저자: ChamathPalihapitiya · 외국인
요약 보기
딥시크 효과로 AI 학습비용 논쟁이 과장됐다고 봄. 제약이 혁신을 만들었다는 견해로, 딥시크의 GRPO·PTX 등 기법을 언급. 오픈AI 기업가치(3400억 달러) 투자 논의와 MS·메타의 AI 전략 변화를 언급하며 엔비디아 CUDA 독점 위협 가능성을 지적. 엔비디아 주가 전망에 대해서는 구체적 방향 발언 없음.
다듬은 글
so I just think it's like worth pointing that out that I think that part of the story got overhyped it's hard to know what's fact and what's fiction everybody who's on the outside guessing has their own incentive right like so if you're a semiconductor analyst that effectively is massively bullish Nvidia you want it to be true that it wasn't possible to train on $6 million obviously if you're the person that makes an alternative that's that disruptive you want it to be true that it was trained on $6 million all of that I think is all speculation the thing that struck me was how different their approach was and TK just mentioned this but if you dig into not just the original white paper of deep seek but they've also published some subsequent papers that have refined some of the details I do think that this is a case and sax you can tell me if you disagree but this is a case where necessity was the mother of invention so I'll give you two examples where I just read these things and I was like man these guys are like really clever the first is as you said let's let's put in a pin on whether they distilled 01 which we can talk about in a second but at the end of the day these guys were like well how am I going to do this reinforcement learning thing they invented a totally different algorithm there was the the Orthodoxy right this thing called po that everybody used and they were like no we're going to use something else called I think it's called grpo or something it uses a lot less computer memory and it's highly performant so maybe they were constrained Sachs practically speaking by some amount of compute that caused them to find this which you may not have found if you had just a total surplus of compute availability and then the second thing that was crazy is everybody is used to building models and compiling through Cuda which is nvidia's proprietary language which I've said for a couple times is their biggest moat but it's also the biggest threat Factor for lockin and these guys worked totally around Cuda and they did something called PTX which goes right to the bare metal and it's controllable and it's effectively like writing assembly now the only reason I'm bringing these up is we meaning the West with all the money that we've had didn't come up with these ideas and I think part of why we didn't come up is not that we're not smart enough to do it but we weren't forced to because the constraints didn't exist and so I just wonder how we make sure we learn this principle meaning when the AI company wakes up and rolls out of bed and some VC gives them $200 million maybe that's not the right answer for a series A or a seed and maybe the right answer is 2 million so that they do these deep seek like Innovations constraint makes for great art what do you think uh freedberg when you're looking at this well I think it also enables a new class of investment opportunity given the low cost and the speed it really highlights that maybe the opportunity to create value doesn't really sit at that level in the value chain but further Upstream bology made a comment on Twitter today that was pretty funny or I think ref this about the rapper yeah he's like turns out the rapper may be the the mo the the mo which is true at the end of the day if model performance continues to improve get cheaper and it's so competitive that it commoditized much faster than anyone even thought then the value is going to be created somewhere else in the value chain maybe it's not the rapper maybe it's the user and maybe by the way here's an important point maybe it's further in the economy you know when electricity production took off in the United States it's not like the companies are making a lot of money that are making all the electricity it's the rest of the economy that Acres a lot of the value well you're about to see a big test of this because if open aai raises 40 billion at 340 billion that just hit the wire the underwriting logic at 340 in exactly what you just said freeberg it is the rapper meaning chat GPT is the next killer app it's getting to a billion plus ma hundreds of millions of Dow it's competing for Consumer usage that's that's the model that's the model is like consumer usage which puts them on a on a collision course with meta it's the only company that could really impact that because the only company right now that has billions of eyeballs of daus per day and who and by the way zck said this in his earnings release he's like there's only going to be one company that brings AI to a billion plus people and it will be us some version of that quote in his earnings released yesterday and then Microsoft showed weakness in the in their cloud and then Microsoft's down 6% today and you know I think it's a a window for open AI to say we're going to go up against meta this is it we're going to be the players chth Palapa recently went on the all-in podcast and attempted to analyze the Nvidia stock after deep seek AI was unveiled shath came to a very shocking conclusion about the stock doing something that it has never done before and it seems that deep seek was not the only factor in all of this let us take a deep dive into his analysis and what shocking truth it had uncovered about Nvidia welcome back to invest ocracy what do you guys think is happening right now between open Ai and Microsoft because if it's true that this distillation thing actually happened well there's only one place where you could have distilled the o1 model but it's on aour so what the hell is going on over there well and their R1 is supported on explain distillation real quick yeah so when you have a big large parameter model the way that you get to a smaller more usable model along the lines of what saxs mentioned is through this process called distillation where the big model feeds the little model so the little model is asking questions of the big model and you take the answers and you refine and by the way you can see this Nick I sent you a clip you guys can see this I mean it's there's clearly distillation happening Nick can you show the C of the of the deep seek run where it shows the China answer and then deletes it what was Winston's job in 1984 right and it sort of starts to go through this whole summary and then the person says are there any actual states that currently do that hold on here it goes it says North Korea wait it goes China and then wait watch this boom so the reason why this is happening is like you're seeing this Chain of Thought you're seeing the several layers and then it's catching it after the fact so we know that this is distilled from some other model and my only point there it's a little tongue and cheek is right now when you go and use open aai you're using it sitting in an Azure instance somewhere right so this is Microsoft's Cloud infrastructure that runs it so it begs the question it's not that it's 0's fault or open ai's fault that this distillation happened and I'm not trying to assign blame but typically if this were to happen you'd look your cloud provider and say how are you letting this happen and I don't think anybody's had a good answer for that well and the cloud provider is hosting R1 now so they're literally undercutting chat GPT and open aai at the same time just to clean that up they're they're hosting their own copy of it right because r1's been when you say say they who are you who are you referring to sxs Microsoft yeah Microsoft is hosting their version of R1 which means they are actively subverting their partner open Ai and and pushing people to a cheaper model well whatever I mean look Amazon's going to host their own version of R1 grock has a version of R1 have one just red out it's open source now have a bud who has R1 on his laptop you know yeah exactly yeah but if it was if it was stolen as um and the IP was stolen as Sam is claiming that would be like You' think he'd be able to call up Sak and say hey can you not put the stolen IP on your server and promote it to everybody at a lower cost it just shows Microsoft has no loyalty to open AI yeah and they have but you think they would have loyalty they have no loyalty what it would take to distill 01 like brute force it wouldn't be like oh geez I can't believe it was distilled it would be like such a massive number of calls against an API or against something something that it wouldn't be unnoticed oh they did actually came out and said they blocked some suspicious activity recently in the wake of deep See's announcement Invidia stock experience a notable downturn capturing the attention of Market enthusiasts worldwide now if one were to recount nvidia's recent stock performance it would resemble a roller coaster rather than a steady Ascent historically when nvidia's forward price to earnings ratio hovered around 30 the company was at a market cap of around $1.5 trillion half of what it is today yet here we are with invidious value soaring matched consistently by an unwavering forward PE hinting at a robust earnings forecast in wall Street's eye by the way you have to be sympathetic I think to open AI in this because if you're building a startup you're trying to raise money we've all gone through this cycle guys where it's like there's momentum we celebrate internally the momentum that's what gets you the energy to push your team even further and harder and then all of a sudden it turns out that some portion of that like it Travis said it well like you're there's probably a chart inside of open ai's offices where you're showing how many times these apis are getting hit right you know how many times these end points are getting hit it all looks positive and then you realize that some portion of it was actually bad and trying to undercut your value it's a hard pill to swallow and then you have to course correct very quickly you have to lock down this is one area where security exactly we have not talked about this like you have to lock these models down now you have to lock the end points down look in the Biden Administration if this had happened the first conversation would have been we need to kyc the people that use these models and it's like what are you talking about we don't kyc the low if you're trying to use like an ec2 endpoint or an S3 bucket you don't have to all of a sudden prove who you are you just use a credit card and go that's the whole point of why proliferation can happen so quickly but if we take the wrong takeaways from this period there's going to be a bunch of people that'll clamor to like lock these folks down and make Innovation go much slower I don't I think that that would be bad here's the other side and totally agree Chim but here's the other side you go through the white paper you see what it is they did what they innovated on the science behind it the thorough and you're like these guys are bad ass it it does not feel or sound like somebody who took something just when you get through it it could be that it could be that open AI wrote the white paper for them just putting it out there but it's real Innovative I agree with that real Innovation strong Tech you're like this is legit I agree with that but in that paper they're very hazy about where the data is coming from and they're they're fairly transparent about everything else they did but they're not really clear about the data and specifically they say that to get from V3 which is the base model to R1 which is the reasoning model they had about 800,000 samples of reasoning they were quite unclear about where those reasoning samples came from by the way it is remarkable that you can get from a base model to an R1 with just 800 100,000 samples but this is the problem like we meaning like the Western AI Community we've been trudging around on this path where we've been very we had a very Orthodox approach the only way you can do reinforcement learning is through po okay but is that true and it turns out that if you're like a really smart team that has no other choice you move away and you invent your way out of it and so we have to get that example too I think it's technically brilliant some of the things they've done but they also use constraint as a very much a feature not a bug and the the Western AI economy has been the opposite so far I think the best part of this is the fact that Sam mman was supposed to be doing open source he made it a closed Source company he stole everybody's data every he got caught red-handed he's being sued by the New York Times for all that and now the Chinese have come and open sourced all the stuff he stole and he's got a real competitor on the original Mission of what open ey was supposed to do so I have zero sympathy for him or the team over there I'm glad that this is all going open source it should have been open source and it's better for Humanity and the fact that the Chinese did it to Sam wman has come up in for him stealing everybody else's content that's my okay there you have it but I don't have strong opinions on it however the smoothness of stocks belies the potential ripples initiated by Deep seek as analysts scramble to re-evaluate their earnings projections for NVIDIA the moot question becomes what does this mean for nvidia's GPU dominance especially in the context of evolving AI infrastructure demands while history offers a sentiment of nvidia's potential to double its value again this year the astute investor must tread cautious meta has to embrace and extend everything that these guys have shown meaning like meta is buying tens of thousands of Nvidia gpus great but what did this show this shows that actually Cuda high level language is in general I think we've all known that they suck okay and so we've all been going through it thinking that it's like the right thing to do deep seek throws it out the window they use something called PTX what does meta do is critical now to understand they need to embrace this stuff and this is where I think again apologies to the Invidia but it's going to create a more head genus environment and the reason is because there's too much money and risk on the line to go through a single point of failure a chip a highle framework to get to that chip that's nuts so I think like that kind of like Emperor has no close moment is upon us currently Nvidia retains its position as a front runner in AI technology yet shifts and adaptations are constant in this rapidly advancing sector the stock market much like a story unfolding is filled with surprises some exhilarating and and others cautionary as you Ponder this intriguing conundrum we'd love to hear your thoughts what is your take on nvidia's future in the AI Tech landscap
Comments 0