Ilya Sutskever Just Revealed The Next BIG THINGS In AI (Superintelligence Explained)

The race towards artificial superintelligence isn't actually sci fi anymore. It's happening right now. And Ilyas Hutskever, one of the co founders of OpenAI, just dropped some jaw dropping insights about where we're headed now. You're going to want to stick around because what Ilya said might actually redefine everything we know about intelligence, humanity and the future of our species. Now, one of the first things that Ilya Sutskova actually talks about is the current paradigm that we're in and why it's coming to an end. He speaks about the death of pre training and essentially he says that pre training as we know it will unquestionably end. And this is important because the current AI paradigm relies heavily on massive text based pre training. But this statement predicts a fundamental shift in how we're going to be training models and it raises questions about what new models will replace today's techniques and how the AI field will adapt once we exhaust all the readily available data.

Age of pre training and the age of pre training is what we might say the GPT2 model, the GPT3 model, the scaling laws. And I want to specifically call out my former collaborators, Alec Radford, also Jared Kaplan, Dario Mode for really making this work. But that led to the age of pre training and this is what's been the driver of all of progress, all the progress that we see today. Extra large neural networks, extraordinary large neural networks trained on huge data sets. But pre training as we know it will unquestionably end. Pre training will end. Why will it end? Because while compute is growing through better hardware, better algorithms and larger clusters, all those things keep increasing your compute. All these things keep increasing your compute. The data is not growing because we have but one Internet, we have but one intranet. You could even say, you could even go as far as to say that data is the fossil fuel of AI. It was like created somehow and now we use it and we've achieved peak data and there'll be no more. We have to deal with the data that we have now. It still let us go quite far. But this is. There's only one intranet and he's right.

We've basically exhausted all forms of available data. Now we're going to need to do is we're going to need to figure out new things. This is where he talks about three key things which are coming next, which I'll break down for you in a second.

So here I'll take a bit of liberty to speculate about what comes next. Actually, I don't need to speculate because many people are speculating too, and I'll mention their speculations. You may have heard the phrase agents. It's common and I'm sure that eventually something will happen, but people feel like something. Agents is the future, more concretely, but also a little bit vaguely synthetic data. But what does synthetic data mean? Figuring this out is a big challenge and I'm sure that different people have all kinds of interesting progress there and an inference time, compute or maybe what's been most recently most vividly seen in O1, the O1 model. These are all examples of things, of people trying to figure out what to do after pre training. And those are all very good things to do.

So one of the first things he actually mentioned there was agents. Now agents are being widely discussed as the next big step in AI. And while current systems like ChatGPT are great at generating responses, they aren't truly agentic. And this just means that they lack the ability to independently set goals, reason about their environment and take meaningful actions in the world. And now he did hint that future systems will be genuinely agentic, capable of reasoning and acting autonomously in ways we've only seen in science fiction. And this basically just means that, you know, future AI will evolve from being passive tools to become active participants. And they're going to be making real decisions, they're going to be adapting to new situations, and they're going to be collaborating with humans in meaningful self directed ways. And this is going to present a massive leap in functionality as agentic AI could perform tasks without needing constant prompts or even supervision. And this is important because agentic AI is going to fundamentally change how we interact with technology. Imagine your personal AI assistant that can proactively manage your tasks, conduct complex research, or even solve real world problems without being micromanaged. Now the second point that he spoke about was synthetic data. And he said that this is one of the potential solutions for the limitations of real world data sets. He described it as both promising and a challenging area of research. And while he didn't go into, you know, all the technical details, he emphasized that figuring out how to create high quality synthetic data could unlock new capabilities for AI. Now this basically just refers to artificially generated data sets that mimic the properties of real world data, and they can be created using simulations, generative models, or even algorithms designed to replicate natural phenomena. And as we hit the limits of the real world data, which is of course, as he referred to in the beginning, the one Internet problem, synthetic data could provide New training material for these AI systems. The ability to simulate edge cases or rare events, for example, for self driving cars or even potentially medical diagnoses. This is important because without synthetic data, the progress of AI could stagnate due to data scarcity. And mastering synthetic data generation, we could train on AI scenarios that don't yet exist or are too rare to capture in the world.

You've got these mammals, right? All the different mammals, then you've got non human primates, it's basically the same thing. But then you've got the hominids. And to my knowledge, hominids are like close relatives to the humans in evolution, like the Neanderthals. There's a bunch of them, like it's called Homo habilis, maybe there's a whole bunch and they're all here. And what's interesting is that they have a different slope on their brain to body scaling exponential. So that's pretty cool. What that means is that there is a precedent, there is an example of biology figuring out some kind of different scaling. Something clearly is different.

Essentially, what we do have here is where he's talking about the graph that illustrates how in biology there is a known relationship between an animal's body size and its brain size. Most mammals follow a predictable pattern. Bigger bodies tend to come with bigger brains in a more or less steady ratio. But human ancestors, the hominids, they actually broke that trend somehow. And evolution produced creatures whose brains grew faster than they expected for their body size. And this shift suggests that at some point, nature figured out a different scaling rule for intelligence in humans. And now, by mentioning this, Sutskova is drawing an analogy. Just as biology found a new way to scale intelligence in human evolution, giving us much bigger and more capable brains relative to our bodies, the AI field may need to find new methods to move beyond its current scaling limits. Right now, AI depends on huge amounts of data and ever larger neural networks. But just as humans evolved a new approach to intelligence, AI researchers will need to find equally clever next steps that just don't follow the old scaling patterns, which is, you know, what we're currently doing.

So right now we have our incredible language models and they're unbelievable chatbots and they can even do things, but they're also kind of strangely unreliable and they get confused while also having dramatically superhuman performance on evals. So it's really unclear how to reconcile this. But eventually, sooner or later, the following will be achieved. Those systems are actually going to be agentic in real ways, whereas right now the systems are not agents in any Meaningful sense, just very, that might be too strong. They are very, very slightly agentic, just beginning. It will actually reason. And by the way, I want to mention something about reasoning. Is that a system that reasons? The more it reasons, the more unpredictable it becomes. The more it reasons, the more unpredictable it becomes. All the deep learning that we've been used to is very predictable because if you've been working on replicating human intuition, essentially it's like the gut feel, if you come Back to the 0.1 second reaction time, what kind of processing we do in our brains, well, it's our intuition, so we've endowed our AIs with some of that intuition. But reasoning, and you're seeing some early signs of that reasoning, is unpredictable. And one reason to see that is because the chess AIs, the really good ones, are unpredictable to the best human chess players. So we will have to be dealing with AI systems that are incredibly unpredictable. They will understand things from limited data, they will not get confused. All the things which are really big limitations. I'm not saying how, by the way, and I'm not saying when. I'm saying that it will and when all those things will happen together with self awareness, because why not Self awareness is useful. It is part your ourselves are parts of our own world models. When all those things come together, we will have systems of radically different qualities and properties that exist today. And of course they will have incredible and amazing capabilities. But the kind of issues that come up with systems like this, and I'll just leave it as an exercise to imagine, it's very different from what we are used to. And I would say that it's definitely also impossible to predict the future. Really, all kinds of stuff is possible.

So this is by far the most interesting part of the talk because this is where Ilya Satskova actually focuses on where AI research might be headed in the long term. And this, this is packed with some juicy information. So he actually talks about super intelligence. If you remember, that's the company that he's actually building right now. Now he first talks about, you know, it being agenting. And this is where instead of, you know, passively waiting for instructions, a future super intelligent system might take the initiative and act on its own. And it's going to have its goals, it's going to have its purposes or sets, or, you know, it's going to redefine itself and it's going to be able to seek out opportunities, tools or information to achieve those specific goals. And this is going to be insane because this is A big shift from the current models which just respond when prompted. But these models are going to be, you know, probably even reflecting from the environment and learning from the environment, which is something that we haven't really had yet. Now of course, today's AI models often give the impression of thinking and reasoning, but they're mostly pattern matches with a robust step by step reasoning process. But in contrast, super intelligence would be able to genuinely reason and carefully consider problems, weigh the different possibilities, work out logical steps and make sound decisions even in complex or unfamiliar situations. Now the crazy thing about all of this as well is that this AI is going to understand us. Current AI basically mimics what we do by producing human like text or answers. And sometimes it even stumbles when dealing with logical or tricky problems. But a superintelligent is going to have a more profound grasp of concepts. It's not going to know what words tend to go together. It's truly going to understand the meanings and the relationships and be able to interpret everything quite like humans do. But probably beyond that. Now the last thing that I found to be truly, you know, fascinating was the fact that they speak about, you know, this thing being self aware. And this is the most speculative quality. This means that the AI wouldn't just process information externally. It's going to have an internal sense of self, its own goals, its own reasoning process and its place in the world. And self awareness would let it reflect on its own actions, potentially leading to more deliberate improvements, moral considerations and responsible behavior, which are qualities we often associate with constant beings. And essentially this slide is absolutely incredible. It paints a picture of a future where AI is not just an advanced tool, but actually something much more akin to a new form of intelligence, one that acts on its own, one that thinks deeply and coherently, one that truly understands complex ideas and is going to have a sense of itself which is going to be incredible.

I have a question for you about sort of autocorrect. So here is, here's the question. You mentioned reasoning as being one of the core aspects of maybe the modeling in the future and, and maybe a differentiator. What we saw in some of the poster sessions is that hallucinations in today's models, the way we're analyzing, I mean, maybe you correct me, you're the expert on this, but the way we're analyzing whether a model is hallucinating today without, because we know of the dangers of models not being able to reason that we're using a statistical analysis, let's say some amount of standard deviations or whatever Away from the mean in the future, wouldn't it? What do you think that a model given reasoning will be able to correct itself, sort of autocorrect itself, and that will be a core feature of future models, so that there won't be as many hallucinations, because the model will recognize when. Maybe that's too esoteric of a question, but the model will be able to reason and understand when a hallucination is occurring. Does the question make sense?

Yes. And the answer is also yes. I think what you described is extremely, highly plausible. I mean, you should check. I mean, for. Yeah, I wouldn't rule out that it might already be happening to some of the, you know, early reasoning models of today. I don't know. But longer term, why not?

Yeah, I mean, it's part of like Microsoft Word, like Autocorrect. It's a, you know, it's a. It's a core feature.

Yeah, I just, I mean, I think calling it Autocorrect is really doing a disservice. I think you are. When you say Autocorrect, you evoke like, it's far grander than Autocorrect. But other than. But, you know, this point aside, the answer is yes. Thank you.

And that was one of the most fascinating parts of this interview because he actually responds to someone that asks about hallucinations and whether or not these models are going to be able to solve them. Now, recently we did get some information from Sam Altman and some people working at OpenAI and Microsoft talking about how hallucinations are a very tricky problem to solve. But Ilya Sutskova seems to believe that in the future these hallucinations are going to no longer exist. Which means that this is going to be a remarkable step because one of the main problems that stops AI from being widely used is the reliability problem. You have to understand that some industries cannot afford hallucinations. The kind of software that they're using, for example, in planes, the failure rate for Boeing parts is like 0.000001%, which means that if you have an AI that fails at around 98%, you can't really deploy that safely. So this is going to be something that if it does get solved, which Ilya Sutskever is saying that it is going to get solved, means that, well, we're going to be living in a very different future. The next part here is, once again, I say everything is interesting, but this one is genuinely super interesting because this is where we get systems that can do everything a human can do and Basically everything the human hasn't even thought about yet. This is where Elisa Svar is responding to a question that asks about in distribution generalization. And this is just basically where if the model is going to be able to solve problems that it hasn't seen before. And this is something that humans do very regularly. You know, you have maybe 16 to 20 hours of driving lessons, but you can realistically easily drive all around the world based on driving in a small city. This is something that current autonomous cars can't really do. You have to map out the city, huge training data. This is something that is kind of an issue. So if we can actually get AI systems that can generalize outside of their training data, that is going to be genuinely game changing because you're not going to need huge data, you're not going to need large data sets. It's going to be a situation where you only put it in for 10 hours or whatever and it's able to look at new scenarios and it's going to be able to generalize out of distribution. And those implications are pretty incredible because it means that the AI is going to be able to solve things it's never seen before or never encountered, which is just incredible.

The question should not be answered with yes or no because what does it mean? Auto distribution generalization, what does it mean, what does it mean in distribution and what does it mean out of distribution? Because it's a test of time talk. I'll say that long, long ago before, before people were using deep learning, they were using things like string matching, N grams for machine translation. People were using statistical phrase tables. Can you imagine, they had tens of thousands of code of complexity which was, I mean it was truly unfathomable. And back then generalization meant, is it literally not in the same phrasing as in the data set? Now we may say, well sure, my model achieves this high score on, I don't know, math competitions, but maybe the math, maybe some discussion in some forum on the Internet was about the same ideas and therefore it's memorized. Well, okay, you could say maybe it's in distribution, maybe it's memorization. But I also think that our standards for what counts as generalization have increased really quite substantially through dramatically, unimaginably, if you keep track. And so I think the answer is to some degree, probably not as well as human beings. I think it is true that human beings generalize much better, but at the same time they definitely generalize out of distribution to some degree. I think it's a useful topological answer.