Are We at the End of Ai Progress? — With Gary Marcus

Great to see you.

Thanks for having me.

Well, I mean, I have to laugh because I wrote a paper in 2022 called "Deep Learning Is Hitting a Wall." And the whole point of that paper is that scaling was going to run out, that we were going to hit diminishing returns. And everybody in the field went after me. A lot of the people you mentioned, I mean, LeCun did, Elon Musk went after me by name, Altman did, and they all like Altman said, "Give me the strength of a mediocre deep learning skeptic." So, people were really pissed when I said that deep learning was going to run out. So it's amazing to me that a bunch of people have conceded that these scaling laws are not working the way they used to be, and they're also doing a bit of backpedaling. I think that that Mark Chen interview, I can't quite remember the details, but I think it was a version of backpedaling and redefining things. So if you go back to 2022, there were these papers by Jared Kaplan and others at OpenAI, and they said, look, we can just mathematically predict how good a model is going to be from how much data there is. And then there were the so-called Chinchilla scaling laws. And everybody was super excited. And basically people invested half a trillion dollars assuming that these things were true. You know, they made arguments to their investors or whatever. They said, if we put in this much data, we're going to get here. And they all thought that here in particular was going to mean AGI eventually. And what happened last year is everybody was disappointed by the results. So we got one more iteration of scaling after 2022 that worked really well, and we call that GPT-4 and all of these models that are sort of like that. So I wrote that paper around GPT-3. We got another iteration of scaling. So right, three was scaling compared to two, it was much better. Two was scaling compared to one, it was much better. So much more data meant much better. But what is much better? Well, I mean, one way to think about it is you didn't need a magnifying glass to see the difference between GPT-2 and, we didn't call it GPT-1, but the original GPT. And you didn't need a magnifying glass for GPT-4 as opposed to GPT-3. It was just obviously better. A lot of people thought that we would pretty quickly see GPT-5, and a lot of people raced to build it. So, OpenAI tried to build GPT-5, and they had a thing called Project Orion, and it actually failed and eventually got released as GPT 4.5. So, what they thought was going to be GPT-5 just didn't meet expectations. Now, they could slap any name on any model they want. And in fact, lately, nobody understands how they're naming their models. But they haven't felt like any of the models that they've worked on since GPT-4 actually deserve the name GPT-5, and it didn't meet the performance that these so-called mathematical laws required. What I said in that paper is they're not really mathematical laws. They're not physical laws of the universe like gravity. They're just generalizations that held for a little while. Like a baby may double in weight every couple of months early in its life. That doesn't mean that by the time you're 18 years old that you're going to be 30,000 lbs. And so we had this doubling for a while and then it stopped. And we can talk about why, but the reality is that's not really operative anymore. So there's been efforts to kind of misdirect and shift direction. So I think everybody in the industry, quietly or otherwise, acknowledged that, hey, we're not getting the returns that we thought anymore. And nobody's been able to build a so-called GPT-5 level model. That's a big deal, right? I'm a scientist, and as a scientist, or was originally a scientist. As a scientist, we have to pay attention to negative results as well as positive results. So when 30 people try the same experiment and it doesn't work, nature is telling you something. And everybody tried the experiment of building models that would 10x the size of GPT-4, hoping to get to something they could call GPT-5 or that was like a quantum leap better than GPT-4. They didn't get there. So now they're talking about scaling inference time compute. That's a different thing.

So let's clarify two things. One is that what people talked about about scaling originally was a mathematically predictable relationship between performance and amount of data. You can go back and look at the Chinchilla paper, the Jared Kaplan paper, and lots of things that were posted on the internet. There were papers saying, or T-shirts saying, "Scale is all you need." You looked at that T-shirt and it had equations from the Jared Kaplan paper and it said, you know, here's the exponent. You can fit the equation. If you have this much data, this is the performance you're going to get. And there were a bunch of papers, a bunch of models that actually seemed to fit that curve, but it was an exponential curve. And what's happening now is, yeah, you add more data, you get a little bit better, but you're not fitting that curve anymore. We've fallen off the curve. That's what it really means to say that scaling isn't working anymore, is you don't, you know, if I drew a curve for you, it was going up and up and up really fast, and it's not going up as a function of how much data you had, or how much compute you had. So you added a bunch of compute and you got this much better performance. And this is how people justified running these experiments that cost a billion dollars, is they're like, I know what I'm going to get for the billion dollars. And then they ran the billion-dollar experiments and they didn't get what they thought they would. Yeah, you get a little bit better, but that's what diminishing returns means. Diminishing returns means you're not getting the same bang for your buck as you used to. That's where we are now. So anytime you add a little piece of data, the model is going to do better, excuse me, on that piece of data. But the question is, does it generalize and give you significant gains across the board? And we were seeing that, and we just aren't anymore.

Well, let's look at what Elon got for his money, right? So, he built Grok 3, and by his own testimony, it was 10 times the size of Grok 2. It's a little better, but it's not night and day, right? Grok 2 was night and day better than the original Grok. GPT-4 was night and day better than GPT-3. GPT-3 was night and day better than GPT-2. Grok 3 is like, yeah, you can measure it. You can see that there's some performance, but for 10x the investment of data, compute, and not to mention cost of energy to the environment, it's not 10 times smarter by any reasonable measure. It just isn't. Okay? And so this would be the point where I say, "Well, then this entire AI moment is done." However, well, it's this moment. There will be other AI moments, but this one I know I'm setting it up to say that it's not, because like you mentioned, you're talking about test time compute. That's another way to say reasoning, I think, which is these models... Well, I'm going to give you a hard time about people do do that. But with reasoning or test time compute, you'll help me figure out the finer details. What these models are doing is they're coming to try to find an answer and they're checking their progress and deciding whether it's a good step or not and then taking another step and another step. Yeah. And we've seen that they have been able to perform much better when you put that reasoning capabilities on top of these large models, which has enabled these research houses to continue the progress and give you... but it's not really you, it's these companies.

So it is true that you can build a model that will do better if you put more compute on it, but it's only true to some degree. So then I'll get to whether it's actually reasoning or not, but it turns out that on some problems you can generate a lot of data in advance, and for those problems, adding more test time compute seems helpful. There was a paper this weekend that's calling some of this into question.

That's right. So, you have these models now like o3 and o4 that will sometimes take like 30 seconds or 5 minutes or whatever to answer a question. And sometimes it's absurd because you ask it like, what's 37 times 11, and it takes, you know, 30 seconds. You're like, my calculator could have done it faster. But we'll put aside that absurdity. In some cases it seems like time well spent, sometimes not. But if you look carefully, the best results for these models are almost always on the same things, which are math and programming. And so when you look at math and programming, you're looking at domains where it's possible to generate what we call synthetic data. And to generate synthetic data that you know are correct. So for example on multiplication, you can train the model on a bunch of multiplication problems and you can figure out the answer in advance. You can train the model what it is that it should predict. And so on these problems in what I would call closed domains where we can do verification as we create the synthetic data, we verify that the answer we're teaching the model is correct, the models do better. But if you go back and you look at the o3, sorry, the o1 paper, even then you could already see that the gains were there and not across the board. They reported that on some problems o1 was not better than GPT-4. It's only on other problems, these cut and dry problems with the synthetic data, that you actually got better performance. And I've now seen like 10 models, and it always seems to be that way. We're still waiting for all the empirical data to come in, but it looks to me like it's a narrow trick that works in some cases. The amazing thing about GPT-4 is that it was just better than GPT-3 on almost anything you could imagine. And GPT-3, the amazing thing is it was better than GPT-2 on almost anything you can imagine. Models like o1 are not systematically better than GPT-4. They're better in certain use cases, especially ones where you can create data in advance. Now, the reason I wouldn't call them reasoning models, though you're right that many people do, is what I think they're doing is basically copying patterns of human reasoning. They're getting data about how humans reason certain things, but the depth of reasoning there is not that great. They still make lots of stupid mistakes all the time. I don't think they have the abstractions that we think, for example, a logician has when they're reasoning. So it has the appearance of reasoning, but it's really just mimicry, and there's limits to how far that mimicry goes. I give you just one more example is o3 apparently hallucinates more than the models that came before it, which is stunning. Like, how does that happen?

Well, I was going to give you a different one, which is black box.

But they're closely related, those two terms. You need interpretability to figure out what's going on in the black box, if you can at all.

Well, that's a different thing, right? So a black box in a plane is actually a flight recorder that records a lot of data. But what we mean in machine learning by black box is you have a model where you have the inputs and you have the outputs. You know how you calculate them, but you don't really understand how the system gets there. So in this case, you're doing all this matrix multiplication. Nobody really understands it. And so nobody can actually give you a straightforward answer for why o3 hallucinates more than GPT-4. We can just observe it. That's what happens with black boxes is you empirically observe things and you say, "Well, it does that," but you don't really know why and you don't really know how to fix it either. Another example just in the last couple days is apparently Sam Altman reported, I forget, the new model is stubborn or what was it? I forget.

It's a bro, but that's GPT-4o. It's just like it became very fratty.

Yeah, exactly.

No, no. Sam said it's now the latest iteration. It's been completely rolled back.

So, right, that was what I would call again empirical, like they tried it out and it didn't work or it worked in the way that it irritated people. Right. And so we don't know in advance. Like there's a lot of just like try it because that's how black boxes work and we have some things but those things are not very strong. So the scaling quote laws were empirical guesses about how these models worked and they were true for a little while which was amazing and they're not true anymore which is also amazing in a way. So we don't know what's going to happen from the black boxes, right? Okay. So, but let me now sort of, and sorry, let me come back to one other thing quick, which is interpretability. So, that's a very closely related notion. So, let's say you look at a GPS navigation system. That's a piece of AI that's very interpretable. So, you can say it is plotting this route. It says, you know, you can go this way, you can go that way. This is the function that it's maximizing. This is the database it's using. This is how it looks up the data. We don't have any of that in these so-called black box models. We don't really know what the database is that it's consulting. It isn't exactly consulting a database at all, and we don't know how to fix it.

And I think he's right. I've said this too myself. Like in my last book, I talked about interpretability being really important. The only difference between Dario and me on this point is we both think that we're screwed as a society if we stick with uninterpretable models. He just thinks that LLMs will eventually be interpretable. And his company, to be fair, has done the best work on interpretability of LLMs that I'm aware of. Chris Olah, I think, is brilliant, but they haven't got that far. They've gotten further than anybody else, but I don't think we're ever going to get very far into the black box. And so I think we need to start over and find different approaches to AI altogether.

So, I mean, there's a couple different ways to think about that. So one is on a lot of benchmarks there's improvements, but there's also issues of data contamination. Alex Reisner wrote an excellent piece in The Atlantic about the issues of data contamination, and we've seen a lot of studies where people are like, "Well, we tried it in my company, it's not really that much better." So they're better on the benchmarks. Are they better in general? Not so clear. There was a new benchmark released by a company called Val AI or something like that, Val AI, that The Washington Post talked about yesterday, where they looked at things like, can you pull out a chart based on a series of financial statements, SEC statements from a bunch of companies, and these systems all claimed to do it, but accuracy was under 10%. And overall on this new benchmark, accuracy was at 50%. Would these new models be better than GPT-4? Maybe. They weren't that good. So I think people tend to notice when they do well, they don't notice as much when they do poorly. And although I think there's been some improvement, there has not been the quantum leap that people are expecting. We have not moved past hallucinations. We have not moved past stupid reasoning errors. If you go back to my 2022 paper, "Deep Learning Is Hitting a Wall," I didn't say there'd be no progress at all. What I said is we're going to have problems with hallucinations, we're going to have problems with reasoning, with planning, until we have a different architecture in some sense. And I think that that's still true. We're still stuck on the same kinds of things. So if you have, you know, deep research write you a paper, it's going to make up preferences. Okay? It's probably going to make up numbers. Like, you know, did you actually go back and check? So for example, with, I think it's called, they all have similar names now, whatever Grok's version is, deep search or deep research.

Deep research. I'll be convinced that we have AGI until these companies learn how to call deep research something other than deep research. They all use the same exact name. It's really bizarre. So whichever version Grok has, I asked it, for example, to list all of the major cities that were west of Denver. And to somebody who wasn't paying attention, it'd be super impressive. But because I really wanted to know how well it was working, I checked and it left out Billings, Montana. All right. So, you got a list that looks really good and then there are errors. This often happens. And then I had a crazy conversation with it after that. I said, "What happened to Billings?" And it said, "Well, there was an earthquake there on February 10th" or whatever. And I looked up the seismological data. I use Google because I want to have a real source, or DuckDuckGo, and there was no earthquake then. And I pushed it and it said, "Well, I'm sorry for the error" or whatever. So, we're still seeing those kinds of things. We may see them less, but they are still there. We still have those kinds of problems. So, I don't doubt that there's been some improvement, but the quantum leap across the board that people were hoping for is not there. The reliability is still not there. There's still lots of subtle errors that people don't notice. And then, you know, if you want to talk to me about retirement calculators, there are a lot of those on the web. So, the easy cases for these systems are the ones where the source code is actually already there on the web. Like Kevin Roose talked about this example of having, he quote vibe coded a system to look in a refrigerator and tell them what recipe to make. But it turns out that app is already there on the web. And there are demos of that with source code. And so like if you ask a system to do something that's already been done, that's always been true with all of these systems. That's their sweet spot is regurgitation. And so, yeah, they can build the stuff that's out there, but if you want to code things in the real world, you usually want to code something that's new. And these systems have a lot of problems with that. And another recent study showed that they're good at coding, but they're not good at debugging. And like coding is just the tiniest part of the battle, right? The real battle is debugging things and maintaining the code over time. And these systems don't really do that yet.

I mean, good on you that you go to the primary sources. I worry the most about people who don't. And we've seen countless lawyers, for example, get in trouble using these systems.

Oh, no, no, no. There's many more than that. There's some in the US. There's some in Canada. I think there was just one in Europe. I mean, it's not really countless. One could sit there and count, but it's got to be at least a dozen by now.

Well, Microsoft did a study, in fact, suggesting that critical thinking was getting worse as a function of them. And that wouldn't be too surprising. We have a whole generation of kids who basically rely on these bots and who don't really know how to look at them critically. You know, in previous years, we were starting to get too many kids relying on whatever garbage they found on the web, basically. And I mean, chatbots are basically synthesizing the garbage that they find on the web. And so, we're not really teaching kids critical thinking skills. And nowadays, like the idea for many kids of writing a term paper is I typed in a prompt in ChatGPT and then maybe I made a couple edits and I turned it in. You're obviously not learning how to actually think or write in that fashion. A lot of these tools, I think, are best used in the hands of sophisticated people who understand their limits. So, you know, coding has actually been, I think, one of the biggest applications, and that's because coders understand how to debug code, and so they can take the system, basically it's just typing for them and looking stuff up, and if it doesn't work, then they can fix it. Right. The really dangerous applications are like when somebody asks for medical advice and they can't debug it themselves and, you know, something goes wrong.

It's not maxed out, but there's diminishing returns.

Well, wait a second. I'm not saying people are rational. I think that people will probably try at least one more time. They'll build things. You probably, Elon will build something that's 10 times the size of Grok 3, which will be huge, and it will, you know, it will have a serious impact on the environment and so forth. I just don't think...

Let's come to the data separately in a second. So I think people will actually try, right? I think Masayoshi Son has just bankrolled Sam to try. I just don't think they're going to get that much for it. I don't think they'll get zero. I mean, there will be tangibly better performance on certain benchmarks and so forth, but I don't think that it's going to be wildly impressive, and I don't think it's going to knock down the problems of hallucinations, boneheaded errors.

Not the kind of reasoning they're doing. No, there's no emergent coding.

I mean, look, I put out some predictions last year in March that people can look up, they're on Twitter, and those predictions include, I said there'd be no GPT-5 this year, or if it came out, it would be disappointing.

Well, this was last year. So I said in 2024 we won't see this. And that was a very contrarian prediction at that point, right? This was a few weeks...

After people had said, 'Oh, I bet GPT-4 is going to drop right after the Super Bowl, like right after the Super Bowl. Won't that be amazing?' So people really thought it was going to come last year if you go back and look at, you know, what they said on Twitter, etc. And it didn't. And I correctly anticipated that it wouldn't. And I said we're going to have a kind of pileup where we're going to have a lot of similar models from a lot of companies. I think I said seven to 10, which was sort of roughly right. Um, and I said we were going to have no moat because everybody's doing the same thing and the prices were going to go down. We'd have a price war. All of that stuff happened. Now, maybe we get to so-called GPT-5 level this year. Keeps getting pushed back. Um, I don't know if we'll get much further than that without some kind of genuine innovation. And I think genuine innovation will come. But what I think is we're going down the wrong path. Yann LeCun used this notion of, you know, we're on the exit ramp. How do you say it? Large language models are the off-ramp to AGI. You know, they're not really the right path to AGI. And I agree with him. Or you could argue he agrees with me because I said it, you know, for years before he did, but we won't go there. The broader notion is sometimes we make mistakes in science. I think one of the most interesting ones was people thought that genes were made of protein for a long time. So the early 20th century, lots of people tried to figure out what protein is a gene made of. It turns out it's not made of a protein. It's made of a nucleic acid that everybody now knows called DNA. So people spent 15 years or 20 years like really looking at the wrong hypothesis. I think that giant black-box LLMs are the wrong hypothesis. But science is self-correcting. In the end, if people put another $300 billion into this and it doesn't get the results they want, they'll eventually do something different, right? But what you're forecasting is basically an enormous financial collapse because... That's right. I don't think LLMs will disappear. I think they're useful, but the valuations don't make sense. I mean, I don't see OpenAI being worth $300 billion. And you have to remember that venture capitalists have to like 10x to be happy or whatever. Like, I don't see them, you know, IPOing at $3 trillion. I just don't.

They're a genuinely good company. They have a wonderful ecosystem. They're worth a lot of money. I mean, I don't I don't want to put an exact figure, but I'm not surprised that they fell and I'm not surprised that they're still worth a lot.

So, they have multiple problems, both OpenAI and Nvidia. So, one is it does look to me like we're hitting diminishing returns. It does not look to me like this inference-time compute trick is really a general solution. It doesn't look like hallucinations are going away. And it does look like everybody has the same magic formula. So everybody's basically doing the same thing. They're building bigger and bigger LLMs. And what happens when everybody's doing the same thing? You get a price war. So DeepSeek came out and OpenAI dropped its prices quite a bit, right? And so every because everybody I mean not literally everybody but you know 10 20 different companies all basically have the same idea and are trying the same thing. You have to have a price war. Nobody has a technical moat. OpenAI has a user moat. They have more users and that's that's the most valuable thing they have like for them. API, I would say the API is close to worth. I don't know worth worthless is the right word but it's worth it's not worth very much. It is the thing. ChatGPT is the thing that that really has that it's the brand name that is most valuable. I also think it's the best bot right now. It might be, I mean, I think people go back and forth. Some people someday say it's Claude. I've been on the Claude train for a long time. Um, now you're on the and I'm on ChatGPT. I think what I think is going to happen is you have leapfrog, but the leaps aren't going to be as big as they were. So 4 was a huge leap. I mean, this is a different way of saying what I said before was huge leap over 3. You know, let's say I can't even keep up with the naming scheme. GPT-4.1, right? Let's say is better than Grok 3 or or Claude 3.7, let's just say hypothetically. And so people run to this side of the room and then, you know, Claude whatever 3.8.1 or whatever will be a little better and then some people will run to that side of the room. Um, but nobody's going to be able to charge that much money because the the advances are going to be smaller and people start to say, well, you know, I use this one for coding and this one for brainstorming and whatever. But nobody anymore says this is just like dominant. Like GPT-4 was just dominant when it came out. There was nothing as good as it for anything. If you wanted this kind of system, you used it, right? I mean, that's my memory of it. I don't hear any of the the ChatGPT or whatever, I can't even keep up with the names anymore, any of those products, any of the OpenAI products being referred to in the same kind of hushed tones like they're just better. And like, you know, Google's still in this race and they may undercut on price. Meta is giving stuff away. People are building on it. DeepSeek, I hear, has something new that's going to, you know, be better than ChatGPT. Um, and you know, maybe it's true, maybe it's not, but we we were we're in this era where the the differences between the models are just getting really small.

Which things? Which things?