Unlocking AI's Potential: Chamath Palihapitiya on GPT-3 and Beyond 🤖

요약 보기

AI(GPT-3/LLM)는 냉장고처럼 거대 가치 창출 기술이나 돈은 인프라 승자(코카콜라 비유)가 가져감. 독점 데이터·RLHF가 차별화 요소, 정치 성향별 모델 분화 예상. 2021~22년 VC 자금은 시간 분산 부족으로 압박받고 수익률 1.6배 대비 과대평가(1.6조 vs 52조 마크) 우려. 엔비디아가 곡괭이 장사로 가장 큰 돈을 벌었고 AMD도 수혜 가능성 언급. 자막이 중간에 끊겨 발언 일부 누락 가능.

다듬은 글

※ 자동 자막 원문(다듬기 전)

01. let me make a prediction it is the most important thing in the next 25 years so I don't care what you like or don't like there's a company in Silicon Valley called open AI it's incredible incredible technology gpt3 where you can have a language model now where it can Auto write a book and you will not know that it was not a human the problem is in these products I think that is the big existential risk with this entire part of computer science that entire infrastructure can be made much much better with AI right a robot that does laser guided Precision surgery characterizing tumors 100 with 100% accuracy so you always get 100% of the cancer out when you go and get surgeries done all these things are possible now and you know what chat GPT shows you is just the amazing value in allowing computers to assist you in doing work it's like a you know calculator replacing the Abacus replacing a pen and paper what's important though is this Buffet quote um friend of mine told me this yesterday which I loved he told this story about refrigeration and the story he tells is that the people and the person that invented refrigeration made some money but most of the money was made by Coca-Cola who used Refrigeration to build an Empire and I view these large language models as Refrigeration will there be some money made in it I think so but the Coca-Cola has yet to be built and those are the companies that are really going to monetize it and in order to monetize it well here's a basic thing about machine learning that's worth knowing which is if you take 1,000 of the same inputs and give it to Facebook and Microsoft and Google and Amazon they'll all come up with the same machine learning model but if you have one extra thing one little ingredient that all of those other companies don't have your output can be marketly different it's like giving two great chefs three ingredients but you give the third Chef one extra one that person has the ability to do something very special so right now we are in the world with where everybody is crawling the open web we're going to move to a world where as everybody gets sophisticated enough where when Refrigeration is widely available somebody's going to say you know what this site I'm not going to allow anybody else to access it's only me only for my models and those models will become better and so we have to let that play out a little bit and so it's going to be a little bit of a really interesting arms race so the the next wave of m&a for example could be companies like Google and Microsoft and Facebook looking at these companies saying can they be viable inputs to my large language models or to my other machine learning and AI models so you could see m&a activity that drives that differentiation before anything else so lots of really so then as a result for guys like us early stage investors we may want to invest in companies that have zero viable Public Market potential whatsoever but is building a data repository that's so unique that we know that it will feed one of these bigger companies and their efforts in Ai and that could very well justify making an investment that we would otherwise not make today knowing what we know what will make all of these systems unique is what we call reinforcement learning and specifically human factor reinforcement learning in this case so David there's an engineer that's basically taking their own input or their own perspective now that could have been decided in a product meeting or whatever but they're then injecting something that's transforming what the Transformer would have spit out as the actual canonically roughly right answer and that's okay but I think that if this is just a point in time where we're so early in this industry where we haven't figured out all of the rules around this stuff but I think if you disclose it and I think that eventually Jason mentioned this before but there'll be three or four or five or 10 competing versions of all of these tools and some of these filters will actually show what the political leanings are so that you may want to filter content out that'll be your decision I think all of these things will happen over time the problem is then these products will fall flat on their face and the reason is that if you have an extremely brittle form of reinforcement learning you will have a very substandard product relative to folks that are willing to not have those constraints for example a startup that doesn't have that brand Equity to perish because they are startup I think that you'll see the emergence of these various models that are actually optimized for various ways of thinking or political leanings and I think that people will learn to use them I also think people will learn to stitch them together and I think that's the best better solution that will fix this problem because I do think there's a large PO non-trivial number of people on the left who don't want the right content and on the right who don't want the left content meaning infused in the answers and I think it'll make a lot of sense for corporations to just say we service both markets Google did not want to release this for years and they they sat on it because they knew all these issues are here they only released it when Sam mman in his Brilliance got Microsoft to integrate this immediately and see it as a competitor Advantage now they've both put out products that let's face it are not good they're not ready for prime time don't you think a lot of this gets solved in a year and then 2 years from now like you said earlier like it's accelerating at such a rapid pace is this sort of like are we making a mountain out of a mole hill sacks that won't be around as an issue in a year from now I think that is the big existential risk with this entire part of computer science which is why I think it's actually a very bad business decision for corporations to view this as a canonical expression of product I think it's a very very dumb idea to have one thing because I do think what it does is exactly what you just said it increases the risk that somebody comes out of the you know the third Act of rebirth and says wait a minute this is not what Society wants you have to stop and that risk is better managed when you have filters you have different versions it's kind of like Coke right Coke causes cancer diabetes FYI the best way that they manag that was to diversify their product portfolio so so that they had Diet Coke Coke Zero all these other Expressions that could give you cancer and diabetes in a more ctitious way I'm joking but you know the point I'm trying to make so this is a really big issue that has to get figured out there is not a single human being on Earth if given the chance to found a very successful tech company would do it in a nonprofit way or in a commoditized way because the fact pattern is you can make trillions of dollars I think it's fair to say a couple of things that there was probably 2 and a half 3 years of capital raised in the industry that's going to get really put under pressure and the reason is that there is not a lot of time diversity in that money meaning people got it and they put it into the ground right away and one of the principles of having a more predictable return set of returns over time is that you leverage time right so if you had $100 and you wanted to have a diversified stream of returns you're much better off spending a dollar a mon month for 100 months versus $10 a month for 10 months so just that thing will cause a lot of impact in headwinds for a lot of the capital in 2021 and 2022 then the other thing you have to keep in mind is that over many cycles where we've had high rates and low rates and medium rates our industry typically returns a160 for every dollar it raises and that's over many cycles and so if you believe that we're going to revert to the meme out of the trillion dollars we've raised maybe we'll return 1.6 trillion now that sounds good except the problem is that 1.6 trillion is marked at 52 trillion let me make a prediction all of the things that you guys said I think are incredible consumer surplus business opportunities which means that the ultimate Winner Is Us incredibly incredibly productive and more leveraged in how we spend our time which will allow us to do all kinds of other interesting things with all the time that we save that I think is almost a certainty the problem with consumer surplus businesses is often times there is no money made in the funding of them and really where the money is made isn't enabling it so for example so far what I would say is there's very little money that has been made in AI there's been an enormous amount of money that's been made by Nvidia and the reason is because they are the pick and shovel provider in the into the industry and so that's an example AMD I think can also benefit so the Silicon players pretty obvious here maybe some of the cloud players um the problem is the cloud players are trapped inside of other big companies with many other business model