Hey, this is Steve Eisman, and today we're going to have an interesting interview with Gary Marcus about AI. But before we get to the interview, I want to talk for a while and kind of lay the groundwork, because this is really not the issue du jour, it's the issue of the age. So I'm going to do an intro on the entire AI landscape. Think of it kind of as a mini master class on all of AI: who's benefited, who's gotten hurt, and what are the risks involved, and whether it's a bubble or not. You know, it's unusual in terms of markets where you can actually point to a date where a story actually began. And the AI story, from a markets perspective, began on May 25th, 2023, which was when Nvidia reported its first quarter numbers for 2023. The actual results were very strong. That's not what was important. What was important was that Nvidia predicted second quarter revenue of 11 billion, which was 50% higher than the expectations.
And as a result, on that day, the stock was up 24%, which for a market cap of that size was pretty astonishing. And that began the AI story. As you all know, Nvidia manufactures GPUs. But the ramifications for what Nvidia was talking about with respect to AI were far broader than just its own GPU chips. So the first group that everybody focused on were the hyperscalers: Meta, Google, Amazon, Microsoft, Oracle. These were the companies that were buying the chips from Nvidia and creating data centers for the future of AI. And the amount of money that is being spent is astonishing. Just as an example, Google spent $90 billion in 2025 and is going to spend $180 billion in 2026. Amazon is going to spend over 220 billion. And so for a while, at least for this part of the story, every time one of these companies reported and they would increase their capex budget, the stocks would go up, until the third quarter of last year when Oracle reported and announced that it had a backlog of 550 billion. And a few days later, everybody realized that at least half of that backlog was from OpenAI. And the stock, which had gone from 230 to 330, came down to below 200. The hyperscaler story now is much more complicated, depending upon the company. I think the market still tolerates Google and Amazon spending a lot of money. It has much less tolerance for some of the relatively smaller companies like Meta and Oracle increasing their capex budgets. So that story, which for a while was extremely positive for everyone, is much more mixed. But the story is even bigger than just Nvidia and the hyperscalers. Then you have companies like Arista and Cisco, which provide all the networking equipment that go into the data centers. Those have been huge stocks for the last several years, and other companies like Amphenol have benefited as well. I call these tech AI-adjacent, and those stocks still seem to be doing well. Another group which has done very well, and maybe the group that has done the best over the last six months to a year, are the CPU manufacturers, memory chips, the Microns of the world. Every data center that buys a GPU has to probably buy at least one CPU. So the demand for CPUs has gone up enormously, the prices have soared as there have been a shortage, and these stocks, which really make a commodity-type product, have gone insane. How long that's going to last, I don't know. But what's interesting is that the chip sector is now about 16 to 17% of the S&P 500. It's about half of info tech. A few years ago it was single digits. So a tremendous rally in these chip stocks. And then we come to the binding constraint: power. Data centers need power, and it is not easy to find it. And so on the one hand you have utilities, some of which have done well, which are providing the power to the data centers. But the more interesting stocks would be like a GE Vernova, which manufactures gas turbines. There are only three companies on planet Earth that actually make gas turbines: GE Vernova, Mitsubishi, and Siemens. The charts of all three look exactly the same. Then you have a company like Quanta, which builds new utility plants. You have a company like Bloom Energy, which is an alternative energy company, and you have nuclear energy, which has done exceptionally well over the last couple of years. And then there are some industrial companies that are involved. I'm thinking of an Eaton, which is an electrification company, Rockwell, which is an automation company. You know, whenever a data center gets built, they need the equipment from those two companies to build out the data center. So, you know, parts of the industrial sector are also AI-adjacent. So those are the positives, or mostly the positives. Now let's turn to the negatives: software, otherwise known as the SaaS apocalypse. So the software business model has been a wonderful business model for a very, very long time. It's actually very elegant. It's a SaaS model, software as a subscription service, where they charge quarterly, and so the revenue increases with every year. They raise prices a little bit and they increase the number of seats. However, AI has decreased the cost of software creation enormously. The moats around these companies are much lower, and people who own these stocks are petrified, justifiably or not justifiably. You know, right now it is so early in the story that you could say anything, and I don't think we'll have any real clarity in terms of the impact of AI on software for at least a year. Between now and then, you could say anything. And what's interesting is that when software companies report, even when they report good numbers, people sell the stocks. The joke I have is that software stocks go down on good news, bad news, and mediocre news. So software is just a group that some people are trying to bottom fish right now. The group has had a little bit of a rally off of the bottom. My personal taste is this narrative is very tenuous, and it's just a group that I would stay away from until there's a lot more clarity. And then there are the implications for private equity and private credit. Between 2018 and 2023, roughly, private equity went on a buying binge of software companies, and those deals were all funded by private credit. Companies like ServiceNow are down over 50% from their peak, which means that all those software companies that private equity bought are at least cut in half. And so as that stuff needs to get refinanced, their debt needs to get refinanced, private equity is either going to have to pony up more equity or walk away. And we won't know that for a couple of years now. But what's been happening with respect to private credit is retail investors have been redeeming because they've been freaking out. There's been some pushback of late politically on the entire data center story. The AI data centers are much, much bigger than your traditional data centers. They eat electricity. They eat water. They eat everything. And you're getting pushback from communities that, you know, do not want this in their backyard. You know, this whole AI story is only going to work with all these data centers being built. If these data centers don't get built or they get delayed, the whole AI story is going to slow pretty dramatically. And let me just finish up with: there is something of a circularity in the entire AI story, which is that when you look at the hyperscalers, a very, very high percentage — it's not clear, but probably at least 50% — of their future revenue is coming from just Anthropic and OpenAI, that are not public, that lose money, that have to keep funding themselves with venture capital. And as long as they can keep funding themselves with venture capital, the story will go on. But if venture capital ever pulls back at all from these companies, or if one of these companies fails or stubs a toe, that would be really bad because the entire ecosystem is really at this point dependent upon these two companies. So that's my summation of the kind of the AI landscape. And now we're going to talk to Gary Marcus, who's really an expert in this area, to find out what's really going on.
Hi, this is Steve Eisman. Welcome to another episode of The Real Eisman Playbook. Not a day goes by that someone doesn't have a comment — positive, negative, neutral — about AI. It dominates the economy. It dominates... it just dominates the markets. It's been dominating the markets for several years now. And today, to talk about all things AI, is someone that we've interviewed before: Gary Marcus, who's a retired professor from NYU of psychology and neuroscience. Gary, welcome back.
Thanks for having me again.
So let's just review what we talked about the first time, which is — why don't you just remind our viewers about your thesis about what I would call the limitations of LLMs.
So that actually goes back to my graduate work before LLMs existed. I was looking at how children learn language and looking at the neural networks that were popular at the time. They were ancestors to today's systems. But that got me interested in, like, how do these systems work? What are their strengths? What are their limits? And I found that the ancestors — and we can talk about the current systems — were good at memorizing things, good at kind of pattern matching, and not very good at generalizing more broadly to things they hadn't seen before. Later, in 1998, I did some, I think, foundational work describing what we now call the problem of distribution shift, which is to say you train a neural network on some data, and if the world it's immersed in remains similar, it does fine. But if you ask it something that's really different from what's there, you find out that its comprehension is limited, that it has trouble generalizing. And I've been thinking about that kind of stuff ever since. So when neural networks made a comeback in 2012, I said, 'Hey, you still got these problems to deal with.' And people just kind of blew me off, and I said, 'You're going to have problems.' In fact, in 2001, I pointed out the problem of hallucinations. And that has never been solved. And what I've heard over and over again for now a quarter century is, 'Oh, Gary, give us a little more data and we'll solve all these problems.' There's actually a whole bunch of different problems around reasoning and so forth, but it's been this constant dialectic for 25 years: 'Oh, just give us a little more data and it'll be solved.' And then they just repeat that claim and don't actually solve the problems.
So the people who run these massive LLM models, they basically what they do is they just keep making them bigger and bigger and bigger with bigger and bigger data sets, and it's supposed to get better and better and better.
Yeah. And that was sort of true until like December of 2024.
And what happened in December? What came out in December 2024?
A bunch of companies recognized privately and to some extent publicly was that they were no longer getting the returns that they expected from investing in much bigger models. So a few things have happened since then. One is that, for example, GPT-5 was disappointing and late, which is something that I predicted. And that was because just adding more data didn't really solve things. Another thing that's happened is some people have started to take different approaches. They're not loud about it, but they are. So Claude Code is actually an example of something that I've been advocating for a long time called neuro-symbolic AI. It takes a mixture of the kind of neural network approach that LLMs and other systems are, and adds to it some classic symbolic code, some classical AI.
I mean, good old-fashioned computer coding.
Exactly. Good old-fashioned AI is the famous phrase for GOFAI. So it's actually a mix of GOFAI, which is just really just computer programming, and the neural networks. What the big companies wanted the investors to believe was, we have this single magic bullet that is going to get us to AGI. We're just going to add —
Artificial general intelligence.
To artificial general intelligence. It's going to be able to solve everything. It's going to stop making the stupid errors that Gary Marcus says they make, and so forth. And it didn't really work. So some people are trying other approaches. Some people are still kind of banging their head against the wall and trying to find nice demos when it does work, and so forth. But the scaling isn't really carrying the progress anymore. What's really carrying the progress mostly right now is what we call tools, or what they call tools and harnesses. Those are actually symbolic AI. And this matters partly for, like, technical scientific reasons, like you just want to know what the architecture is, but it also matters for an economic reason. The economic reason that it matters is that scaling is this thing that seemed to be predictable, and it's where all the money is going, but it's not really where the results are coming from right now. And so if you're a VC, you want to say, you know, 'Give me this massive amount of money to invest. I'll get my 2% on that and everything will be great.' But really, that's not what's actually driving progress anymore. That makes things less predictable.
Why do you say it makes things less? I mean, I understand if the model is just scaling, then the predictability is you make it bigger, supposedly it gets better. So now you make it bigger, but now you got to add some extra sauce to it. But why does that make things less predictable?
I mean, first of all, the pure scaling isn't helping that much, right? And we may enter a new regime — and I'll give you a hint of why this might already be happening — where just building these massive data centers doesn't really pay off and isn't really the right way to go. So everybody has their bet kind of specifically on scaling more chips, etc. So here's an example: CPUs are actually making a little bit of a comeback. Intel's doing all right because agents actually need a lot of CPU time in addition to the GPU time. All right, here's something else that might happen: is people might say the real action is on these tools. They don't really need the best, biggest, fanciest model for this. And then you wind up with these data centers maybe people don't need. You know, maybe, for example, the answer turns out to be that, like, medium-sized models plus a bunch of symbolic AI actually is what you need rather than some giant model. And you know, if people were only putting modest amounts of money on these bets like they used to be, maybe it just wouldn't matter. But when people are making, like, literally two-trillion-dollar bets on the GPUs as scaling being the answer, if it turns out to be anything else, a lot of people are kind of screwed. Then you have these massive data centers that nobody uses. You have people that are, you know, making loans to these companies that might not get paid off, and so forth.
The fact that what's been created so far is not AGI doesn't — you know, from an investor perspective, my perspective is, you know, like, you know, so okay, so we didn't get Terminator yet, so thank God. We don't want Terminator. My issue with all this stuff is this is not my area of expertise. So, you know, when I was dealing with subprime and the financial system, that was my area of expertise. I knew what to look for to say this is bad. The fact that this stuff has not created AGI — let's get that, that's a given. But people are using these products right now. So my question for you is, as someone who really knows the field, like, what is this stuff good for right now? What are its strengths and what are its weaknesses as it exists today? Because if you would listen to CNBC, they're waxing poetic about, like, every single innovation that comes out that gets announced, whether it's good or bad, because they don't know. They're not involved.
Well, they were nice enough to have me on earlier today. But so the first thing to note there, before I kind of dig into your question, is your concession that we have not reached AGI is actually a huge zeitgeist shift. Right? We had for several years people insisting that AGI was about to be here tomorrow.
And people finally backed off on that. You know, especially people thought, I think, that GPT-5 was going to be AGI. Did they really think that?
I think a lot of people did. I mean, at least if you, like, read the social media and stuff like that, people really talked about that. There was a figure that Kevin Scott, who's the CTO of Microsoft, showed where there was, like, this little — what was it? — like, tiny little whale was GPT-3, and then a, like, a little bit bigger whale was GPT-4, and then this giant whale, GPT-5.
So his notion was this is going to be a tipping point. It's going to be, you know, hugely better.
And it wasn't. And Altman had talked about AGI constantly, including in January of last year when it finally came out. He was like, 'We now —' I can't quite quote from his blog, but I can come really close. He said something like, 'We now know how to build AGI in a conventional sense.'
And when did he say this?
This was in January of 2025, before GPT-5 came out. About eight months before GPT-5 came out. And I gave him a hard time on Twitter about it, and so forth. And I gave Kevin Scott a hard time. I said, 'This is not really true.' And it wasn't true. But a lot of all of the hype came around the idea that we were going to get to AGI. And part of that connects to what we really need right now is maybe not magical infinite intelligence, but we want intelligence we can trust that does not hallucinate, that you know the answer is right. You know, you type something into a calculator, you expect the answer to be right.
You know, there was once a scandal when Intel made a chip — I think it was the Pentium 4 — had a bug in its arithmetic. Never happened in integer arithmetic. It only happened, like, once in 10 trillion times or 10 billion times in floating-point arithmetic. It was very rare, and it was a scandal that, you know, nearly broke down the company when it cost them $500 billion back when that was real money.
Right.
The current systems aren't anywhere near that reliable. And now that allows me to circle back to your question.
Okay.
Hi, Steve Eisman here. If you've been a Real Eisman Playbook follower for a while, you know that I recently finished treatment for cancer, and I feel very lucky to be in recovery. A cancer diagnosis is a lightning bolt I did not see coming. Breast cancer as a man was definitely not in my life's bingo card. As a podcaster and a cancer patient, I have learned that too many people affected by cancer don't have access to the care they need. This is why I feel very strongly about the most recent sponsor for our show, Color Health. Color has created the first ASCO-certified virtual cancer center of excellence that is available all over the nation. Certification by the highly esteemed American Society for Clinical Oncology matters a lot. Color's goal: reinvent how we all manage cancer, from screening to diagnosis, through treatment, and all the way back to the life you thought you lost. You no longer have to live near the world's top cancer centers. Color has created the first cancer center of excellence that is available all over the nation. The best possible care for everyone, everywhere. Color works with employers, unions, and health plans to lead the way through the human, clinical, and financial impact of cancer. You've all seen some of my recent episodes about the state of American health care. Getting the care you need is too often unaffordable. Almost half of all cancer patients lose their entire life savings within two years of being diagnosed. Catching cancer early saves lives and money. This is why Color's work is so important. The facts are that Color delivers the speed needed to get from suspecting cancer to treating a cancer 66% faster than the average in-person clinic visit. As a virtual cancer clinic, it's accessible across all 50 states, 24/7, working with local providers. If your rising cancer costs for your company feel out of control, you have to check out Color, because there is a better way to get strong member use, better health outcomes, and real control of cancer costs. Learn more about Color as a partner to overhaul your cancer strategy at color.com/eisman. And while you're at it, get screened.
So again, I'll repeat the question, which is: okay, it's not AGI, but people are using it. So from what you've — I'm sure you've used it — and from people you've talked to who use it, like, what's it good for and what it's not so good for?
I don't use it that much. My famous line on Twitter is, 'I don't pay for sex and I don't write with ChatGPT.' But many, many people use it for — I mean, I just don't happen to have a lot of uses for it. But look, there are some things it's great for, and especially Claude Code is really great for coding.
Yeah. Now, I have coded often in my life. I don't happen to have a need for it right now, so I haven't used it for that. But I've seen demos, and you know, it's clearly quite good. And coding has been a strength of these systems all along. Claude Code is really, I think, a quantum leap in the ability to use these systems for coding. There's still things to worry about. We could talk about that, but it's real.
Coding good. Not perfect, but good. And maybe, you know, oversold a little, but really genuinely good. Brainstorming has been a strength of these.
What do you mean? What does that mean?
So, you're like — especially maybe if you're not an expert in a field, you're a non-expert, and you want, like, 10 ideas about a marketing pitch I can do for my little small business, and you don't have the money to go to Madison Avenue to get the best, but you just want some ideas. Like, it's pretty good for that.
Sometimes it can be pretty good for, like, teaching a student a new concept and things like that. Though there are some errors and there are some things to worry about. Customer service — I think there's some data that shows that it's better than the worst people at customer service. It's not as good as the best people. You can imagine a situation where, like, some customer service is better than none, and you don't really have the money to do it right. Unfortunately, I think we live in that world, and there's a lot of bad customer service now, some of which is done by AI. But you know, it's at least a possible application.
You're not going to get the best out of it. And now it turns out that in domains where you can generate a lot of data because they're very formal, that these systems are somewhat useful. So —
Like, give me an example.
If you have something that can be formally verified, which includes mathematics and to some extent coding, then you can do two things. One is you can generate a lot of synthetic data. Sometimes people call it fake data, but it's real. You can generate new data, not just the data that you observe in the natural world. But for example, if you wanted to train one of these things on multiplication, you could use a symbolic system to make a bunch of multiplication examples where you know the answers are right, and then you enhance the training that way.
