Intelligence + Continual Learning = Expertise

Talk transcript

1. 0:00

   [upbeat music] All right. Uh, I understand that I'm standing between you and the lunch, so I'll try to be quick.

2. 0:20

   Uh, my name's Yu Su. I'm a professor at Ohio State, The Ohio State, and, uh, I also have another job, which is, uh, CEO at a company called NeoCognition, and we focus on agents and continual learning.

3. 0:34

   So today's talk, um, it won't be too technical, but I would- it will be mainly a, a conceptual one. But I think it's a very important, uh, conceptual distinction that I will try to make between what is intelligence and what is expertise.

4. 0:50

   And through this I will try to answer some of the, uh, very bothering questions for me that, um, like why we are so successful at the coding agent, but, uh, so terrible at anything else, right?

5. 1:04

   Why, uh, the current agents are so token inefficient, uh, like to the degree that, uh, every company right now is like coming out and try to curb their, uh, their token maxing efforts in the company.

6. 1:18

   Um, so hopefully this will provide some food for thought before lunch.

7. 1:25

   Right. First, a bit of a history. Um, so AI agents are not a new thing, right? It's, uh, we have been trying to develop agents throughout the whole history of AI.

8. 1:35

   Um, but the problem is that in the early stages, uh, let's say, um, in the ni- in the 1960s to, uh, '80s when we developed these expert systems or logical agents, or like in the, uh, 2010s when we developed these people RL-based, uh, neural agents, we were only able to capture some very limited facets of human intelligence,

9. 2:00

   right? Whether it's like logical reasoning or it's like, uh, perception in single mo-modalities to decision. Um, only recently with multimodal LLMs and the language agent built on top of them, for the first time we have a neural model that is able to encode multi-sensory inputs into a, uh, unified neural representation

10. 2:25

    that is also conducive to symbolic reasoning and communication, right? So that was a trait, uh, unique to humans. Now, uh, AI agents finally have the same thing. So that drastically, uh, improved their expressiveness, their reasoning ability, and adaptivity.

11. 2:43

    So that's why, uh, I think we have really entered a new evolutionary stage of machine intelligence.

12. 2:52

    And, uh, it didn't take long for these language agents to find their first mass markets, which is coding. And the best way to illustrate this is probably through the, uh, revenue graph of Anthropic's, right?

13. 3:06

    In just under two years, their revenue has grown four hundred times, uh, to, uh, forty billion. I think the newest number is maybe sixty billion, uh, analyzed runway. And it's largely driven by coding and coding-related productivity, uh, capabilities.

14. 3:24

    But if we think about it, right, coding is the really the ideal market for these language agents because code is already a language native world. Everything is already represented symbolically and like, uh, recorded, uh, in a very structured way.

15. 3:41

    And you get your rewards, you get your, uh, like tests all in place in symbolic ways.

16. 3:48

    So then what happens when we leave the privileged world of code? Well, not so well. Um, we are running into a lot of challenges deploying these agents in enterprise settings and the, uh, uh, also in personal settings, like the, uh, like open CL constantly make this, uh, like, uh,

17. 4:13

    quite brittle and silly, uh, errors. And then to the extent that Andrew Karpathy, uh, said that, uh, it's not gonna be the year of agents, it's gonna be the decade of agents because they, they cannot do computer use.

18. 4:26

    They don't have continual learning. Um, I don't know how much Andrew's, uh, thought has changed, uh, since like, uh, last time because of the coding agent e-everything. But I think the difficulties with computer use, with continual learning, still largely the same right now.

19. 4:42

    So how can something be so smart but also so brittle at the same time?

20. 4:49

    Here's my thesis around it. I think we are actually witnessing a modern version of the Moravec's paradox, right? So the para, uh, Moravec's paradox essentially says that, uh, uh, for AI, uh, hard things are easy, easy things are hard.

21. 5:06

    So the modern version here is that, uh, we are very good at these symbolic reasoning tasks like coding and math, which were considered crown jewel of, uh, of intelligence, uh, earlier.

22. 5:20

    But then we still struggle with this everyday digital work because they really require quite different set of cognitive competencies to excel as them.

23. 5:31

    And more specifically, I think modern society is really not just one unified world. It's millions of these micro worlds. Like every domain, every, every profession is different, every company is different.

24. 5:44

    Even if you're using the same software, every company, a company configure it differently. So it's extremely idiosyncratic, especially in the digital world. It has these unique local physics, like different structures, constraints, affordances, and dynamics that you have to learn.

25. 6:02

    It's just, like, too heterogeneous and dynamic for any, uh, monolithic model to try to compress it into one static representation. So agents must continually learn on the job to acquire what I call specialized expertise for each specific micro-world.

26. 6:24

    The second part of the talk, I will try to establish the differences between intelligence and expertise. Here are the working definitions. For intelligence, it's the capacity to reason through unfamiliar problems from available context, right?

27. 6:41

    This is what the frontier models are increasingly good at. Um, you give it the problem, statement, the context, the tools, and it can reason through this even if, uh, it's seeing them for the first time, uh, and they can do a great job.

28. 6:56

    Every episode is more or less, like, independent from each other 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 super real, uh, performance in a particular domain, right?

29. 7:20

    So this is in stark contrast with intelligence.

30. 7:25

    And to, uh, show what does ex- expertise actually contain, I think, uh, the, the key idea from cognitive science is that experts don't just know more facts. They actually see the world differently, right?

31. 7:42

    So, um, the, uh, expertise allows you to do different pattern recognition, so you see through the surface patterns. Like, if you're looking, an expert is looking at, like, a gigantic bug report, they can immediately locate, like, the most plausible places where things could go wrong.

32. 8:01

    Um, and they think about the problem with, like, a very deep structure, right? Y- when you are scheduling a meeting, you know that it's not just, like, finding the shared slots on everyone's calendar.

33. 8:15

    It's actually a constrained optimization problem over everyone's authority, the priorities, the urgency, and everything. Um, and we don't, expert don't just operate with a set of rules or set of facts.

34. 8:29

    We know that every single thing is conditional, right? Every rule has, like, the preconditions where it applies, but we also know when we can bend the reality, we can bend the rules when exceptions happen, right?

35. 8:44

    And finally, that also give us judgment and taste, is importantly, what's, like, high quality and, uh, very importantly, when to stop, when is good enough. Um, so all this together, I think experts effectively has- have built a world model of their environments, right?

36. 9:04

    That's, it's a generalized notion of world model that captures how that micro-world works, and that becomes the basis for all of our perception, uh, reasoning, decision-making, uh, and judgment.

37. 9:21

    So intelligence and expertise are really quite different across many dimensions, uh, but some of the impor- uh, interesting ones here are, like, intelligence is about, hey, when we have the context, uh, how to solve the problem through the context.

38. 9:36

    But expertise actually will bring you the c- the right context, right? Given any problem, we know what context bring into are important for this problem and bring it in to solve the problem.

39. 9:48

    And because of that, uh, intelligence tend to expand our search. Like, every problem-solving is a search problem. So intelligence tend to brute force it, try to, uh, try to spin up, like, one hundred different, uh, like, uh, parallel ways to, to try to solve the problem, while expertise will actually try to compress the search space because expertise

40. 10:12

    has constructed this, has learned these sh- essential shortcuts for the problem space, so that whenever you have a problem, you know the most plausible ways to solve it. Um,

41. 10:27

    and then I also think the final part here is that I think 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.

42. 10:42

    Um, and, and Jack just, uh, gave some definition earlier, uh, with, like, ten different names. Um, but here's my-- the definition I work with. I think continual learning is adaptive compression of experience into reusable structures for future behavior.

43. 11:02

    So all of these four elements here are very important. For experience, we need to, uh, answer the question, like, what kind of experience we're talking about? It's more like episodes of experience, or it's like, uh, these semantic facts or procedures or feedback from human or en- environments.

44. 11:20

    And how do we compress that? Uh, so in, we embed them into vectors, or we index them into some symbolic structure. Uh, we, uh, distill them into model parameters or do some, uh, kind of reinforcement learning.

45. 11:33

    And it's not just, like, one-time compression. It needs to be adaptive compression. Like, what you have learned, what you have compressed so far should, uh, largely, uh, influence how you compress further.

46. 11:48

    And what kind of structure we're looking at? It's just like parameters, like adapters of your, uh, language models, or it's vectors, graphs, skills, or even word models. And then how to use these reasonable structures.

47. 12:01

    It's like, uh, you use it just to recall these facts or use it for prediction of like future states. You use it for, uh, for better planning, for, or even for the control, like actuation layer of agents or as a value function for potential states, right?

48. 12:19

    So it's because of this, uh, the continual learning problem is so rich, like it has these four different aspects and if e- different aspects can be instantiated in different ways that makes this field so confusing.

49. 12:33

    But hopefully this is a definition that, uh, encompasses, uh, most of the, uh, versions of continual learning.

50. 12:42

    Then I think the, uh, this is maybe the most important figure in this talk. Um, if we put raw intelligence as the X-axis and, uh, expertise as the Y-axis, I think we'll find that they are largely orthogonal to each other.

51. 13:02

    If you don't have continual learning, uh, y- all you do is scaling your model to, to get better, like raw intelligence, then what we will get is what I call the world's smartest novice, right?

52. 13:14

    Super smart. It can try to, uh, try to attack at any problem, uh, provi- uh, given to it, but it doesn't accumulate expertise, so it end up by just like brute forcing its way at every problem.

53. 13:27

    Then if you have continual learning, uh, like different continual learning algorithms will essentially set the slope of your learning, uh, curve here, right? If you have a sloppy CL algorithm, maybe some kind of simple in-context learning, then, uh, with like, uh, increasing e- intelligence, then your expertise will increase like a little bit, but you have a really

54. 13:51

    strong, uh, continual learning algorithm then, uh, the expertise will, um, increase like, uh, rapidly. Of course, this is assuming like a given time horizon and experience horizon.

55. 14:06

    And then among all of these potential futures that, uh, good continual learning will bring us, I think this is prob- probably the, the one I like the most or I think it is the most interesting, which I call the unbounded expertise from bounded intelligence, right?

56. 14:24

    What if we can, uh, come up with a continual learning algorithm such that, um, given up-- once the raw intelligence has crossed a certain threshold, we don't need stronger intelligence anymore, right?

57. 14:38

    Continual learning will bring us like unbounded expertise once we have like a reasonable level in, of intelligence, right? Then we can call this the Escape intelligence, and if this is indeed true, then it will have a lot of implic- uh, implications for the whole ecosystem, right?

58. 14:57

    Do we need to continually training to train these larger and larger models? Or like these models like, uh, meet us, uh, maybe they're already good enough. What we're missing is just like better continual learning algorithms.

59. 15:14

    So the two be a little bit more concrete, I think, uh, to provide more food for thoughts, uh, here are some open questions, I think, uh, in this space.

60. 15:24

    The overarching question is like, given any domain or environment, right, how can an agent continually learn to specialize and reach expert-level competency? But to do that, you need to answer many other questions, right?

61. 15:39

    How do you even measure, uh, define and measure expertise? And this is probably, uh, environment specific. And how to handle the trade-off between reliability and plasticity, right? Um, we want these agents to be both reliable and plastic, but they are inherently conflicting with each other, right?

62. 16:02

    Reliable systems or s- stable systems, they resist change, but the plastic systems likes change. So how do we reconcile that? Um, but fortunately, we do have a living existence proof, which is us, ourselves, humans, uh, that we are incredibly, uh, plastic, but also manage to be dependable most of the time.

63. 16:25

    Um, then, uh, from a technical perspective, like when we talk about learning, largely, uh, there are like this, uh, two forms of learning, parametric or nonparametric. So how-- and, uh, my, uh, belief here is that both are really needed for, uh, these type of continual learning to, to actually work, but how do we synergize the two?

64. 16:49

    And finally, even though we are focusing on specialization, I think there is a great potential for specialization to actually gener- to lead to like better generalization. You know we are, we have exhausted the public data for training LLMs, but the next stage of training, the next internet skill data opportunity is actually in all of these different, uh,

65. 17:14

    like private worlds. If we can make this specialized agent work, they can learn in situ and, uh, channel back the learning to the general model, then that may be the next, uh, internet skill, uh, data opportunity.

66. 17:30

    Okay, so finally, a call to action. Um, I think let's start scaling expertise. This will be a new dimension for us to scale because intelligence is already becoming abundance.

67. 17:43

    The frontier models, they are probably smarter than average humans, um, but expertise is still scarce.

68. 17:53

    And we want to build a world where expertise becomes abundance, where everyone can get expert, uh, support because in an ideal world, everyone can, uh, can have their personal healthcare, personal financial advisor, and, uh, personal tutors, and so on and so forth.

69. 18:13

    And then every company can build their, uh, their own learning loop, right? They can, as Satya said, uh, uh, two weeks ago, like we want to enable this human-AI learning loop, uh, at each company that turns into institutional memory and for every company to build their, uh, own moats and to, uh, to still be in charge of

70. 18:35

    their means of production. And finally, um, I think with abundance o- of expertise, uh, we will actually see more types of work become possible because there, uh, right now there are still a lot of opportunities that, uh, that are locked up because the friction is just so high to make them, uh, econ- economically

71. 18:59

    viable. But with abundance of expertise, I think that we will be able to lower the friction and, uh, make many of the new type of work across the threshold of worth doing.

72. 19:11

    So this is the future we're building, uh, towards at NeoCognition and happy, uh, to share this with you and, uh, uh, thanks for the attention. [audience applauding] [upbeat music]