Hello. Can everybody hear me? Can you text me a signal? I couldn't see the response. Are you guys seeing my computer and also hearing my voice? Well, please speak up. Sorry, I didn't hear anything. Anything wrong with it? We can hear you. Oh, you can hear me. Okay. Great. Thank you. Thank you very much. Well, thank you for the introduction, Eric. It's a great privilege and honor for me to be a speaker in your program. Unfortunately, you know, I was originally planning to come in person, but due to the situation in the Middle East, I have to do it remotely. So bear with me if there are any audio or visual issues. So the title of my talk today, you know, I have two titles. Broader speaking, I'm going to talk about our recent work toward the next generation AI systems beyond what I call book intelligence. You will see what that means in a few minutes. And then, of course, the focus will be on the Pan world model, which is a stateful, manipulatable, and long-horizon world model for reasoning, planning, and thought experiments. In fact, the background movie that brought many famous artworks alive was actually the simulation from the Pan model, because as a world model, its main function is to simulate real and imaginary worlds, and that's what you actually see from those simulations. Now let's go back to the background of this body of work. So why do we even care about building more foundation models and more architectures for artificial intelligence? Because in my opinion, the current models are often not powerful enough. For example, let's ask this question: how well do current large models understand and reason about the world? Right, so this is a picture that I like to share. I ask you, is this person, this guy on the right-hand side, trying to help that lady or not? I guess as a normal human being, we have no problem in figuring out, in fact, it takes no time for us to figure out whether he is helpful or not. But if you pose the same question to, say, the advanced large language models like ChatGPT-4, for example, here is the answer you'll get. It believes that this person actually is helping rather than playing some bad things against that girl. Well, maybe ChatGPT-4 is not the most powerful model, or maybe I'm not giving it more time to think. Let's ask a more advanced version called ChatGPT-4o, which is a specialized GPT model for reasoning. After spending 16 seconds reasoning about the situation, he still concludes that he appears to be helping. So that's kind of an anecdotal instance about how powerful our current models are in solving real-world reasoning problems. There are other more prominent problems that you can use to test the capability. For example, there is a problem set called the detection of motion processing, and you can ask whether these two objects are moving at the same trajectory, and the answer from GPT-4o is actually yes, which is obviously a wrong answer, right? And you can also ask a spatial understanding question, say, are these purple objects and the red objects of the same size? Or you can ask a positional question like, are the left and the right of the same size? Again, you don't always get the right answer. In fact, you can systematically test many of these so-called world model automatic benchmarks to test the ability to perceive the world or to predict the world. And by and large, many of the state-of-the-art models are still quite far from what humans can do in those spatial and physical reasoning simple tasks. Now, given that kind of testing results, why do we still believe that modern artificial intelligence systems are very smart and very intelligent? For example, they can solve a math problem and win an Olympic gold medal in the IMO. Right, in a sense, I want to first break down the definition of intelligence because this is a very subjective characteristic of a human being rather than an objective metric of a system. Therefore, whether it is a useful metric to measure the quality of the system is in question. For example, we believe Einstein is very intelligent, right? Why? Because you can say that he has very deep insight into the universe. He's very good at mathematics and so forth. And we all agree that he has a very bad memory, but he's very good at reasoning. So bad memory isn't indicating an intelligence issue, but reasoning is a hallmark of intelligence, right? But on the other hand, we can also ask, is Mozart a composer intelligent? Obviously, we all believe he's very intelligent, but why did he become famous in the first place? Well, the story goes that in his young age, maybe at the age of seven, he went to the Vatican and heard from the cathedral, a basilica, a piece of music, and that music was sacred, was forbidden to be copied outside of the basilica. But he actually, just by listening once, he could come out and reproduce the score with very good memory, and then he was declared a genius and even got a knighthood from the Pope. So good memory is sometimes celebrated as good intelligence. Therefore, intelligence is something that is not necessarily objective in different people's eyes. So maybe there's a need to even begin our quest by redefining what is intelligence. For example, here I want to propose that we define intelligence in terms of capability, measurable capability, and utility. We use purpose, not sensation or stance, to define the degree of intelligence. Right? And this spirit has now carried a lot of weight. For example, recently, a post from the famous mathematician Terence Tao, he studied the intelligence systems, the GPTs, in a sense using a lens from a utility standpoint. He said that the AI assistant was very intelligent, maybe for some small scale, very useful for some scale, and somewhat helpful at a medium scale, and becomes useful again and helpful again in larger scales. So the definition is really on utility rather than really sensation and the feeling, right? And utility is really a very defining feature. We define our own needs and also our capability. We're all familiar with this Maslow's pyramid of capability, and using that kind of philosophy or mindset, you could easily define the utility of an individual or other system using such a similar hierarchy pyramid, such as the ability to read and write, and the ability to listen and react, doing logical reasoning, planning, execution, and so forth, right? So you could use this as a benchmark to measure not just people but also systems. And here is how I see the problem start to break down. For example, at the bottom of this pyramid, we see read and write, and maybe see and draw. And obviously, our GPTs and our current large language models and vision models seem to be doing quite well already in those spaces. And at the top, you see something called prophecy. Maybe that's something never achievable by any system unless you have a crystal ball. But what about the middle, right? The middle basically is a space. Some say it's already covered by GPTs and the LLMs, and some disagree. So here is how I see the problem start to break down into different tiers of intelligence. What we see from LLMs and large language models is, in my opinion, a limited form of intelligence because they only operate in the space of language. You could call them book intelligence because even though they can solve very advanced math problems, they cannot really tell you how to swim and how to interact with the physical world because that requires embodied knowledge and the physical knowledge of the universe. And that's actually where another level of intelligence, such as displayed in this simulated video, could be reflecting, which actually draws upon knowledge of space and time, of the physical space, and that leads to obviously a model that is not an LLM. If you want to put an individual in such a real world, you start to deal with agents and agencies, and that requires the system to be able to communicate, interact, understand self and others, and therefore develop a new level of intelligence which I would call social intelligence. And above all, you also want to understand nature, study the science, develop new drugs, and so forth. And that obviously requires you to have an even extra degree of intelligence, which means curiosity-drivenness and maybe the ability to ask the right question and so forth, which obviously builds on all the lower levels of intelligence: social, physical, and lingual and textual book intelligence, and so forth, right? So I think this is the way I see intelligence breaking down into different tiers, and that's helpful in defining where the model can be positioned against. Let's start from the ultimate intelligence and see what might be a good approach to achieve them. Right? So this is one of my favorite sci-fi classics called Dune. I suppose many of you probably are familiar with this book, right? It basically played a major role in inspiring many of the science fiction legendary movies and so forth that we've seen so far. But in this movie, in this fictional novel, one of the central figures is a legendary individual called Muad'Dib, and he has a supernatural capability which can allow him to, say, command armies to win the galactic war or oversee global-scale projects like turning a desert planet into a green paradise, and so on and so forth, right? So these are really complicated and big tasks, and which could be the epitome of superintelligence. But what's interesting in this book is how he achieves and solves these very difficult problems. It turns out that he has this amazing mental capability of simulating outcomes of all possible plans before he even executes. And this simulation capability actually comes from his mental world, which was built from inheriting all the knowledge from their ancestors. So that's actually a very interesting thing because the methodology sounds surprisingly similar to how we build computers and how we build repositories of knowledge, right? So the question I want to ask next is how to build such a system, a superhuman. Maybe building a superhuman of Muad'Dib is too far a reach, but maybe we could build a system with a similar behavior and going up the hierarchy of capabilities like what I saw I wrote in this pyramid from low to high. And then that basically leads to a design of a system which is a goal-oriented agent acting in the environment and achieving goals in a Dune style, which is simulate the plan before you act. And that basically is the setup of my entire presentation in terms of the technical architecture, which starts from defining an optimal agent that can solve more and more harder problems. So in an optimal agent, the agent can generate an action upon receiving or seeing a state from the real universe. And then optimal agent means that you want to achieve a goal optimally by maybe maximizing the cumulative reward function, and that leads to a solution which is an optimum policy that maximizes the cumulative rewards. So what's the problem with this system? What this system basically means is that you need to, among many things, have direct access to the real universe, which is the world, and that's actually a very difficult thing. For example, how can we train an astronaut or maybe even a robot to go to Mars before even on Mars? Obviously, you need to obtain some capability before you have access to the universe, and that leads to our second goal, which is that you can maybe build a model which can simulate the universe as a surrogate, and that is basically where the world model arises. It is a simulation model that can allow you to simulate the environment so that you can do optimal reasoning. And therefore, my definition of the world model is a generative model that simulates possibilities. These possibilities are not limited to only physical possibilities, such as pouring the water, throwing the objects, the physics of the objects, and so forth. It could also include embodied experiences and emotional experiences, and maybe social experiences, and maybe also mental worlds, maybe even counterfactual worlds, and many other things, right? So imagine such a world model can be built, then how does it change your way of behaving, planning, for example, strategizing? So this is a proposal: you can reason by simulation rather than solving optimization. What you have from the world model is a generative model of the next state given the previous state and the action. And then once you are given a goal, for example, you can basically just use this simulator, this world model, to simulate multiple trajectories and then look ahead and see which trajectory gets you to the goal and also at the lowest cost, and then you execute. So this is a unique way of reasoning which is different from what we designed many computers to do, which is to do inductive reasoning and solve a big optimization problem. In here, you reason by simulation. There are books in fact about this behavior, such as this one, Hypothetical Thinking, and in our common language, we call them thought experiments. That's actually how most common individuals or maybe even great thinkers are used to solve problems: thought experiments rather than solving a big linear system, a big mathematical equation system, and so forth. And this approach is highly flexible and repurposable. You may know the system AlphaGo. It is actually a world model plus a Go-playing agent, where the world model is the Go board and also the play rules, and then you can simulate upon a board position multiple outcomes and then using a Monte Carlo tree search algorithm to extract the best. The same algorithm or the system, once you replace the world model with a different one, such as a self-driving world model, a battleground world model, and maybe the Mars model, you actually can use the same simulative reasoning algorithm to solve problems in those spaces. And the experience learned from such an approach with a world model is also highly transferable, as we also experience in our daily life. For example, we train the astronauts who can walk on the moon by letting them do scuba diving, right, to experience that loss of gravity. A mountaineering athlete could of course be a soldier to command ambush in the mountain regime. And we also know that nowadays, a good game player could be a good combat soldier because he may be very good at manipulating weapons like drones and other things. So now let's go back to the core question about how to build such a general world model. And here are the four metrics or maybe elements that I found to be key: representation of the space of latent concepts and the world states, architecture to effectively allow you to reason, and also objective function for you to train the model, and of course how to use it. So I will break down each of these elements more carefully, but because we are also set up a discussion, a debate with Professor Yann LeCun, I also want to present a few alternative approaches such as I listed here in terms of addressing all the design questions of a world model. I just listed here a proposal in this particular school of thought. It was mentioned that the data to use to train a world model should be images rather than text because text data is too small. Image data are bigger and richer, and also the world state should be continuous embedding rather than discrete because the continuous representations are very friendly to mathematical manipulations like gradient operations. They're efficient. We should use an autoregressive model. We shouldn't use an autoregressive generative model according to this school because they are error-prone due to the intractability of the complex outcomes that they have to generate. And also, the loss function shouldn't be data reconstruction loss as in many generative models, but should be more like a latent representation prediction problem. And then how to use it, there are proposals like MPC, model predictive control. So this is a proposal which I found to be interesting, especially this one, which is that the loss function should be built on latent state prediction rather than data reconstruction. And among many of the recent efforts, there are different implementations of world models that I listed here, which kind of are not all in compliance with the proposal I mentioned earlier, but they also share some commonality, which is that it seems like what they care about is to generate good videos and pictures rather than focusing on additional functions of a world model such as deep reasoning through simulation. So in that context, I want to present a body of work that my group has done recently. The idea comes from this paper, Critical World Models. If interested, you can go and read where I systematically break down the foundational theoretical insights behind our design. Again, the key considerations for building a world model is data, representation, model, loss, and usage. And let me discuss each of them one by one a little bit. So on data, you know, text data is very small, as argued by Yann and many others. The data that we see as a child could be already way exceeding the volume of all the data in text form used to train large language models. However, I want to point out that when we talk about data, it is not just about data volume, it's also about information density. I would say that text data really gives more dense information than pixels because language is kind of a human embedding of what they see in the real world. And there are complex contents, information such as abstract concepts like regrets and justice that is very hard to extract from video. And also, you use language to enable natural interface of interaction, which is also difficult through visual cues. Therefore, our proposal for the training data is to use all modalities rather than only text data. And that's the first question, hopefully rather straightforward. Representation is another issue. The proposal from the JEPA school proposes or prefers a continuous latent representation because they are gradient-friendly. However, we also know that continuous sensors can be very variable, and also if for every continuous input you compute an embedding, these embeddings are stateless because they are highly variable and unstable. Therefore, it is not easy to extract semantic meanings out of those embeddings. Just like what I show in this figure, different image patches can hardly lead to a consistent semantic embedding that can be stable upon different appearances of the same content. And therefore, our representational choice is to use a stateful and identifiable semantic latent space to anchor the reasoning in both long-term and short-term memory. And for example, we use the words, as I mentioned, the words are human tokens for embedding already what people see, but also even for new things that we don't have a word, you can use techniques like VQ-VAE to convert sensory signals into tokens and so forth. And these representations really enable a more powerful way of capturing information both in continuous embedding and also in discrete tokens. So we keep both of them together, and we also have a theory to prove that in fact you have the flexibility of building new vocabulary to capture information even in continuous signals as close as possible by playing a trade-off of the vocabulary size and also encoding length. So for the sake of time, I'm going to skip the technical theoretical details of that argument and jump to the next dimension, which is architecture. So the architecture debate on world models is the most interesting one. For example, the JEPA architecture is like what I draw here. It is an encoder-decoder architecture which basically encodes the inputs from both the previous time point and next point into latent space, and then you build a world model which makes prediction of the latent states from the previous embedding to next embedding. This is a very interesting architecture which is computationally very, very efficient because the loss function is rather easy to minimize. However, they have some flaws as I'm going to discuss in the debate. So for example, there are issues about identifiability, right, and also there are issues about collapsing, which I'm going to mention in a second. So our approach is called a generative latent prediction approach, which is a closed-loop architecture with an encoder to begin with and then carried through a reasoning backbone and then a decoder back into the visible space. Of course, they operate on a mixed latent representation. Therefore, in the backbone, you can have an LLM-style transformer reasoning backbone and also a diffusion model-like backbone to drive the numerical signals. And this hierarchical representation, each level can have a proper level of abstractness, and that's basically an architecture that we propose to use. And one of the main reasons, other than the capability of reasoning in mixed representation space, is also the more principled trainability and the loss function definition of this architecture. Because when you use JEPA or a latent prediction architecture, you can define a loss function that is more optimizable and stable. It turns out that this loss function which you defined on a pair of latent states is mathematically ill-behaving. It can easily collapse. For example, you can imagine that if I just encode both the previous observation and the next observation into a constant, then you have a trivial predictor that always does the same, right? And there are other conditions which can compromise the utility of the encoder and also the predictor as I will discuss later in the debate. So what we use is a generative loss which directly measures the difference between the generated observation from the decoder against the actual observation that you receive, and it is a well-defined probability-motivated loss function. And we can also prove that this generative loss is strictly an upper bound of the latent loss. Therefore, by minimizing this generative loss, you also can minimize the latent loss. But the reverse is not true. When you minimize the latent loss, you do not necessarily minimize the generative loss. And on the usage, once you have a world model, how to use it is another interesting topic that I want to say a few words. The JEPA school proposed an approach of using the world model only at inference time called model predictive control. This is a well-known technique that is used in factory industrial control and so forth. The idea is that once you have a world model and when you are doing planning or predicting actions, you can always use the world model to simulate a few steps, and then you can solve a bounded optimization problem by finding the best path among the world states and also the simulated actions, and then you pick an action sequence. Well, this is mathematically an elegant approach, but it is very expensive because in many real-world problems, you may be simulating or solving a plan that involves hundreds of steps, and in each step, hundreds of thousands of action options, and that's becoming an intractable optimization problem under MPC, and also it has to be done in real time when you actually solve problems. So our approach in using the world model is a combination of MPC but also an offline RL algorithm where when you are not even doing anything as an AI agent, you can still simulate the world model just in your system and learn policies of possible simulated world states and world spaces. This is more like a Go player not only learning during the actual play but also doing simulation or self-play when he's not in a match and still gains more skill. And we found that the world model gives you that flexibility, and we actually can have a theorem proving that the RL algorithm that you learned and then stored in the associated memory as a policy won't compromise your actual action when you deal with a real-world situation because you always have the other option of going to a real MPC-based optimization approach if need be, but you don't have to always go that route. So to summarize, our world model architecture, which is called the generative latent prediction architecture, is a system like this. It does consistent understanding and prediction of the world states, and it is a generalist high-level world dynamics plus a low-level perception details. So two types of reasoning backbones enabled by a large language model pre-trained and also a multi-modal decoder, and then you can also incorporate instruction tuning and action encoding through the spectable. So based on this framework, we recently implemented and released an instance of the world model called Pan. And just as you can see, Pan is an initial version implementation of this architecture which uses next latent state prediction but also an integrative diffusion fusion architecture for the reasoning, and it actually predicts multimodal observations which allow you to communicate with other agents and also perform generative supervisions. And also the building blocks in each of the components and modules here is actually compatible with other foundation models that you can obtain from the community, such as the existing language models, VLMs, the nodes, and so forth, with a canonical interface. The learning is based on next prediction and the generative reconstruction and also reinforcement learning as I mentioned earlier. Against a particular technical issue raised by the JEPA approach, which is that reconstruction of complex visual signals can be expensive, we actually also have some additional techniques to mitigate that computational cost. For example, we can design a particular video diffusion decoder which does partial reconstructions of the outcome focusing on only the important elements through a queuing kind of mechanism that allows you to do reweighting of different contents. Recently, we also released and developed a new technique which is working at the infrastructure level to speed up the very computation of LLM reasoning and also of a diffusion model based on diffusion distillation, sparse attention, and also fast LLMs. It turns out that this approach is highly efficient that it can generate, for example, the video 10 times faster than the play-out of the video. So for the first time in the community, we can actually produce video contents faster than what the video content itself takes. And this type of technique again can reduce the cost of the JEPA architecture of the world model to make it more viable and more able to leverage on complex data. So let me show you some demonstrations of this Pan model. To qualify as a world model, you must go beyond just the image generation but enter the world of visual understanding and also visual and world simulation. So what does understanding mean? It's very hard to really articulate very clearly the property of understanding. So here we take a more operational definition, which is manipulability. You can actually manipulate content through text prompts continuously to drive and steer the production of the contents. For example, here I have a current state. If I want to go one step forward and say move forward, you see this happening. If I say go lift off, you see this happening, which kind of tells you that the model actually understands what is in the picture, and depending on what you want to see happening, they will branch out into different next states. You can go further and put some weather conditions. Likewise, you see a nice simulation in compliance with your prompts. So the Pan model is not a task-specific model that only focuses on a narrow type of contents. It is very, very diverse and universal, and you can see many of the interesting scenarios simulated from here. For example, this scenario caught many people's attention because it is simulating a car behavior. Let's see if I can play it or not. Yeah, this is a video which leads companies like Waymo to be very interested in because to train autonomous driving cars, you need to have enough video signals to really start illustrating and training behaviors of rare accidents when you run into that. But such videos are very, very rare if they exist in the real world, and therefore our world model can help to even do data generation of those rare scenarios. Here is another video showing you step-wise manipulability of the possible worlds. Again, the video itself is not fascinating like what you see in Sora or Gemini and so forth, but it is a different type of video which is not pre-scripted or bulk programmed in a single passage. You can actually go into the video and in any second type in a new action, a new direction, and the model actually doesn't know what's going to happen next. Only after you type in, they are going to start to simulate what is going on next, which again demonstrates that the model knows what is in the video in terms of micro with the oven, the balls, and also the bottles and so forth. Again, we use this type of task to demonstrate understanding of the possible worlds. Again, coming to evaluation and demonstration, evaluation benchmarks of world models is actually a problem by itself because the current visual-based so-called world models really are testing something not so fundamental to a world simulator. For example, they test short-horizon metrics like frame-level fidelity or motion realism and so forth, which of course is cute and useful for high video quality, but they have very little value to generate long-horizonally consistent scenes for reasoning and planning. So we come up with a new set of benchmarks which measures long-horizon consistency, high-level controllability, and also planning qualities. And this is basically a set of more extensive definitions of those metrics: action simulation, long-horizon forecast, and scene planning and reasoning. Again, around these tasks, we see very few models able to achieve very good scores because that's exactly what the current models have been largely ignoring. So our paper did a very systematic comparison of, for example, action simulation fidelity, long-horizon forecast, and simulation reasoning. Especially the simulation reasoning, you simulate to solve a problem rather than to generate a pretty video stream, and those kind of functions are rarely if ever considered in the past as a quality measure of the world models. So let me just give you a quick kind of demonstration about how this task each look like. For example, here is a simulation of what we call long-horizon and branching possibilities. You start from somewhere and then you can enter different action prompts and steer the movie multiple steps into different next worlds or next states. For example, you ask the flower to bloom by itself, or you do a golden light spell, you light a fire, or you maybe enter a tunnel, and so forth. You can really drive the whole possible world further, maybe even indefinitely into the next world, and actually can come back to its original world because the states are stateful simulation, therefore the past, the history will actually be remembered. So this is useful for you to do simulation of planning and thought experiments. Here is another instance about how to use the world model to solve a problem. So here, for example, the goal is to start from this particular scene where you have three bins, and the middle bin has the four yellow objects, the four yellow cans, and then you have the blue tray on the right and the middle tray in white, and you want to now achieve an equal split of the cans between the blue tray and the white tray, right? So this is a simple task. How can you actually do simulation reasoning to find the right movement to achieve that? And this is basically what the world model is able to demonstrate. It basically can simulate multiple things, picking the can up, put it right or left, and then next put it back or continue to pick the next can, and so forth. So there is one path which leads to an equal split of the cans through a simulation of possible actions and next worlds, and others failed. So before you do this action physically using a robotic arm, you could already simulate the whole thing very quickly in your brain, and then the arm can now get a proper sequence of actions to execute this goal. So that's basically how we see the world model to be able to get used, such as leading you out of a complex maze if you get confused or lost your way. Right? So there are a lot of use cases that a Pan model could offer, and here are just examples about how they can be used to simulate sports, industrial, robotic manufacturing environments, physics, or even social training. And I think this is really a very promising direction for world models to pursue beyond just generating nice movies. Okay, so I have a few minutes left. I want to go back to our original goal of superintelligence and super agents. And so now we have a world model, and our goal is to build a goal-oriented agent acting in the environment and achieve goals in a Dune fashion. Now the environment is given by the world model already, and what's next? It turns out that this definition of the agent is just as confusing, if not more confusing, than the world model out there. Right? For example, there are many agent systems these days from different companies. I have to say that it may be a little bit inappropriate to really call them agents rather than pseudo-agents because the main focus of these tools is really automation, finishing a job using different tools and so forth. Does this really suggest agency? Well, what is human agency? Well, human agency is a very, very different property. First of all, humans are very versatile. They use one cognitive system to solve all these tasks which involves physical intelligence, social intelligence, emotional, and many things, and therefore functionally it's more versatile than any of the existing agent systems that you see in terms of adaptation, innovation, communication, and continuous learning, and so forth. But even more so, human agency, a biological agency, has a very, very interesting cognitive element such as the definition and awareness of self, the boundary between self and others, so that you can define collaboration and teamwork, and also the clear recognition of a goal and accordingly self-regulation and self-organization and reflection, and so forth. None of the current agent systems had anything to do with these qualities. But I think this is really needed if you want to define a truly autonomous agent that can rely less on external instructions and still be able to accomplish complex tasks. Therefore, we define the agent model with a set of characteristics and capabilities closer to human agency, such as actively perceiving the goal, reasoning about the purpose, and reasoning about purposeful actions, and learning from experiences, being able to adapt and generalize, and so forth, right? And so going back to the elements of the agent model, again, here is what we see to be useful: goal, belief, world models, and planning. And I'm going to propose again a draft architecture for these types of agent models. So I want to start from what's out there already in terms of empowering agent models. There are now two conventional schools of thought for agent models. One is a policy-based kind of approach, which is a fast reaction and built on end-to-end policy models that can output action directly, right? They have the advantage of being very, very simple and also being very, very computationally efficient and rapid. They don't maintain a model of the world; it's purely reflexive based on perceived states of the world and then roll out action using a policy model. The other approach is a model-based approach, which is actually closer to what Professor Yann was also advocating, right, using a world model to really instruct and guide the agent and to come up with rational actions. Again, the advantage is that it really opens the possibility of long-term reasoning and data efficiency without necessarily to train a complex policy model. But on the other hand, obviously, it is expensive on the other end because you need to always do simulative reasoning or MPC. Right, I would classify these two approaches into what we know as the
System one and system two cognitive reasoning: one is based on policy learned from reinforcement learning, a subconscious, reflective kind of activity. The other is more like inference-time planning based on MPC or MCTS using a model-based approach. I don't think these two should really be put against each other. In fact, I would propose in our model what we call a system three, which not only can do both but also has an additional function which we call a regulator or a deliberator, which can instruct the system when to use system one, which is cheaper but short-term, or when to use system two, which is more expensive but longer-term. And that's basically a proposed architecture which includes a world model that is pre-trained and a policy model that can be trained previously or can continuously be trained, and also a system three reasoning which is a configurator for when to plan, deliberate, or when to stop planning and act directly. So the architecture is a little more complicated. You see the configurator which will deliberately choose a policy model or a world model depending on the distance or the quality of the effects coming out of the previous action and so forth. And there are other designs that I'm going to skip the details just to wrap up. So to validate this hypothesis, we actually recently also built a prototype of a self-regulatory agent model called SLAM, which is presented on the right. On the left is a classical LLM agent which is really a black box that you prompt and it will basically do reasoning for you using an unknown thinking path or chain of thoughts. On the right you have a system that is a white box where the configurator will allow you to deliberately choose when to plan, how much to plan, and also using what intermediate results to replan or to evaluate. And the results shown here show an interesting contrast. We found that using this system three architecture, we were able to achieve very complex planning and reasoning behavior with a relatively smaller model at a capacity comparable to a model that is one or two orders actually bigger. As you can see from here, this graph on the x-axis shows many models of different parameter sizes and the y-axis is the score of the performance. You can see that the SLAM model is comparable in performance to the best models out there such as DeepSeek 2 and Qwen 2.5, which are literally hundreds of times bigger. And this model is only 20-30 billion parameters. And the way to achieve that is because there is a cost-effective planner in the system that allows you to do more smart planning. And this is another demonstration about the comparison of our agent model compared to other models of comparable sizes, apple-to-apple comparison. You can see that the reasoning per token is much smaller and the overall score is strong. Here is an insight about what's happening inside. So in a self-regulatory agent model, you actually start with a much richer initial plan which allows you to deliberate and replan when necessary, even for a very complex goal. But the overall steps required to arrive at the right solution, in this case, it's a pretty difficult question: how many non-indigenous crocodiles were found in Florida from some years between. It's a very interesting problem that requires you to do planned search and literature review and analysis of data in multiple steps. So our model had a couple of different plans and ultimately reached the right answer in 45 steps. Whereas a model that is not pre-trained with reinforcement learning for the policy for the configurator takes many more steps and gets the wrong answer. And the external benchmark model that we use also takes longer steps and gets the wrong solution. So this is basically what we like to see: how the world model in the future can go hand-in-hand with the agent model to really go toward a higher level of intelligence. The world model here is really used for perpetual retraining of the agent system in an infinitely possible world. And this is actually a realization of such a game plan. We already used the Pan world model to build an initial version of the game model where you can simulate long-horizon, high-fidelity, interactive simulation experiences. Long horizon here is important because that's where you really can learn quality policies and also models with enough complexity closer to the real-world rollout of complex scenarios. And the world model you can see is really flexible in responding to complex action prompts and come up with different outcomes. And accordingly, agent-to-auto benchmarks can also be built on such a world model which really measures the system against the task completion. In this case, for example, opening the correct door to go in, or find the parking spot for your vehicle, or find the ice cream selling station out of many possibilities and so forth. So you can use the world model to simulate the scenario and the outcome and then let your agent be trained. So to wrap up, this is basically just the beginning. With the world model and the agent model, you can always compound it with some more complex scenarios. For example, a regular world model when additional agents are added into it becomes a super world model of a physical world plus many agents. And then the super world model can lead to a super agent model which really thinks about the world in certain abstractions. For example, economists think about the world as a collective population rather than every individual in terms of financial and economic behaviors. So this type of nested structure of a super world model and super agent model can allow your planning and reasoning to go to arbitrary levels of complexity and granularity. So to sum up, what I presented just now is a roadmap to go beyond book intelligence toward physical intelligence and social intelligence and philosophical intelligence. And I believe that a simulation of reasoning with the world model is a very promising path toward building true intellectual intelligence and problem-solving capabilities in very complex scenarios because they are scalable, transferable, and also very flexible. They actually can come up with very complex scenarios and generate very complex possibilities even from simple root models. And that actually gives the world model a lot of potential in achieving high performance even with limited amount of resources. And of course, I talked about the issue about representation learning and the architecture, which is still an ongoing debate, which I again look forward to an exchange with Yann in the next few hours. That's all for my first talk. Thank you very much.
