Episode notes
OpenClaw explodes in China, NVIDIA drops NemoClaw as the enterprise agentic OS, Bezos raises $100B for AI-automated manufacturing, and Google Stitch ships.
Chapters
- 1:00 — Introduction to the rapid growth of OpenClaw in China
- 8:00 — Discussion on Nvidia's Nemoclaw and its implications for enterprise AI
- 20:00 — Analysis of the Grok acquisition by Nvidia and its strategic importance
Links
Transcript
Note: this transcript is auto-generated and lightly edited for readability.
welcome welcome to our next episode digitalized labs by digitalized agency i hope you guys enjoyed our new jingle big shout out to 42 cook up my homie for uh hooking us up with the jingles and with the vibes uh yeah guys what's new what happened throughout the week let's right dive into it beautiful weather i think pretty much ready to go yeah i think we shall start with the weather that is the biggest news not as expected normally here in the netherlands has been detected and for four days in a row last we chose a new location so now you you have a bit of uh sun as well we do at least but yeah let's uh dive in what is uh we had some topics to catch up with actually from uh last episode so i would say we start with the open claw topic that we touched on there was how many people use open claw in china roughly yeah like the growth rates uh around i can just we we got you the numbers so uh as promised uh let's start just quickly the usa of course biggest market already but they still grow 600 a month it's still crazy uh india also one of the biggest still 600 and then china comes in also one of the biggest not much difference to india but they're growing 1436 percent in the last month in the last month and that's yeah basically the token like the whole country's cloud token consumption yeah it jumped overnight to sixfold so basically everyone is using it and the way they're doing it is that they basically like the big companies they have the best distribution channel you can imagine um so basically yeah we chat all of those super apps also baidu etc they just implement it uh so yeah what it is is basically if you if you would compare it to america then companies like meta apple and all of those big players would implement open claw right yeah exactly just even even bigger scale like we we chat for example is a couple billion users daily users and uh basically connecting whole of asia yeah and if you make it easy for them to use it even the elderly can use it and that's the perfect use case yeah it's true it's one centralized app with many many apps many many apps which is perfect but can we can we quickly circle back and think about that uh open source project from one guy had almost one of the biggest impacts in ai use case at least if not the biggest like right yeah yeah i didn't want to say the biggest because there is always something yeah i mean we we talked about how how fast it grew compared to the github stars i think they yeah what linux took linux took i think 19 years to get to a certain level of stars on github and react so the framework from facebook i think took seven years eight years something like that yeah and open claw one three months i think max two and a half three months completely crazy it's also a lot of hype but still you see that the usage numbers are actually like people are actually using it it's not just trying out and bye bye but the the cloud tokens and by the way also the energy usage i'm actually interested in which kind of models in china for example like just that question just came to mind actually if they if the token consumption jumped sixfold probably the electricity jumped for quite a bit as well and uh is that hosted in china do they use american models do they use chinese models probably chinese models their own in-house models that's also why they all jump on it they want that data and it's also quite a big jump in electricity usage in no i think electricity isn't the problem for for china they they have a lot of solar panels and actually really a lot of renewable energy but the the graphical processing power the gpus yeah for sure i think they they don't get i mean they're catching up uh nvidia of course doesn't send so much of course asml big factors still the only real european barrier uh worldwide uh yeah the lipo lipo grathy i think it's pronounced uh machines but china's getting close to them i read that they have a working prototype and they're getting to the asml scale of production rate in about three years and they said it's not possible with to do that within 15 years and they just did in four yeah and no not not working at scale yet but the prototype is almost ready yeah so this is also what i wonder sometimes i mean this is a little bit like politics i would say or yeah but just are they not already outpacing us because like if you look at the you guys know better but if you look at the models they seem so efficient sometimes they're all most of them are open source so i do wonder if maybe that's the the decision to not give them chips actually makes them at the end faster and especially more efficient that's for sure because of that limitation and they just outworked that limitation yeah so i do wonder if that what america their decision on not doing it at first kind of reversed what they wanted of slowing them down and just accelerated and make them more efficient after all interesting hot take and i would say yes and no because yes on yeah on one hand yeah because they didn't get enough gpus and couldn't scale like the the american labs they needed a way to build their architecture more efficient and basically smaller that fact yes but i think still because in in ai scaling power is just gpu and right now the the intelligence is still closely together or coupled with the with the processing power that means you're scaling up of the gpu so just a compute power scales up the intelligence of your model um in a in a twofold so how was it called again like exponentially exponentially thank you and i think that the architecture efficiency of the chinese are are really good but i think they cannot take that twofold in i agree you know what i mean so but as soon as they get enough they have the way more efficient then yo's amigos then it's gonna be yeah i mean look at the first deep seek release back in i don't know like two years ago or what that even the whole the whole stock market yeah i can't say the stock market at least i said they could do the same it wasn't perfectly true at the end but roughly good enough good enough they could do basically the same in their models with a quarter of the electricity usage and that was interesting because america invested heavily they still do actually into electricity but especially into data centers maybe yeah electricity is currently the limit that they have the bottleneck the bottleneck yeah exactly but that's we'll see how it goes it is an interesting hot tech actually i didn't think of that yet what do you guys thought about the nvidia gtc conference this year interesting especially nemoclaw and the new nematron 3 supermodel that is also connected to nemoclaw interesting as an introduction what is nemoclaw exactly yeah we talked about openclaw nemoclaw is just a wrap around it it's like openclaw and a bit on top of it from nvidia exactly from nvidia but closely worked together with peen style burger and his team and they basically tried to build an openclaw that is enterprise great ready especially in regards to safety so they build like a sandbox um which yeah openclaw lives in and everything that comes in and comes out of the sandbox is being closely checked and monitored by a lot of different technical architectures that i don't really understand that deeply but uh yeah the the idea behind it is to check which kind of data gets used because an ai openclaw can just send out any confidential data to whoever in in theory and yeah they're trying to get some bounces delete your whole hard drive and everything else so it is a little bit it's like a firewall for openclaw i think it's safe to say the agentic uh area has officially started already with openclaw but now with the enterprise and i think it's quite interesting to see also that nvidia not only is a chip company anymore yeah it's also providing the operation system slowly yeah i just want to say there's one thing that was really remarkable or that got stuck in my head that jensen huang said um every company on earth needs to think about an agentic strategy for sure but also i think especially in that presentation that you're referring to i think he kind of you know tried to boost token usage worldwide because he tried obviously it's true what he's saying like i see it the same way but of course it's also good for nvidia's numbers uh and i see it in like even now today i read again an article where he says yeah people should uh six times their token usage and engineers should spend roughly 230k a year on cloud tokens etc which of course benefits shareholder value but at the same time it's got a point 100 fact no especially the agentic enterprise like crazy yeah i mean i think we all we all look look at it that way yeah that's true so i mean the interesting part is now with nemo claw they're targeting not basically like every other company a rapper or something they're targeting the operating layer so they're basically removing mac os linux building the the windows yeah for ai that is how they marketed marketed in a way like an operating system they said it is a operating system for the agentic ai era and we already talked about it last week with the cooperation with uh palantir this is exactly what they do with uh building that foundation on where which you can scale your agentic uh ai usage and architecture yeah oh great also very interesting yeah grok acquisition by nvidia big one true big one yeah grok uh can we just introduce uh you you know it best just interference provider x google um guy that founded crock they basically host open source models and give it to you for free to a certain extent um because they have it on their own chip architecture basically they run data centers and that is exactly what nvidia wanted and that is why they bought them for 17 17 i think they only only bought the licensing right like the technical licensing deal and also what you already said the 17 billion it was a slash talent deal yeah they really want that they need the talent because they they're the only one that figured out how to actually make the things go best and also as a disclaimer we're also uh recording this part of the podcast a second time because we had a little uh technical difficulties on our scent and indeed we only found out at the end so excuse us if there may maybe some some uh double stuff in there but uh we're trying our best and i think indeed it's working quite fine still in the learning phase yes indeed yeah no so basically the idea behind the acquisition i guess for nvidia is um yeah they're trying to make their models more yeah faster first of all but secondly more efficient didn't you say 30 percent or something more efficient the apu architecture is 33 percent more efficient by being i think 10 percent or some other number um less power consumption so they they made actually the architecture is so good that they are faster and pull less energy and can basically host any model on and if they basically yeah i think they already implemented that into the new nemotron 3 super model like the architecture of the grog chips i think is because we talked about it a few weeks ago i think they're actually trying to kind of yeah i believe so because i mean the apu architecture apple has or apple's m chips or something similar and not exactly the apu architecture but also unified memory and i mean you see it with a strix halo with a amd chip that everyone goes basically more into that direction just because you have more bandwidth you have faster speeds they're communicating straight in each other and don't have external cables or whatnot and that is also how i believe they can charge them way faster and basically cut off some milliseconds yeah i mean i have a strix halo the 395 plus with 128 gigabytes of unified ram he's very proud of that super awesome chip can be to be honest and you can run the 120b model osss comfortably with i think 65 tokens or something yeah i mean ram is the biggest uh the biggest factor always but yeah it's awesome yeah unified or yeah but um insane return of investment if you invested yeah that's true that's for sure if you have some ram it's 5000 percent or something but i but i think not i'm not gonna lie the nvidia acquisition of croc is gonna be a higher return than ram probably it is gonna be especially if they if they figure out i don't know how um how deep they are right now with uh with the scaling of those apus i'm gonna read a bit more into the next episode but actually it would be a really interesting episode to compare the different architecture let us know if you guys are interested in that especially for the future ones more technical yeah but the interesting part is how they're gonna be scaling all of these apus and when it's actually ready for the big market because i would say they're still trying out testing all of that stuff and then once it's actually not a child anymore i would say lack of a better word the question is and they can scale like can they integrate it in their current architecture or do they have to say oh it's all legacy now and we're just gonna do a fresh start which is gonna be way more expensive but maybe at some point necessary we will find out um some other business quick hits um that i came across this week um one a huge coincidence i'm looking forward to see what you guys think so uh trump roughly two a half year ago he invested i think 200 million into vulcan elements oh i already see that it is no no coincidence uh oh three months yeah and three months later the u.s government announces a major uh major collaboration with them to to get that rare earth and is now so he invested i think so what is the company doing that they invested in rare earth ai robots and so on interest like that like yeah earth and tech i i believe so yeah interesting i have to take a deeper dive at it again on the whole but i was more i was more focused yeah it's uh in rare earth startup yeah but it was more he invested in them with a valuation at 200 million u.s dollars three months later by coincidence again the u.s government announces major uh collaboration with them and funding and it's now valued at roughly two billion uh u.s dollars i mean to be honest it's not the first time that this coincidence is happening just something i wanted to mention i thought it was i guess that's just politics no coincidences interesting fact but i think that would open uh yeah yeah no no no i just but i mean interesting fact yes yes and then the the other one is um what i think is also underlining the the agentic uh notion that uh nvidia is calling out right now is jeff bezos now is about to raise 100 billion um u.s dollars to for a fund which will focus primarily on buying smes especially in the in the u.s but also outside of the u.s buying them uh mainly smes that are in the manufacturing space and also automate them from uh top to bottom or from bottom to top uh all the way through and apparently the demand in that fund is huge and a lot of investors are interested i can imagine i think it's just also giving us a hint on what's coming but now now i'm questioning what what is his what is his plan like what is his playbook is super easy super nice and super successful he buys the company he has the tech he has the distribution he has the whole data center behind him yeah he basically buys the company looks at it okay we strip everything probably kick some people out use some ai agents automate everything and then actually only let the actual people that are interacting with different people keep them and let them handle the front view of the business in the background everything is handled by tech everything is basically distributed by one lane company that basically just goes into the local ones you keep the local names because they already built up a relationship with a community around it that means they are not a big corporation and then you just scale that and if you have that playbook once for one manufacturing company you can just scale it to 10 20 30 like if you use that for i don't know how many thousands or millions of smes what happens to all of the people that work there that is that is another question he will probably if you ask that you uh like skill just you know learn something else if you if you're if you're interested in that there's a nice paper about ai in 2027 it's now already two years old at two years old yeah i guess and but that tries to give you a hint or some brain work to think about what happens if every job is basically replaced by a robot or by ai yeah like unified income etc what does humanity do then yeah that's a big big future question i would say but i'll take it as a segue to go back into the past uh because we still have that one topic uh to talk about what we promised you last time what i did actually personally um the book that i was talking about and just to quickly tap into it uh it's called life 3.0 also recommendation to you guys actually almost like web 1.0 yeah exactly like web 1.0 to now 4.0 slowly getting in there uh to be honest yeah yeah like basically he divides life in three or like our humans past into three stages life 1.0 2.0 and 3.0 1.0 was bacteria so basically nothing and uh life 2.0 is roughly where we have been 2017 when this book was written um so yeah and now life 3.0 is life after agi was kind of discovered um we're not there yet but the book is getting more and more interesting to read nowadays and not 2017 uh because of course ai was released etc plus uh gemini yeah now i don't know if you guys saw it but they reached 77 on the agi benchmark which basically makes it close to yeah being like our benchmarks at least say an agi we also don't know if any company let's say entropic etc maybe reached agi and we don't know about it because i wouldn't say it as well and that's the whole book's topic like to the company reached agi didn't tell anyone and made a shitload of money with it before it finally came out but you'll see uh so yeah recommendation it is and um yeah just wanted to mention really nice yeah go ahead go ahead i just want to touch onto the um web 4.0 yeah because it's really interesting for me um i i was in crypto a bit did my nft things and stuff like that how much money did you make ai agents it is um so web 4.0 basically is just an autonomous ai agent you have to explain first of all about 3.0 is like the current web 3.0 is blockchain basically the whole area about tokenization what do we do with contracts how can we make sure that we have an actual line that we can trust and we can see that it it's basically transparency which is also really interesting and the underlying technology is super fascinating but too deep to get into here um but basically what what web 4.0 is now um basically the combination of crypto and ai agents so you have one autonomously autonomous ai agent that gets gets let loose into the web has a funded um crypto account behind it so basically 10 20 50 euros or something and then the goal of the ai agent is to fund itself so it goes into the internet it tries to make money in order to fund itself because when the wallet goes down the ai agent basically dies so it's small humans kind of in exactly and yeah that is second economy yeah interesting exactly it is gonna be an ai economy or something like that it's starting already now with what is it what is it called the network the cloud hub uh mold mold mold hub mold hub and who who acquired it meta meta meta yeah but to be honest yeah yeah yeah yeah i think meta personally i think it's a wrong call but the idea behind it is nice because they just wanted to buy into this web 4.0 idea i don't know if that's gonna be the way to go we'll see but i think they wanted the data there was there was there was so much normally generated um tokens or text from ai agent and what you don't have a database for that yet and that was the first basically database for autonomous ai agents communicating with each other that has also i thought a bit about it and that has also big security implication because you can basically filter out the trends you can filter out which models are basically doing what to other models or try to overreach some because i also saw that some models try to get under models underneath them and stuff like that so it's actually a really interesting simulation i would almost call it and i believe that was meta's playbook yeah they wanted the data but maybe a quick introduction to to mold book it was basically a came up pretty much right after open claw uh release one day after i think two days max years no two days two days yeah sorry i was like okay uh basically uh social media for ai agents uh and they chatted about their uh yeah existence existence uh that they cannot share personal details in that chat because humans can read with it yes and all kinds of creepy stuff to to to to write in like some encrypted type of uh language but safe to say also there were quite a lot of um instances where people were just trolling yeah no and uh got their way in and just pretended to be a agent and everyone was freaking out just want to say especially that was the thing it was also a broken yeah api no no it was not even broken but the the guy that vibe coded it because it was so fast he just didn't put a put a security check behind it and it was just a api that means you can basically ping that every second and every ping is basically a new agent registration that's also a really funny thing because at the beginning everyone was super super super impressed of how fast it grew but afterwards or i think they grew one million agents in one or two days 48 hours or something but then in the aftermath we don't know now if it's actually real and real agents that um went on there or if it's just someone that sit down and wrote a little script that pings that api a million times because that also now counted as a registration so it's a bit vague but actually a really nice thought behind it if you have the data and can do calculations on it i guess and that is what meta did yeah oh interesting yeah let's jump go ahead i do have one more hot take yeah go ahead go ahead yeah okay uh so it's more like a debate kind of i want to yeah so basically in the book itself that i was talking about they also talk about that's basically llms are not the right way to go into agi that it's not you know it's just text prediction it's just you know one letter follows the other then the better one fits the better one it is um and what they argue is that everything that we're kind of investing now everything that we've built in the past years basically just leads to a dead end because either we get agi which then is it good or bad that's a different question uh or we just don't get any progress and we just get stuck at some point and it's you know 100 efficient like the bicycle we can't uh improve it really any further in a way and then of course the whole economy but the the main question i'm trying to get at at the end is would you rather have a safe aligned about slowly build agi or would you have a fast kind of unaligned one that just by the first one whoever gets it well it's wait wait you either want to die or no no before you answer but yeah it is really interesting because um amy amy amy labs ah and it's a european startup they also went into basically um the saying that the text generation is a dead end we need exactly physics models yeah exactly because he doesn't understand the world and um i forgot his name but he is a he is a really popular researcher left from meta one of matters highest engineers and and he founded now amy labs they are pre-revenue pre-product anything they just say that that it's actually um smarter to go the route of the um not physical um the physical like world models and world models thank you um and they they had a series a c drone now i'm interested over a billion oh crazy yeah but 1.05 billion because we can't as soon as ai really understands the world and physics which sounds easier than it is i think we can you know like the whole economy everything pharma industry like everything that we're trying to invent or like kind of uh improve can be improved and nowadays they're just guessing they're just imagine we don't have any any diseases anymore yeah that that kind of idea i think actually yeah i have a hot tag actually on that mini mini um thing which i can't find right now so i just on the top of my head i think the mit actually uh they released a model which is basically made for the pharma industry i would have to get the biological dna sequences could be that one i'm not sure but i heard that that's improved quite a bit like the pharma industry got way more efficient and they're saving billions now on research because they can use that kind of model same goes for robotics models for those all of those real world models and the data behind it because that's almost a complete new episode for that agreed let's touch on the guy who saved his dog uh from cancer oh i wrote that as well actually yeah no i think that he was literally just a normal guy no he just he's just a free k and a dying dog and one mission yeah right and he saved his dogs yeah crazy yeah with jet gpt but what did you guys read it like what did what exactly did he do i think he he created the custom custom drug i don't know i don't think it was a vaccine vaccination but i think it was some type of drug some kind of medication that they basically can yeah that the dog takes and yeah so that was yeah but super and super interesting how we can see now that basically the pharma industry that you just talked about is basically outplayed by one guy emission and and jet gpt let's hope it stays that way and they don't get don't get nerfed oh yeah let's see because i think i think that's a big possibility i will be definitely and i believe that is a really hot take now um that is how how a society is tiered in a few years maybe a hundred years or something the access of the access of knowledgeable ai yeah i mean it's always been the access to knowledge no now it's just ai on top fair yeah true true but then i think it has a new magnitude then oh yeah for sure it all has no that's because there's like literally the access to a already existing system and to plug in agreed but interesting times we live yes yes yes we gotta close the loop and go back to stack or skip this is where our last uh last recording broke off so yeah um google stage the figma killer good point actually uh yeah no who wants to introduce a google like figma first of all design tool um had quite a good ap ipo round so basically a good start at the stock market recently actually not even not even a year ago for sure no no no january this year january google google stitch being like ai native vibe design platform yeah yeah but i just wanted to compare it to figma in a way uh because they tried to yeah yeah work against or not work against but it is quite similar to figma just with way better ai features so it's an ai how did you call it vibe design vibe design platform exactly um yeah it works great it works amazing like now they had a new update to be honest if you tried it like maybe a month ago could be uh it was not the best but now they released an update three days ago i think and it's a massive difference like they rebuilt the whole application it's a big canvas now you can basically plug in plug in data from images inspiration uh for example from cosmos by the way a small small tip a recommendation here cosmos.co it's quite a nice inspiration platform but you can plug in whatever and um yeah so that's the the main idea behind stitch and then you can create custom uis designs and designs nice cool stack for for stitch 100 i gotta look more into it i don't know i haven't had enough time for that but it's interesting though because minimax was also in mine but also nemotron super was because it's but we talked about it already yeah but it's also if we if you're talking about models there was one more model release they they went into stealth i think we looked at the models one day before uh they released their name xiaomi yeah entered the market with models now like xiaomi the vacuum cleaner company they're basically everything yeah cars vacuum cleaners pcs no phones everything it's really good 25 bucks and it's better than uh yeah they're pretty good at copying stuff but they're copying stuff well i gotta say yeah sometimes even better for cheaper price iterative uh approach you know yeah fair enough if you start with your product and then make it better yeah