then you're targeting these non-crypto customers was that the right move in the first place but then there's a trade-off there and so I think that we're early there are some things that are working some infer is working but I do think the applications is where most of the value is going to acrew here using crypto properties helping us do things across

software in this open way that we can't do right now with centralized incumbents that we're still only just really scraping the surface for. Uh so that's what I'm very excited about right now. >> Yeah, I think I don't have a ton to add on the on the deep end side. It's really just a question of like are these products competitive? Um I am much more fascinated I think like in this general you know future where it's like okay every company in the world does you know AI transformation top of mind whether that's their internal like automating their internal workflows and like full-on jobs being automated to AI right or they're you know making their actual products more like AI native to their

customers like having a chat interface for you know when you go on like Amazon, right? And now an agent's going to like give you like shopping recommendations, maybe even buy [ __ ] for you, right? Um, and that's where like I think uh crypto and AI both like reduce costs a ton. Um, whether that's making like you know able to do more with less headcount or whether that's reducing transaction costs. Um, at some point in this transformation, it's very clear that like the agents that are you're using to automate jobs or to, you know, um, create new user experiences, right, are going to need to hold money um, and do things with with money, right? And that

can that's gonna have a huge spectrum from stuff that's super low stakes like me telling an agent to go buy me like socks right on Amazon to something that's like you know automating like your company's treasury ops right with an AI that is going to keep your you know everything optimized at all times right that that's pretty scary if anything goes wrong um but I think you know at some point over time like it's going to happen regardless. I have very high conviction in that and a lot of the transaction or like most natural way for an agent to handle like value is digitally native, you know, tokens,

right? And so that's kind of where like um crypto will I think play like a big role. it'll just happen very much so in the background and like the more it'll be abstracted away from like the end you know companies that are using like automating things you know um or like the end users that just trying to have like you know a normal like shopping experience and yeah it's like using stable coins in the back um you can kind of see how this is coming together right like it's Stripe is a great example it's like you know a checkout option on every single you know on most websites um eventually you know They're going most of those like e-commerce sites are going to have some sort of chat interface.

Then it's going to have an agent that's going to move money. They're probably going to use like stable coins, right? Um but we'll see. There's there's a lot. It's it's still sort of TBD on at least the timeline of this. >> I think uh maybe for my take to close, I think on the inf uh the lower level infrastructure side, it just remain pretty bearish on it. Um, I think one thing that we didn't even talk about really was latency. And it's kind of it's kind of like quite silly to look back on some of the distributed training pitches. And you know, when you like think about coordinating training across clusters which are all over, you know, the United States or the world versus like these insanelyworked up um, you

know, Nvidia style data centers. I mean, yeah, no wonder the AI people were laughing at us. I mean my gosh that that just never really made a ton of sense. But I think my hope would be that okay we ran some of the you know most obvious and there's there's a knee-jerk reaction here which happens in crypto which is take a thing which is working and wouldn't it be better if it was decentralized and forecasters maybe not to you know poke holes at that but like this there's always my challenge with like the decentralized social media stuff. It's like exactly X but it's decentralized like no one it doesn't matter. You you haven't innovated on the form factor in in any way. So I think we're about to run that same experiment moving up up the stack towards areas which are again less about coordinating

billions of dollars of or you know hundreds of billions of dollars of compute build out which is so highly specialized and definitely done better by a central like centralized entities at least today um towards more of like the use cases and yeah I mean we're going to spend an entire episode on agents and you know agentic uh you know workflows and uh payments and trading and finance and all of that stuff. And um again my hope there is that we can get into some of the nuance around like you know are we built as like betterment you know raia style easy investing versus you know alpha trading spotting arbitrage opportunities like there's a huge spectrum even when you say like agents are going to invest on your behalf what that means and where that

might work. So, I think we should dig into that. Um, you know, we can dig into that next time. But this was a good first run and intro at this uh thorny thorny problem. Are the AI coins going to run? What the audience want to know is are the AI >> I'll run, mate. Don't worry about that. I'll run. >> You know what? I actually This is not investment advice like at all. You know, I was looking at some of these, you know, I I uh one angle that no one really I think everyone just writes off and hates, but I remain very long-term bullish on is like AI like kind of like AI XBT like experiments. Um because I'm very bullish long term on AI driven creators. Um, I can go and look at Japan and they're already uh, you know, before

all the LLMs kicked off like holograms selling out stadiums and yeah, I think as LLMs get better and content gets better and you know it can produce 1,000x uh, better than you know a human creator. I don't know. I think it would make sense that you having a big audience like that should work. I think that should work for what it's worth. >> When my hologram >> if I could do it now I would. Yeah, I'm an That's how I know I'll be successful when there's a Mike LLM that um that you guys don't even recognize the whole like, "Wow, your tanks are really good today." Wow.

>> Keep it rolling. I don't know what you drank this morning, but can I have some? Um >> All right, guys. >> Sorry. >> Coin it. You know, >> you were always a skeptic, Miles. You're always a skeptic of um Zor will have its day. It'll have its day. Um all right, guys. This was a really fun one. I'll see you here back for uh you know, round two on the AI saga next week.