저자: ChamathPalihapitiya · 외국인
요약 보기
AI 버블 우려 속에서도 AI는 초기 과열 후 옥석 가리기(winnowing) 단계에 들어서는 중이라는 진단. 백오피스 AI 적용은 정확도가 높고 유효하나, 기존 기업 도입 과정에서 저항이 크다. 다만 현재 LLM 중심 접근은 초기 단계이며 새로운 접근이 등장할 여지가 있다. 정치 쪽에서는 금리가 2026년 경기 흐름의 중요 변수라고 봄. 인터넷 버블 사례처럼 단기 과열이 있어도 장기 변화는 이어진다는 취지.
다듬은 글
they're in charge of a process and they're they're dealing with edge cases. And I think that you can get extremely high rates of accuracy if you implement AI correctly in back office tasks. I think the real question is like what happens to all of this revenue that has been generated. You're seeing companies generating $50 million of ARR in a matter of months and then raising huge rounds.
I think what we haven't seen is whether there'll be any sort of either logo churn or dollar churn as new companies come in with even cheaper solutions, the foundational models move up the stack and just absorb capability or things just don't work and they get abandoned. All of that churn. Welcome to Tesla hype everyone.
Artificial intelligence has seen tremendous growth over the past 2 years. Both Chat GPT and Generative AI were viewed as novel and disruptive technologies when they were released in late 2022 and 2023, respectively. Startups quickly made large sums of money. Large corporations like Microsoft and Meta heavily invest in AI.
But like any excitement wave, questions are beginning to surface. Will AI truly develop at an unstoppable rate? Or is an AI bubble just getting started? This past week, it was more difficult to ignore these concerns. The CEO of Open IIIA, Sam Elma, noted that bubbles form when people become overly excited about a fact and likened current state of AI to the dotcom boom of the late 1990s, even though artificial intelligence, AI, is the most significant new technology in decades.
He claimed that investors might be becoming overly enthusiastic about it more quickly than the tools can actually handle. According to an MIT study, the majority of AI projects aren't even close to being completed, which supports this warning. The stock of public AI fell roughly 10%. Additionally, after Meta's rapid growth, Mark Zuckerberg stopped hiring AI specialists.
David Sax's David Friedberg and Chimath Palahapisha all discussed the most recent Olan show. They discussed whether the AI bubble is starting to burst and referred to it as a reality check moment. It offers insightful predictions about the future and one of the most significant developments of our day.
First wave was just a lot of boards who read the words AI somewhere in an article and then went to a board meeting and turned to the CEO and said, "What's your AI strategy?" Mhm. And then the CEO turns around and sends that down into their org and it eventually hits the CTO's desk. And I think the first wave is mostly people just spending money because they had large existing budgets and so they were like, "Let's just go and try a bunch of different things." And I think now we're going through the sorting function of realizing that there's a big difference between probabilistic software and deterministic software.
That's probably the biggest reason why you're seeing so many failure modes in sales and marketing. It's very hard to codify sales and marketing into a set of heruristics that never change. But back office processes, why they're so good and a great target for AI is ultimately you have so many people that have been hired to deal with edge cases, right?
I think like that's what people do in most companies is they're in charge of a process and they're they're dealing with edge cases. And I think that you can get extremely high rates of accuracy if you implement AI correctly in back office tasks. I think the real question is like what happens to all of this revenue that has been generated.
You're seeing companies generating $50 million of ARR in a matter of months and then raising huge rounds. I think what we haven't seen is whether there'll be any sort of either logo churn or dollar churn as new companies come in with even cheaper solutions, the foundational models move up the stack and just absorb capability or things just don't work and they get abandoned.
All of that churn happened in social. I remember when I was in the middle of helping build Facebook, we went through that whole cycle. There was seven or 8,000 social companies and within 6 years there was five of us left. It happened in SAS when I was investing in SAS. There's a couple of very early and important successes like Yammer which Sax started which I was very lucky to be an investor of but then it took many years for the handful of winners to get really sorted out and I suspect we're about to go through that same cycle in AI.
So I think that article basically paints a very accurate picture. There's been a lot of triing and experimentation. We now need to go through a sorting and a cleansing and then we'll rebuild. Is it's concerning when models successively don't get meaningfully better especially because there's a tendency for incremental models to get overfit to benchmarks.
That's another concerning thing that I think people were like scratching their heads about. And if you look at the last big set of model updates, X of Grock, they didn't really meaningfully improve their separation from the last generation model on the big benchmarks that matter. And that's also very concerning because they tend to overfit to begin with.
How at some point like you get so invested in a way of doing things that when things change, it's almost too hard to change the way your company orients. And you see it in every single phase of these big innovation cycles. The question is, is there a point at which these foundational model companies have invested too much money in LLMs where so many of their employees only understand that and all of a sudden there are completely different ways of building these models.
State space is one example. But then you know some new entrant decides to build a completely different kind of representation. They may also then build custom silicon for it. All of those things have yet to happen. And the reason is we're really only a few years into what should be a multi-deade innovation cycle.
So those are like some really important questions because then you have this problem of the sunk cost fallacy of well we spent all this money on this LLM based approach and now all of a sudden there's this thing that comes around the corner. That's the thing that I'm really curious about and waiting for is will there be these small teams and it may take advantage of what Freeberg said.
uses an LLM as a jumping off point, but then builds an entirely different approach, is able to generate entirely different and cheap silicon. That hasn't happened yet. And the reason it hasn't happened is not because it doesn't make sense, but it's because we're so early into it. Yeah. And then what do these big guys do?
The big guys have spent tens of billions of dollars. Yeah. According to there are many parallels between the current AI wave and earlier tech trends like social media and SAS. Only a small percentage of the thousands who initially entered each eps one made it to the end. According to him, AI is now entering a stage known as winnowing, whereby businesses that are founded on hype and excessive spending without a clear business plan will disappear and be replaced by businesses that are sustainable and have practical applications.
David Saxs responds that while he acknowledges the industry is undergoing a shakeout, he does not believe it is collapsing. He views it as a constructive correction instead. claims that making untrue statements about a GI such as claiming that computers will surpass humans in intelligence in a few years feeds both fears of the end of the world and utopian fantasies.
He makes it apparent that rather than making big abrupt leaps, the field is advancing slowly and steadily. His main argument was that while artificial intelligence AI is still very popular, it appears to be more of a technological race than a magical race to become extremely intelligent. All back office type stuff.
But our first customer where we delivered enormous gains, we were fired. And the reason we were fired was not because the technology wasn't working, but it's because as it goes lower and lower from the board and the CEO into the bowels of the organization, you actually run into huge push back. And I think that there's going to be that cycle that we have to deal with as well, which is in all of these AI businesses, there's all kinds of stuff that scare people.
And there are many people in decision-m roles at many of these companies that will have to sort out where's their comfort level. And it was like one of these things where I was like, well, what are we supposed to do? We escalated and blah blah blah and it got resolved. But my point is that I think is a cycle that has yet to play itself out in all of the Fortune 2000, global 2000.
I would not expect that the AI boogeyman that has been created doesn't have some negative ramifications over the next couple of years as well as this stuff gets sorted out. The bull case. I'll make the bull case. Yeah, please. Yeah. I think the simplest way to make the bull case is if you look at the kager on their mouse and the conversion from ma to dows.
You essentially take a small minor percentage of the Facebook or Google terminal arpoo and apply it to some number of ma dows 3 4 years from now. So if I had to guess if you take 4 500 million ma growing at I'm guessing let's just be conservative probably doubling every two years. So 500 goes to a billion, billion goes to two in four years and at two billion DAO they generate a tenth of Facebook's revenue just to be very conservative and you probably get to a trillion five valuation then
yeah midterms might be easier to think about 500 midterms are going to be about the economy but the economy is going to be a window into a lot of these other issues the safety security issue the immigration issue the tariff issue rates. I think rates are going to be an enormous swing factor here. I think Powell is really holding on white knuckled to a position that probably doesn't underly the the data.
So, I think in the short term it's just going to be more of an economic up or down vote for the Trump administration. Many things I think are positive and in their control, but there are a few things that are not in their control. In 2028, I don't know. The one thing I would say about the candidacy though is like I I don't know if you guys remember, but when the Republican primary process started, the person at the top of the polls at the time was Ronda Santis.
And I said, "You do not want to be ahead early. These people never win. They never win. They take all the arrows and peak too early." No, it's it's not even that. It's just like they never win. So, if I had to apply that rubric here, what I would say is the practical choice that the Democratic voters will have is to actually ask whether they want a replay of California on an American 50-st state scale.
And if you look at it through that lens, I think Gavin Newsome is going to have a very difficult path because of what has transpired over his tenure as the governor and he's going to have to explain away a lot. So I think that in the absence of real choices, land of the blind, the oneeyed man is king sort of a thing.
I think when you actually go into the election cycle starting after 26 in 27, honestly, it's going to be up in the air. What I will tell you is that the way that the electoral college is set up and the voting dynamics for the presidency is set up, a socialist cannot win. But the way that the Democratic primaries are set up, a socialist can absolutely win.
So this is the sort of Saxis point which is I think for the Republican side the best outcome is a very radical agenda that isn't really supported by substantive facts and is more supported by vibes. This is what I feel. I just feel it should be right. And I think that those things may win a primary but will just get destroyed in a federal election.
debate the impact to the broad population, but there are enough examples where actual take-home wages are higher than what they would be in a counterfactual. The technology behind the internet advanced far more quickly than the number of users in the late 1990s. The doc crash, which was brought on by that hole, destroyed billions of dollars worth of wealth and left behind overly busy networks and abandoned buildings.
However, those same networks sparked an unprecedented wave of expansion and innovation in the years that followed. What appeared to be a bubble at first glance was actually the beginning of a transformation. All right, the cycle displayed a persistent trend. Although short-term market overshoots are common, the additional funding nevertheless creates the frameworks necessary for long-term change.
There was more to the internet story than mere conjecture. It focused on how the digital economy was created by investments made in the wrong areas. What went wrong wasn't the concept per se. Investors lacked sufficient patience. The true good was visible to those who looked past the crash. As Chimoth and others have noted, the dotcom era demonstrated how overbuilding can have negative short-term consequences while laying the groundwork for the subsequent wave of innovation.
The main takeaway is that investors typically aren't able to observe how rapidly new technologies are embraced. History demonstrated that the internet continued to exist after the crash. Rather, it was the tenative beginning of a shift that altered global business, communication, and society that now raises a more significant query for us.
Are new tools evolving? Similarly are the platforms that appear to be overloaded today sewing the seeds for the major innovations of
댓글 0