What Went Wrong with Crypto AI Agents (and what's coming next) - Tarun Chitra & Wei Dai

Transcript

You know the LLM of the last few years were very good at like giving you an answer but they were not very good at like correcting themselves. Are you actually convinced that the underlying mechanism is is actually quantum this actually quantum computer under underlying it you know up to 20 30 cubits we can always simulate right so so there's no way of knowing the intersection is lagging by like 3 to 6 months on the actual progression be like the deepseek moment on steroids where it brings down like it just crashes like the market for compute because now we can tap into this like supercomput welcome back to the rollup your home for technical talks and open conversations that shape the modular future Nothing on today's show is to be construed as financial advice. You can see our full disclosures list at

therollup.co/ventures. Sit back, relax, and enjoy today's episode after a quick word from some of our awesome partners. The intersection of AI and crypto is moving fast, and soon an agent will be managing your onchain transactions for you. But without guardrails, it's a disaster waiting to happen. Newton locks agents to your intent with cryptographic proofs and real penalties if they slip. Learn more at Magic Newton onx. What's going on, guys? Welcome back to the rollup. Uh, today we have the two good guy VCs of the space. Here we've got Wei and Tarun, our chief researchers. Tarun, a man whose hair makes other men's hair look normal. And Wei, a man's whose brain makes other men's brains look normal.

Guys, welcome to the show. Hey, it's good to be back. So, I was I was saying before we jumped on that uh the Rollup Ventures, which is soon to be redacted name and kind of disassociated from the rollup has been has been quite slow on on the VC side. The tax season amongst just valuations being quite high still. Would love to just start off with a kind of rundown from your perspectives as to what you guys have been seeing in the in the landscape o over the last two to three months particularly which sectors have been hot or even just like the research topics that you guys have been looking at that have been particularly interesting from a a first principles view. Um I think it'd be super helpful for our audience just to get a rundown

as to as to the latest infra and app side of things. So um yeah, we why don't you kick us off and again welcome to the show guys. Yeah, it's it's good to be back. I think you know obviously I think venture continues as usual. Um typically right in the early stages we're still seeing builders building. We're still seeing new ideas. We're still seeing uh experimentation and but there's definitely like a you know slowing down of of like memecoins and like you know especially on the AI agent side. I think um that market we're still seeing a lot of builders but right there's definitely uh there's like a cooling down of of all these um hey pick a button launch a coin. Um it's still going to be there. I think there's still going to be you know maybe more more capabilities to be built

there. before we just saw let's launch uh you know just AI agent posting on Twitter and with a token right and I I don't know quantum coins maybe is another form but but really like I think there's more and more like real content actually so wait wait wait wait are you investing in postquantum stuff are you are you did you alpha leak something that you're telling us um I actually curious to hear what you think sort of I think um I still think it's early but I think there's a lot of potential long term yeah I I also think it's early But, you know, I think in uh 2010, I almost did a PhD in quantum computing, and I definitely don't regret not doing it.

Let's put it that way. I mean, quality has been 10 years away for 30 years. Yeah. It's like it's like I I just think it I'm not I I'm not I'm not I'm not that convinced. Although I think the Microsoft chip, the idea that we actually have an ANO and a topological quantum computer which has very low error correcting requirements would be cool, but they haven't really demonstrated that they can factor anything very big with it. So I'm not I'm not that worried. I'm not worried in the sense that like I don't think like hey Satoshi's coins are moving tomorrow.

You know what I mean? Can you break that down? What like anon is that? Sorry. Sorry. Yeah. So, so, so there's two types of quantum computers broadly speaking. Um, there are, you know, your classical quantum computers which are you have a bunch of cubits. So, maybe you have, you know, some kind of uh the classical, you know, when you're in in, you know, a physics lab. The classical example is like you take a uh either like an argon atom or a trying to remember what the other noble gases that you use, but you put in what's called a penning trap, which is like you put in this really cool thing.

The molecules really tightly packed. You shoot lasers at it and you get it to change its state. And like this is like the like class experiment of like how do you make a cubit that lasts for, you know, micros secondsonds. So it's not like a stable cubit but it's like that's and that type of cubit is usually like single atom or single particle. Um and one problem is that that you know that that historically has happened is you know if I try to get many of those cubits because I need more than one to actually do real computation. Um those cubits interact with each other and they interact with the environment. So they're unable to like hold their state for a long time. So like you know maybe

you have three cubits and you know one is in this state, one is in this state and one is in this state. Well the problem is due to the environment and interaction excuse me like you know even just like a random electron coming in from you know a light that you have on it may lose its state. So it's like as if you had memory but it was like randomly changing itself and you couldn't keep it stable, right? So it's like that's sort of and and so the biggest problem for quantum computers historically has been what's called error correction. So error correction is like a little like error correcting codes that you see in data availability in other places but it's sort of more complicated in the quantum world where it's how do I you know even if it is noise even if there's like

noise like changing my memory randomly I've encoded enough signal such that I I can always redecode it. Now historically the problem has been either the decoding algorithm is more complicated than the algorithm itself. So you spend more energy trying to decode it. So you don't get the quantum speed up anymore. Uh or it's just like very complicated to work. There's this other type of quantum computer called topological quantum computer which Microsoft research has been working on for 25 years. Actually I was going to do my PhD at Microsoft Research um which was in Santa Barbara actually um and uh like Microsoft research opened this kind of department Santa Barbara called station Q and

topological quantum computers don't you don't have to have error correction but they rely on this very weird state of nature called an anon which is multiple particles kind of stuck together in in sort of a weird configuration and because of that they're very resistant to noise. So like you can perturb them, you can like shoot lasers at them and they'll just stay coherent. They won't like and you that that becomes good memory, right? Because it's it's it's robust to things. The problem with that is like no one knows how to make those particles like last longer than you know again microsconds before they like break up. Um and so Microsoft actually ironically has been the number one research entity for this for for a long

time. I think in like the 80s and 90s it was more popular in academia but then people gave up in academia. Um and this year Microsoft announced that they had made their first topological quantum computing chip a mayorana named after Mayorana who invented the anon which is this type of particle Italian physicist from the 60s 50s or 60s and um you know they haven't shown that it does anything yet though so uh so there's no gate being applied well they they've shown that they can run very low depth circuits like factoring 52, you know, it's like state of the art, right? Cuz state of the art.

Exactly. But that's my point is like it's just I just don't feel like we're on the growth curve where like I'm I'm really that worried like about, you know, breaking RSA or elliptic. Anyway, sorry that was a very long detour into this stuff. Two points on quantum computing, right? I think uh one is right if if you see a number um that right any of these articles give out like memor cubits it's always there's always a caveat right is it logic logical cubits meaning error corrected cubit that persists for a long longer longer time so you can do larger computation on it or is it really right like a more like a lousy cubit that that that's um you know not not really good for like longer computation and typically the number is right okay this

is not really a logical cubit I I don't think we really systems with logical cubits yet where you can really have um you know more logic gates being being applied right um and then and the second is like you know there's there's um the breaking cryptography right breaking breaking private keys um and like basically sending the world uh you know going haywire versus okay let's actually use them for some useful things right so there's postquantum cryptography which is hey let's actually upgrade Bitcoin upgrades our chains to be quantum resistant right that actually is should be that should come now but then the actual like quantum computer that can do that will come later right where um quantum computers can actually be useful before they actually break cryptography.

So like they could you know for example we could do quantum key change we could do um maybe useful types of like utilizing right the the noncoin principle to to build useful things before they're actually powerful enough to break break you know crypto uh as we know it. So like so I think like it's very interesting but not in the sense that you break things you you actually make things more more interesting. Yeah. So so the expectation is that first we'll upgrade some of these systems to something that's quantum resistant or quantum secure and then we'll start to have the things uh that really have quantum applications and at some precipice there there is uh a time at which it becomes investable. right now

it still seems too early um to invest in anything that that is quantum uh like a quantum you know kind of system. Um but it's not too early to start to upgrade some of these systems to be quantum secure. Um I guess one thing that I'm actually curious about that that I've seen a lot more about is you know kind of these these marketplaces that pop up around compute. This isn't just quantum but just a little bit more generalized around compute and the financialization of compute like you see you can you can trade you know resources in cloud you can trade GPU resources um how do you guys think that quantum will change like the compute marketplace

if we had a quantum computer tomorrow like assuming that or just are you asking that from the perspective of like if everything moved to like quantum resistant cryptography. What would change? Because like quantum resistant cryptography can be implemented maybe not tomorrow but like you know sooner than quite quantum computers. No no no like we know how to do proof of concepts. We've just never kind of made production ready. We have standards actually right? Yeah. Yeah. Exactly. It's just like no one has built all the you know rebuilt to the Bitcoin client around that or rebuilt a ETH client. You know what I mean? But it's engineering it's not science. Whereas the building a quantum computer is still science. It's not engineering. Yeah, if that makes sense. Like it's like it so they're two very different worlds. Like that's why

I'm kind of clarify like which one which one you want. Recent theoretical results on like you can actually verify that there's a computation that's actually quantum, right? This um I think a 20 2019 2018 results, right? That's like it used to be you had to have a quantum computer locally to verify that hey this remote thing like this entity trying to claim that it's doing quantum computation is actually doing a quantum computation where we have access to like two quantum computers right but like we know theoretical results that hey you know it's like $0 proof right like um you know zero proof says does the prover is claiming a true statement can I verify that right without knowing the underlying thing uh it's similar here right you can you can actually verify that a piece of computation is done in a quantum manner without actually having

access to quantum computer itself. Uh but you know, but the construction is very theoretical. Um and but so like because that's possible, we could potentially see a world where you can have actually quantum computers that's remote and and this is actually what some of these quantum computers are, you know, if you like I think Microsoft or IBM had this like service available for for the longest time where you could actually like instruct their quantum computer. But but the question is you know like do do you actually like are you actually convinced that the underlying mechanism is is actually quantum? this actually quantum computer under underlying it you know up to 20 30 cubits we can always simulate right so so there's no way of knowing u but but maybe we can get to a stage where right just remote quantum computers and you can kind of verify that hey this is

actually quanting computer and I can trust the results and then there's a question is is this actually going to be enough enough number of cubits for this to make sense yeah I think that's more more to the root of the qu like my question is that you know suppose there are there is a quantum computer made up you know, an arbitrary amount of cubits and it lives in like Microsoft headquarters somewhere. And for instance, like suppose that we could tap into the resources that that are made available through that quantum computer. Well, that basically like be like the deepseek moment on steroids where it brings down like it just crashes like the market for compute because now we can tap into this like supercomputer or do you guys see

that playing out differently? Well, that's go first. I uh I would say I don't think I see that happening to be honest because like realistically and and and maybe this is because I'm being smallrained about this but like there's not really that many kind of edge applications of quantum computers that I can think of where people are like yes we really need to use them for some type of speed up like there's a lot of theoretical stuff that's like oh you can do certain types of learning theory computations faster or whatever. But like realistically there's only two use cases of quantum computers that I think are like 100% you know if you had one tomorrow you would

spend time on them. One is simulations of physical systems. So like simulating you know like biological systems simulating physical systems simulating nuclear fusion stuff like that where like you can do it much faster than a classical computer and you can show that it actually does that significantly faster. And then second like the cryptography breaking type of stuff like there there's really no nothing else that like has like extreme value that tomorrow people Oh third one uh I think third one is really the non-cloning um right this is what we don't get if you think about why we have consensus in blockchain systems right it's like double spending but we'll have to make sure that I see you're you're a Daniel Kane quantum

money per uh person well not exactly that for I think it's you know that's one property of of quantum systems, right? That is not possible uh under a classical system unless we have right consensus. Yeah. Um, a fun fact is one of the an uh anthropic co-founders wrote a a paper on that a long time ago, but uh uh from this no from 2012 on quantum money on quantum money on like just like the a kind of equivalent of a blockchain, but you it's much cheaper than having to use consensus. I think like no one has proven that you can even reach the coherence level for that. I think that

physical systems you don't need to be perfect which is why I think it's like the you know in fact a lot of the existing quantum computers the things that have the most cited research papers on them are like using them to simulate other physical systems uh more than the cryptography stuff like you know there's like nature papers about like using the like 10 cubit IBM thing to simulate some weird physical system. So I all I'm saying is like I think like there's a lot of theoretical things you could do. No one has proven that any of them are actually faster. So I think it's very different than um the deepseek type of thing um where you know I think in the deepse case there's like something profound

that they found that is not just purely compute like it's clearly an algorithmic thing that underlies it that is the reason that you have so much less compute usage and I think in general uh I don't know if you uh you guys have looked at like the kind of like a lot of the new AI evals like I mean the fact that Gemini 2.5 is still like 10 times cheaper than all the open AI models and gets like better performance should already tell you that I think everyone is moving to this like lowcost world except for open AI and anthropic. There's like a two-part world. There's Gemini and Deep Seek and

then there's Open AI and Anthropic. And I guess we'll find out who wins in the long run. I mean the lowerc cost model should really be run client side, right? Um hopefully I mean I I think like we're still a little aways from like the average per you know you can definitely run deepseec on your home home computer but we're not at we're not we're not quite at running on your phone yet. I guess one one thing that I I think like is lost in that in the client side like world of running models locally is like is the why like and there's all this conversation especially in in like the crypto niche about verifiable AI

that I just don't think it just kind of gets lost in translation to the end consumer like how important do you think that really is to to people that are using LLM to write emails and generate pictures of Gibli that they run their their models client side. Why why would they someone do that? I I I think okay there's really like a couple different questions here, right? I think for this particular use case, no, you don't you don't really all you care about, right, is hey, like it's writing my emails correctly. It is, you know, recalling the right information. It is doing the capabilities I needed to do, right? And there's there's kind of a lower bound on like how much that needs to be the performance-wise for it to be useful uh for those use cases. And really, you know, I I think it's if you

really have sensitive like information like in business setting like okay, you know, this organization does not want it data to be to live in a cloud that's you know possibly um easier to be compromised, right? And I think you know for the personal data maybe not so much um but you know but for likeization like like let's say you know uh we all saw the signal gate right for the US government right for that for that type of use cases maybe you know you want that model to run on on a client side or like on prime setting and I think back to your original question I'm a little disturbed that we is suggesting that um a war in Yemen is is decided by a client side LM but I guess that's That's what they do with 1KX, man. So

were the tariffs. Tariffs were just AI generated, too. I mean, certainly the formula. Yeah. I I do think like there's, you know, different use cases of AI. For example, I mean, if you look at um I I don't know um for for example, loan underwriting, right? That's currently done uh somewhat manually with automatic uh processes. Um if you think that's going to move to AI, then the question of fairness uh really it, you know, and the question of like, hey, is this really done correctly? or just some human really like filling with the numbers. uh then then there's really right value in making that verifiable that it's actually coming from a model that's certified with certain certain properties right so it's not I don't think it's really for like personal like use case of generating image then you need that to be verifiable but it's for

these mission critical uh systems that over time will get replaced by AI maybe not state-of-the-art so that there's not that much computation required right but the output is really really significant and um you know consequential right it decides like humans like for example like uh admission process to colleges for example right and and and that you make it verifiable and with with with properties you can certify that's um you know just much better right than maybe a committee of human uh like designing on these things yeah have you guys used uh notebook LM yeah although it kind of feels like it got like dated really quickly because like everything else suddenly got better

very right like right after they released it I feel like a lot of the other voice AI stuff like 11 Labs got like 20 times better. And so I I feel like what's 11 Labs? They're like similar to Yeah. or like they're like a voice generation AI company. So you can you can use their APIs to make podcasts in the same way out of text. Um, I just basically think the the the reasoning models are just like in a different category at this point like than everything else.

And I think yeah, notebook LM is kind of like the on the last cusp of like the 2024 stuff. The 2025 stuff is like totally, you know, uh, totally different. I mean I yeah I I I would say like most of my time nowadays in terms of like research or thinking about things has just been like in that world. I have very little in crypto cryptography crypto because like I just feel like there's not that much happening right now. Yeah. So this next intersection that we you know because we kind of had that first that first wave of like as as we was describing it was like agent with a meme coin attached and people are

excited. I is is there something to that that you're seeing will that will be uh more prolific in terms of its staying power that actually will be very helpful for users or or developers in crypto or are you just fully focused on the AI side? Uh, no, no, no. I think there is basically I think all of those agents were like kind of too early. Um, I think the, you know, the LLM of the last few years were very good at like giving you an answer, but they were not very good at like correcting themselves and like being more stable. And the current set of reasoning models are like, you know, the reason they're getting superhuman

performance on all these human tasks like math contests and research stuff and like I mean I use them in research like all the time now um is they are able to correct themselves in a way that like you know you not only are they giving you a rationale but they're like playing a game against themselves to give you the answer. So the way the way to think about how these things work um is if you think about the models that like beat humans at Go or uh Starcraft or uh Diplomacy, what they did is like you have in in those models you have like a game, you have a game board and you have

the model play against itself many rounds until it gives you know figures out what its strategy is, right? Um what's happening in reasoning models and this is like a simplification but reasonable one is you give a query and you can kind of think of the the the like policy models as generating a game right there's a game with a set of rewards that is constructed from your query and then the model plays against itself. So it it's like trying to collect the rewards in the game um against itself uh until it gets to an answer and it if you design the rewards correctly it like should kind of take you to to

hopefully somewhere better. Um and this idea of like combining this gameplay aspect with like a single LLM is like why they're getting this performance. Um, I feel like I was uh I I spent like February being in like real doomer land. I was like, "Oh man, like everything is I'm g there's no point in writing anything anymore." Like everything's, you know, going to get replaced. And then I just spent the last like month like working on a paper on on reasoning model stuff which was sort of inspired by crypto things. And I after spending a lot more time, I would say the places that'll be better is a These

models are much better at planning um in the sense that they can like plan long-term tasks. They're not just like it's not like AXBT where I'm like, "Hey, uh you know, what do I think the price of TIA is in the next month?" And it gives me some oneshot answer. Uh I can now say like hey can you construct a strategy based on that uh your your belief on the tea price that you can that will like automatically update me on like when I should be buying or selling and can you plan how my portfolios are like you can do a lot more long-term planning things rather than single one action queries and I think that is a huge step

And so I think like all the people who worked on a agents were just like too early. They like built everything around these old models that couldn't really do much, right? They could like analyze data and give you like a summary, which is kind of useless. Like that from the perspective of like a financial product, sure, that's good if you're like deciding whether to invest, but it's not really good of like it's not doing actions for you. It's not like giving you actual like long-term planning and like doing a job task for you. Whereas now I think they actually are at that point. They're just straight up doing large scale tasks for you. It seems like it's like behind. It's like cryptos the the intersection is lagging by like

three to six months on the probably more like a year. Probably more like a year. So I think there'll be another agent boom. I think it's just going to be like we need a model better than deepseek that is good enough basically and maybe deepseeek R2 will be that right like maybe the second deepseek model will be good enough for the crypto use cases. Do you guys think that MPT like the standardized MCP? MCP M MCP. Yeah. Um like this protocol is going to be like pretty instrumental in like in kind of like integrating these two worlds and and allowing actions to be taken by these agents and perhaps you explain MCP as well. Yeah. So go for it.

Yeah. I got I I can take a stab at it. I think I mean the most succinct way of approaching it it's like MCP is like HTTP for uh for agents right if you think about um right what OM is is doing locally like when you want to use tools currently right before MCP it's all like war garden you can do this in each AI provider but there's no interpability um you know if you make a plugin you have to make it for open AI or anthropic um but now these days right if if you make a MCP tool then it's accessible by any agent uh capable of using MCP I think right I think it's similar to HTT HTTP for for for uh for information and and MCTP is really designed around two use and um it's certainly plausible that like hey this this is going to transform

basically how um how like two marketplaces are built right if you if you build a two marketplace in crypto like for example with prices or with um you know access to like uh social data right for for for crypto projects you should probably think of how you can plug this into um you know be MCP compliant so that you can actually have this be used by by by other agents. Um I think but you know it's certainly not I think opens up the playing field but it's certainly not clear whether um like how how things are going to shake out on top of that you know similar to how you know we have the internet HTTPS um and then right how how how like browsers um you know kind of entire web two revolution happened on top of that right that required another like decade and a

half to sort out I think the same thing is going to play out on top of MCP I would say the one thing that is still unclear to me and may actually require, you know, protocol changes to MCP is the security standards are like really horrible. There's no isolation and like every day I wake up and find 500 exploits that people are talking about that they've performed.

Um, it sort of reminds me of like early Ethereum where every contract was getting hacked every 3 seconds in 2015 and like it kind of feels like it's not ready for crypto. Like it's not ready for onchain usage yet. Let's put it that way. But like could it be? I kind of suspect there will be a protocol that's like a fork of it that is like a an isolated secured version that like has key management separate so that like there's no way of like a context accidentally revealing a private key or revealing a a witness or some secret that it shouldn't and the current protocol has no isolation right it was like clearly written because it's like

hey you know I don't think people really kind of were assuming assuming that these things would be having that as an assumption, right? It's a very thin layer, but it's a little bit like SSL, right? Like you had to add the secure transport thing on top of HTTP. No, this is like cross cross-ite scripting. Not sure if you you know, this is a huge thing in in web two, right? It's like you current MTP is like your Python script that evolves like external code basically. Um, and you can do whatever and it can read whatever. Yeah, exactly. So, I guess my point is like I don't think it's there, but I do think it's a challenge for crypto people. Well, first off, I think it's kind of embarrassing.

Crypto people didn't come up with this because like isn't the point of blockchains like building communication protocols like this? But, but the saving grace is that this one's kind of shitty because it it literally leaks everything and like it's not secure in any way, shape or form. Um, is ZK the solution to that? Like, hey, I can prove that like, you know, this model has this output and is going to perform this action, but I don't give up my private key. I think theoretically, yes. From an engineering standpoint right now, though, I think like if I look at the size of the things people are trying to prove, it's like not it's like much bigger than an Ethereum block. And Ethereum blocks are

already slow at least in the current MC, you know, if you look at some of the current MCP servers. They're like they're trying to prove like I mean obviously very inefficiently because everyone's just like writing this like raw Python code and I'm proving properties about the text instead of the bite code, right? It's like it's in it's inherently inefficient, but the ZK stuff has someone has to build the adapters to make it more efficient. And I'm not sure like there's a cat-and- mouse game of like people are going to start building the AI stuff to be like way bigger really fast. So like if you're trying to optimize certain a certain subset for ZK, the AI stuff might have grown so much that like you kind of get beat. So, I'm a little I'm a little nervous, but I I think the more basic things are

like encumberment for keys, like having some type of like isolated area that can't be touched by the context and can only be almost like a virtualized TE of like there's an enclave that the models can't touch, but they can like do certain types of requests to. Um, and like inside that unit, you could it could be CK, it could be T, whatever. But like my point is that the protocol doesn't have any a way of isolating things so that you can generate those proofs right now. I I think it's you know it's architected to be that but there's basically certain things that's leaking right like the the the the attack that actually got circulated on

Twitter was actually basically it's not the actual well it's really the description of what this tool even is and that description is getting leaked to uh the model interpreting it which is then right like um right you can basically inject in the to description to to then like poison kind of what what the interpreting LM is doing um so I think I mean the But problem that I see there is it's actually like how do you you know have like two selection right um to be generalized like currently you know you just access like a HTTP uh URL and right like there's basically the same thing that goes on but like which website do you actually visit right like in in the right in the early days we had

like search right then you you know okay this is how I how I'm going to like sort out which website I'm going to go to now the question is if I have all these tools right? MCP servers like how am I going to know which one is relevant for me, right? And this type of search problem is drastically different for for MCP I think because currently like you need to we need to parse the descriptions, right? To even know like which one is relevant for me. Um and and and and but to parse that effectively I'm using M to parse it, right? And that's where the vulnerability is. Um but there's also a routing problem, right? where like I if I if another MCP client routes to another server I can't say hey I don't want to trust this other server because it's security properties are less than X

like there's a it it has too much composability if that makes sense like it has no way of me as a user being like actually I only want queries that touch these these types of language models which maybe are like the secured or you know whatever ones that I would want and these other types I don't want to touch right like the user doesn't even have a way of like expressing something as simple as that right now. Uh it so like it's very primitive in that sense like there's not there's no access control fundamentally, right? So you can't even like imagine a public or private key separation if there's not even basic access control. I guess that's that's more the point I'm trying to get at which is and I think that's like the the challenge but also kind of like the

opportunity for crypto team is building like a secure you know layer on top and then being able to like correctly do all this onchain stuff and and have much longer term applications like I I sort of think the future of telegram bots like whatever last cycle telegram bot is going to be this type of stuff but it needs to be like you know remember when the telegram bots first started and they were constantly getting hacked and then eventually they moved to NPC wallets and they eventually you know like they eventually you know like that's that's the type of transition we have to see here for I think their like adoption if that makes sense. I I think yeah I agree the room I think um another thing I like to add on top of that is you know if the

structure is going to be right we'll have personalized agents using MCP or A2A to communicate with like service agents uh like huge unlock here is actually payments microp payments u because right like currently what happens is like every single payment effectively there's a human deciding it right uh on the consumer side and that's it's a huge overhead right that's is why subscription I I think it's like the the default for the past decades. But once actions are delegated to agents in a way that like you know humans can trust there could be a lot more microp payment um ecosystem like just infrastructure there so that right my agent can make payments on on my behalf and if this is built directly into MCP right you know I

I can be ordering like my lunch and and just you know by by by asking a message right uh as if had a year doing it for me rather than like doing that myself on on door dash and that could have a lot of unlocks right and that that agent could then try to find a find a restaurant for you know exactly the the food that I like. I think I think that's going to be huge a lot and that requires right everything that you said right cultural information making sure that things are secure but as well as like payments and discoverability and those things are basically not really there um in this web two form of MCP but I think right it's it's something that crypto is uniquely positioned to hey like this is exactly the thing that we want to build on top of yeah actually um to to your

point I like a company I would want to invest in but have not seen anyone good at and It's not really a crypto company, but it it like could become one. Is basically there's there's no like search engine crawling for these MCP servers, right? Right now it's like I go to GitHub and I like look at the link that someone sent and it's like it feels very like 1997 like piracy, you know? It's like I have to go follow this link to like download the movie to whatever, you know what I mean? But like clearly there will be someone who aggregates and like is the Spotify that recommends to you which server to use. and that there's like clearly a thing for that. I just don't you know it's like it's probably like 1997. So it's like Spotify wouldn't have

survived if it started in 1997. Um however someone will sell a company like Mark Cuban that is a fake Spotify if we take the analogy to the 90s too far. Um, but my point is like I think there's the Yeah. to this discoverability point. I think in a world of like these agents at being able to like fake reviews and fake, you know, like be like my server is the best to try to route people to them. I actually think like yeah having to compete on chain for for like you have like provable recommendations where people like oh I know that this MCP server like generated this much

income it there is a lot of value in that and I think that's sort of like you know you were talking earlier about this like loan agreement credit score stuff I actually think that's better for this for how do you rate these servers these MPCP servers you need to like kind of know how do I know like how good the server is at doing what it claims it does Right. Like two hard to tell. One is the actual like underlying model and one is hey how does it actually perform? Uh like it's a reputation system versus kind of evaluation system. Right. Yeah. Yeah. Yeah. So have you guys um theorized bit more deeply then on the second and third order consequences of this of this world and then effectively how that aligns with the way that you're thinking about allocating into whether

it be like L1's L2s infra apps like like has this has this came into how much has this influenced your decision- making on okay there's going to be trends that are will be in effect as a result of this kind of agent to agent communication and like where do you see this trending for for the infraight of of the space innovating finance global institutional and onchain deep liquidity to drive capital efficiency now maximize yields on institutionalgrade liquid staking with me unlock trillion dollars of capital via Fitcoin

All of this secured by Mantle's treasury of billions in stables and ETH capital pools. This entire ecosystem is coordinated by a vast network of seasoned VCs that are aiming to create the next generation of banking and financial technology. This is Mantle. I guess I'll go first. I think you know in always goes first um when it's a new a capability, right? But I think a lot of times you will see the infra team also being the app. I would say right open AAI for example was the model itself was infra but trypbt was the the you know the breakout app. Um and I think we probably also see

this in crypto right I think uh if you look look at you know all the agent coins for of of you know last year right that those are really infra apps together. Um but then you open up putting more apps on top of your infra once you have a breakout app. Um I think you know another instead of doing this app versus infront uh what I like to do now is really you know what is like what is fundamental value add uh value created here right based um as a you know first order second order effect of that technology and you know what are the downstream changes um and I think info teams should think about apps apps should think about you know hey you know is this what is the actual like tech we're building on um but you know another another point I like I like to

add is um like another I I also spend a lot of time thinking of like what is exactly the crypto unlock right there's a lot of uh ways that a technology could just change the world but you know the key question is like where does crypto come in and I think if a team in crypto can answer that correctly and you know actually be early there's a lot a lot to be um to be done like I think you know if you look at agents right the way that web two companies talk about agents is very much like representing an intention of a human uh of a human behind it, right? Either you know personal agents work for for a company and often time there's friction to how actions are generated, right? Um there's always a

prompt. Hey, do you actually mean this? Do you want me to do this thing? Whereas when a crypto project, hey, a type of agents, it's all about fully autonomous agents that have its own independent decision-making, right? I think that type of instrumentation is is what crypto is uniquely good at. and that these web two kind of person are thinking about, right? Um, yeah, I guess I think of it a little bit differently in one way, um, which is I'm in general like there's been a little too much like copy trading in crypto where everyone's like, "Oh, this thing works in centralized land." And so like we're going to try to like do something very similar but like a little at toss in a token somewhere, maybe make a decentralized thing. But if you think

about the things that are actually really in the crypto ethos that have worked really well, there's things like Freysa.ai where there was a security contest, right? Where it's like, hey, if you can figure out how to convince the model to tell you it's private key, then you've basically you basically made a bug bounty, right? And like stuff like that is actually closer to the crypto ethos and like mechanisms than kind of like here's an agent that's going to buy every other agent memecoin that looks like it, right? which is like if you go look at all the portfolios of the virtuals and Eliza things, all they did was like buy each other's memecoin, but like that can only work if like the AI6Z and virtuals coins themselves go up.

When those crash, everything's dead, right? Like that's kind of exactly what happened. Um, and I guess like in the spirit of things that crypto is good at, first, let's start with what crypto is bad at, right? It's never going to be like the best frontier model is ever going to be a decentralized one. There's just no way. like compute and communication wise just I just not gonna happen. The second um is uh coordinating compute actually is quite expensive especially when there's extremely scarce resources and like there's this kind of like hoarding of uh compute resources. So it's like it's actually going to be very hard to like

have a pure decentralized uh network. So you should instead think about things where crypto has a competitive advantage. And one competitive advantage is generally uh speaking I think I think things in crypto that work well as applications are things where in the centralized or traditional economy you're forced to buy a bundle of things and you can't break it up into like sub things. And so like the simplest example of that is DeFi. So D5 protocols are almost like individual desks at a investment bank or like an asset manager, right? They're like individual like you know and there's no protocol that aggregates all of them, right? Like

we've found in the last five years there's no like one super app, right? There's no super DeFi app that's like does everything. And the reason for that is because there's the benefit of composability is you can unbundle things. you can separate them into components where they get really specialized and go to the component and then you can mix and match them together as you need and that gives you a cost saving because instead of having to buy the whole thing where you don't use half of it, you're only buying the subset you need right? So crypto is very good at optimizing cost when you can break something down into components that you can then price separately and then com you know build a bundle you want as opposed to take the bundle the market force you know the

company forces you to have and in AI I think the deepseek moment was the first moment people were like oh actually maybe there is a way of unbundling these AI models right there's a way of unbundling the pre-training art from the inference time compute from the u you know actions that it takes and I think that deepseek moment where costs became much more important than just like pure performance and I think like until that point it was just like everyone was in La La Land right like we're just going to have infinite VC subsidy because every VC fund was just like leverage long open AI Anthropic and like three other people right

um that game changed And this gets to your point about the compute, right? Which is like once you actually start competing on compute cost downwards, then you become able to do this unbundling like crypto type of thing where I can offer you the same component. I can offer you the reasoning part, maybe not the pre-training part for way cheaper on chain. And you know why? Why is that? Um, so actually, and maybe this is like I got too much in this this rabbit hole the last few months is like I spent a lot of time trying to understand why Deepseek was really able to do this so much more cheaply, right? Like everyone

talks about like, oh, they use less compute, whatever, right? Like inference time costs are cheaper than 01. You look at you look at all the the uh evals that people run, they get like top defile performance, but they're paying like a hundth of what the the Frontier models have, right? Like if you look at Ader. So like do you guys mind if I show share the screen of this? Yeah. Yeah, I see. We're looking at the uh LLM leaderboards here. Yes. For those who are listening on Spotify. Yeah. So basically, you know, you have the first one which is OpenAI 03 high which was released yesterday. So, you know, it gets the highest score, but it cost $69 to run the eval, right? So, each time you you give it the the set of queries, it cost $69. 03 alone without the GT4.1, it cost

$111. Gemini 2.5, 72%, $6. You already see this like order of magnitude cost. We go down to DeepSeek, 64% $13. Deepseek alone 56% $55% $1. So like you see these multiple order of magnitude differences in costs for like 10 or 20% degradation in quality and that type of thing is exactly this type of like unbundling trade-off that I'm saying you could make like right like are there people in the world willing to pay for a reasoning model at onetenth the cost but maybe they're willing to trade some accuracy and like what's the market price for

that and that is exactly where I think crypto stuff is actually useful is like figuring out. So I went down this deep rabbit hole of like how do you you know what is the secret or I mean obviously there's many secrets that many are not known obviously but what is one of the secrets at least theoretically that we understand about why deepseek was able to do this for so little and one of the things that's kind of interesting is there is this concept of reward optimization which is very much similar to like crypto incentive optimization and choosing kind kind of like how you think about inflation curves in crypto which is based on how what rewards you're giving

the model you will get you will change how long it takes to converge to an answer for you and like how you choose those incentives will choose how much compute you use is this like good dog bad dog like like tangly what's the what's the rewards like is it so right now the rewards are like synthetic rewards right so it's like the person who's training the model is like hey Okay, I I'll give you like a concrete what my best and and then you know I just wrote this paper that I posted like two days ago which kind of has has this example but um suppose I give you a task that is the following. Ride a bike 100 ft break U-turn and ride the bike back. So

there's one way and like suppose you're just like a kid who's never ridden a bike, right? There's one way to teach you. One is to give you like a sequence of rewards. Like I give you like one candy, one lollipop if you can balance on the bike. Then I give you another lollipop if you can pedal. I give you another lollipop for breaking. I give you another one for riding one foot. Another one for riding two feet. Another one for riding three feet. Right? So I can give you these like rewards and like you try to like optimize to do that. So that's what you can think of these reasoning models as doing, right? You're giving there's this like kind of and the rewards right now are just metaphorical.

They're just like it's like points. It's actually it's it's synthetic. It's like it's not tradable or anything. It's like but it's like points. They're just trying to maximize their points. Um now the problem is those rewards are sequential. Like I have to give you every single sequence of operations and tell you like, "Oh no, no. If you learn how to break too late, then you won't be able to do the task, right? Because like you'll have ridden too far and then you couldn't U-turn, right? So maybe I only give you like five lollipops if you make it one direction, but you don't U-turn because you didn't learn how to break and you fell, but I give you 20 if you do the whole task. So what Deepseek did was they act, but the thing is that's very expensive.

So you need a lot of human feedback. So that's like scale.ai, all these companies that are like doing tagging, that's like what they're doing is this type of stuff. And it's also computationally more expensive because you have to actually like follow the steps throughout each of those rewards. What DeepS realized was that actually if you have very low computational budget, you can't afford to do such rewards. You can't reward each. Instead, what you do is you just say, I'm going to give you a million points if you reach the end point and just search and here's a compute budget and just search as opposed to like me trying to give you the step-by-step instructions for how to do the task.

Does that make sense? Yes. Yeah. So, one one question about the points like these are synthetic points that represent kind of units of value, but there's some inherent notion of scarcity that the scarcity here is just like how you choose the rewards in a very nonlinear non-obvious way affects how much compute time you spend. So, it they implicitly they're kind of implicitly guiding how much compute you have to spend. So there that's where their real value is. They're sort of like a and and that's I I got the idea for trying to model this in crypto sense because they like resemble a lot of these crypto systems. Um and you can imagine a world where people are

competing to give you the like cheapest inference on chain by choosing these rewards by choosing these and like to your question about like where does where does the compute cost stuff come in I think that is where that's where something like crypto actually is like useful for for this type of stuff and and so just to close this loop it turns out that deepseek saved all of their rewards until the very and rewarded just in bulk at the end. And that turned out to be more efficient than sequentially rewarding along the way. Yeah. Yes.

So they that's so counterintuitive. It's very counterintuitive. And historically there have been tons of papers that have shown that you can't really like it's very you these outcome rewards like hey I just tell you to go to the end point usually doesn't work because it's like too noisy. It's like there's too many possible paths it could take and like it gets lost and it fell off the bike and it maybe it went and figured out how to what what's there's a thing called reward hacking where it like it finds it found the box of lollipops so it's just going to eat all the lollipops and not actually do the task blew right through the break and U-turn it just just kept riding the bike all the way to the ice cream shop or or it's like a kid who like found the lollipop box and is like

wait why am I going to do this ride learn to ride this bike I'm just going to eat the lollipops right like like that that's kind of the and and that's sort of this thing called reward hacking, but they figured out how to do it by like adding certain constraints to the reinforcement learning algorithm such that it like made it fast. And so like my point is there's still a lot of these algorithmic improvements that are on cost savings, right? Like I'd say the last five years of AI stuff has been there more compute, the more compute and like not caring about cost savings. But the thing that is closest to crypto is when there's cost savings because you can now start unbundling and having people compete on giving you a lower price

fundamentally. And I I think that is where I see stuff that could be very interesting. I just haven't seen anyone who's done it. I think there's one thing I saw recently um which is from Prime Intellect uh which I'm I'm not an investor in but I kind of had sort of vaguely been paying attention to them where they had been trying to replicate doing the the DeepC RL algorithm asynchronously. So that means like you can distribute it to multiple people. Uh and they're getting like very similar performance to the raw one which kind of says like most of machine learning stuff is very sequential. But this this deepseek thing that where I only give rewards at the

endpoint is actually very paralyzable. So you could imagine doing it in a decentralized manner. All of this being said, like this is where I think you're going to find the the like new network new networks is like this type of stuff of like how do I make it cheaper to give you these um things. Sorry, I know it's very long rant, but you know way anything to add there as we kind of come to a wrap here? Yeah, I think um right like this idea that drone talked about which is really just commoditizing um compute that is not frontier right like it's it doesn't require really a state-of-the-art cluster and right and

then you really open up like who can serve this model if it's open source um and I mean I think this type of paradigm like you know happens repetitively right and um but really like there's other ways that crypto could add value besides just commod of commoditizing this compute and then opening up the competition. Um, you know, like like I said before, right, we're two AI agents are very much um, you know, they're very much not autonomous right now and and really represent human human intentions, there's really ways to build something that's completely different in crypto uh, that's like fully autonomous, right?

Like if you think about currently a lot every everything happens because of human intention, right? when where you know if there's you know companies generating value right for shareholders like you know you know individual humans making making making decisions there could be for example like agents that's actually generating economical value uh and that's an idea I think it's it's very fascinating to think about like what if actually a piece of uh states and compute right when it's when it's cheap enough when it's um smart enough that it can actually generate value autonomously over a long period of time right what will be the uh effect of that if you know if we can just ramp up um that type of uh systems. I think that's something that I I don't think we will see in new web 2 but we'll see

incrementation of that in in um in crypto. I think the first ramification of this is like trading trading bots, right? Like can can these things actually work out over a longer period of time? Probably not. But you know but there's make other ways um um that you know other value could be generated on top of just you know uh like zero sum trading um and you know there's other properties for example like permissionlessness right uh and you know can you actually like currently we're seeing this commodification of open source models but so so before this right the the entire question is like can we actually get these models uh in a way that's right that's accessible by a banning one. But maybe at some point

this going to shift the other way. uh where so because model commoditized now but the the question is really about data right can can we actually have these more like cleaned up uh state-of-the-art data sets to also be commoditized and and that's a part like these when I use AI right I I often use you know is maybe a little bit different than terms um I use deep research a lot where you know it's about I bring a lot of data and and just sifting through it and trying to make making sense of it and a lot of times the data that you can access it's very very um limited a lot of Wikipedia sources um you know not enough kind of searching through and and and categorizing I think um so that's that's

another aspect um I'm not sure where crypto come in but it's you know it's it's a huge issue uh if this data sets become more walled off um you know maybe maybe crypto can can help there to like unlock access but yeah it's it's definitely a not an easy to answer question like where crypto still can add any value. Well, this uh I feel like we started off by talking about the most forwardlooking thing that is completely uninvestable because it's so far out, which is quantum, and then just slowly worked our way backwards at to like bring more so

the present into like what's happening here and now from quantum to AI and then ultimately to like what is what is happening in the world of compute and this transition from like traditional cloud to as like the basis of compute to like compute potentially being done in a decentralized way uh with proper incentives. Um so guys, thanks so much for joining. It's been an absolute pleasure as always and uh we'll see you next time on the rollup TV. Looking forward. See you. Cheers, guys.