Why Decentralized AI Needs Cosmos: Greg Osuri of Akash Explains

Transcript

cash plays a very interesting role for for a company like Nvidia and they understand the value so they integrate F tuning as a service product to demonstrate that look you can actually use underutilize gpus and that's working out really well jgb own employees admitted they look at user logs so which is stupid which doesn't give me enough confidence that they're going to keep my data private in the future do we have a generalized model that is good at everything or do we have a model that's very good at utilizing specialized model there's a billion dollar question everybody's trying to answer I want to see human civilization reach type one status in my [Music] lifetime hey Greg hey SE how's it going

great uh how is your cosmoverse day two productive you just gave day1 was really productive D2 was very productive I didn't anticipate I came with very few expectations but my outcome the outcomes are exceeded expectations by by March yeah so yeah I think my expectations were met I mean there's been a lot of Cosmos fud yeah sure and my expectation was that the community would be kind of feeling that fud a little bit and kind of coping with it and dealing with it and uh and that's that's definitely the

case but I think it's positive uh in the end because what what we have here is there's really the people here are the people who are really building and like who are going to you know move Cosmos forward in a positive way and when I say Cosmos I mean like the infrastructure the ecosystem IBC the cosmos CK um and and also some the cosmos Hub so yeah I think it's it's positive it's like this is the bottom like the um AA has a t-shirt can it get any worse or can things get any worse yeah so uh no I like it I like it's only up from here I like how I mean it's not very small yet it's not very large it's

like you know I think thousand people or 800 people came yeah yeah so I like these small inter conferences more than the token 29 or to 2049 or whatever right like crypto Carnival massive like you don't get to do much I mean you I guess you do a lot but you don't get much done yeah is my feeling when when you go to the large conferences but this format I could meet people that I wanted uh run into people on the on on on the floor have amazing conversations and Cosmos is still very pure you know this I that we feel pure it's still very pure in the sense there's no mean coins and there's no none of this like let's go

launch a token and pump yeah there's still a conversation coming I hope not I mean it's still pure right I I have I have some insights and just some pretty sure it's coming but you know it's it's still very uh it's still very like how do we build things that work and and in all honesty I think like Cosmo stack is very mature right like yeah uh I mean not because we use Cosmo stack that that um bias but in actuality Akash launched four years ago and hasn't gone down a single time yeah we saw what happened with Tera I mean Tera work for all the misgivings ter worked phenomenally under under under stress yeah right the cosmos stack really works technology is is very mature and uh now we're seeing Cosmos

being adopted at least Cosmos design being adopted by pretty much everyone that you know has an rchain thesis right so I I was I was talking about a panel earlier one of the one of the takeaways was that and uh out of context this is going to sound really weird but Cosmos is the Isis of crypto and what we meant by that is that like Cosmos uh is like super resilient and like through thick and thin you know the there's this very like resilient ideology and uh once you're sort of in it like go and you don't stop right and and and and like conquer it's a cockr of crypto it's cuz could also be the Cockroach of crypto yeah that's another way to put it uh yeah in context that

made more sense of course but and then nuclear G nuclear so um wi yeah you guys have been up to a lot like I feel like I I was using AOS uh with this like really kind of clunky terrible client uh in 2021 because I wanted to host some websites and it was like I was I think I was texting you at that time or someone else on your team and man like this thing like just drops every couple of days like my website's not working anymore really and and uh this was like three years ago right I mean uh and and but but I I just saw the the potential for for this platform to you know be used for other use cases than just kind of

Hosting Docker containers uh which is still a cool use case um so yeah what's uh what's happened since then I mean give us sort of a well last three years quite a lot but particularly last two years we um we um we launch gpus yeah now fortunately coincided with Chad gbt launching yeah uh but we had gpus on the road map since 2018 right just took a long time to to get a decentralized GPU uh Network working in fact similar code we wrote for building aach was used by Nvidia to build a community driver which we we in turn used that code so our code went full

Circle back to us uh and that's what it took to really enable gpus on a KUB setting right so is in KU so depends on kuity on on the back back end yeah and um and so it's been it's been great it's you know gpus now account for the most of usage aash today we hit alltime high like me at a conference right yeah uh we slowly rolled out a new feature uh that enables user on boarding fairly quickly without having to need crypto so I think a lot of the uh in the in in the last one year we got a record growth I mean we grew uh quarter to qu I mean uh year-over-year quarterly growth was

about 1700% in Daily fees in in user fees uh that's a whole other magnitude right yeah in in a year and uh and uh and and we're able to you know attract really major players and Nvidia now uses aach I don't think Nvidia uses any other crypto protocol um so what are they using it for I was to ask you like who's using a c i mean because like we see the token yeah but I mean there's actually people using it oh yeah and people talk about on Twitter quite a lot too I know how much you see but uh Nvidia now is the biggest user for to a degree uh we have uh Nvidia has a product called brev which is a uh fine tuning of the service product so it lets you uh fine tune

fairly quickly I me set up all the necessary software and and give you a container with with the py Dodge instance you can just run your fering job uh and that runs uh and they use the cash for it okay it's very interesting because Nvidia you know I mean of course they all the gpus I they have a lot of gpus last to check but their gpus they have Hardware I mean they have the devices but they don't have the cloud capability they're not a cloud company they're a hardware company right they're a chip maker um so they but the customers uh the gpus because they sell these gpus and a lot of times the customers don't use the gpus right so um you know may they between trading runs

or maybe they have an older version of gpus that they upgraded to a newer version there are lots of underutilized gpus out there and Akash has a capability to to Federate them and offer them at a significantly lower price than what you would otherwise pay for a new GPU right sash is a use GPU market so Nvidia uh you know eBay or gpus it is EB it's gpus or eBay or gpus it's without a physical delivery right so Nvidia you know because of the shortage of Nvidia gpus because of the demand they're seeing other lgp providers be it AMD be it tpus be it even Intel to a certain degree get prominence because now

researchers that don't have too much money are the ones that are innovating to move to Nvidia or or to sorry to move to AMD so we see the rise of AMD 350 is coming up and that's threatening to Nvidia dominates right yeah so they want to keep their researchers on Nvidia chips the only way now is to to go in look into under utilization because they cannot produce more chips fast enough yeah so Akash plays a very interesting role for for a company like Nvidia and they understand the value so they integrated F tuning as a service product to demonstrate that look you can actually use under utilize GPU and that's working out really well uh so NV is a big user uh second big user uh uh one of the big users is um um Venice

AI is new company yeah uh they use quite a lot in fact their entire infrastructure runs their whole infrastructure run with theh I would say uh not the web infrastructure butru that's cool the GPU infrastructure I would say like 90% or so and they built their own model or they're using they use open source models but they also have a an amazing way that that ties these models together okay multimodal cap right so they might be using like I don't know llama llama 3.1 uh they use new research version of llama which I think is much better uh they use um a whole lot of model they use flux for for image generation they use a whole range of Open Source models and they give you a privacy optimized or

privacy preserving rather sensorship resistant chat interface it's beautiful it's I use Venice as my primary chat interfase more than chat GPD I switch away from chat GPD it's a pro version of w which gives you incredible super powers okay and and it's private because I know they they use a car so I have confidence that they're not like listening to me and they're not recording Vlogs you know so they um I use benit for my health data like I don't give Char any Health Data because it remembers yeah if you know but charity I mean their own employees admitted they look at user logs so which is stupid which does enough confidence that they're going to keep my data private um so I like I like when it's a lot uh and then we have a whole lot of

researchers a lot of universities uh because a lot of them are priced out UT is a big big user unx is Rochester in technology is a big user in fact they're they're developing uh they they training new AI to be a u uh uh what do you call office hours for the new class that's coming up I think RIT is the first first first University to do so so the new undergrad uh the computer science undergrads are going to have a AI tutor uh to complement the classes which I think is phenomenal and each of them are going to get uh Akash counts to to to run their training jobs on what what what does this kind of like custom tutor model have more than like gp01 or like some so they finetune the

model on the class curriculum so it can address questions very there's more it'll take an ex Mar and they'll find tun on some content I don't know exactly what model they use I mean they haven't disclos those detail to us uh or maybe they have but somewhere in the logs but but it's it's a great use because they're priced out of of the cloud University that you know they you know can have too many funds uh and the students don't necessarily have the money so it's aach great alternative to them uh and then uh we have a I mean so another big uses of aach right now is decent trust training so your companies like news research news big us of companies like private intellect that have cracked the nut on De training which I think is phenomenal so for

context training uh so far has been predominantly High latency sorry uh low latency high band with uh odl in the sense to do any meaningful training be it gbt 3 or four or in know Gro as a matter of fact required uh high density compute to be collocated and connected with the very low low nency so you need really interconnect like um you know U interconnect uh 3.1 or like these are very very fast connections that have chipto chip connections that don't go through the motherboard right so uh because the way Neal networks work yeah

uh so distributed training was not possible even though they there gpus spread everywhere you couldn't leverage them for doing training yeah you could use them for inference um maybe find Jing but training has been predominantly very very localized R clusters yeah now that changed so news research was the first company to do a billion parameter model fully decentralized and they invented a framework called drol uh I think d stands for decentralized training um on on the Internet or something like that what does that mean a billion parameter model like like compared to like a GPT or or some other You Know M that people are familiar with uh the non AI nerd yeah so

well Falcon GPT is can't remember I think GPT is a I think gpt2 is 100 million parameters yeah I think GB3 of the billion parameters if I remember don't quote me I don't remember not top of my head right so the parameters is the the the scope of the data on which how many tokens yeah how many tokens okay sorry yeah yeah how many tokens have been trained right yeah how many different combinations right so uh how much input was given so it building parameters is a big model it's it's not a small model right it's a very good model it's considered llm at a billion parameters right yeah I think less than a billion I think it's called slm or small language

models right but interesting thing is uh so they are the first company to do a bilding parameter model and then you have Prime intellect which is uh the first company to do a 10 billion parameter model 10 billion that's have big step it's like an order magnitude better order it's one order magnitude yeah better than a billion what is that like what practically speaking I mean again like I'm not an AI nerd so like these things are I'm sort of learning about these things will you step from a billion parameter model in GPT 4 which is already pretty good I mean like a lot of people use gp4 most most people like I know we use gp4 and you two like gpd1 is also very good now going to a 10 million parameter model

like how what will be the noticeable difference that people will oh quality of results accuracy accuracy yeah so you have less illustration you highly accurate like output it's better model yeah um so and also the Precision matters too so the higher the prec more accuracy you get in general mathematics right like the longer you have after the do you have higher quality right so that how deep you can go into the into into the search Precision so FP 16 or the FL Point 16 is highly precise yeah but it takes significantly larger amounts of uh uh data so that your weights are

significantly larger than let's say a a a eight Precision right or four Precision or even one in so if you want really fast response you want to go with like one in like you know uh but if you want really highly accurate response you want to go with highly uh like fb16 or something but if you're doing so depending on what the use case is if you're doing self-driving you probably want like you know low Precision because the reaction time has to be good so it's it's trained a less amount of data but only driving data yeah it's not it doesn't know the rest of the wall it doesn't need to right yeah so so so depending on how you're using them all find some weird vulnerability that like exploits that yeah blind spot right yeah

so I I hope that a car that's driving doesn't need to know anything else beside how to drive a car or something that's performing a very specific Tech that's why you have small models that are better than large models because they can work perform much better because they have less data and more focused data to be trained in particular so you can find a model that can only that right that's way better than gbd4 yeah cuz gbd4 is a general purpose model yeah it's so large and so big so the higher the Precision the the the slower the response time comes and and bigger the the weights are um so so today we have Prime intellect that's doing 10 billion parameters right um and then potentially you have 100 billion

parameters and the 10 billion parameters I think they're already like 15% complete and they only launched last week so people are yearning to give compute because they want to be part of this Evolution and Prime intellect is very exciting because again I'm not they just use a Kos they don't I'm not not not an investor or anything right um I just love the fact that they've cracked and not on decentralized training yeah so they're using aach to build this temp yeah both news as well as private use a um so so talking about demand where is demand coming from is one of the big demand drivers because it's training is happening quite a lot and so who's provid like where is the supply coming for Akash yeah so a big supplyer for Akash is Foundry so we have about 70 I

think 70 a little over 70 providers on aash and growing set of providers uh a big one is Foundry digital it's owned by dcg uh so Foundry uh or you know Foundry bought a bunch of a 100s 100s for crypto mining sometime back and they just showed up on day one when we laed with like a lot of GPU and I was like wow I mean I don't know whether we have demands for them but they're somehow stuck around and they they really like uh the whole like deepend you know Arena so they have a whole page called Foundry accelerate that's dedicated aash and dener so they they Supply quite a lot of aash uh you know uh notes from um we have oblas we have U valdi we have there a lot of like what happened last

year was uh so there was a huge GPU crutch right it's not secret people it was impossible to get views on Lambda I mean you need to get into the large contracts a lot of players uh everybody wanted gpus you know must going purchasing 100,000 gpus and you know like or h100 rather and like now we saw meta purchasing like 300,000 uh you know h100 so there were a lot of demand that created incredible Supply crunch for gpus so that led to a lot of opportunist like looking at potential ways to make money so a lot of these business people that know how to get financing are you know started getting these and how how

to do business deals started getting allocations for like small quantities about, 2,000 4,000 what not to supply to the AI companies right so and lot of them ended up you know creating these companies with extreme like saturation and uh fragmentation that and and the demand quite didn't catch up to the expectation so now we have an over supply of gpus in a weird way um and um but these GPS are not very usable because you need like large interconnected clusters to do meaningful training and not everybody's doing training it's only like large companies like open AIS and anthropics of the world are doing training so what happens

to all these gpus now so now there brings the price down mathematically it just brings the price down it does bring the price down if the supp available so so the price is very interesting so if you have large interconnected cluster the price is still up but you have a bunch of loose gpus uh the price down so it's very hard to commoditize H 100s because the dimensionality is so huge that you know differentiates pricing like you can't can't have nh100 uh so I mean you you have the interconnect you have the uh you know you need to have the networking you need to have a lot of other components that price in a 200 cluster not the chip alone the chips can

be like well commoditized I mean a chip is Chip but how do you commoditize the service right so it's the energy needs too so it's quite a lot of like you know commoditize on hourly basis you have energy cost right so in US you have 20 cents an hour I mean 15 cents an hour things the average uh for a kilowatt hour basically and each consumes about 1.2 kwatt hour for both for the chip as well as the cooling in generally speaking general terms so and then you have the chip cost alone right so uh you can say if one wants to advertize their chip within two years I think the price is around $2 an hour roughly including the energy cost right cuz na 100 is like what 40 40 Grand 32 32 and if you like

buy wume I they can get so if anything less than $2 an hour you're losing money I go we have chips for dollar an hour too the people that literally literally are like well I just need some money yeah so that tells you the market is not Market there's over supply of chips now okay interesting so there there was a really high demand or was I mean that like two years ago was that demand real or was just like people stockpiling gpus because they anticipated in high like an increase so the demand was real yeah but the stock piling was also real the hoarding was real too and a lot of times what happened was these companies got ahead

of themselves and oh it was impossible for aash in for us to go get these GBS because people were like oh we don't want to just give you a GB oh you're crypto we want to touch you yeah it's still coming up FTX right yeah yeah and now every everybody's like begging us to like take the gpus you know it it's kind of crazy like Lambda laps wouldn't talk to anybody they were so up their noses right and now they're struggling to like sell the GP I think they're selling the gpus for like $150 an hour or something and still they can't sell them right so so yeah so but we still need a network that can aggregate GPS because there a lot of GPU so they're taking whatever they can get uh fortunately we're seeing a an interest in h100s now because of

llama 3.1 or uh and it's a big b people want H 100s to host them um and but the only problem is are these h100s in like right location the right bandwidth uh so a lot a lot lot happening but I guess like the distributed trading is going to be very very promising so if you say the marginal cost of chips are going to come down especially with over saturation the cost of decent training is going to come down too right so which is amazing so llama 3.1 for example was train on 30,000 h100s so can we get 30,000 h100s for a dollar that's like $30,000 an hour it's not bad if you think about it right I mean sure it's a lot of money but yeah

if you're getting me a 30,000 plus h100 cluster for $30 an hour oh $30,000 an hour uh that can create a 400 billion parameter model how many hours do the trade cost I think two months or something okay yeah not bad so like the total cost is uh yeah so but do it yeah the ma do the math we'll do the math and we'll somewhere yeah that bad I think I think I did the math uh I think you can train if I remember correctly don't quote me on it I think you can train a good llama equivalent for about $50,000 sorry $50 million apologies $50 million uh like that's a that's not a

bad price if you think about it like you know like decentralized networks have the money but T loan is like 4 billion uh market cap right there's a there's a lot of money in in crypto that that if we all pull together we can trade the next you know a trillion perameter model yeah fairly well so let's let's go into this then I mean I want to ask you also about like we've been talking about the the decentralizing the marketplace for producing AI models and for inference now like what's the next step after that it's like building a model building a decentralized model that is not controlled by anyone company that uh doesn't fall into the hands of like some you know big corporation that uses our

data and that can use it against us perhaps I mean there's also these things that are at stake yeah which overlap with a lot of the uh concerns that a lot of folks in crypto have about you know the way technology is get uh is progressing so yeah talk a little bit about like what the vision so I so yeah yeahi is such transformational right I mean it in for the first time you have a technology that influences how we think because you don't question the result sometimes may but most of the time you don't and it's very dangerous because now we are entering in Era of digital feudalism right especially as we get towards AGI Whoever has AGI control controls the world because it's I mean we already seeing some variations of AI

in the crypto itself we saw what happened with goat the the meme coin uh so New Meme coin was launched last week uh by truth terminal uh I think Mark andreon serated about $50,000 it went and created a meme coin uh called goat and the goat is about $300 million market cap now uh and it's actually making money people are selling the stuff it's tweeting it's created old Community it writes all the wicked tweets um but that's a real impact on real Commerce and and and and in the real world I think the first time you see an example of crypto of AI driving a a a an asset price right right and

that's just the tip of the iceberg imagine and it's not complicated to build something like that right so it's fascinating to terminal ont follow it you'll you'll see the tweets actually pretty good it's very engaging tweets too so uh you're going to see things like that as we get towards AGI uh now the question I mean there's going to be whole prolification of mem points now ai ai M points and as sad as it sound but that's a truth the reality you're going to face right yeah uh so you can imagine what a a more advanced AI can be in terms of playing the stock market playing like you know doing all kinds of things that are Impossible by humans right so we are entering a a an era of digital feudalism the question is who

are these like the the Lords and who's going to be peasants I mean is it going to be like few people that are going to control this technology or is it going to be the people that use the technology the you know that they want to control the technology right so I form in the ladder and obviously we Building decentral Systems to to avoid something that that could deliver digital funeral fistic Society but the challenge right now is so so far the challenge was like how do you get enough computer to build now we have a problem solved and second is how are the technologies that can leverage this Compu to actually do training that is Sol now the third problem is what is the incentive for someone to fund the development of in to train AI like why would I give my money to this

whatever decent Network so what's incentive like if I get some degree of profit sharing yes that's a good incentive so now there there's a lot of work happening um and some companies claim to have solved this I'm not going to name them uh because I'm still doing my my my due diligence on these companies but uh there's a lot of work on fractionalizing ownership of a model like if I contribute compute to the model I should get a share of the future revenues from inference because inference is how you make money and training is how you how you invest right yeah it's like organizing real estate exactly any so is this is easy model but but the challenge is how do you keep the

model open source but not weights open source not open weights but open code yeah uh in fact how do you even fractionalize I mean fractionalization is a layer on top that's easy uh but but the challenge is how do you keep the weights private but but keep the model open uh and I think that's being solved and once you solve that now you have an incentive uh mechanism to incentivize ownership of a model and I think that's going to kick start a whole new Revolution all kinds of different models uh that are better than others and there's going to be a lot of competition uh and networks like bit tensor is creating this modernization too which is phenomenal right so I think B tens what is doing I'm a big fan uh I think they're the seaing model development uh and they're attracting a lot of

traditional AI folks to do model development crypto and that's going to be very very important important I think we're in this really exciting time for decent AI I mean there's a lot of like scammers out there but there's real work happening with training with inference with we distributed training we distributed inference and model model ownership but where we going and that's the exciting part right what is it unlock like what does decentralized training really unlock that centralized training doesn't have so the main thing for me for decent plus training is open biases what does that mean so so today gbd was or open AI mostly open AI training happens through uh through reinforcement learning right so so you

have fults you have you know basically uh situations where a human has to intervene tell what's right to the AI so depending on who this human being is the AI is going to behave that way so we saw what happened in San Francisco especially with like with like CH the earli version CH and Google Gemini it's heavily left M right and heavily like trained by the people that are that aligned to left ideology so it's famous he wouldn't say anything bad about Biden but he would do say bad about Trump right like these weird tweets that you see where like it's asking you know why are you proud to be white or why are you proud to be black

and it gives you like these different answers and so goog Gemini was particularly bad because it was like you would ask it to create um what do you call uh image of founding fathers it would be like black women you know like I mean it's just really bad because that really shows you the biases that people that train so main important thing is open biases and if we want to have a say so open biases means that everybody gets to contribute to reinforcing the model and so therefore the biases are in theory supposed to be more balanced unless the people that are tra unless there's a bias unless there's a bias towards people who are like for

people who are training this model should be either one or another kind of geopolitical I mean do you I I think that what's what's going to happen is we're going to have like models that are like either sort of leaning towards one sens one type of sensibility or political ideology and models are going to lean towards another and like the existing Echo Chambers in which we're in with catalyzed are going to be catalyzed by the models that we're using I that's how I like logically but at least a free market that determines right at least we know we know where the buyers right now we don't right now we don't know who who train them we don't know where the biases are we don't even know what system process better does it doesn't make for a better

world but at least in cre for fair world yeah right like it's open world at least we'll see if it's fair but there also like to your point right so there is examples of like this might Microsoft AI I think of Sophia like that was launched in Twitter and become super racist yeah yeah like with like a week or two you know so there could be that outcome too I mean I'm not say I mean and the world could be the world I mean depending on how you look at it other world is a really bad place or world is a really good place but we'll see right but at least it's but at least we'll know and it's not controlled by a few that is in a closed room uh and second is the data and the quality of the data right like

big thing today is um most of the data is not on the Internet it's actually on our phones m maybe Apple has all the data I don't know yeah but at least it's in people's hands and how do you reach that data um while preserving privacy is a critical aspect of of training like in particularly medical data right so if you want meaningful medical diagnosis systems uh right now Big Challenge to get access to high quality data is that's anonymized right so if there's an incentive for me to share Mone data paper Pro uh yeah electronic yeah scan but if there is some incentive mechanism that's going to incentivize me to provide my medical data and guarantee my

privacy yeah I would contribute because I mean it's great good like if it's getting me better diagnosis yeah there's an example recently which was amazing they detected uh a a a lump in someone's breast a breast cancer patient 5 years ahead of time before they could develop into a a bigger that's life saving technology and imagine how so what AI does it's not going to replace doctors but it's going to make doctors do more with less time yeah like imagine most of the world doesn't have access to Quality Healthcare right even C does all right you developed World it it it takes a lot it's very expensive you can reduce the cost of healthare and improve access to the healthcare and that's something that

I want to be part of right and if you incentivize and make it fun even better like World coin showed that giving identity can be fun right like you know it's giving some incentives back yeah with crypto has a phenomenal incentive Network you're really good at that so if you can bring the crypto crazy D incentive mechanism to healthare and train a decentralized healthcare system or a medical diagnosis system uh that's fair in private I think that that I'm not going to trust open ey with my Health Data no way in hell is the sure um like in in the future do we have a generalized model that is good at everything or do we have a model that's very good at utilizing specialized model that's a billion dollar question

everybody's tring to answer I mean that's a that's a big question I just have a question right now must be Prett yeah that that's a big question right so if we don't know right and so there are different theories depending on who you talk to right so either you're at the big model camp or in the small model Camp right okay uh I like to believe I I it's my personal opinion right so I think I think it's going to be a lot of small models that going to come together to to essentially create a multim moral mechanism small mors are inherently good at doing small tasks so if you if you consider that Humanity evolv by like groups lots of groups of people contributing uh small amounts of uh skills that they're good at yeah that's

how Humanity grew as a big feels more intuitively correct yeah like it feels intuitively correct to think that specializ is what moves us forward in in any in any kind field you need specialization in order to move forward and so even for an AI like if you consider that an AI wants to make progress it has to be Specialized or at least I think with my limited knowledge here is like it has to be specialized in a specific task in order to make progress and then report back say like to some kind of you know General model about like I don't know knowledge but but this is also how like yeah Society works and how information uh

transmission works and how knowledge growth works and or you can say like small models can reinforce a large model yeah essentially yeah right so there's that theory as well so it's like so we don't know what what like but it's still a large model but but this is exactly how Society progresses like you have like Industries or like you know people who specialize in a specific field say I don't know quantum physics uh and and and then um you know they go on and like build a plasma uh reactor yeah right and it's like highly specialized and then it comes back and like that that progress comes back is shared and improves the general kind of nature of humanity right because we have abundant

energy and then we're able to build upon that yeah the fox 25g and the one mutation that happened uh that that made a Humanity retain knowledge that chipon that that basically created Humanity right so that that that sort of like I mean it's very parallel to how humans evolve humans didn't evolve because of bigger brains I mean me and are bigger brains technically but Homo sapiens evolve because of diversity and mixing lot lots of lots of mixing for throughout thousands of years yeah and uh I think we we are the only ones that can like share the knowledge and we have knowledge retention sharing mechanism so I like to believe yes I like to believe that I'm part of that camp uh but also like so where we're going it's very it's

a crazy world right so today we saw so gbd3 was trained on uh a 10 megawatt data center right so 10 you can get 10 megawatts hard to get but you can still get a 10 megawatt data power capacity yeah um I think gb4 um or the next version uh is going to be trained um the grock at least is is being trained on 50 megawatt data center the Memphis data center you data center the grock is Twitter's or X yeah X's uh AI model so their new model is getting trained on I think Rock 3 was tring on 50 50 I'm Su a little hazy on the details but Rock 3 is like not great

it's it's okay it's actually pretty good uh it's except for if you remove all the system prompts the open source version is pretty good right um I mean it just stes to be too funny and like so that doesn't really come off well yeah but it's a big model um and it was trained on um is it only Twitter no it's trained on like general information not just Twitter uh I I don't know the details I mean they haven't disclosed the training set yet uh or they may I discl but it's larger model right so it took about uh 50 uh megawatt data so so the notion is that every year the amount of power we need is growing by half an order magnitude okay right so gr was 50 uh

megawatt Center you don't have 50 megawatt Supply in a single source so they literally are burning diesel like talk about El musking Diesel and people are concerned about crypto being yeah know the environment no one is talking about crypto not being bad for the environment I think people are falling on AI now they're not writing it yet because everybody likes AI yeah people didn't like crypto but no one's talking about like hey they're burning diesel to power in the AI data center anyway the the next moral is going going to be 250 billion uh 250 megawatt right if you go by this like PR yeah and a model after that and the next year is going to be 1.2 gigabyte uh so gwatt capacity to get one to get anything

beyond a gwatt you need a nuclear reactor yeah uh you can go solar but solar a 2X two so a mile a square mile of solar panels generate about 2.2 gaw capacity so it's possible yeah but you need continuous energy so you need to have batteries to store them right well to to store 2.2 gaws of batteries you need in the sun they get really hot I mean lithium ion batteries tend to to explode you need cooling yeah so the amount of energy you need for cooling the batteries may not be yeah it's detrimental to the amount of energy you generate so it may not it's not actually

possible to do solar panels only way to get that much amount of energy is to go solar also nuclear yeah so Microsoft bought a nuclear uh reactor the three three long three mile long in Pennsylvania to power the new data centers Oracle is building new data centers uh new nuclear reactors to build the next uh a data centers so I think us has approval for 90 new nuclear reactors now Christ Us in total produces uh 4.4 trillion uh uh trillion watt uh capacity now we did not need to produce more energy till now but now we are seeing an entire revolution in like energy so uh mean China has a a double double

capacity yeah so we need to like ramp up our energy and there's a whole nuclear Revolution now happening silently we not talking about I think it's a good thing because uh because energy costs in us are extremely high I mean we I mean on average I think we're doing like 15 uh cents a kilow hour uh California is like 30 cents a kilowatt hour so nothing is happening in California Texas is 10 10 cents a kilowatt hour uh there used to be a Time in America uh we used call too cheap to meter so we need to get back to that so I'm really like excited for lower energy cost because with lower energy cost you can Sol a lot of worse problems we can solve desalinization of water we can you can solve the water problem you solve the water problem we

can solve a lot of problems right so I think this whole AI Revolution is going to end up actually great for Humanity in general because it's going to bring the marginal cost of energy significantly lower that presents another problem it takes about 10 years to build a nuclear data center I don't think we have gotten a nuclear data center yeah sorry yeah so basically a nuclear a nuclear reactor to power that c right so it takes about 10 years I mean I don't think we've gotten faster because there a lot of regulations there a lot it's quite a lot that goes into building a nucle act right uh so we don't have 10 years to wait to build the Next Generation AI so I think what's going to happen is there's going to be a lot of the small data centers like I talked about they're

everywhere that need to be connected and and if dial or like you know distributed training ass all we can actually use a highly distributed Network to do to to devel the next am model yeah uh and that's why I'm very like uh excited and very bullish on decentral zi because considering the energy constraints that we have uh we're going to continue building nuclear reactors there's a lot lot investment is going there's a lot of finances happening I don't think that's going to stop because the energy needs are just going to be So High by 2030 we going to consume AI is going to consume 20% of US Electric grade right and that's a given right but I think in the meanwhile we having incredible

opportunity for for deepend to really Thrive the decentralized GPU networks to really Thrive to take advantage of this nuclear uh you know speed to be a nuclear reactor but also puts you a dystopian setting where the nuclear reactors powering AI data centers if you picture that in a movie that is creepy right like you have literally urania powering AI so it's like I think it was it was terminat that we that we had this like nuclear reactors whatnot so we're going to that future very soon like it or not yeah this whole conversation has just made me a lot more bearish Europe because of the whole like it's sad kind of you know debate around nuclear France is doing

great I mean France is doing okay they're building your reactors Germany is not um I mean I you know from what I've seen and what I've heard about like Fusion reactors I think like that's a long-term future but Fusion is always 10 years away yeah it's very longterm right but but I think that like if you know if you look at like the next you know so Millennia of human civilization I think probably that's one of the ways that we're going to be able to generate yeah you and it's already starting us I mean so Investments are being happening I mean I saw some banks that are like financing quite a lot if you talk to a lot of Wall Street folks they're like buying a lot of I mean when Microsoft announced the the the what do you call

the three M three three long Tre myON I think it's Island see my Island it's called in Pennsylvania uranium stocks went up like yeah quite a lot so money flowing into building nule nuclear reactors and I think we're going to uh really reduce the price of elect to a point that going to be to CH me I think that's going to be a reality I'm really excited about that because I mean I want to see human civilization reach type one status in my lifetime right you know so m i mean I think it's going to happen like and the only way type one is defined by uh type one effect civilization that can consume all the energy resources on planet Earth yeah uh

to a point that we can exert Cate Humanity to to explore space and you can do a lot of things right if we reduce a marginal cost of energy that is going to be fundamentally great for Humanity to to progress right the problem is alignment like the the the antagonist to this that deth folks you well the degrowth I mean we could go on for whever it is but I think I think that at the same time the technology is moving at an incredible Pace um and and AI is also like in moving at an incredible Pace I worry that the social division that exists that that is perpetuated by algorithms like unless we fix the algorithms but I

think that we can't fix the algorithms unless we change the incentives yeah we continue to be like divided yeah and and I mean just like you know I'm oversimplifying here but I mean if you look at like the Chinese model versus uh sort of what's happening in the US and Europe and and the level of sort of leadership and alignment and long-term thinking that goes into you know that that particular society versus um sort of progress in in the western world uh I'm I'm not hopeful that like I think we can't like problem is we can't get our act together we can't like get people to have a common project and that makes me a little bit more pessimistic

about uh the future of humanity I have a different thought process I mean historically he humans were I mean us Has Come Together in Manhattan Project and in so many different cases I mean there is of course you can't compare to China in terms of how fast they can build infrastructure right they can speed up a nuclear reactor in matter of year I mean like three years that yeah they already have like Double C they're building like for got statford but they're building like every year or every three years the entire capacity of Europe or something that like the energy right all Europe is declining Europe is declining I mean Europe is a retirement home at this point right uh you know

Europe already always had it ups and downs so it's it's historically like very align like in easn always had this like terms of I mean like in pre- colonialism China and India arguably had like 50% of global GDP right so historically they have been like ups and downs but right now what what China is doing it was very interesting things a geopolitical situation so anything beyond I believe 7 nanometers uh you know is bad is export controled to China yeah so you cannot send them chips so they're developing indigenous chips now uh we talk about Huawei and Huawei is

actually incredible company for all the things be be you know being controlled by China's Communist party but they have their hands tied on the back and they're making incredible progress right they have such an amazing like market share I forgot what it was but majority Market CH all Telecom chips right and they are developing uh they have 7 n capacity and the rumored is they have 5 Nom capacity um and even us doesn't have 5 Nom capacity only Taiwan has right tsmc that is yeah tsmc has 3 n capacity right so we are finding it very hard to export the technology back to us even though we invented it I mean Intel invented that stuff right like it's sad to see Intel

going down um it's sad to see like I think West has lost his way in issues that are depending on you would talk to but you know I don't want to take a political stance but issues that may not be contributing to human progress in the way IC progress in terms of technological exploration yeah right uh so like I mean like when like two weeks ago like we you know the super heavy booster you know that that whole thing I mean most I think most people in you know in my surroundings and certainly the people like in my neighborhood don't even know or care yeah I thought like this is this is an amazing Fe of human progress and

engineering and I turned on the news next day and like not the president didn't congratulate he like well in France right I mean it was so but yeah not a peep and and possibly even people who didn't know about it were probably like how you know complaining about it about how it's like you know elon's latest whatever so yeah this mindset is a little bit um I think China competition is a good comp I think free market the is playing out right China is out accelerating and they have their own chips now some of the models are way better than the Western models right so the AI is very different it's very local it's very indigenous to China uh they're very good at image models they're very good at image recognition why because surveillance you know you know uh so

they're very good at video models so I think there's going to be like a very fascinated East Ai and West AI yeah um I'm obviously bullish in West AI because we're more free people in general I think freedom in thought is very important to to for Innovation me West West innovates China replicates right and Europe regulates um so uh I I'm generally bullish I think like we as a free market is coming together and there's there's also notion of like we need to accelerate in California especially right there obviously there's this degrowth mindset but also e mindset like we need to accelerate as fast as possible AI is going to let us accelerate like you yourself you told me earlier that you're coding uh this this

app yeah imagine doing that without AI right so you as CEO and as a leader of a company that doesn't have time to write sit and write code is able to do more more so AI is a time machine it lets you buy back time yeah that's why I'm like super bullish it's a huge time buy backer exactly so what happens so the way the two ways to look at AI right the degrowth people the way you look at AI is going to take our jobs yeah but the growth oriented people that look at AI is going to be like they allows us to do a lot more with a lot less time yeah so I belong to that c I always have milon things to do like if you give me time I have I can do so many things

time is my biggest challenge right if AI lets me get more time I can accelerate myself I write code too I mean I was not writing much code like last few years but now I'm writing a lot more code because with the AI can do more you everybody can do more I think people are going to realize that they're going to be of course classism where people that know how to use AI are going to out explo people that don't use AI so then people that don't know how to use AI will have no other option to use AI right like back in the day internet was the same thing people that use internet knew how to use internet people didn't know how to use internet kind of were left back and they had no other option ex to use internet so a free market

system will eventually work it may take longer it may take a few uh events that may not be ideal in terms of progress but those events are going to make us realize and America's always had this ups and downs and like so it's you know historically a a big big big uh you know student of history and a big F of History so historically we look at at we had so many it was worst right we had Civil War we had so much misalignment um but eventually we all came together in the West in general always you know figured out how to do these things uh in know in a well and East you know we look at East as this this lens of Communism but it's a lot

more complex like if you go to China absolutely yes it's communistic but it's a lot more free than what you would otherwise think uh it's just different and I think the moment we look at it from that lens uh I think uh it's we're going to have a better better uh understanding of the world so I'm you know overall very very bullish to human progress lfg that's iting go thanks man appreciate it [Music] [Music]