Intelligence + Continual Learning = Expertise — Yu Su, NeoCognition

>> All right. >> All right. I understand that I'm standing between I understand that I'm standing between I understand that I'm standing between you and the lunch, so I'll try to be you and the lunch, so I'll try to be you and the lunch, so I'll try to be quick. quick. quick. My name is Isu. I'm a professor at The My name is Isu. I'm a professor at The My name is Isu. I'm a professor at The Ohio State, The Ohio State, and I also Ohio State, The Ohio State, and I also Ohio State, The Ohio State, and I also have another job, which is COO at a have another job, which is COO at a have another job, which is COO at a company called The Neo Cognition, and we company called The Neo Cognition, and we company called The Neo Cognition, and we focus on agents and continual learning. focus on agents and continual learning. focus on agents and continual learning. So, today's talk, So, today's talk, So, today's talk, it won't be too technical, but I would it won't be too technical, but I would it won't be too technical, but I would it will be mainly a conceptual one. But, it will be mainly a conceptual one. But, it will be mainly a conceptual one. But, I think it's a very important conceptual I think it's a very important conceptual I think it's a very important conceptual distinction that I will try to make distinction that I will try to make distinction that I will try to make between what is intelligence and what is between what is intelligence and what is between what is intelligence and what is expertise. And through this, I will try expertise. And through this, I will try expertise. And through this, I will try to answer some of the very bothering to answer some of the very bothering to answer some of the very bothering questions for me that um questions for me that um questions for me that um like why we are so successful at the like why we are so successful at the like why we are so successful at the coding agents, but they're so terrible coding agents, but they're so terrible coding agents, but they're so terrible at anything else, right? Why at anything else, right? Why at anything else, right? Why uh the current agents are so token uh the current agents are so token uh the current agents are so token inefficient uh inefficient uh inefficient uh like to the degree that every company like to the degree that every company like to the degree that every company right now is like coming out and try to right now is like coming out and try to right now is like coming out and try to curb their curb their curb their uh their token mixing efforts in the uh their token mixing efforts in the uh their token mixing efforts in the company. Um company. Um company. Um so so so hopefully this will provide some food hopefully this will provide some food hopefully this will provide some food for thoughts before lunch.

for thoughts before lunch. for thoughts before lunch. Right, first a bit of a history. Um so, Right, first a bit of a history. Um so, Right, first a bit of a history. Um so, AI agents are not a new thing, right? AI agents are not a new thing, right? AI agents are not a new thing, right? It's uh we have been trying to develop It's uh we have been trying to develop It's uh we have been trying to develop agents throughout the whole history of agents throughout the whole history of agents throughout the whole history of AI. AI. AI. Um but, the problem is that in the early Um but, the problem is that in the early Um but, the problem is that in the early stages, uh let's say stages, uh let's say stages, uh let's say um in the night in the 1960s to uh '80s, um in the night in the 1960s to uh '80s, um in the night in the 1960s to uh '80s, when we developed these expert systems when we developed these expert systems when we developed these expert systems or logical agents, or like in the uh or logical agents, or like in the uh or logical agents, or like in the uh 2010s, when we developed these deep 2010s, when we developed these deep 2010s, when we developed these deep RL-based uh neural agents, we were only RL-based uh neural agents, we were only RL-based uh neural agents, we were only able to capture some very limited facets able to capture some very limited facets able to capture some very limited facets of human intelligence, right? Whether of human intelligence, right? Whether of human intelligence, right? Whether it's like logical reasoning or it's like it's like logical reasoning or it's like it's like logical reasoning or it's like uh perception in single modalities to uh perception in single modalities to uh perception in single modalities to decision. Um only recently decision. Um only recently decision. Um only recently with the multi-modal LLMs and the with the multi-modal LLMs and the with the multi-modal LLMs and the language agent built on top of them, for language agent built on top of them, for language agent built on top of them, for the first time, we have the first time, we have the first time, we have a neural model a neural model a neural model that is able to encode multi-sensory that is able to encode multi-sensory that is able to encode multi-sensory inputs into a uh unified neural inputs into a uh unified neural inputs into a uh unified neural representation that is also conducive to representation that is also conducive to representation that is also conducive to symbolic reasoning and communication, symbolic reasoning and communication, symbolic reasoning and communication, right? So, that was a trait uh unique to right? So, that was a trait uh unique to right? So, that was a trait uh unique to humans. Now, uh AI agents finally have humans. Now, uh AI agents finally have humans. Now, uh AI agents finally have the same thing.

the same thing. the same thing. So, So, So, that drastically uh improve their that drastically uh improve their that drastically uh improve their expressiveness, their reasonability, and expressiveness, their reasonability, and expressiveness, their reasonability, and adaptivity. So, that's why I think we adaptivity. So, that's why I think we adaptivity. So, that's why I think we have really entered a new evolutionary have really entered a new evolutionary have really entered a new evolutionary stage of machine intelligence. And uh it didn't take long for these And uh it didn't take long for these language agents to find their first mass language agents to find their first mass language agents to find their first mass markets, which is coding. And the best markets, which is coding. And the best markets, which is coding. And the best way to illustrate this is probably way to illustrate this is probably way to illustrate this is probably through the uh revenue graph of through the uh revenue graph of through the uh revenue graph of Anthropic's, right? In just under 2 Anthropic's, right? In just under 2 Anthropic's, right? In just under 2 years, their revenue has grown 400 times years, their revenue has grown 400 times years, their revenue has grown 400 times uh to uh 40 billion. I think the newest uh to uh 40 billion. I think the newest uh to uh 40 billion. I think the newest number is maybe 60 billion uh annualized number is maybe 60 billion uh annualized number is maybe 60 billion uh annualized runway. And it's largely driven by runway. And it's largely driven by runway. And it's largely driven by coding and coding-related productivity coding and coding-related productivity coding and coding-related productivity uh capabilities. uh capabilities. uh capabilities. But if we think about it, right? Coding But if we think about it, right? Coding But if we think about it, right? Coding is the is the is the really the ideal market for these really the ideal market for these really the ideal market for these language agents because code is already language agents because code is already language agents because code is already a language-native world. Everything is a language-native world. Everything is a language-native world. Everything is already represented symbolically and already represented symbolically and already represented symbolically and like uh recorded uh in a very structured like uh recorded uh in a very structured like uh recorded uh in a very structured way. And you get your rewards, you get way. And you get your rewards, you get way. And you get your rewards, you get your uh like tests all in place in your uh like tests all in place in your uh like tests all in place in symbolic ways.

symbolic ways. symbolic ways. So, So, So, then what happens when we leave the then what happens when we leave the then what happens when we leave the privileged world of codes? privileged world of codes? privileged world of codes? Well, not so well. Well, not so well. Well, not so well. Um we are running into a lot of Um we are running into a lot of Um we are running into a lot of challenges deploying these agents in challenges deploying these agents in challenges deploying these agents in enterprise settings. enterprise settings. enterprise settings. And the uh also in personal settings And the uh also in personal settings And the uh also in personal settings that the uh that the uh that the uh like open class uh constantly make this like open class uh constantly make this like open class uh constantly make this uh uh uh like uh quite brittle and silly errors. like uh quite brittle and silly errors. like uh quite brittle and silly errors. And then to the extent that the Andrew And then to the extent that the Andrew And then to the extent that the Andrew Ng Ng Ng said that said that said that it's not going to be the year of agents, it's not going to be the year of agents, it's not going to be the year of agents, it's going to be the decade of agents it's going to be the decade of agents it's going to be the decade of agents because they they cannot do computer because they they cannot do computer because they they cannot do computer use, they don't have continual learning. use, they don't have continual learning. use, they don't have continual learning. Um I don't know how much Andrew Ng's uh Um I don't know how much Andrew Ng's uh Um I don't know how much Andrew Ng's uh thought has changed uh since like last thought has changed uh since like last thought has changed uh since like last time because of the coding agent and time because of the coding agent and time because of the coding agent and everything, but I think the difficulties everything, but I think the difficulties everything, but I think the difficulties with computer use with continual with computer use with continual with computer use with continual learning still largely the same right learning still largely the same right learning still largely the same right now. now. now. So how can something be so small but So how can something be so small but So how can something be so small but also so brittle at the same time?

also so brittle at the same time? also so brittle at the same time? Here's my thesis around it. I think we Here's my thesis around it. I think we Here's my thesis around it. I think we are actually witnessing a modern version are actually witnessing a modern version are actually witnessing a modern version of the Moravec's paradox. of the Moravec's paradox. of the Moravec's paradox. Right? So the Right? So the Right? So the uh Moravec's paradox essentially states uh Moravec's paradox essentially states uh Moravec's paradox essentially states that uh uh for AI, that uh uh for AI, that uh uh for AI, uh hard things are easy, easy things are uh hard things are easy, easy things are uh hard things are easy, easy things are hard. hard. hard. So the modern version here is that we So the modern version here is that we So the modern version here is that we are very good at these symbolic are very good at these symbolic are very good at these symbolic reasoning tasks like coding and math, reasoning tasks like coding and math, reasoning tasks like coding and math, which were considered crown jewel of uh which were considered crown jewel of uh which were considered crown jewel of uh of intelligence uh earlier. But then we of intelligence uh earlier. But then we of intelligence uh earlier. But then we still struggle with this everyday still struggle with this everyday still struggle with this everyday digital work because they really require digital work because they really require digital work because they really require quite different set of cognitive quite different set of cognitive quite different set of cognitive competencies to excel at them. competencies to excel at them. competencies to excel at them. And more specifically, I think modern And more specifically, I think modern And more specifically, I think modern society is really not just one unified society is really not just one unified society is really not just one unified world. It's millions of these micro world. It's millions of these micro world. It's millions of these micro worlds. Like every worlds. Like every worlds. Like every domain, every profession is different, domain, every profession is different, domain, every profession is different, every company is different. Even if every company is different. Even if every company is different. Even if you're using the same software, every you're using the same software, every you're using the same software, every company a company configure it company a company configure it company a company configure it differently. So, it's differently. So, it's differently. So, it's extremely idiosyncratic, especially in extremely idiosyncratic, especially in extremely idiosyncratic, especially in the digital world. It has its unique the digital world. It has its unique the digital world. It has its unique local physics, like different local physics, like different local physics, like different structures, constraints, affordances, structures, constraints, affordances, structures, constraints, affordances, and dynamics that you have to learn.

and dynamics that you have to learn. and dynamics that you have to learn. It's just like too heterogeneous and It's just like too heterogeneous and It's just like too heterogeneous and dynamic for any dynamic for any dynamic for any monolithic model to try to compress it monolithic model to try to compress it monolithic model to try to compress it into one static representation. into one static representation. into one static representation. So, agents must continually learn on the So, agents must continually learn on the So, agents must continually learn on the job to acquire what I call specialized job to acquire what I call specialized job to acquire what I call specialized expertise for each specific microworld. The second part of the talk, I will try The second part of the talk, I will try to establish the differences between to establish the differences between to establish the differences between intelligence and expertise. intelligence and expertise. intelligence and expertise. Here are the working definitions. Here are the working definitions. Here are the working definitions. For intelligence, For intelligence, For intelligence, it's the capacity to reason through it's the capacity to reason through it's the capacity to reason through unfamiliar problems from available unfamiliar problems from available unfamiliar problems from available context. Right? This is what the context. Right? This is what the context. Right? This is what the frontier models are increasingly good frontier models are increasingly good frontier models are increasingly good at. Um you give it the problem at. Um you give it the problem at. Um you give it the problem statement, the context, the tools, and statement, the context, the tools, and statement, the context, the tools, and it can reason through this even if it's it can reason through this even if it's it can reason through this even if it's a scene done for the first time, a scene done for the first time, a scene done for the first time, uh and it can do a great job. uh and it can do a great job. uh and it can do a great job. Every episode is more or less Every episode is more or less Every episode is more or less independent from each other here. independent from each other here. independent from each other here. But expertise is different. Expertise is But expertise is different. Expertise is But expertise is different. Expertise is really accumulated and situated really accumulated and situated really accumulated and situated competence.

competence. competence. It's the ability to act reliably, It's the ability to act reliably, It's the ability to act reliably, efficiently, and with judgment to efficiently, and with judgment to efficiently, and with judgment to achieve reproduce pretty super real achieve reproduce pretty super real achieve reproduce pretty super real performance in a particular domain. performance in a particular domain. performance in a particular domain. Right? So, this is in stark contrast Right? So, this is in stark contrast Right? So, this is in stark contrast with intelligence. And to And to uh show what does ex- expertise actually uh show what does ex- expertise actually uh show what does ex- expertise actually contain, contain, contain, I think uh I think uh I think uh the key idea from cognitive science is the key idea from cognitive science is the key idea from cognitive science is that experts don't just know more facts. that experts don't just know more facts. that experts don't just know more facts. They actually see the world differently. They actually see the world differently. They actually see the world differently. Right? So, Right? So, Right? So, um the uh expertise allows you to do um the uh expertise allows you to do um the uh expertise allows you to do different pattern recognition. So, you different pattern recognition. So, you different pattern recognition. So, you see through the specific patterns. Like, see through the specific patterns. Like, see through the specific patterns. Like, if you're looking an expert is looking if you're looking an expert is looking if you're looking an expert is looking at like a gigantic bug report, they can at like a gigantic bug report, they can at like a gigantic bug report, they can immediately locate like the most immediately locate like the most immediately locate like the most plausible places where things could go plausible places where things could go plausible places where things could go wrong. Um wrong. Um wrong. Um and they think about the problem with and they think about the problem with and they think about the problem with like a very deep structure, right? When like a very deep structure, right? When like a very deep structure, right? When you are scheduling a meeting, you know you are scheduling a meeting, you know you are scheduling a meeting, you know that it's not just like finding the that it's not just like finding the that it's not just like finding the shared slot on everyone's calendar is shared slot on everyone's calendar is shared slot on everyone's calendar is actually a constraint optimization actually a constraint optimization actually a constraint optimization problem over everyone's authority, the problem over everyone's authority, the problem over everyone's authority, the priorities, the urgency, and everything.

priorities, the urgency, and everything. priorities, the urgency, and everything. Um and Um and Um and we don't experts don't just operate with we don't experts don't just operate with we don't experts don't just operate with a set of rules, a set of facts. We know a set of rules, a set of facts. We know a set of rules, a set of facts. We know that every single thing is conditional. that every single thing is conditional. that every single thing is conditional. Right? Every rule has like the Right? Every rule has like the Right? Every rule has like the preconditions where it applies, but then preconditions where it applies, but then preconditions where it applies, but then we also know when we can bend the we also know when we can bend the we also know when we can bend the reality, we can bend the rules when reality, we can bend the rules when reality, we can bend the rules when exceptions happen. exceptions happen. exceptions happen. Right? Right? Right? And finally, that also give us judgment And finally, that also give us judgment And finally, that also give us judgment and taste. It's importantly what's like and taste. It's importantly what's like and taste. It's importantly what's like high quality and uh very importantly high quality and uh very importantly high quality and uh very importantly when to stop, when it's good enough. when to stop, when it's good enough. when to stop, when it's good enough. Um so, all these together, I think Um so, all these together, I think Um so, all these together, I think experts effectively has have built a experts effectively has have built a experts effectively has have built a world model world model world model of their environments. of their environments. of their environments. Right? That it's a generalized notion of Right? That it's a generalized notion of Right? That it's a generalized notion of world model that captures how that world model that captures how that world model that captures how that microworld works, and that becomes the microworld works, and that becomes the microworld works, and that becomes the basis for all of our perception, uh basis for all of our perception, uh basis for all of our perception, uh reasoning, decision-making, uh and reasoning, decision-making, uh and reasoning, decision-making, uh and judgment. So, intelligence and expertise are So, intelligence and expertise are really quite different across many really quite different across many really quite different across many dimensions, uh but some of the impor- dimensions, uh but some of the impor- dimensions, uh but some of the impor- interesting ones here are like interesting ones here are like interesting ones here are like intelligence is about, "Hey, when we intelligence is about, "Hey, when we intelligence is about, "Hey, when we have the context, uh how to solve the have the context, uh how to solve the have the context, uh how to solve the problem through the context." But, problem through the context." But, problem through the context." But, expertise actually will bring you the expertise actually will bring you the expertise actually will bring you the the right context. Right, given any the right context. Right, given any the right context. Right, given any problem, we know what context bring into problem, we know what context bring into problem, we know what context bring into are important for this problem and bring are important for this problem and bring are important for this problem and bring it in to solve the problem. And because it in to solve the problem. And because it in to solve the problem. And because of that, uh intelligence tend to expand

of that, uh intelligence tend to expand of that, uh intelligence tend to expand our search. Like every problem solving our search. Like every problem solving our search. Like every problem solving is a search problem. So, intelligence is a search problem. So, intelligence is a search problem. So, intelligence tend to brute force it. Try to uh try to tend to brute force it. Try to uh try to tend to brute force it. Try to uh try to spin up like 100 different uh like uh spin up like 100 different uh like uh spin up like 100 different uh like uh parallel ways to to try to solve the parallel ways to to try to solve the parallel ways to to try to solve the problem. Well, expertise will actually problem. Well, expertise will actually problem. Well, expertise will actually try to compress the search space because try to compress the search space because try to compress the search space because expertise has constructed this has expertise has constructed this has expertise has constructed this has learned this essential shortcuts for the learned this essential shortcuts for the learned this essential shortcuts for the problem space. So, that whenever you problem space. So, that whenever you problem space. So, that whenever you have a problem, you know the most have a problem, you know the most have a problem, you know the most plausible ways to solve it. Um And then, I also think the final part And then, I also think the final part here is that I think continual learning here is that I think continual learning here is that I think continual learning is the important bridge is the important bridge is the important bridge from intelligence to expertise. from intelligence to expertise. from intelligence to expertise. But first, let me try to define But first, let me try to define But first, let me try to define continual learning because it's such a continual learning because it's such a continual learning because it's such a confusing term. Um and and Jack just uh confusing term. Um and and Jack just uh confusing term. Um and and Jack just uh gave some definition earlier. Uh it was gave some definition earlier. Uh it was gave some definition earlier. Uh it was like 10 different names. Um like 10 different names. Um like 10 different names. Um But here's my the definition I work But here's my the definition I work But here's my the definition I work with. with. with. I think continual learning is adaptive I think continual learning is adaptive I think continual learning is adaptive compression of experience into reusable compression of experience into reusable compression of experience into reusable structures for future behavior.

structures for future behavior. structures for future behavior. So, all of these four elements here are So, all of these four elements here are So, all of these four elements here are very important. very important. very important. For experience, we need to uh answer the For experience, we need to uh answer the For experience, we need to uh answer the question like what kind of experience question like what kind of experience question like what kind of experience we're talking about. Is it more like we're talking about. Is it more like we're talking about. Is it more like episodes of experience or is like uh episodes of experience or is like uh episodes of experience or is like uh these semantic facts or procedures or these semantic facts or procedures or these semantic facts or procedures or feedback from human or in environments? feedback from human or in environments? feedback from human or in environments? And how do we compress that? And how do we compress that? And how do we compress that? Uh so, in we embed them into vectors or Uh so, in we embed them into vectors or Uh so, in we embed them into vectors or we index them into some symbolic we index them into some symbolic we index them into some symbolic structure. Uh we uh distill them into structure. Uh we uh distill them into structure. Uh we uh distill them into model parameters or do some uh kind of a model parameters or do some uh kind of a model parameters or do some uh kind of a reinforcement learning. And reinforcement learning. And reinforcement learning. And it's not just like one-time compression. it's not just like one-time compression. it's not just like one-time compression. It needs to be adaptive compression. It needs to be adaptive compression. It needs to be adaptive compression. Like what do you have learned, what do Like what do you have learned, what do Like what do you have learned, what do you have compressed so far should uh you have compressed so far should uh you have compressed so far should uh largely uh influence how you compress largely uh influence how you compress largely uh influence how you compress further. And what kind of structure we're looking And what kind of structure we're looking at? Is just like parameters like at? Is just like parameters like at? Is just like parameters like adapters of your uh language models or adapters of your uh language models or adapters of your uh language models or is vectors, graphs, or skills, or even is vectors, graphs, or skills, or even is vectors, graphs, or skills, or even word models? And then how do you use word models? And then how do you use word models? And then how do you use these reusable structures? Is like uh these reusable structures? Is like uh these reusable structures? Is like uh you use it just to recall these facts or you use it just to recall these facts or you use it just to recall these facts or you use it for prediction of like future you use it for prediction of like future you use it for prediction of like future states? You use it for uh for better states? You use it for uh for better states? You use it for uh for better planning, for or even for the control planning, for or even for the control planning, for or even for the control like actuation layer of the agent, or as like actuation layer of the agent, or as like actuation layer of the agent, or as a value function for potential states.

a value function for potential states. a value function for potential states. Right? So, it's because of this uh the Right? So, it's because of this uh the Right? So, it's because of this uh the continual learning problem is so rich, continual learning problem is so rich, continual learning problem is so rich, like it has these four different like it has these four different like it has these four different aspects, and if different aspects can be aspects, and if different aspects can be aspects, and if different aspects can be instantiated in different ways, that instantiated in different ways, that instantiated in different ways, that makes this field so confusing. But makes this field so confusing. But makes this field so confusing. But hopefully this is a definition that uh hopefully this is a definition that uh hopefully this is a definition that uh encompasses most of the uh versions of encompasses most of the uh versions of encompasses most of the uh versions of continual learning. Then, I think that uh Then, I think that uh this is maybe the most important figure this is maybe the most important figure this is maybe the most important figure in this talk. in this talk. in this talk. Um Um Um if we put if we put if we put it raw intelligence as the x-axis and uh it raw intelligence as the x-axis and uh it raw intelligence as the x-axis and uh expertise as the y-axis, expertise as the y-axis, expertise as the y-axis, I think we'll find that they are largely I think we'll find that they are largely I think we'll find that they are largely orthogonal to each other. orthogonal to each other. orthogonal to each other. If you don't have continual learning, uh If you don't have continual learning, uh If you don't have continual learning, uh all you do is scaling your model to to all you do is scaling your model to to all you do is scaling your model to to get better like raw intelligence, then get better like raw intelligence, then get better like raw intelligence, then what we will get is what I call the what we will get is what I call the what we will get is what I call the world's smartest novice. Like super world's smartest novice. Like super world's smartest novice. Like super smart, it can try to uh try to attack at smart, it can try to uh try to attack at smart, it can try to uh try to attack at any problem uh provide given to it, but any problem uh provide given to it, but any problem uh provide given to it, but it doesn't accumulate expertise, so it it doesn't accumulate expertise, so it it doesn't accumulate expertise, so it end up as just like brute forcing its end up as just like brute forcing its end up as just like brute forcing its way at every problem.

way at every problem. way at every problem. Then, if you have continual learning, Then, if you have continual learning, Then, if you have continual learning, uh uh uh like different continual learning like different continual learning like different continual learning algorithms will essentially set the algorithms will essentially set the algorithms will essentially set the slope of your learning uh curve here. slope of your learning uh curve here. slope of your learning uh curve here. All right, if you have a sloppy CL All right, if you have a sloppy CL All right, if you have a sloppy CL algorithm, maybe some kind of simple algorithm, maybe some kind of simple algorithm, maybe some kind of simple in-context learning, then uh with like in-context learning, then uh with like in-context learning, then uh with like uh increasing intelligence, then your uh increasing intelligence, then your uh increasing intelligence, then your expertise will increase like a little expertise will increase like a little expertise will increase like a little bit. But, if you have a really strong uh bit. But, if you have a really strong uh bit. But, if you have a really strong uh continual learning algorithm, then uh continual learning algorithm, then uh continual learning algorithm, then uh the expertise will um increase like uh the expertise will um increase like uh the expertise will um increase like uh rapidly. Of course, this is assuming rapidly. Of course, this is assuming rapidly. Of course, this is assuming like a given time horizon and the like a given time horizon and the like a given time horizon and the experience horizon. And then, among all of these potential And then, among all of these potential futures that uh good continual learning futures that uh good continual learning futures that uh good continual learning will bring us, will bring us, will bring us, I think this is prob- probably the the I think this is prob- probably the the I think this is prob- probably the the one I like the most or I think it is the one I like the most or I think it is the one I like the most or I think it is the most interesting, which I call the most interesting, which I call the most interesting, which I call the unbounded expertise from bounded unbounded expertise from bounded unbounded expertise from bounded intelligence. intelligence. intelligence. All right, what if All right, what if All right, what if we can uh come up with a continual we can uh come up with a continual we can uh come up with a continual learning algorithm such that um learning algorithm such that um learning algorithm such that um given up Once the raw intelligence has given up Once the raw intelligence has given up Once the raw intelligence has across a certain threshold, across a certain threshold, across a certain threshold, we don't need a stronger intelligence we don't need a stronger intelligence we don't need a stronger intelligence anymore.

anymore. anymore. Like, continual learning will bring us Like, continual learning will bring us Like, continual learning will bring us like unbounded expertise once we have like unbounded expertise once we have like unbounded expertise once we have like a reasonable level of intelligence. like a reasonable level of intelligence. like a reasonable level of intelligence. All right, then we can call this the All right, then we can call this the All right, then we can call this the escape intelligence. escape intelligence. escape intelligence. And if this is indeed true, then it will And if this is indeed true, then it will And if this is indeed true, then it will have a lot of impli- uh implications for have a lot of impli- uh implications for have a lot of impli- uh implications for the whole ecosystem. All right, do we the whole ecosystem. All right, do we the whole ecosystem. All right, do we need to continually training to train need to continually training to train need to continually training to train these larger and larger models? Or like, these larger and larger models? Or like, these larger and larger models? Or like, these models like uh meet us uh maybe these models like uh meet us uh maybe these models like uh meet us uh maybe they're already good enough. What we're they're already good enough. What we're they're already good enough. What we're missing is just like better continual missing is just like better continual missing is just like better continual learning algorithms. So, the two So, the two To be a little bit more concrete, I To be a little bit more concrete, I To be a little bit more concrete, I think uh to provide more food for think uh to provide more food for think uh to provide more food for thought, uh here are some open questions thought, uh here are some open questions thought, uh here are some open questions I think uh in this space, I think uh in this space, I think uh in this space, the overarching question is like given the overarching question is like given the overarching question is like given any domain or environment, right? How any domain or environment, right? How any domain or environment, right? How can an agent continue to learn to can an agent continue to learn to can an agent continue to learn to specialize and reach expert-level specialize and reach expert-level specialize and reach expert-level competency? competency? competency? But to do that, you need to answer many But to do that, you need to answer many But to do that, you need to answer many other questions, right? How do you even other questions, right? How do you even other questions, right? How do you even measure, uh define and measure measure, uh define and measure measure, uh define and measure expertise? And this is probably uh expertise? And this is probably uh expertise? And this is probably uh environment-specific.

environment-specific. environment-specific. And how to handle the trade-off between And how to handle the trade-off between And how to handle the trade-off between reliability and plasticity, right? Um we reliability and plasticity, right? Um we reliability and plasticity, right? Um we want these agents to be both reliable want these agents to be both reliable want these agents to be both reliable and plastic. and plastic. and plastic. But they are inherently conflicting with But they are inherently conflicting with But they are inherently conflicting with each other, right? Reliable systems or each other, right? Reliable systems or each other, right? Reliable systems or stable systems, they resist the change. stable systems, they resist the change. stable systems, they resist the change. But the plastic systems likes change. But the plastic systems likes change. But the plastic systems likes change. So, how do we reconcile that? Um So, how do we reconcile that? Um So, how do we reconcile that? Um But fortunately, we do have a living But fortunately, we do have a living But fortunately, we do have a living existence proof, which is us ourselves, existence proof, which is us ourselves, existence proof, which is us ourselves, humans, uh that we are incredibly uh humans, uh that we are incredibly uh humans, uh that we are incredibly uh plastic, but also manage to be plastic, but also manage to be plastic, but also manage to be dependable most of the time. Um dependable most of the time. Um dependable most of the time. Um then uh then uh then uh from a technical perspective, like when from a technical perspective, like when from a technical perspective, like when we talk about learning, largely they are we talk about learning, largely they are we talk about learning, largely they are like this uh two forms of learning, like this uh two forms of learning, like this uh two forms of learning, parametric or non-parametric. So, how parametric or non-parametric. So, how parametric or non-parametric. So, how And my uh belief here is that both are And my uh belief here is that both are And my uh belief here is that both are really needed for uh this type of really needed for uh this type of really needed for uh this type of continual learning to to actually work.

continual learning to to actually work. continual learning to to actually work. But how do we synergize the two? But how do we synergize the two? But how do we synergize the two? And finally, And finally, And finally, even though we are focusing on even though we are focusing on even though we are focusing on specialization, I think there is a great specialization, I think there is a great specialization, I think there is a great potential for specialization to actually potential for specialization to actually potential for specialization to actually generate to lead to like better generate to lead to like better generate to lead to like better generalization. You know, we all we have generalization. You know, we all we have generalization. You know, we all we have exhausted the public data for training exhausted the public data for training exhausted the public data for training LLMs, but the next stage of training, LLMs, but the next stage of training, LLMs, but the next stage of training, the next internet-scale data opportunity the next internet-scale data opportunity the next internet-scale data opportunity is actually in all of these different uh is actually in all of these different uh is actually in all of these different uh like private worlds. If we can like private worlds. If we can like private worlds. If we can make these specialized agent work, they make these specialized agent work, they make these specialized agent work, they can learn in situ and can learn in situ and can learn in situ and channel back the learning to the general channel back the learning to the general channel back the learning to the general model. Then that may be the next model. Then that may be the next model. Then that may be the next internet scale data opportunity. internet scale data opportunity. internet scale data opportunity. Okay, so finally a call to action. Okay, so finally a call to action. Okay, so finally a call to action. I think let's start scaling expertise. I think let's start scaling expertise. I think let's start scaling expertise. This will be a new dimension for us to This will be a new dimension for us to This will be a new dimension for us to scale because intelligence is already scale because intelligence is already scale because intelligence is already becoming abundance. The frontier models becoming abundance. The frontier models becoming abundance. The frontier models they are probably smarter than average they are probably smarter than average they are probably smarter than average humans.

humans. humans. But expertise is still scarce. But expertise is still scarce. But expertise is still scarce. And we want to build a world where And we want to build a world where And we want to build a world where expertise becomes abundance where expertise becomes abundance where expertise becomes abundance where everyone can get expert support because everyone can get expert support because everyone can get expert support because in ideal world everyone can in ideal world everyone can in ideal world everyone can can have their personal health care, can have their personal health care, can have their personal health care, personal financial advisor, and personal personal financial advisor, and personal personal financial advisor, and personal tutors, and so on so forth. tutors, and so on so forth. tutors, and so on so forth. And then every company can build their And then every company can build their And then every company can build their their own learning loop. I think their own learning loop. I think their own learning loop. I think as Satya said as Satya said as Satya said two weeks ago like we want to enable two weeks ago like we want to enable two weeks ago like we want to enable this human AI learning loop at each this human AI learning loop at each this human AI learning loop at each company that turns into institutional company that turns into institutional company that turns into institutional memory and the full every company to memory and the full every company to memory and the full every company to build their build their build their own modes and to own modes and to own modes and to to still be in charge of their means of to still be in charge of their means of to still be in charge of their means of production. production. production. And finally, And finally, And finally, I think with abundance of expertise we I think with abundance of expertise we I think with abundance of expertise we will actually see more types of work will actually see more types of work will actually see more types of work become possible because they are become possible because they are become possible because they are right now they are still a lot of right now they are still a lot of right now they are still a lot of opportunities that that are locked up opportunities that that are locked up opportunities that that are locked up because the friction is just so high to because the friction is just so high to because the friction is just so high to make them make them make them economically viable. But with abundance economically viable. But with abundance economically viable. But with abundance of expertise, I think that we will be of expertise, I think that we will be of expertise, I think that we will be able to lower the friction and make many able to lower the friction and make many able to lower the friction and make many of the new type of work across the of the new type of work across the of the new type of work across the threshold of worth doing.

threshold of worth doing. threshold of worth doing. So this is the future we're building uh So this is the future we're building uh So this is the future we're building uh towards uh Neocognition and happy uh to towards uh Neocognition and happy uh to towards uh Neocognition and happy uh to share this with you and uh uh thanks for share this with you and uh uh thanks for share this with you and uh uh thanks for the attention. the attention. the attention. >> [applause]