Microsoft CEO's Stunning Statement "AGI Is Nonsense"

So Satya Nadella actually made a statement on the Dwarkesh Patel podcast where he said something that ruffled some feathers. He actually stated that the definition for AGI is one that is quite useless, and he thinks the real definition is a little bit different. So in this video, we'll dive into exactly what he thinks about AGI and what the entire industry has to say in reaction to that. So let's take a look at what he says in this first clip here, where he basically says that the definition for AGI is essentially one that would be constantly changing in practice.

One thing I'm not sure about hearing your answers on different questions is whether you think AGI is a thing in the sense of will there be a thing which automates all, at least starting with all cognitive labor. Anything that anybody can do on a computer.

See, this is where my problem with the definitions of how people talk about it is cognitive labor is not a static thing. There is cognitive labor today. If I have an inbox that that is managing all my agents is that new cognitive labor. And, and so today's cognitive labor may be automated. What about what is the new cognitive labor that gets created? Both of those things have to be thought of, right? Which is the shifting. So that's why I think this distinction, at least in my head I make is don't conflate knowledge worker with knowledge work. The knowledge work of today could probably be automated. Who said, my life's goal is to triage my email, right? Let an AI agent triage my email. But after having triaged my email, give me a higher level cognitive labor task of, hey, these are the three drafts I really want you to review. Like, that's a different abstraction.

And so if you didn't understand what he said there, he basically said that, look, this AGI benchmark thing doesn't really make sense. One of the key things he actually talks about here is that cognitive labor isn't something that is, you know, static. So cognitive labor is something that is constantly dynamic. And what that means is essentially the cognitive labor of today could be very different from the cognitive labor of tomorrow. And that also means that essentially we need to be analyzing what kind of tasks need to be automated. And of course, that basically means that if AGI is something that automates all tasks, that is going to be something that is constantly changing. So the definition of AGI is also going to be constantly changing. And I think it's really important to make that distinction because AGI is genuinely a vague term. And I would say that across the Board. We have different definitions for different people now. And I will say that Sam Altman does have his definition here, which is pretty similar to Satya Nadella's in the sense that it needs to provide some real world value.

I think at this point we're kind of close enough that the precise definition of AGI matters. You know, by the time you have a world expert in every field working together tirelessly, I think that's beyond what most people would consider AGI.

So.

So I think rather than talk about, you know, when we're going to get there or whatever, I, or yeah, I. Sorry, that was really inarticulate. I would say. I think we'll get to something in the next couple of years that many people will look at and say, I really didn't think a computer was going.

To do that now. Satya Nadella wasn't done. We also have another area which Satya Nadella spoke about and he spoke out. The fact that if we want to look at AGI as an actual definition, his definition for AGI is one that is rather different. He basically says that, look, if we want to call ourselves the country or the company that created AGI, we'll know we have AGI when we have 10% economic growth year on year. And I think the reason he's made that distinction is because so many times AI companies are focused on benchmark hacking. And of course there are incentives to do that so that people can see how good the model is. But what companies aren't focused on is of course bringing real world value to, to real world customers and real world companies. And if you can do that, I then think those metrics are the ones that matter more. And that's why this benchmark of having 10% economic growth, I think that one matters more than just a vague definition of something that can automate a lot of tasks.

Us self claiming some AGI milestone, that's just nonsensical benchmark hacking. To me, the real benchmark is, is the world growing right at 10%, the.

World economy is 100 trillion or something. The world grew at 10%. That's like extra 10 trillion in value produced every single year. If that is the case, it seems like 80 billion is a lot of money. Shouldn't you be doing like 800 billion? If you really think in a couple of years we could be really growing the world economy at this rate? And the key bottleneck would be, do you have the compute necessary to deploy these AIs to do all this work?

The classic supply side is, oh, Let me build it and they'll come, right? I mean that's an argument and you know, after all we've done that we've taken enough risk to go do it. But at some point the supply and demand have to map. You know, you can go off rails completely when you're like all hyping yourself with all the supply side versus really understanding how to translate that into real value to customers. And you're not going to say oh, they have to symmetrically meet at any given point in time, but you need to have existence proof that you are able to parlay yesterday's, let's call it capital into today's demand so that then you can again invest maybe exponentially even knowing that you're not going to be completely rate mismatched.

And so what he was actually talking about there is the fact that of course right now these companies are spending billions and billions of dollars on AI infrastructure. And he's arguing that, look, whilst yes it makes sense to spend money on AI infrastructure, we have to be wary of the fact that maybe, just maybe we could potentially be overspending. And we need to at least have some proof that the AI that we currently have is going to provide value to the customers of tomorrow. And so that when we're building out all of this infrastructure, we're not going to be in a situation where companies have overspent and we have too little demand for what we have. I think it does make sense for companies to have the build out now. I think it does make sense considering the implications of AGI and esi. But, but it will be interesting to see exactly how things are moving forward with regards to investments.

I wonder if there's a contradiction in these two different viewpoints. Because look, I mean one of the things you've done wonderfully is you make these early bets when there's, you know, you invested in OpenAI in 2019 even before there was copilot and any applications. If you look at the industrial revolution, these, you know, 6, 10% build outs of railways and whatever things, many of those were not like we've got revenue from the tickets and now we're going.

To a lot of money lost.

That's true.

The so if you, if you really think like this, there's some potential here to 10x or 5x the growth rate of the world. Shouldn't you just like let's go crazy, let's do the hundreds of billions of dollars of compute.

It's not about building compute, it's about building compute that can actually Help me not only train the next big model, but also serve the next big model. And you understand until you do those two things, you're not going to be able to really be in a position to take advantage of even your investment. So you have to have a complete thought, not just one thing that you're thinking about.

Now, if we do go back to the AGI definition, I do think it's always good to refer to OpenAI's internal documentation that was leaked through Bloomberg. Essentially, this was what they have. Of course, I'm pretty sure you're familiar with this, but currently they are at level three, which is, of course, agents and systems that can take actions. And I think by next year, the agents are going to get really, really good when we have a bunch of different workflows. And that's something that I'm personally working on in my AI academy. It's going to be linked in the description, but I'm working on tons, tons of AI agent workflows that actually allow you to automate many different tasks. But the point here is that we have five levels, and the level four is of course, Innovators AI, the Canadian invention, and of course level five, which is probably going to be many, many AI systems together doing the work of an organization. And you might be asking, okay, so we have the AGI definition. How do other CEOs think about that and what do they think about AGI? Well, recently on a podcast, the Anthropic CEO actually spoke about this, and he said he thinks that AGI will be here by 2027.

Extrapolate the curves that we've had so far, right? If, if you say we extrapolate the curves that we've had so far, right? If, if you say, well, I don't know, we're starting to get to like, PhD level. And last year we were at undergraduate level and the year before we were at like the level of a high school student. Again, you can, you can quibble with at what tasks and for what. We're still missing modalities, but those are being added. Like computer use was added. Like image in was added. Like image generation has been added. If you just kind of like, and this is totally unscientific, but if you just kind of like eyeball the rate at which these capabilities are increasing, it does make you think that we'll get there by 2026 or 2027. Again, lots of things could derail it. We could run out of data. You know, we might not be able to scale clusters as much as we want like, you know, maybe Taiwan gets blown up or something and you know, then we can't produce as many GPUs as we want. So there, there are all kinds of things that could, could derail the whole process. So I don't fully believe the straight line extrapolation, but if you believe the straight line extrapolation, you'll, we'll get there in 2026 or 2027.

Now interestingly, while doing research for this video, actually found some people who are very different in terms of their opinion. And I don't want to say this to sort of provide any kind of divisiveness, but I do find it very interesting that companies that are raising significant amount of money often have shorter timelines than the ones that are more established. So for example, companies like OpenAI are saying AGI is nearer and not that I think AGI is far. I do think the incentive for them is to say that AGI is close. But of course people like Andrew Ng, the American computer scientist, he actually says that AGI is quite far away.

Regarding AGI, the standard definition of AGI is AI that could do any intellectual tasks that a human can. So when we have AGI, AI should be learn to drive a car or learn to fly a plane or learn to write a PhD thesis in university. For that definition of AGI, I think we're many decades away, maybe even longer. I hope we'll get there in a lifetime, but I'm not sure. One of the reason that there's hype about AGI in just a few years is there's some companies that are using very non standard definitions of AGI and if we redefine AGI to be a lower bar, then of course we could get there in one or two years. But using the standard definition of AGI of AI to do any intellectual task a human can, I think we're still many decades away, but I think it'd be great if we managed to get.

I remember how I was talking about what other CEOs do talk about AGI. We have Demeter's Harvest saying it's actually three to five years away.

We've been working on this for more than 20 plus years. We sort of had a, we've been working on this for more than 20 plus years. We sort of had a consistent view about AGI being a system that's capable of exhibiting all the cognitive capabilities humans can. And I think we're getting closer and closer, but I think we're still probably a handful of years away.

Okay. And so what is it going to take to get there?

So the models today are pretty capable. Of course, we've all interacted with the language models and now they're becoming multimodal. I think there are still some missing attributes, things like reasoning, hierarchical planning, long term memory. There's quite a few capabilities that the current systems I would say don't have. They're also not consistent across the board. They're very, very strong in some things, but they're still surprisingly weak and flawed in other areas. So you'd want an AGI to have pretty consistent, robust behavior across the board, all the cognitive tasks. And I think one thing that's clearly missing, and I always, always had as a benchmark for AGI, was the ability for these systems to invent their own hypotheses or conjectures about science, not just prove existing ones. So of course that's extremely useful already to prove an existing maths conjecture or something like that, or play a game of Go to a world champion level. But could a system invent go? Could it come up with a new Riemann hypothesis or could it come up with relativity back in the days that Einstein did it with the information that he had? And I think today's systems are still pretty far away from having that kind of creative, inventive capability.

Okay, so a couple of years away till we hit AGI, I think, you.

Know, I would say probably like three to five component of an AGI system, but probably not enough on its own. I think there's still two or three big innovations needed here to we get to AGI. And that's why I'm a more of an on a 10 year timescale than others. Some of my colleagues and peers and some of our competitors have much shorter timelines than that. But I think my view is that.

Demis Hassabis actually said that his time is 10 years. So for the companies like Google they have a longer time frame. But what about those who aren't at Google and those who think that AGI is something that is so far off? So I'm always bringing up this interview clip because at the AI Action Summit, Yann Lecun basically says that we are missing a key component of AGI. And he kind of shares the same opinion that Demis Asaba says and that we need certain breakthroughs in order to actually get there. Now of course this isn't about 10% economic growth or not, but if we're actually going to look at what a system is going to be able to do, I think he provides the most realistic understanding of the difference between AI systems and what humans can do. And he talks about how a cat can plan complex actions. A 10 year old can clear up the dinner table and fill up the dishwasher without learning. A 17 year old can learn to drive a car in 20 hours of practice. And we have hours and hours and billions and billions of minutes of data for Tesla's full self driving. And yet it isn't perfect yet. And when we think about comparing that to what a human can do, like a human can learn how to drive a car in a few hours, but having an AI system that has years worth of data and it still is not perfect, that should tell us that we're clearly missing a key component of something.

We're missing something really big. Because you know, never mind trying to reproduce human intelligence, we can't even reproduce cat intelligence or rat intelligence, let alone dog intelligence. They can do amazing feats. They understand the physical world. You know, any house cat can plan very highly complex actions and they have causal models of the world. Some of them know how to open doors and taps and things of that type of. And in humans, a 10 year old can clear up the dinner table and fill up the dishwasher without learning. Zero shot. The first time you ask a 10 year old to do it, he or she will do it. Any 17 year old can learn to drive a car in 20 hours of practice. But we still don't have robots that can act like a cat, we don't have domestic robots that can clear up the dinner table, and we don't have level 5 self driving cars, despite the fact that we have hundreds of thousands, if not millions of hours of supervised training data. Okay, so that tells you we're missing something really big.

Now, something that is probably a controversial opinion and you guys might not understand what I'm saying, but I do think that we will have several areas of ASI before AGI because I think AGI is a lot harder than people think. And although OpenAI have said that they know how to build it, I want to quickly talk about why. I believe we'll probably have many areas of ASI before we even get to AGI. Because the thing is, for ASI you only need one thing to work extraordinarily well. But for AGI you need an entire system that works so well. It can work that well across anything that humans can do well, which is, when you think about it, pretty difficult. Now if we refer back to the paper done by Google DeepMind, we can actually see here that what I'm talking about the fact that we're probably going to get ASI before AGI. It's actually something that has already happened in a variety of areas. You can see here that they talk about performance, generality. If we look at the narrow area, which is where we have a clearly scoped set of tasks, we already have level 5 superhuman AI and that outperforms 100% of humans. For example, we have AlphaFold, which is of course, you know, protein folding and understanding the protein structures. And then of course we've got AlphaZero, which is that game that is able to play Alphago basically better than any human has ever played it. It's superhuman. Then of course we have Stockfish, which is of course the best chess player in the world. And that is essentially ASI for narrow AI. Now, narrow AI is of course, you know, something that is much easier, like I said, and I do think that certain companies are probably going to focus on the applications this first, because it seems like it's much easier if we go over to general AI. At the time of the paper was writing, we can see that at least 50% of skilled adults, we still haven't even achieved competent AGI just yet. So it seems like this is something that is extremely hard for current AI systems to manage. And I do think the promise of AGI might be a little bit weaker than we do think. Now, like I said before, you have to remember that OpenAI actually said we are now confident we know how to build AGI as we have traditionally understood it. Which is of course a very, very big statement. If you are saying that you know exactly how to get to the architecture that can basically replicate human intelligence, that is a bold claim that you will have to back up with, of course, some kind of product or some kind of proof. And OpenAI have also said, which is why I said I think companies might start to focus on this, which is why I don't think GPT 4.5 was that big of a deal, is that they're going to focus on superintelligence instead. So you can see right here, it says we're beginning to turn our aim beyond that to superintelligence in the true sense of the word. We love our current products, but we're here for the glorious future. And of course, with superintelligence, we can do anything else. So for me, I do think that these companies that are trying to get to AGI, I don't know, maybe I am on the wrong take for this, but I do think that it wouldn't be surprising if we get some superhuman AI systems that are able to just do medical research like a tool that's able to do specifically medical research or a tool that's able to do certain pieces of scientific research, maybe biology before we actually get an AI system that can drive your car, be in a humanoid robot, give you the best advice, write the best poems, control the computers, maybe, who knows, we'll get it all at once. But it's definitely super, super intriguing. And for those of you who are skeptical of the progress, it's definitely worthwhile to look at how certain benchmarks are being crushed. I do feel like any benchmark we've had that has been a struggle has pretty much been crushed by AI in two years. So let me know what you guys think about the AGI benchmark. Do you guys think that, you know, 10 economic growth is better or do you guys think that the other thing is better? With that being said and let me know what you guys think.