원문 보기
in the middle of helping build Facebook. We went through that whole cycle. There was seven or eight thousand social companies and within six years there was five of us left. It happened in SAS when I was investing in SAS. There was 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.
I think that over the past week or so there was a correction in sentiment towards AI, but I think it was a healthy correction. I don't think this was the beginning of a bus cycle or something like that. I I still think that we're in a boom. I still think we're in a investment super cycle. Artificial intelligence has surged into the mainstream over the past 2 years.
Since the launch of chat GPT in late 2022 and the rapid expansion of generative AI throughout 2023, the technology has been hailed as both revolutionary and disruptive. Startups are raising massive funding rounds at record speed while tech giants from Microsoft to Meta are going allin on AI development.
But with any period of intense hype, questions arise. Is AI on an unstoppable trajectory? Or are we entering the early stages of a bubble? OpenAI CEO Sam Alman recently compared the current AI moment to the dotcom boom of the late 1990s, warning that even smart people can get carried away by a kernel of truth during bubbles.
He stressed that while AI is one of the most important technological developments in decades, investor enthusiasm may have outpaced the actual capabilities of the tools. Supporting this caution, an MIT study went viral showing that most AI projects aren't yet making it into production. Public AI stocks corrected roughly 10% and Meta CEO Mark Zuckerberg froze AI hiring after a year-long blitz.
A clear reality check. In the latest episode of the All-In podcast, billionaire venture capitalist Chamath Palipidia, entrepreneur, and angel investor David Friedberg, and David Saxs, President Trump's AI and crypto adviser discuss what many are calling the AI bubble burst. They provide informed, nuanced perspectives on the likely trajectory of what could be one of the biggest technological shifts of our lifetimes.
We're not here for small gains. We're here for life-changing wealth. Crypto is the frontier of the digital revolution, and most people aren't paying attention yet. If you're aiming for massive gains, why wouldn't you do it tax-free? With iTrust Capital, you can buy, sell, and hold over 80 crypto assets inside a tax sheltered IRA.
That means no taxes or capital gains on trades ever. It's not just about what you invest in, it's how you invest. With iTrust Capital self-directed IAS, you can grow your portfolio in the most tax advantaged way possible. Sign up today with the link and use code do7 for a $100 funding bonus and start stacking taxfree gains now.
Mania hit a bit of a detour this week. all over the, you know, three or four days, AI stocks were down across the board because of this MIT study that went viral as well as Sam Alman's comments about a bubble and Zuck instituting a hiring freeze in AI after going on a complete blitzkrieg. So, let's get into it.
I think what I would tell you is that I think the 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 or 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 I think we're seeing a lot of interesting AI applications being developed but it's still very early days and I think that over the past week or so there was a correction in sentiment towards AI but I think it was a healthy correction.
I don't think this was the beginning of a bus cycle or something like that. I I still think that we're in a boom. I still think we're in a investment super cycle, but I think there was a healthy dose of skepticism applied to some of the more fantastical claims that have been made about AI. And I think this is why you saw there was like roughly what, like a 10% correction in public AI stocks.
And there was that MIT report that said that 95% of projects and companies are are not making it to production yet and so forth and so on. So, I feel like we're getting in the weeds a little bit here and what we should be talking about is just sort of where we are in this um in this AI super cycle. And where do you perceive us at?
We're in the experimentation phase. We're in the pilot phase. But this issue around probabilistic versus deterministic makes it hard to trust the software. Is that what you think the key issue is? Well, let me tell you why I think that this correction is actually healthy. Is that after Chat GBT launched at the end of 2022 and then throughout 2023, the dominant narrative in AI is that AGI was just two to three years away.
And everyone kind of had their own definition of AGI was, but it was kind of this idea of smarter than human super intelligence and kind of magic AI. AI would be able to do everything. And as a result of that, you kind of had both utopian and dystopian narratives really proliferated. And so, you know, you started getting this like job loss narrative that within a few years, 50% of knowledge workers would be out of jobs.
You got this rapid takeoff narrative that basically the leading AI models would be able to turn their intelligence towards improving themselves towards recursive self-improvement and therefore within a couple of years the leading models would basically achieve super intelligence and leave everyone else in the dust and then capture all the value of humanity and then based on that narrative which again it was the same underlying narrative that fueled both utopian and and dystopian or doomer takes on AI.
You got, I think, a huge backlash which has already been forming where you have a thousand bills running through state legislatores right now and you have all this AI safety legislation. You got bills like in California the SB 1047 which would have applied tremendous amount of of new regulation to AI.
So you saw this policy backlash happen as well and it was all based on these fantastical and kind of magic views of what AI was going to do in just the next 2 to 3 years. And I think that the reason why this recent skepticism is healthy is because I think it's rebutting all of that and it's showing that, you know, AI is a powerful tool.
I mean, I I definitely think it's a new and important form of computing and it is going to unlock tremendous value in the economy, but it's going to take us a while to get there. I mean, you can't just tell the AI, you know, be a sales rep, be a customer service rep, and kind of throw it over the wall and expect that it's going to replace a human.
It takes a lot of prompting and iteration and validation to make the AI work to make it generate business value. And if we were on a path towards rapid takeoff, then what you would see is that the leading AI models would be increasing the distance between like the top one or two models would be increasing the distance between, you know, the rest of the models.
And instead, what we're seeing is a clustering of model performance around the same performance bench. Incremental, right? They're incrementally the progress to be a little bit more incremental. It's more evolutionary rather than revolutionary. And I I think this really crystallized around the launch of chat GPT5 where a lot of people were expecting GPT5 to be this huge breakthrough.
Sam Alman was sort of teasing this concept by posting photos of the Death Star that the idea that this model was going to blow everybody else away and the reviews ended up being very mixed and then we saw that on the performance evaluations. It's not that the model didn't represent progress, it just fell short of these lofty expectations that have been created.
Chamomath draws a clear parallel between today's AI boom and previous tech waves like social media and SAS. In each case, thousands of companies flood the market, but only a few emerge as lasting winners. He sees AI entering that same sorting phase where companies riding hype or unsustainable spending fall away, leaving those with durable business models and practical applications.
David Saxs adds to this perspective, agreeing that the AI sector is undergoing a shakeout, but he frames it as a healthy correction rather than the start of a bust. The frenzy around AGI with claims that machines are just years from super intelligence creates unrealistic expectations and fuels both utopian and doomer narrative.
Sax emphasizes that what we're actually seeing is incremental progress, not runaway breakthroughs. his larger point, AI is still in a boom, but it's unfolding like a normal technology race, not a sudden leap to super intelligence. I think that what people can now see is that we're not in like a a loop of recursive self-improvement.
We're seeing that there are a handful of of great model companies, but the development of this technology is going to be a more normal technology race. It's not like the leading players just all of a sudden get achieve AGI just very quickly. And as a result of that, I I think because it is a more normal technology race, I think we can apply a more normal logic to it from both an investment and a policy standpoint.
And I think that a lot of the narratives that were hyped up about imminent doom or imminent utopia, depending on what side you were on, were just massively overhyped. And this is why I think it's just a very healthy thing. By the way, the other concerning thing 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.
Yeah, and I would just say that, you know, another just piece of evidence here is that the model companies are still leaprogging each other to some degree with their latest versions. And again, that should not be possible if one model achieved rapid takeoff. The other thing you're seeing is that it's not like one model is dominating all the other models.
You're seeing that models are developing areas of competitive advantage. They're becoming increasingly specialized. They have different personalities. If Chat GBT has been accused of being sick of fantish, Grock is maybe more based. I mean they have different modes that they excel at. Google I think has gotten really good at video for example.
Anthropic is is really good at coding. The specialization really belies this idea of one model becoming all knowing and all powerful. We're seeing a proliferation now of specialization. I think one of the key things that's being understood so you know the internet was being connected. there was connectivity and and fiber being deployed in a rate that exceeded the economic value creation by the end users on the internet during the dotcom boom.
And so as a result, the dot boom was a became a bubble that burst. But the ultimate promise of the internet was realized many orders of magnitude greater than I think you know most folks declared and proclaimed at the time. It was just delayed by a few years. I think it's likely to be the same here.
Whether this particular capex cycle is going to have that ROIC or not, I think is is a bit TBD. But I think there's kind of three big trends underway that are worth noting. The first is that there's a realization now that you need to pair humans and human engineering with the generative tools that these generative AI systems are providing.
So you can get a bunch of code written, but you still need to debug the code. You still need to get it into production. You still need to get it to integrate. So this kind of pairing of humans with AI is something that's being learned. What's the right model? It's not that you just turn on generative AI and it runs your business for you.
It's like where does it fit in the org? Who are the people that run it? How do they use it? Etc. And then similar to that, I think we're seeing a lot of partner models being developed. So rather than have like a Sora or a video model just generate a movie, it turns out that you can use the generative AI to generate the objects and then use an existing rendering engine like Unity to actually make the video.
So those sorts of pairing models where you use what's being called a deterministic system or an existing piece of software that integrates with generative AI can accelerate outcomes. So filmmakers can now make AI movies by coupling the AI generative tools with an existing offtheshelf kind of rendering system.
And that works a lot better than just telling the AI make a bunch of video where then all the people look different. You can't control the objects, you can't change the camera angle, you can't change the framing, the color, the lighting, all the features that you may want to otherwise control. So layering AI on top of the existing system seems to be the right answer.
And I think we're seeing that not just in film making or digital content creation, but in other areas as well where the AI provides almost like a substrate that can sit on top of existing tools and just be like a marionette and do a better job. And I think I think the third key thing that we're seeing that's causing a bit of the observed hiccup is this movement to SLMs.
And you guys have seen a bunch of these papers from Nvidia and others that have shown performance improvements by actually moving away from one large model that does it all to much smaller specialized models that work in a specific task in a specific application or that work together as a network of models.
And my personal belief on this is actually that these models are going to move towards SLMs that we don't understand how and why they work the way they do in a network fashion because today we're designing the SLMs in a deterministic way. We're saying this is an SLM that does X, Y, or Z. But if the models eventually get developed by other models, those SLMs will become really, you know, not well understood, but they'll work in a network.
The reason that's important is because this rearchitecturing of these systems dramatically reduces the energy cost and the dollar cost per token generated. And so it brings down the cost and the runtime of these systems. So to your point JCL, the capex investment suddenly gets much higher ROIC rather than the challenged ROIC that it's dealing with today because the systems are so energy and cost inefficient today.
So if the architectural redesigns work, then these systems become much more efficient, 10x, 100x more efficient, and I actually think at that point token production will go way up, not way down. The internet's infrastructure boom in the late 1990s far outpaced real world adoption. That mismatch sparked the.com crash, wiping out billions in paper wealth and leaving empty offices and overbuilt network.
Yet, in the decade that followed, those very networks fueled an explosion of innovation and growth that few had anticipated. What looked like a bubble at the time became the foundation for a technological revolution. This cycle revealed a pattern that has repeated in other emerging industries. Markets often overshoot in the short term, but excess investment builds the groundwork for long-term transformation.
The story of the internet wasn't just about speculative frenzy. It was about how misplaced capital created the infrastructure powering the digital economy. The crash didn't fail the idea. It failed those who lacked the patience or perspective to see beyond short-term losses. The long game rewarded those who understood that the true value of a paradigm shift often takes years to materialize.
As Chamath and others have noted, the.com era showed how speculation and overbuilding caused massive short-term pain while delivering the infrastructure for the next generation of innovation. The lesson isn't merely about bubbles. It's that technology adoption rarely aligns with investor timelines. History proved the crash was not the end of the internet story, but the messy beginning of a transformation that redefined commerce, communication, and culture.
Those who stayed the course captured the real upside, the bigger question today. Are we seeing similar patterns in emerging technologies with today's overbuilt platforms laying the foundation for tomorrow's breakthrough
Comments 0