AI Arms Race Heats Up: China’s New Agent, Grok’s Crypto Moves, and MCP

welcome bankless Nation to the AI rollup where we say up to speed with the emerging Trends and developments in the crypto AI space I'm David hoffen here with my co-host jaw jaw how's your week um how do I describe that I I think there is currently a massive dichotomy between like the advances being made in Ai and the markets pricing this stuff in like both markets crypto markets and TR markets I mean it's all One Market at this it's just all one it's it's Trump

it's Trump at the end of the day right but like in recent weeks David on these episodes we've always noted some sort of like New Frontier AI model being released like literally every week it's like oh open air has dropped this model claud's dropped this model right there's a new leap frogging every single week yeah and that's like an insane crazy rate of innovation but this week David it's all been about Frontier AI agent breakthrough specifically oh back to oh wait oh AI agents not necessarily crypto AI agents agents okay in the traditional

AI World um agent breakthroughs and David it's happening in this country um where you know most innovation has uh been Frontier for the last couple of decades and you're probably thinking it's America right but no I'm talking about China baby it's either one of those two either the Innovation is coming from America or it's coming from China it's never anything else so a Europe yep China's back um and they've dropped this agent called Manus DAV so if you pull up this demo um what

you're seeing on your screen is this AI agent that's performing essentially 50 tasks simultaneously at once right okay so for the listeners we are looking at a single desktop screen one of those like Ultra wide curved Ultra wide monitors and then there are like what it looks like iPhone screens or smartphone screens and that and there's just like it's just one by one by one by one and it looks like there's 20 on the top row and then there's three rows do you know what it reminds me of David do you remember those like uh Those computer

Farms or mobile Farms that we typically see in some I don't know some third world country and it's like just some dude that's just like scrolling or there's a robo scrolling literally aund actual devices but this seems like there are like a 100 virtual phones and they're all like kind of scrolling I think they're all scrolling through X maybe there's other apps there but there's a lot of Twitter going on and there's just a thousand schizophrenic interactions happening on X in these like virtual mobile devices there you go David Innovation um but

what what you're looking at here is an AI agent that can be kind of thought of as a combination of open ai's deep research agent so you know deep research is an a product they launched I think literally three weeks ago and it's able to do PhD level thesises uh in a matter of like 10 minutes which is just insane right it's a combination of that uh open ai's Agent product which is essentially um their agent that can like Leverage your computer or desktop and do a bunch of different things um and then there's

Claude um which is anthropics model which can like kind of navigate your own computer tabs and do a bunch of things it's as if all of them just had a baby essentially just they're just bundle together one big bundle of these capacities it's it's really smart it can control a device like a human can control a computer exactly and what what else and well actually if you open up this thread um that I I just sent the next one um it actually gives you a few examples of what it can do um a few examples that they do here is he he kind

of like goes through a progression of different tasks um one of the tasks that he does is like hey find me the top rental spots in San Francisco that is close to almost plugged into the AI Community you know so if you imagine if you're someone that wants to move to SF because you know you want to kind of like create a new tech startup or get funding from someone or be tapped into the AI Community you kind of want to be in the right spots at the right time this agent can do all the research for you it can suggest a bunch of different things it can create a website to display what those spots are it'll take

you through a booking flow or it can just do it all completely for you so it gets kind of like progressively more complex as he goes down into this thread and the bit that I found the most interesting is not only can It perform things like deep research for example like a nuanced ask like hey find these rental spots um but it can also kind of show you its Chain of Thought whilst it's thinking through this and I think what's really cool about that is if you look through its Chain of Thought and you're like hey you went wrong here or I think you should do something different here you can pretty much just amend it which I think is like pretty cool and

when you compare this agent to other state-of-the-art agents David Mana beats the compet competition flat which includes open ai's deep research agent it's pretty insane to see wait okay so I understand deep research in like chat gp4 and whatever all these models I understand them more as models not agents can we can we Define that difference here because you're saying that om Manis is this agent and it's doing things which is cool yeah um and I think we're bullish on specifically that sector because having a useful informative PhD level model is great to

access intelligence but what what I think really excites us and why we're doing these episodes just like okay can we apply that to this autonomous being that has goals so like maybe maybe you can CU it doesn't seem like it's just a Ono one comparable to open ai's deep research it has deep research knowledge but it also has it has it has agentic thoughts and goals and motivations right yep and and you actually hit on a really really important Point David um actually if you bring up this tweet by Philip Schmidt um he unpacks something which I

think is really important to get into which is David this isn't just an AI model that can do all of these different things at once it's actually a component of different AI things that will allow it to do it so you it is the model which is Claude Sonet so it doesn't have its own Frontier llm you know this team that created Manis didn't just create their own AI model they leverage Claude Sonet 3.7 and they so the model the model is Claude the engine of the agent is CL the brain is the Claude okay yeah you know

I'm not a car guy David but it's as if like some of these other sports cars you know when they plug in some other engine from another one of and then they've plugged it 29 other tools right which allows it to do a bunch of different things so one of those things is this browser use which is like an open- Source browser control situation that allows it to take over your desktop David and like eval the Brows use I'm going to call that a driver a like a software driver that allows agents to understand well okay no

a driver is actually a piece of software in your computer like I have a GPU I need to download a driver this is like the wheel browser agent is like the wheel connect the agent connects the AG into the browser and and and it connects that form factor yes K is punching the air right now yeah yeah so essentially like it is a model connected to an array of different tools which allows it to do a few different things if I would to summarize it it is evaluate what the user is asking for so it's thinking

brain right and that's like within the model itself but it's also like okay if I use this tool does this tool make sense to use or should I use slack or should I use email or should I use something else right and then it has its navigation system its wheel which is essentially this browser use kind of like controlling figure right and then it has its memory and data context and all that kind of like unsexy stuff so basically model plus 29 different tools gives you Manus and I think what's blowing people's minds here is well you have two two sides of of it right the

critics are kind of like this is just another GPT rapper this thing sucks and I actually think that argument is incredibly mid the reason why is if it's producing net new value for the person that's using it why on Earth would you just discredit it because it's not this perfect instantiation of an AI product that you thought it was oh it's not a new model it's not does it need to be a new model I would argue that all the tools are already there David and that's what we actually have really good models already yes

yes okay so can I try and like Place us into history as I think what's What's Happening Here do you remember when chat GPT what was the chat GPT 3 that came out and then that's that's what what like everyone understood chat was that's three and I remember using chbt and it would do these hallucinations you would you would ask it a question and it would it would every once in a while actually pretty frequently come up with a complete hallucination and you really had to be careful about understanding or believing what it would it would sh you kind of had to fact check chat TPT and

um I'm I'm now getting into the world of understanding like popping let's go with another car metaphor popping open the hood of AI models and like seeing how it works and so now I understand that models like Claude or Chachi BT now 4.5 uh there's pre-training post-training right and once you have a base model which is at the end of the pre- stating pre pre-training phase the base model all it is is a word predictor it just predicts the next word and then post training is where that turns into

like a useful assistant product that understands how to structure everything so that when you ask it a question because if you ask it a question like what is two plus what does What's 2 plus two equal and then it it's not going to say the answer is for it's actually just trying to predict the next word it's like you could you could that could be a part of a philosophical like essay right uh and so it needs post trainining to understand what the actual correct way to respond to that user input is and that's post training so we have this base model which should predicts next

word you have the post training which turns it into a ual assistant now it seems to be like this agent thing is another layer on post training that we're still kind of like working out the Kinks that are like similar to our hallucinations that we had back in chat GB chat tbt3 where now like it's still doing weird agent things I talked to hi about this on the weekly rollup where he like yeah like we have these agents and they're getting things correct like three out of four times which is a ter terrible success rate like 75% success rate is terrible you actually need as a

product 99% correct like correct rates uh and so right right now the agent race which like I think we are still in this like um model race but the models are are actually PhD level intelligence right now and so as far as like being useful models like they're now useful they are now viable products and so now the the race is Shifting over to agent Frameworks which is kind of just like an extension of post training as to how do we like work out the Kinks in the agentic side of things as we like

request it to build ourselves a a 7-Day vacation and buy all the flights and get all the hotels and plan everything and make sure that that actually works out and so we're still it's an extension to the post training phase that we're like smoothing over right now yeah I think that's a really good way to describe it David the way I kind of have it in my head is we created this like new magical power source but it's very generalizable right it's a power yeah we can kind of like throw it at random things right but it doesn't really it's like it's like getting discovering oil for the first time or

fire and then just like throwing it at a Clank of metal and expecting it to become a steam engine it it doesn't really happen we we're we're now like trying to connect the dot that would a combustion engine brother not that sorry yeah would be Engineers punching the air right now discredit discrediting a lot of professions right now but um yeah the the point being is I think we we've discovered fire entally and we're trying to figure out what to do with it is yeah and agents is just another way to kind of figure out how or what we

can do with this how we can mold or Manifest this energy into into something else right I think it's kind of similar to the same phase of development that robot Robotics are in where we have these like llm models and now we're figuring out like the physical Hardware side of things and the physical Hardware side is is kind of clunky but actually getting really good really fast yeah and and if we're being honest like this agent stuff has popped up or innovated pretty massively over the last couple of months David so can you imagine where

we're going to be at the end of the year I mean it's it's pretty insane and so like kind of like going on with this thread right um is this just located in China with all this agent stuff or is there other stuff going on and and the answer is like well open AI actually is rumored to be launching their own set of Agents David um and it's going to come in three different categories it's going to going come in a $2,000 a month agent a $10,000 a month agent and a $220,000 a month agent right it'll be split across

a range of these different price levels um the $20,000 a month agent David is supposed to act as a PhD level employee so it's as if you were hiring I don't know a PhD level um you know compi or mathematics graduate into your Quant fund or whatever that might be and this agent you know paying being paid significantly much less than an actual PhD level candidate human would be can now just like operate 24/7 at your business and kind of like run that particular role for you right the the

10K agent is then supposed to be a software engineer expert level at that and then at the $2,000 and maybe this is a little more insulting is meant to be it's actually termed a high paid knowledge worker AG which you high paid knowledge worker you can Outsource a high paid knowledge worker's job for the low low cost of $2,000 a month exactly which I thought was I thought was pretty hilarious and

of course this is all just hearsay right now we have to see what these agents look like um I found it interesting that they're going down the subscription model so very much like a hey you now have a new employee and look you're saving so much money if you just took our agent and replace your employees with that kind of vibe that I'm getting from there so that was one interesting update from the agent site of open AI but they also did something kind of less sexier but equally as cool this week David which was they released something that they calling their agent SDK or their agent software developer kit um

I'm not going to get into like you know the Deep complexities of this but tldr is it allows you to strap your AI model with any kind of tool and get these agents to interact with one another one another pretty easily and this is a trend that we're seeing with the likes of like how Manis that we just spoke about is constructed um and how mCP which is something else that we're going to talk about later on this episode is constructed we're basically seeing okay we have these llms we have these really powerful AI models what happens if we

give it a knife what happens if we give it scissors What happens if we give it um a mouse and control of your desktop what could it do let's give it let's give it some weapons you know let's give it some weapons basically let it let it do something and I this is something that I've been advocating quite a lot for obviously not the knives and scissor analogy but it's like I you get to choose yeah we need to give these things something to be able to act to be able to do something in the world versus just be glorified chat Bots right and the number one concern that people have given David is well these things are

just going to run rampid and I actually think we've got the opposite effect of that now where we're being too conservative one thing I like about open ai's Agent SDK is it allows you to operate within a Sandbox so they've only released a I don't know five to features right now and they do those features really really well right so you with slack um email you know just basic Integrations for software tooling that you and I maybe use every day or everyone kind of like interacts with their Gmail account or they go on slack or they go on their messenger or their

text or whatever that might be right and it allows the agents to basically kind of like you know organize this book this ping you about this set an event in your calendar really basic simple um you know events or actions which can't really call catastrophes or cause catastrophes as far as I'm concerned but you know it allows us that further level of experimentation David imagine a world where your day-to-day banking runs on a blockchain that's exactly what mantle is building powered by a $4 billion Treasury and poised to become the largest sustainable onchain Financial

Hub as part of their 2025 expansion mantle is introducing three new core Innovation pillars that bridge traditional Finance with decentralized technology first is their enhanced Index Fund aiming for $1 billion in AUM by q1 it provides optimized exposure to bitcoin e salana and USD complete with built-in yield opportunities next mantle banking promises to revolutionize Global value transfer through seamless blockchain powered banking services bridging crypto into your daily life finally mantle X Blends AI with defi to deliver an intelligent userfriendly

experience for everyone and the best part is that this is all in addition to their already launched products like mantle Network me and fbtc ready to step into the future of Finance follow mantle on X Atmore official and join the on train Revolution today in the wild west of defi stability and Innovation are everything which is why you should check out fra Finance the protocol revolutionizing stable coins Defi and Rolex the core of fra Finance is fra USD which is backed by black Rock's institutional bidle fund fra designed fra USD for best-in-class yields across

defi T bills and carry trade returns all inone just head to fra.com then stake it to earn some of the best yields in defi want even more Bridge your FRA USD over to the fraxel layer 2 for the same yield plus fraxel points and explore fractal's diverse layer 2 ecosystem with protocols like curve convex and more all rewarding early adopters frax isn't just a protocol it's a digital Nation powered by the fxs token and governed by its Global Community acquire fxs through fra.com or your go-to decks stake it and help shape fra nation's future ready to

join the Forefront of defi visit fra.com now to start earning with fra USD and staked fra USD and for bankless listeners you can use fra.com r/ bankless when bridging to fraxel for exclusive fral perks and boosted rewards C is transitioning from a mobile first evm compatible layer 1 blockchain to a high performance ethereum layer 2 built on op stack with igen da and one block finality all happening soon with a hard fork with over 600 million total transactions 12 million weekly transactions and 750,000 daily active

users cell's meteoric rise would place it among one of the top layer twos built for the real world and optimized for fast lowcost Global Payments as the home of the stable coins cell host 13 native stable coins across seven different Curren IES including native usdt on Opera minay and with over 4 million users in Africa alone in November stable coin volumes hit $6.8 billion made for seamless onchain FX trading plus users can pay gas with erc20 tokens like usdt and usdc and send crypto to phone numbers in seconds but why should you care about cell's transition to a layer

2 layer 2's unify ethereum l1's fragmented by becoming a layer 2 cell leads the way for other evm compatible layer ones to follow follow cell on X and witness the great cello happening where cell cut Cuts its inflation in half as it enters its layer 2 era and continuing its environmental leadership each we talking about um knowledge workers for $2,000 when talking about AI agents just becoming better at things that they do I'm going to read you a DM that I got I'm not going to share all the details because I'll let that just be a surprise when it actually happens I'm just going to read you a DM if if

blank name if if blank name of an agent was a digital representation of itself that could go on podcast and conversate with you about vision and motivations and its goals would you have it on Bank list and that was just a DM that I got from some guy who works at like works with a AI agent project in the space and I just immediately respond absolutely I would totally do that uh and so I think we might be ezz in the category of high paid knowledge

workers 100% so just let you know that like our jobs might be under threat right now yeah I'll be out of a job next week yeah um that's so I I would actually love to get one of these agents on the show one week David I would love this person claims that they can have there's like this AI agent that is out we've we've talked about it on this on the show before um that they they say that they can get there to be a live audio video feed of this agent who I can talk to and yeah you you're going to

bring you on uh and we'll just talk to this agent and see how that goes yeah I can battle for my position that would be awesome to the death oh my God I don't fancy my odds mate in about three months this thing is going to absolutely kill me um wow that's let's keep going with some of the developments in the we'll call it the Trad AI space okay we're kind of like churning through these right I said that there were this was the week for AI agents spe specifically in the traditional AI world and the one final rounding Point here is a set of X Deep

Mind so that's Google's like main AI arm Engineers left and is now creating what they're called or what they're calling autonomous coding agents and they've raised $130 million from the likes of like light speed Ventures Sequoia Reed Hoffman um at a half a billion dollar valuation and this basically tells me that like autonomous agents are going to replace software Engineers um pretty imminently and I know that this has been a rumor that's been floating around for the last couple of years and people don't like to really believe it because

when they see it in production they're like ah this is kind of like whatever but I'm just seeing like these proof points appear now every week David that suggest that it might actually be viable including the most recent one which was anthropic CEO Darius amode said at a recent conference he was speaking and he was saying within 3 to six months agents will be accounting for 90% of code being put out there which is a pretty bold claim but then he takes it even further

and he says in 12 months time it'll be 100% now of course like a lot of this must be hyperbole at some point but it's pretty insane to hear from one of the leading AI creators of this generation I mean anecdotally I the engineers of banket and other Engineers I talked to say that they like agents write most of their like lowl code not solidity because I God that would make me very nervous but like normal JavaScript and like other other like less high stakes code uh I I think that's like already

starting to be the case last week we talked about uh games that AI developers are just excuse me software developers are building with AI and we're looking at one right now Ultra real dog fight simulator 100% AI 0% human coding and it's this is a this a it's a polygon game there's like pretty low number of shapes here but it looks smooth it is incredible that this is all built by Ai and this thread has like 10 examples uh so like I'm I'm just kind of scrolling through and there's like 10 examples of

like games or things uh that have been built by AI I mean okay so I'm going to push back David and I'm going to say the graphics are truly [ __ ] on these ones it's kind of like the graphics aren't the point well look at this one this was this is like oh gosh that's playing in my ears it's okay it's okay but like I think you can create these really good graphics within a game already just using a set of different tools similar to like how Manis agent isn't actually just a new AI model so what you're watching here is someone creates a 3D model using using Claude Sonet which is

like you know it helps you kind of like code at a basic level right and so it creates this code and it creates This Magnificent CLE you're looking at it right now and and it looks like something out of like RuneScape V1 and then what um this Creator does is he kind of like takes this image or this video and he puts this in Runway Runway is like a image to video generator and Runway has gotten really really good over the last year and so what he does is he just like kind of like takes this video clip puts it in Runway and says hey can you make this like you know much

better and High Fidelity and it results in this like really crazy amazing detailed looking Castle right um there's another example of this happening as well David where you know someone um takes a similar route and says hey can you create me an F1 car and uh you know this F1 car is created and then again puts it in Runway and suddenly you end up with this extremely High Fidelity car so I'm imagining what happens when you combine this set of like tasks which can very be an agentic flow with a game that

you just demonstrated you could have something like be created in real time you create like a high fidelity game which typically takes years to make pretty soon yeah I mean maybe just not also not games basically what we're showing is there is this base geometry structure built by Claude and then we're putting it into this app Runway and then the runway adds a Cosmetic Skin onto that base geometry in order to make it seemingly hyper realistic I do not know how easy that is to like make a game out of but what it seems very obvious is for

movie production or video short production that seems very like the final product is what we're watching in front of us like now now there's a mountain on top of a hill and they're zooming over it kind of like some introduction be real into like some movie or something yeah maybe it's like GTA opening scenes that's see that we get I know now now we're beyond our our pay grade all right David I want to move on to uh another China update but this time it comes from Alibaba which is essentially like the amazon.com of China which

released their latest model quen 32b parameter model now for all2 billion 32 billion parameter model seems like a high number not a high number not a high number at all compared to like deep seeks R1 which is like 670 billion right um for all the hardware nerds out there 32 billion means that it's right on the cusp of being able to be hosted on your local device like you know your mobile phone um I think this could run on like pretty high-end laptops right now but even that in itself is like yeah pretty

insane so why is this cool well because it Rivals the some of the top models out there right now including as I mentioned deep seeks 671 billion parameter model which is just insane that's just 5% of the size of deeps R1 model and it's cheaper to run now bear in mind R1 was already cheaper than open ai's model right so we're already like you know factoring down by a massive massive rate a really quick rate um and basically you can have the power of this 700 billion parameter model in the palm of your hand

which means that it can run locally it can integrate with all your apps all your data and the reason why this is so cool is it becomes more personalized and more private for you to use David and the reason why this is important in my opinion is if we are to allow just like a bunch of monopolies to create these huge models and the only way you can access it is over the cloud that means they just own all of your data and that could be a very dangerous presumption uh or proceeding to kind of like go on forth with right so this really benefits

um open source development in particular because it allows all your data to be privatized and all your app interactions to be personalized right now America had a response to this and it came in actually this morning um of the you know this morning that we're recording this episode uh from Google with the release of their latest model Gemma 3 which is a new open 27 billion parameter model that isn't as good as deep seek R1 but is somewhere between that and its previous model V3 so it's still pretty good but

it's not quite there yet so the trend I'm noticing here David is um these models are getting smaller but way smarter and this is net net really really good for us okay so the way that I understand the 32 billion model from Alibaba again 32 billion model compared to to 671 billion models uh I think actually uh Anatoli put it well he like how did we get how did we get a 32 billion uh parameter model that's 20 times smaller more efficient than deep

seek R1 while also performing at par with it how did we get that and basically you use the big models to train the little models and so we make we make a very expensive very large PhD gigabrain model that's you know has has a bajillion parameters out there Infinity parameters uh and then we just use that expensive one to hey you're very smart can you train a smaller model that's very efficient uh and like it's kind of like a nesting doll I guess like use the big model to train the small model and in addition to that we're this

is I think this is like the fourth week in a row we've had some model release that is blowing other previous models out of the water that just came out like one to two months ago so the slack in the system I think is incredible the mechanisms of improving model power and efficiency are still like we're uncovering them week after week after week uh and so there's so much The Innovation curve we've never seen it be steeper before and now we're finally actually figuring out how to apply it into like products and use cases well I

mean I I I would say we're able to apply it to products and use cases because it's become so easy to innovate David uh remember less than two months ago everyone believed that you needed a huge expensive cluster of gpus and CPUs and data to make the best model and then since then we've had reinforcement learning from Deep seek appear and then we've had these agentic workflows be made and now we've had distillation which is the process you described of a much smarter model training a smaller

model to be able to do what it does right so all of that is going to lead to like what I think is going to be an application explosion for AI which is really cool and the Tweet you have pulled up right now David um true to web 3's uh um a bunch of uh decentralized compute networks which are basically networks that can provide compute and access to different models already integrated some of these top models that were released this week so you know that iteration cycle is getting quicker and quicker and quicker which is awesome to see what am I looking at right here uh the Tweet

reads Hunan AI just dropped image to video and open sourced you can generate 2K video with sound they added lip sync and AI motion capture it's crazy what's going on here yeah so you know to stick in line with the trend of China is shipping they're not just shipping agents they're not just shipping groundbreaking new generalized models they're also now shipping text to video models or image to video models so I referenced an americanmade company uh just now David called Runway um runway's been around for like two years now maybe