Intelligence + Continual Learning = Expertise — Yu Su, NeoCognition

I understand that I'm standing between you and lunch, so I'll try to be quick. My name is Yu Su. I'm a professor at The Ohio State University, and I also have another job as COO at a company called NeoCognition. We focus on agents and continual learning. Today's talk won't be too technical; it will mainly be a conceptual one. I think it's a very important conceptual distinction that I will try to make between what is intelligence and what is expertise. Through this, I will try to answer some very bothersome questions for me, such as why we are successful at building coding agents, but they are terrible at anything else. Why are current agents so token-inefficient that every company is now trying to curb their token-mixing efforts? Hopefully, this will provide some food for thought before lunch.

First, a bit of history. AI agents are not a new thing. We have been trying to develop agents throughout the whole history of AI. But the problem is that in the early stages, let's say in the 1960s to the 1980s, when we developed expert systems or logical agents, or in the 2010s, when we developed deep RL-based neural agents, we were only able to capture some very limited facets of human intelligence—whether it's logical reasoning or perception in single modalities leading to decision-making. Only recently, with the advent of multi-modal LLMs and the language agents built on top of them, do we have, for the first time, a neural model that is able to encode multi-sensory inputs into a unified neural representation that is also conducive to symbolic reasoning and communication. That was a trait unique to humans.

Now, AI agents finally have the same capability. That drastically improves their expressiveness, their reasoning ability, and their adaptivity. That's why I think we have really entered a new evolutionary stage of machine intelligence. It didn't take long for these language agents to find their first mass market, which is coding. The best way to illustrate this is probably through the revenue graph of Anthropic. In just under two years, their revenue has grown 400 times to $40 billion. I think the newest number is maybe $60 billion in annualized runway.

This growth is largely driven by coding and coding-related productivity capabilities. But if we think about it, coding is really the ideal market for these language agents because code is already a language-native world. Everything is already represented symbolically and recorded in a very structured way. You get your rewards and your tests all in place in symbolic ways. Then what happens when we leave the privileged world of code? Well, not much goes well. We are running into a lot of challenges deploying these agents in enterprise settings. Also, in personal settings, the open class of problems constantly causes these agents to make brittle and silly errors. To the extent that Andrew Ng said it's not going to be the year of agents, but the decade of agents, because they cannot do computer use and they don't have continual learning. I don't know how much Andrew Ng's thoughts have changed since then because of the coding agent and everything, but I think the difficulties with computer use and continual learning are still largely the same now.

How can something be small but also brittle at the same time? Here's my thesis around it. I think we are witnessing a modern version of Moravec's paradox. Moravec's paradox states that for AI, hard things are easy, and easy things are hard. The modern version here is that we are very good at these symbolic reasoning tasks—coding and math—which were considered the crown jewels of intelligence earlier. But then we still struggle with everyday digital work because it really requires quite a different set of cognitive competencies to excel. More specifically, I think modern society is really not just one unified world. It's millions of these microworlds. Every domain and every profession is different. Every company is different. Even if you're using the same software, each company configures it differently. It's extremely idiosyncratic, especially in the digital world. It has its own unique local physics, different structures, constraints, affordances, and dynamics that you have to learn. It's just too heterogeneous and dynamic for any monolithic model to try to compress into one static representation. Agents must continually learn on the job to acquire what I call specialized expertise for each specific microworld.

In the second part of the talk, I will try to establish the differences between intelligence and expertise. Here are the working definitions. Intelligence is the capacity to reason through unfamiliar problems from available context. This is what the frontier models are increasingly good at. You give them the problem statement, the context, the tools, and they can reason through it, even if it's a scene they've encountered for the first time, and they can do a great job. Every episode is more or less independent from the others here.

But expertise is different. Expertise is really accumulated and situated competence. It's the ability to act reliably, efficiently, and with judgment to achieve reproducibly superior performance in a particular domain. This is in stark contrast with intelligence. To show what expertise contains, I think the key idea from cognitive science is that experts don't just know more facts—they see the world differently. Expertise allows you to do different pattern recognition. You see through the specific patterns. If an expert is looking at a gigantic bug report, they can immediately locate the most plausible places where things could go wrong. They think about the problem with a very deep structure. When you are scheduling a meeting, it's not just about finding a shared slot on everyone's calendar; it is a constraint optimization problem involving everyone's authority, priorities, urgency, and other factors. Experts don't just operate with a set of rules or a set of facts. We know that every single thing is conditional. Every rule has preconditions where it applies, but we also know when we can bend reality or bend the rules when exceptions happen. Finally, this also gives us judgment and taste. It's important to know what is high quality, and, very importantly, when to stop—when something is good enough. All these together mean that experts have effectively built a world model of their environments. It's a generalized notion of a world model that captures how that microworld works, and that becomes the basis for all of our perception, reasoning, decision-making, and judgment.

Intelligence and expertise are really quite different across many dimensions, but some of the interesting ones here are: intelligence is about, "Given the context, how do we solve the problem through the context?" But expertise brings you the context. Given any problem, we know what context is important to bring in to solve the problem. Because of that, intelligence tends to expand our search. Every problem-solving process is a search problem. Intelligence tends to brute-force it, trying to spin up a hundred different parallel ways to solve the problem. Expertise, on the other hand, tries to compress the search space because it has learned essential shortcuts for the problem space. Whenever you have a problem, expertise helps you find the most plausible ways to solve it. I also think the final part here is that continual learning is the important bridge from intelligence to expertise. But first, let me try to define continual learning, because it's such a confusing term. Jack just gave some definitions earlier—there were ten different names. But here's the definition I work with: I think continual learning is adaptive compression of experience into reusable structures for future behavior. All four elements here are very important. For experience, we need to answer the question: what kind of experience are we talking about? Is it more episodes of experience, or are these semantic facts, procedures, feedback from humans, or from environments? How do we compress that? Do we embed experiences into vectors, or index them into some symbolic structure? Do we distill them into model parameters, or use some kind of reinforcement learning? It's not just one-time compression—it needs to be adaptive compression. What you have learned and what you have compressed so far should largely influence how you compress further. What kind of structure are we looking at? Is it just parameter adapters of your language models, or is it vectors, graphs, skills, or even world models? How do you use these reusable structures? Do you use them just to recall facts, or for prediction of future states? Do you use them for better planning, for the control and actuation layer of the agent, or as a value function for potential states? Because of this, the continual learning problem is rich. It has these four different aspects, and each aspect can be instantiated in different ways. That makes this field confusing, but hopefully this is a definition that encompasses most versions of continual learning. I think this is maybe the most important figure in this talk. If we put raw intelligence as the x-axis and expertise as the y-axis, I think we'll find that they are largely orthogonal to each other. If you don't have continual learning and all you do is scale your model to get better raw intelligence, then what you get is what I call the world's smartest novice. It's super smart and can try to attack any problem provided to it, but it doesn't accumulate expertise—it just brute-forces its way through every problem. If you have continual learning, different continual learning algorithms will set the slope of your learning curve. If you have a sloppy continual learning algorithm, maybe some kind of simple in-context learning, then with increasing intelligence, your expertise will increase only a little. But if you have a really strong continual learning algorithm, then expertise will increase rapidly. Of course, this assumes a given time horizon and experience horizon. Among all the potential futures that good continual learning can bring us, I think the most interesting is what I call unbounded expertise from bounded intelligence. What if we can come up with a continual learning algorithm such that, once raw intelligence has crossed a certain threshold, we don't need stronger intelligence anymore? Continual learning would bring us unbounded expertise once we have a reasonable level of intelligence. We can call this the escape intelligence. If this is indeed true, it will have a lot of implications for the whole ecosystem. Do we need to continually train larger and larger models? Or are these models already good enough, and what we're missing is just better continual learning algorithms? To be a little more concrete and provide more food for thought, here are some open questions in this space. The overarching question is: given any domain or environment, how can an agent continue to learn, specialize, and reach expert-level competency? But to do that, you need to answer many other questions. How do you even define and measure expertise? This is probably environment-specific. How do we handle the trade-off between reliability and plasticity? We want these agents to be both reliable and plastic, but these qualities inherently conflict with each other. Reliable or stable systems resist change, while plastic systems embrace change. How do we reconcile that? Fortunately, we have a living existence proof—ourselves, humans. We are incredibly plastic, yet also manage to be dependable most of the time. From a technical perspective, when we talk about learning, there are largely two forms: parametric and non-parametric. My belief is that both are really needed for continual learning to work. But how do we synergize the two? Finally, even though we are focusing on specialization, I think there is great potential for specialization to lead to better generalization. We have already exhausted public data for training large language models, but the next stage of training—the next internet-scale data opportunity—is in all of these different private worlds. If we can make these specialized agents work, they can learn in situ and channel their learning back to the general model. That may be the next internet-scale data opportunity. Finally, a call to action: let's start scaling expertise. This will be a new dimension for us to scale, because intelligence is already becoming abundant. The frontier models are probably smarter than the average human, but expertise is still scarce. We want to build a world where expertise becomes abundant, where everyone can get expert support. In an ideal world, everyone could have their own personal healthcare, personal financial advisor, personal tutors, and so forth. Every company could then build its own learning loop. As Satya said two weeks ago, we want to enable this human-AI learning loop at each company, turning it into institutional memory and allowing every company to build its own moats while still being in charge of its means of production. With an abundance of expertise, we will see more types of work become possible, because right now there are still many opportunities that are locked up simply because the friction is too high to make them economically viable. With an abundance of expertise, I think we will be able to lower that friction and make many new types of work cross the threshold of being worth doing. This is the future we are building toward at NeoCognition. I am happy to share this with you, and thank you for your attention.