The AI-Crypto Arms Race Heats Up: DeepSeek’s Rise, Virtuals Goes Multi-Chain, and More

welcome bankless Nation to the AI rollup where we say up to speed with the emerging Trends and developments in the AI crypto space I'm David hoffen here with my co-host ezz e Jazz man big week in AI this week big big week big week big week how how you doing my man good to see you uh China David China going the dude the CCP is coming for our AI models and I don't know what to do about it um but no okay so let's set the the stage for a second here right um because a completely unheard of

AI model has taken away the Limelight David forget about open AI forget about anthropics clae it's all about this model called Deep seek which is an open-source AI model developed by a team in China and it may surprise you to hear that it cost next to nothing compared to open Ai and it beats open I's Top Model by a decent chunk or at least matches its standard and the thing that's blowing people's mind is aside from the fact that it's you know from China USA's

number one rival um is the fact that it costs next to nothing and they pioneered two groundbreaking research um techniques which seemed to catch everyone else who spent billions and billions of dollars into this off guard did you catch any of this David yeah so I've actually been going down this Rabel hole pretty deeply over the last two days we did an episode with this uh one individual who uh made this article that apparently went around uh both you know Wall Street investors and then then he was watching in like Google analytics

where this article was getting read and it made its way to Silicon Valley like where invia headquarters are uh and and so people think that the market got tanked because of deep seek he thinks and many other people think it's actually because of his article that he he wrote which is not just deep seek but many other companies unbundling in from the periphery and then deep seek was kind of the thing that Lit the match if you will um as I I'll like distill everything that I know about deep seek and I can get you to check to check it

uh deep seek the model is very legit it does in a very legit way pass benchmarks and performance tests that is very competitive with open AI uh the cheapness of how it was trained the $6 million in training costs people are saying that that is much less legit uh that's where uh the cccp China and deep seek has the most incentive to lie and then open AI is saying like well there's extremely strong evidence that they

actually just used chat GPT to train the model so nonetheless there were some very real engineering breakthroughs that changed how llms work in order to produce an output that's much more efficient that uses much less resource cost to produce a very strong output that's that uh is just it's just cheaper to run and so the fact that um deep seek the model is use is charging 95% less than open Ai and for their API calls is just evidence that this is very very real um but uh they also just like rode

on the backs of uh chat BT in order to produce their open source nonetheless uh this thing is going to make waves in American AI models and now uh you just everyone at Facebook at meta and open AI is are now understanding the tricks and bells and whistles that was built into deep seek to incorporate that into their models and I think really people are realizing that this the arms race between United States and China is fully on uh and any Innovation that we make in America is

just going to be copied and improved upon by China but that's a positive feedback loop between development on in the in the west versus the East that's kind of how I understand it yeah I I really good summary David I would just add one more thing which I I don't know if you mentioned it but just in case you didn't they open sourced the whole they open sourced it they open sourced it which is just a crazy move that's like you know um creating a a potion that allows you to live forever and then just telling everyone what the recipe is instead of monetizing it you know which is did you see American Way um but yeah

did you see archival uh tweet I didn't where he says uh it's ironic that we got uh AI that cost $200 a month from a nonprofit and then we got the open source AI from a Chinese hedge fund yeah yeah that was hilarious we should pull that up um okay so I want to I want to touch on a few things that you mentioned David right so you mentioned that they made that it's legit and it made these like groundbreaking achievements what are these groundbreaking achievements well I want to kind of like touch upon like a few things here um and and kind

of like trying to St it for the audience right so the one pioneering thing that they figured out was um how to use data that they train their model with more effectively so typically pre deep seek what everyone thought was the case was you had this model so you kind of designed it to do like super cool things but then you need to train the model so you need to combine data with it you know all this like raw data like hey this is David Hoffman by the way and these are humans

and the sky is blue all this kind of stuff right you need to run it through the model in order to do that you need to combine it with a bunch of compute power right that's why Nvidia Rose to fame cuz it like produces the best chips which allows you to run a bunch of compute which allows you to train these super smart models and then suddenly you end up with magic like chat GPT um and so what that looked like was data model and a ton of compute and it was so expensive to do David it cost billions and billions of dollars to to train these things right um but then what

these guys the Deep seek team figured out was like okay well we don't have a lot of compute that's pretty expensive and you know there's a conspiracy theory saying that they actually did didn't it hasn't been proven yet but the point they we do know that China has less resources compute resources available they need to be they are compute constrained because of the chip band exact so even though they are getting able to get around the band in finicky ways nonetheless that is a constraint that they have yes exactly so they're constrained right and so they have to figure out a smarter way to overcome

this like setup where they need a lot of Compu to train a smart model so what did they do well they designed the model and then they thought hm I wonder if a smart way to train it is to feed it a bit of compute let it come out with an output from the data that it's the little bit of data that it's used and then get it to review the output and figure out why it's wrong or why it's correct or how it might get closer to the truth and then you know rerun that compute but just a little more smarter so think about it

right it's like um learning to ride a bike For the First Time Imagine hypothetically you jumped on a bike you know you sat on the seat but then you put your feet on the handlebars and you fell off immediately because you lost your balance and obviously putting your feet on the handlebar doesn't make the the bike move forward it's like getting off and being like hm let's not do that again yeah let's not do that again I wonder if I put my feet on these little pedy things beneath me maybe it'll Center my gravity and maybe I'll figure something out like maybe if I push on it it'll it'll make the bike go forward so

they pioneered this method David which basically meant that they had to the model had to take fewer steps less compute and data to Fig out what ited to it's kind of like a human you know when you make a mistake you don't just run straight back well some humans do they just run straight back at it but it forces you to think and be like hm okay interesting the way that I heard this is that uh llms and we all I think we all saw this in early models of chat gbt is like hallucinations where you would throw it a prompt and then it would spit you back out a very long answer and the

end of that answer wouldn't really like line up with the first part of the answer because llms are very bad at backtrack and correcting previous responses the metaphor that I heard is like it's kind of like when a child is just like having train of thought they're like they're three years old four years old and their logic isn't so great and they can't stop themselves and consider what theyve said earlier and how does it square with what they're saying now so they kind of just sound schizophrenic and they go off into the wild what and so what this has done

is it's broken up into chunks it's broken up its process into chunks reviewed each chunk and make sure the chunks fit before creating a new answer and so it'll before actually spitting out an answer it is able to make sure that the chunks fit with each other uh coherently and then it produces an answer but this this actually my understanding was that this takes more compute this is more inference time because each chunk has inference costs inference and then sumting all of them is also more inference and so this was

what people this boosted the valuation of Nvidia because because they were like wow the value of inference is just up only like so much inference me makes a better answer right so so there's some Nuance there with your chunk Theory David right so you're right that you need more inference and and for those listening who don't know what what inference is think of it as like making a call to the model like you know a website will make a call to the model to be like hey by the way this user asked this question can you give us the answer for it real quick okay cool thank you right when you are ask when you are

quering chat gbt you are making it do inference that is when when you type in a sentence like hey I want to make my mama's banana bread recipe it's it's it's making a call to to the to the GPT model anyway um inference is cheaper than training full stop right so um was it's making a lot training is cheap inference correct exactly so whilst it's making a lot more calls inference calls um it's getting smarter by using less compute effectively but the point you were making just now David is people

suddenly woke up and was like wait hang on a second if if we apply this inference model to like a bunch of other things right like hey like what if I could just like keep prompting chat GPT in my chat to give me a better recipe or something like maybe I'll end up with the ultimate recipe right and people are now like thinking of how they apply that model you know to a number of different things but I just want to round this up David the second breakthrough that they made which is super important is you know typically how you have like the GPT model right let's take open AI right they have the model when you query it or

when you inference it it queries the entire model so think of this model as the design it hits the entire brain David and every single neuron every single neuron is stimulated it's quered exactly now if we take the Deep seek model it gets a request and it only routes it to the section of the model brain that it needs to hit which is just so much call thise there's something something oft Mi experts mixture of experts right so think of this new model as l a mixture of experts and when you

query it it identifies all right I need to I'm going to ask these three Experts of my 27 experts and I'm and every like 24 of them are going to shut down and not not run inference and three of them are going to answer and that's going to produce 97% of the quality of the output by Saving 97% of the energy required yes yeah exactly and David to be honest we've just like splurged a bunch of stats and a story to folks can you pull up this chart that that I just sent you please so um it basically displays what

we're talking about here so so what are we looking at in this like you know on this website so artificial analysis. a by the way like just basically tracks all the top models whether it's closed Source or open source and ranks them against each other now if you look on the left we look at overall quality of these models you'll notice that 01 is there you know right at the stop right at the front you know it's got you know all the bells and whistles but right next to it is suddenly deep seek R1 out of nowhere and you know it's just a single point below it okay okay but what about the speed you know we're like come on this I

heard this thing isn't very quick well actually if you look at the speed tab it is the slowest but if you pay attention to what's right next to it there's a familiar Brown there it's 01 so it's actually not too indifferent from 01 when it comes to like reasoning and delivery and finally David I mean the point that we're making is this thing is like so much cheaper if you look at the price right we've got like so expensive yeah yeah I mean deep seek is is out of four and um you're looking at 01 which is like over six times as expensive or

or around six times as expensive which is just nuts so clearly like this just captures the Breakthrough here and maybe to get to why this rattled the markets so far it's like so hard on Monday is that the idea here is that through creative use of this like medium of experts and a few other efficiency gains uh this deep seek R1 model is measured be 45 times more efficient than its competitors it's United States based competitors and so what that means is like you're using 45 times less compute

resources in order to get the same response in about the same amount of time uh and as a result people are like well realizing like well this was a major in the tug-of warar between software and Hardware between the AI outputs software just got a big victory and it's diminishing the value of Hardware because we're realizing we can eek more out of our models uh and by doing things other than just throwing Brute Force resources at it uh there's like this ancient not not even ancient

I've just used this metaphor on Bank list probably like 20 times over the years to talk about scaling how how scaling works and I use this in the context of blockchains there's two different ways to get scale out of any any system whatsoever you can write more efficient software and you can just build stronger more performant hardware and I think maybe a metaphor people can relate to is like early Xbox 360 games versus late Xbox 360 games Grand Theft Auto San Andreas or maybe maybe it's Grand Theft Auto 4 came out on the Xbox

360 and so did Grand Theft Auto 5 same Hardware different games and you can just look on screen as to like how you can probably count the polygons on Grand Theft Auto San Andreas and then Grand Theft Auto 5 you might be able to mistake that for like a real photo and so this was just an example of these Grand Theft Auto teams writing better software to use the same amount of the hardware better to produce a better product and then you can also scale hardware and just now we're on Xbox one so the hardware is even better and so

this is this is how blockchains work right we can either scale a blockchain by writing better software and we can also scale blockchain by using better Hardware honestly the answer is always going to be both uh and this is what we're going to see out of uh llm models is we are going to write models that use compute resources more efficiently and with the creation of deep seek R1 it's 45 times more efficient than previous models and then also we're going to be able to use more Hardware as Hardware improves people have been talking about jeevan's Paradox in this like uh Nvidia

market crash Tech crash that's already recovering uh but jeevan's Paradox I think is something like useful to to understand jeevan's Paradox states that uh as technological advancements improve the efficiency of a resource its overall consumption can paradoxically increase rather than decrease this occurs because increase efficiency lowers cost and expands potential use cases driving greater demand AI is now a major economic driver and jeevan's Paradox suggests that as compute gets cheaper AI

adoption will spread out faster expanding demand rather than reducing it so in short as Nvidia and deep seek improve AI efficiency they don't reduce overall compute consumption they make AI more accessible uh leading to Greater Global demand for gpus energy and data infrastructure and so this is what so if I were to left curve thesis what you just said David cuz I think it's very important and I I want people to understand deep seek made models that are groundbreaking much cheaper to make

yes which means that there's this massive surplus of compute and everyone's panicking they're like oh my God we have all this compute what are we going to use it on maybe we've been totally over indexing on comp maybe Nvidia is super overvalued yes and so what you've just said is is hang on a second no there's this Paradox which kind of explains that um this compute will simply just get utilized kind of at the app it's induced demand it's induced Dem it's induced demand so so so this like groundbreaking discovery has basically meant that now more models are

going to exist which means more apps and more cool things are going to get built which will consume that same compute that is the Surplus So eventually we'll still have over demand for the same thing totally and I think listeners are probably looking at the crypto token prices over the last two weeks and they're like oh uhoh uh and and then maybe they're thinking all right like well how does deep seek and Nvidia and all of this like impact my bags like what about my virtuals tokens and so I kind of want to just attack that conversation head on uh deep seek R1 the

introduction of R1 is inherently fundamentally massively bullish for the consumer the AI sector is going to become more useful more efficient more uh just like we're going to be able to build better products faster because of something like deep seek R1 nonetheless Nvidia is still has still like sold off it started to pick back up here is the chart uh I'm looking at the chart right now it started to pick back up a little bit yesterday but it's still down uh 14% since um last Friday which is you know

hundreds of billions of dollars uh and so I kind of want to just talk about the stars that need to align before like our AI tokens go go up in price because AI tokens AI crypto AI agents it's a niche within a niche this is a single sector of crypto and crypto is a niche inside of broader Finance so we need broader we need macro to go well actual macro that's the Nvidia price that is the performance of the stock market that's

also Federal Reserve monetary policy and Trump actions Nvidia is macro right David Nvidia is macro at this point because of how how large it is and so we need we need we need macro we need the tradition stock market to like at least hold on rather than go down we need the Federal Reserve to signal that it's okay to invest in Risk on assets we need we need uh Cuts uh and then Trump can also kind of do that same thing by just spending a bunch of money that's actual macro and then internal to crypto we need AI agent tokens to be the meta and

I think people are going back and forth on whether or not AI agents and AI tokens are actually going to be the meta of 2025 I think if you asked people a month ago or two months ago they're like oh yeah this is going to totally cause a bull market I think now today people are are not so sure and so there's like a stacking of contingencies that we need in order to have an AI agent driven bull market uh and while nonetheless like this deep seek R1 is fundamentally bullish and I think EZ and I are going to say that this makes the AI agent meta

the AI agent part of crypto stronger because we now have another more efficient model to use these agents on nonetheless we still need like a series of contingencies in order to like be bullish about crypto AI so I kind of just wanted to to like lay that conversation out because I think we're both bullish on the fundamentals here but does that actually translate into AI agent tokens unsure yeah I I it's a worthy reminder for the audience David and for us actually that um crypto is still such a young industry it's still burgeoning and blossoming so it is very

dependent on macro the macro people by the way are looking at crypto and they've seen these ETFs and stuff they're seeing Larry talking about tokenizing everything and only now you know it's like oh maybe we'll add a tiny little portion to a strategic reserve and and see what happens but we are still very much the Younglings here and and you know whatever happens to the macro Market at large if they're like a major war or whatever that might be it will have a knock-on effect to more risky assets right um and crypto will

always be the first one to hit um one thing I want to point out as well David is you know we we we said just now that Nvidia is macro and and what we meant by that is it is the biggest company in the world it's the most expensive you know market cap for a single company um and it's AI related so it has a direct correlation or onetoone kind of branding with guess what AI coins in crypto so if you look at like alt caps within crypto you know you know you've got like

Bitcoin coming down a bit it's still above 100 okay which by the way no one's talking about which might be the most bearish sentiment on on Twitter ever um but you know other alts are kind of shaken out but not as much as like crypto AI coins right and I think that there we have kind of like a onetoone kind of relation between Nvidia leading the market on the traditional stock index and then like cryptocoins having a direct branding thing here and the point we're making here is the crypto markets are still very narrative driven fundamentals often like lag a bunch of these things but in a bull cycle narrative often drives like a ton of

these different things now that is good or bad right like if you're focused on the fundamentals and you know you have a thesis that you know crypto AI is going to be pretty huge open source AI um unlocked by this deep seek Innovation is going to be absolutely killer from now going forwards then you should just be chilling right and maybe you would buy the dip again not Financial advice but maybe you would have more conviction in your positions and and see these prices as kind of like a bargain right but it's important to pay attention and understand how this game works right and understand that the price will not

always reflect fundamentals and you got to figure out what that means for your investing strategy with over $1.5 billion in tvl the meth protocol is home to me the fourth largest eth liquid staking token offering one of the highest aprs among the top 10 lsts and now CME takes things even further this restake version captures multiple yields across carak Igan lay symbiotic and many more making CME the most efficient and most composable LRT Solution on the market metamorphosis season 1 dropped 7.7 million dollars in cook rewards to

me eth holders season 2 is currently ongoing allowing users to earn staking reaking and AVS yields plus rewards in cook me eth protocol's governance token and more don't miss out on the opportunity to stake restake and shape the future of me eth protocol with cook participate today at me e.m. XYZ what if the future of web3 gaming wasn't just a fantasy but something you could explore today Ronin the blockchain already trusted by millions of players and creators is opening its doors to a new era of innovation starting February 12th for players and investors Ronin is a

home to a thriving ecosystem of games nfts and live projects like axi and pixels with its permissionless expansion the platform is about to unleash new opportunities in gaming dii AI agents and more sign up for the Ronin wallet now to join 17 million others exploring the ecosystem and for developers Ronin needs your platform to build grow and scale with fast transactions low fees and proven infrastructure it's optimize for creativity at scale start building on the test net today and prepare to launch your ideas whether it's games mem coins or an entirely new web3 experience ronin's millions of active users in

wallets means tapping into a thriving ecosystem of 3 million monthly active addresses ready to explore your Creations sign up for Ronin wallet at wallet. Ronin chain.com and explore the possibilities whether you player investor or Builder the future of web 3 starts on Ronin are you ready to swap smarter Unis swap apps are simple secure and seamless tools that crypto users trust the Unis swap protocol has processed more than $2.5 trillion in alltime swap volume proving it's the goto liquidity hub for swaps with support for growing numbers of chains

including ethereum main net bass arbitrum polygon ZK sync Unis swap apps are built for a multi-chain world Unis swap syns your transactions across its web interface mobile apps and Chrome browser extensions so you're never tied to one device and with self- custody for your funds and me protection Unis swap keeps your crypto secure while you swap anywhere anytime connect your wallet and swap smarter today with the Unis swap web app or download the Unis swap wallet available now in iOS and Android and chrome Unis swap the simple secure way to swap in a multi-chain world let's get into some of the actual fundamentals

that I think picked up this last week because you know usually it's me e Jazz and maybe a hundred other people on crypto Twitter AI agent crypto Twitter talking about AI agents but this last week different people were talking about AI agents can you want to run us through this yeah okay so we're putting a metaphorical Bull hats on now David we're done with the news it's time to like get into like why this is so exciting and why know deep seek and everyone is just pushing this entire AI space forwards okay David not just in a single day within a two hour window two

of the largest AI setups not crypto AI setups AI setups open Ai and perplexity announced their agent products what do you have to say about that so we had open AI demoing or teasing their agent product D this is something that we've been speaking about for months now and we've been kind of teasing with the IDE like open AI is going to launch something blah blah blah um open AI teased their agent and guess what it did David um aside from you know showcasing Sam Alman and and our future overlords

um it showed this agent doing a bunch of shopping it showed this agent ordering pizza by the way you saw it here first on bankless an agent ordering pizza so I'm just saying open source kind of like we've already done that Sam listen Sam just chill we're ahead of you it's fine you can take a leaf out of our book it's it's it's cool but anyway um and the third point which we'll get into later which their agents are using computer use David now remember we've always spoken about apis apis apis and it's important but just an interesting little thing there but let's move on to

perplexity right which has a very slick kind of like apple like demo here which basically demonstrates the same thing so we have two of the biggest companies announcing agents at a time where deep seek is coming out with open source and all these new products are being pushed forwards techniques products whatever it might be just an insanely like opening week uh for us let's uh watch the introduction of the open AI agent just the first 20 seconds of Sam Altman uh as he talks about this good morning we've

got something exciting for you today we're going to launch our first agent AI agents are AI systems that can do work for you independently you give them a task and they go off and do it uh we think this is going to be a big Trend in Ai and really impact the work people can do how productive they can be how creative they can be what they can accomplish okay so my understanding of this e jazz is that whatever AI open AI builds with this agent thing we just get to have that technology in the crypto AI agent sector right like that whatever capacity that they're imbu inside of open AI do

we get that or is it still closed off to us so open AI is still close Source just just just to make things abundantly clear right um but what it does do for us in the short term David is it convinces us that we're either already on the right path or it corrects us to make sure that we're currently on the right path and they've already given us a bunch of information just in this teas um alone right I mentioned computer use I mentioned some of the things these agents can start doing there are already crypto platforms that are doing this and

we'll talk about it later but it's just cool to know that we are at least directionally correct but these guys are still close Source I have a feeling they're going to do something like they've done already which is like you know here's an API you can query it it's still centered you can't open source it blah blah blah but I bet you that meta is probably going to follow up with a model that can do something similar and release their agent product David what if it becomes open source what are you about to say you look excited isn't China just going to do it like I feel like we're going to build all the proprietary like high touch white glove

llms that cost $200 a month and then China is going to be like well thanks thanks for that we're going to spend $6 million more doll by riding on the back of all of your guys' hard work and then we're just going to open source all of your labor yeah well I'm curious like how far this strategy is going to get taken right like meta did it within the us against open AI so that they can kind of Vamp attack them and like you know bring them ahead of of everyone else and you know there's been this like um thought that eventually meta will eventually close Source it you know CU

they're a private company they have shareholders Etc and you know we'll go back to normal but like whilst this we're in this Golden Age like there's a lot to learn from here right now to your point on deep seek right um I think they will end up releasing an agent model and typically and it's not something we mentioned they open source this model which typically means that they've got a better model under closed like wraps that they have been already been working on for a while and I don't know if you saw some U Sam basically mentions that you know R1 is very impressive deep seek is super impressive and he acknowledged

that they have made advancements that they didn't quite figure out and and kind of this has been the take from like a number of open AI employees as well as like you know um I'm pretty sure Jensen um J hang of Nvidia also said something similar being like wow this is groundbreaking blah blah blah so it's put people on their kind of like alert um um kind of scenario right now where they're like okay can we be doing something better here and what does it mean for like the the kind of like wider sense of things part of this is just heavily ripping off the ripping the

realization off that no we are in a deep deep arms race about Ai and the strategy of announcing that it only cost $6 million to train the R1 model which is not true that's like not that's probably the one part of this whole thing that's probably not real uh was an intentional strategy to knock down the valuations of us-based AI companies conspiracy that the hedge fund took a short position as well yes stuff like that and then also like you know China will China has for

decades heavily heavily subsidized their tech industry in order to commoditize some of the value created inside of United States copy it uh and then you know produce it at scale and then uh just have that become highly competitive with United States models so it makes a ton of sense that they are in heavily intented to take the value of United States AI companies and open source it as much as possible because that slows down investment they are trying to slow down investment in United States AI companies and so anything that I think

comes out of China in my mind I'm kind of uh viewing that perceiving that as both trying to bolster their own AI tech industry but also be an attack on American Market values of AI tech companies yeah yeah for sure I mean I think it's a I I I would personally brand it as a splash of very very cold water on Americans America's AI Darlings right they need to figure out you know are they too bloated are they directionally correct with their research and are they building the right