AI Pioneer: The Bubble Is Real And Could Trigger an AI Winter | Andrew Ng

AGI has become a marketing term. If AI could do a useful economic work task as well as a skilled professional human, then that seems to me like a more reasonable definition for AGI. I think we're very far away from that. Maybe more than decades.

Now, just to be transparent, just to be completely truthful, there is a small number of job roles that are fully automated by AI. So candidly, I think excessive hype that leads to disappointment, that leads to the collapse of the so-called bubble, that would not be good for the world and good for the field of AI. A lot of the best open-source models are coming out of China.

For any reasonable definition of AGI, I think the answer is no. We will not get AGI in 2026. The best way to get AGI in 2026 is if someone manages to dramatically lower the bar for what AGI actually means, then maybe some company clears the hurdle.

The definition of AGI that I'm most familiar with is AI that could do any intellectual task that a person can. And today, a human can learn in maybe tens of hours how to, I don't know, drive a truck through a forest, which I've never done before, but I think if you train me for a few hours I could probably do it. A human could learn to take calls for a contact center and answer questions consistent with how certain business needs it answered. And there's certain things a person can learn to do, intellectual tasks as opposed to physical tasks. And a lot of these tasks still feel like it's a ton of work to build custom AI workflows to perform, which we do do and which turns out to be really valuable, but needing an engineer to spend so much time building these things is not how the public typically thinks about AGI. And as much as I hope we'll get there someday, of computers as intelligent as people in every sense, I think we're very far away from that.

I'm excited about this. So, I feel like because of the hype around AGI, AGI has become a marketing term rather than something of precise technical meaning. And the way it's used as a marketing term, it actually misleads a lot of people. But since people want AGI, are excited about it, I'm excited about it. Why don't we come up with a test to see if we are actually, you know, getting close to achieving AGI? So, here's the idea. In the original Turing test, a human judge would get to type text back and forth to either an AI or a human. And the original Turing test was, can AI fool the judge into thinking or being unable to tell if they're talking to AI or to a human. And it was a great invention for 1950, which is when I think Alan Turing came up with this test. But it's not really measuring intelligence the way we think about it today. Here's how I think we can build a Turing AGI test, is what I called it, which is have a human judge design a multi-day experience that could be some onboarding training via a computer, and the test subject will be either AI or a human with access to a computer and normal software like a web browser, maybe Zoom, maybe other types of software. And if over a few days, in a multi-day experience, if AI could do a useful economic work task as well as a skilled professional human, then that seems to me like a more reasonable definition for AGI. And the reason I'm proposing this is because this is actually much closer to what the broader public thinks of as AGI. So if people think AGI could be here, they think, wow, AI will be able to do people's jobs. And indeed, if AI is able to function like a remote worker for multiple days and do productive work, that will be incredibly valuable and I think much closer to what people think of as real AGI as opposed to some of these, you know, alternative standards that some businesses are promoting for PR or policy or fundraising purposes.

One of the challenges with tests or benchmarks of AI is when someone fixes a test set in advance, then you're just measuring one specific dimension of AI. This is why for a lot of AI models, large language models, there are a lot of standard benchmarks, you know, SWE-bench, GPQA, and so on, and it's very difficult to avoid having teams optimize for the known test sets out there, even if they don't do it directly. It ends up... and AI is a jagged form of intelligence, it's great at some things, really awful at some other things. But AGI, people think of it as AI that could do any intellectual task a person can, should cover all intellectual tasks that people can cover. So to me, one big difference between a fixed test set kind of benchmark versus letting a human judge probe in real time is a human judge can take a look at the state-of-the-art and can probe to see where is AI strong, where is AI weak. And for AI to be this AGI, you know, most people think of it as matching humans in every dimension, which means that a human judge should be unable to probe to find places where the AI is materially weaker than what a human can do in terms of these, say, economically valuable job tasks that we want AI to do.

Oh, I think benchmarks are measuring something real, but it's a very narrow slice of the sorts of things we want to measure. One weakness of a lot of benchmarks is we're much better at designing objective benchmarks. So things like math, which is a right or wrong answer, or a factual question, you know, like in this year's Olympics who won this, you know, 200-meter freestyle swimming or whatever. So those questions have objective right and wrong answers, and so a lot of benchmarks is on measuring is the AI right or wrong, it's very black and white. A lot of life, there is, it's very difficult to design one completely objective best answer. So even in a human conversation like we're having, you know, what's the one right thing for me to say? I don't know, but I could probably say stuff that is better or some stuff that is worse. And on average, we've been not very good at designing benchmarks to measure these things that are more subjective or where gradations are good and bad, but it's very hard to write a test set saying this is one right answer, this is one wrong answer. Not that we don't try, but I feel like a lot of the benchmarks we have don't capture those things as well. And the reality is a lot of human work, if you ask me to write a research report, you know, there's no one right report, there are different gradations of how good or bad the report is, and our ability to write benchmarks to capture these shades of gray is much weaker than our ability to write benchmarks to say did it solve this math problem correctly, did it write a piece of code that runs correctly. Those we know how to do better than what a lot of actual useful human work requires.

I think AGI is a distraction. We're very far from it. You know, we're not going to get there anytime soon. But even without achieving AGI, there's so much incredibly valuable stuff that we are doing and that we will continue to do in 2026. I coined the term agentic AI to describe what I saw as a growing phenomenon. And even though I was saying that it's so much work to build AI to capture useful business processes, sometimes when we do that it's incredibly valuable. So in 2026 and beyond, there'll be a lot of exciting work to build AI agents or to build agentic workflows to do a ton of really valuable, economically important work. Here at my team at AI Fund, our teams have been using agentic workflows to write code, like many others do too, but also to look at paperwork to check for tariff compliance, to read tricky legal documents to, you know, help lawyers do their work better, or to do medical assistance tasks, or to support customer service. And I find that taking the mental processes that people are currently going through to do these types of tasks, be it legal compliance or whatever, and coding them into an AI agent or agentic workflow so the AI could do it for you, this I think will go on for a long time. We'll be building these agentic workflows for many, many years to come. But the value will be very large.

So Rich's article on the bitter lesson was very influential and very well written. Just to be clear, I like scale. When I started Google Brain, which later merged with DeepMind to form Google DeepMind and created Gemini and so on, my number one mission that I set to the Google Brain team was, let's just scale, right? There, build really, really big neural networks, throw lots of data at it. So I think I was probably one of the earliest in AI...

So I like scale, and I think it was because the Google Brain team, the way I set it up, its DNA was just scale. That was the number one thing. You know, obviously a team does many things, worry about many things, but the number one mission was scale. That's how I set up Google Brain. That's why, with that DNA, it was the Google Brain team that invented the transformer neural network, that was the most scalable neural network architecture, you know, in history, and that powered the generative AI revolution, right? So, so I really believe in scale. Having said that, at different moments in time, our ability to scale and our ability to inject other forms of knowledge creates a balance where I don't think we'll achieve everything we want only by scale. Agentic workflows are incredibly powerful too. And like many things, sometimes we've seen businesses hype up one phenomenon for fundraising or PR or whatever purposes. And it is absolutely true that because of scaling laws, when you scale systems you can fairly predictably forecast where performance will go. This was shown by my team at Stanford, later by OpenAI, I think. And so this is actually a wonderful argument to raise funds because you can say, give me more money, I will scale up these machines, scale up the data, get better. But because of that huge kernel of truth in scale, I think it has been hyped up beyond what it actually is, which is amazingly valuable, but not as valuable as the hype says it is. But nonetheless still amazingly valuable. And agentic workflows allows us to take advantage of the scale AI models, large language models, and additionally inject other types of knowledge into the system to build more reliable, more performant workflows.

No, I don't think the era of scaling is over. It's been getting harder and harder. Yeah, it's been interesting. AI has been progressing maybe exponentially. It's been interesting seeing that the number of dollars needed to drive this exponential progress has also been exponential. So you spend exponentially many dollars to drive exponentially rapid improvement, it's not bad. It is very valuable because the cost of building it can be amortized over a lot of users. And I think there's still more to be gained by scale, but it's not the only avenue we have for improving AI at this point.

About what?

Oh, what would make me give up on scaling? If for an extended period of time further efforts to scale up models are not paying off, that would make me change my mind. But maybe one important asterisk to that, just as over many decades Moore's Law drove progress, but it took lots of different technologies that kept on changing over time for Moore's Law to keep on making semiconductors work better and better. What we see is in the early days of generative AI, the recipe was get more data, train a bigger model. This is exactly what I, at least at that level of abstraction, exactly what Google Brain was set up to do, right? More than 15 years ago or so. But since then, AI models have read pretty much the entire open internet, so that simple recipe for scaling, it doesn't really work anymore, which is why many teams are now doing much more work on synthetic data generation, which requires much more human engineering, a lot more work on different reinforcement learning recipes. And so even though scaling still continues to pay off, the specific recipes for driving that scaling has definitely shifted quite significantly over these last, you know, two, three years.

Um, yes. I wish it were that easy though. In practice, I think it is... in theory yes, in practice very harder than most people think.

So one thing I think has been fantastic was with more intelligent models, be it, you know, Claude 3.5, Gemini 1.5, GPT-4, we've been increasingly able to give a large language model a set of tools and just have it let loose, right? So just give it a set of tools for reading and writing for a file system, then just tell it to do some tasks like, you know, look for extraneous files and help me clean up my hard disk or whatever. And the level of performance of these is fantastic. It's really, really, really impressive. Having said that, for many workflows, it's just not reliable enough to be production ready. And so while smarter models are great, what I see right now for a lot of practical business use cases is teams thinking through what is the workflow you want, what are the key steps, and implementing them piecemeal so they can get that reliable performance to get this stuff working and have it run 10,000 times and maybe even work, you know, pretty much every single time. And maybe, and as models become more intelligent, we are letting them be more autonomous. We're removing guardrails. One thing my teams often do, stuff that we built six months ago probably had more guardrails, was more scripted, and we routinely end up reducing the scaffolding. Instead of giving very detailed step-by-step instructions, we're more likely to say, you know what, go decide for yourself what you want to do. I mean, one example, we used to build deep researchers where we say to do research on a topic, do a web search, do this many queries, then download this many pages and summarize it, blah blah blah, common research agentic workflow from like a year or two ago. Now AI models are much more capable of deciding do they want to keep on searching the web more, do they want to summarize or not. So I find myself frequently taking systems we had built a prototype a year ago, six months or a year and a half ago, and ripping out instructions to say, hey, prompt the model to just figure out the decision by itself. But still a long ways to go. It works for use cases like deep researchers where it's not taking action, if it misses a citation or misses a reference it's not the end of the world. But for a lot of high-stakes enterprise use cases, I think the reliability gap to letting AI models just let loose and do what it wants to do, I suspect that gap, while it's closing, is still bigger than some people think.

You know, I see no contradiction. Maybe I'm missing something. To me, the amazing thing about the human brain, assuming the human is AGI except it's not artificial, right? It's general intelligence. The amazing thing about the human brain is its plasticity or its ability to learn. And I feel like AGI to me should be less about AI that already knows everything under the sun. That seems very challenging, doesn't seem practical. But what makes a human brain so valuable, or one of the things that makes it so valuable for economic tasks, is ability to just learn to do new stuff, whatever is needed. And it is through learning that we then gain these incredibly specialized intelligences. Like that human brain has learned to solve incredibly difficult math problems by getting a PhD in math. And so, you know, that is very specialized intelligence as the end point, but it was learning. And then, but theoretically that same human brain just given different training could have been a chess master or could have been amazing at, you know, playing tennis. So to me, a lot of what makes the human brain so general is not that my brain or your brain already knows everything under the sun, it's our ability to adapt, to learn a huge range of things.

Yeah. One thing that really motivated me when I started the Google Brain team was this idea that people don't talk much about anymore, was that a lot of human learning may be due to one learning algorithm. What that means, but it turns out that our DNA is not that long. So our DNA contains a very limited amount of information. But somehow our DNA has encoded the biology of the brain, and the brain is a fairly general learning algorithm, which is why the brain could learn to, you know, do a PhD in math, or learn to ride a motorcycle, or learn to type on a computer, or learn to work in a contact center in a call center, is because the brain has this very general learning algorithm that allows it through learning to specialize in a huge range of things. And that's what makes intelligence seem general. It is this learning ability that lets it gain, you know, almost any specialty under the sun.

I mean, the race is on. This is exciting. I think Sam was my student at Stanford, and I have a lot of friends at Google. So I'm very pro-OpenAI, also very pro-Google. If we look back at the history of technological disruptions, whenever there's a tech disruption, sometimes the new entrants won, sometimes the incumbents won. There's actually room for either one to do well. And I think the game is still on. In fact, if we look at the internet disruption, Google was a startup that rose with the rise of the internet, but there's some incumbents like Microsoft and Apple that were founded long before the internet, and they did just fine. So clearly AI is very disruptive for incumbent businesses like Google. I think Google's played its hand well. I think Gemini 1.5 is an incredible model.

I use both Gemini and ChatGPT and Claude and a whole bunch of models quite often.

Can I just say I actually don't like talking about AGI, but others talk about it so much and it's so hyped. I feel like, you know, I feel like as leaders in the field, when there's a massive amount of hype...

It is vastly hyped. At least the public thinks of AGI as AI becoming as intelligent as people in a very general sense. We are so far away from that. I hope we'll get there, I would love to get to AGI, but realistically I think we're decades, you know, maybe more than decades away from that. And so this idea that, oh, we just need, you know, another few quarters to get to AGI, that's just not going to happen unless you redefine AGI to lower the bar and make it much easier to achieve. And candidly, in the history of AI we've seen a few AI winters where well-meaning people overhyped the promise of AI, this led to elevated expectations that were not met, and a collapse of investment and interest, and this was bad for the field. AI works really well right now. It's incredibly valuable. I see relatively few things that could derail AI's momentum. One of the things I actually worry about is excessive hype that leads to disappointment, that leads to the collapse of the so-called bubble, and that would not be good for the world or good for the field of AI. So defusing hype about AGI is an important thing to do to lay the foundation for us to have more sustainable growth.

I don't remember credible teams declaring achieving AGI. I remember teams declaring that AGI was only a short time in the future, but I don't... teams were not three years ago declaring they got AGI. They were saying we could get there really soon. And then, you know, so far no one's gotten there.

I think that if anything, teams have been trying to lower the bar to what it means to achieve AGI. I mean, AGI, AI that can do any intellectual task that humans can, is a very high bar. We have not gotten anywhere near to that bar. But if teams come up with alternative definitions that are easier to achieve, then maybe we can get there sooner. And by the way, I don't mind how we define AGI. You know, we can define AGI however we want. But the problem is most of the broader public thinks of AGI as generally very intelligent AI, basically human-like intelligence. So that's what most people out in the world think. And so the problem with...

Yeah. Basically, so you know, I think because people keep coming up with alternative definitions, it turns out that terms lose meaning when lots of different people use that same term to refer to very different things. And so, for example, we all have a sense of what the word blue means, right? My shirt is blue. But if for some reason society points at all sorts of different colors and says this is blue, this is blue, then the word blue will lose meaning because people don't even know what it means anymore when someone says blue. And that's what happened with AGI. Different teams came up with different alternative definitions. People say this is AGI, and because people are applying lots of different definitions to the term, when someone now says AGI, it's hard to know what exactly they mean. And the problem is the public thinks of AGI as human-like intelligence, and so if someone comes up with some weird narrow technical definition and says AGI will appear in two years, the broader public nonetheless thinks AI will be as intelligent as humans in two years, which just doesn't seem true to me.

I think the harness matters a lot. It's been really incredible watching Anthropic build Claude Code, as well as their SDK, to make sure that they and maybe others have a good harness to use the model powering this. And the details of the harness, how you structure the prompts, what tools you give the LLM, all those details really matter still. Actually, maybe...

One small example: our current models are incredibly intelligent and they're becoming better and better at tool use, making function calls. Having said that, if you give a language model too many tools, it consumes a lot of the input context and it's much more likely to struggle and make the wrong API call, make the wrong tool use call. And so these little kind of maybe it feels like, you know, it's 2026, why do we need to worry about engineering details like these? It turns out we do because it still makes a big difference to the overall performance. And many teams use MCP. I use MCP a lot and one of the practical engineering things is if your MCP server has too long a list of tools, it consumes a ton of your own input context. And it may be too many tools to effectively figure out which ones to use. So you get into context engineering and sometimes a harness can help make these decisions much more smooth.

I look forward to when continual learning will be solved. It would be amazing. It would be delightful if it ends up being completely solved in 2026. I expect we'll make progress. I think continual learning is very important.

One of the challenges of brute force learning is to the extent that intelligence, human intelligence comes from the generality of the learning algorithm, that's what makes a human brain so powerful. A general learner learns new things really quickly and this is why, for example, humans that work with us can do so much. Whereas if you need to spend a long time to brute force AI to do some narrow task, it could still be really valuable in some cases but the case for doing it just isn't going to be there for a lot of tasks. So the ability to say hire a human and talk to them and then figure out what to do and do a job that's really valuable. If you need to hire an AI and then spend a million dollars to train the AI to do this one task, then there are a lot of tasks that it just doesn't make sense to spend that much effort to train AI.