WEBVTT

NOTE Sentence-level transcript of https://www.youtube.com/watch?v=I6aiEf3aEFQ

NOTE One cue per sentence. Cue ids are the line anchors on /transcripts/I6aiEf3aEFQ.html. A cue ends where the next begins, or 2 s after its last word.

s1
00:00:01.309 --> 00:00:03.309
[music]

s2
00:00:12.960 --> 00:00:13.759
All right.

s3
00:00:13.759 --> 00:00:20.160
Uh, I understand that I'm standing between you and the launch, so I'll try to be quick.

s4
00:00:20.160 --> 00:00:21.359
Uh, my name is IU.

s5
00:00:21.359 --> 00:00:34.000
I'm a professor at Ohio State, the Ohio State and uh I also have another job which is a CEO at a company called Neocognition and we focus on agents and container learning.

s6
00:00:34.000 --> 00:00:46.160
So today's talk um it won't be too technical but I will 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

s7
00:00:46.160 --> 00:01:03.600
between what is intelligence and what is expertise and through this I will try to answer some of the very bothering questions for me that um like why we are so successful at the coding agent but they're so terrible at anything else,

s8
00:01:03.600 --> 00:01:04.080
right?

s9
00:01:04.080 --> 00:01:18.320
Why uh the current agents are so token inefficient uh that to the degree that every company right now is like coming out and try to curb their uh their token maxing efforts in the company.

s10
00:01:18.320 --> 00:01:25.280
Um so hopefully this will provide some food for thought before launch.

s11
00:01:25.439 --> 00:01:25.840
Right.

s12
00:01:25.840 --> 00:01:27.360
First a bit of a history.

s13
00:01:27.360 --> 00:01:30.479
Um so AI agents are not a new thing, right?

s14
00:01:30.479 --> 00:01:35.520
it's uh we have been trying to develop agents throughout the whole history of AI.

s15
00:01:35.520 --> 00:01:48.560
Um but the problem is that in the early stages uh let's say um in the n in the 1960s to uh 80s when we developed these expert systems or logical agents

s16
00:01:48.560 --> 00:02:02.479
or like in the uh 20110s when we developed these deep RL based uh neural agents we were only able to capture some very limited facets of human intelligence right whether it's like logical reasoning

s17
00:02:02.479 --> 00:02:07.119
or it's like uh perception in single modalities to decision.

s18
00:02:07.119 --> 00:02:19.280
Um only recently with the mult model LRMS and the language agent built on top of them for the first time we have a neural model that is able to

s19
00:02:19.280 --> 00:02:30.000
encode multiensory inputs into a uh unified neural representation that is also conducive to symbolic reasoning and communication.

s20
00:02:30.000 --> 00:02:30.239
Right?

s21
00:02:30.239 --> 00:02:33.200
So that was a trait unique to humans.

s22
00:02:33.200 --> 00:02:36.879
Now uh AI agents finally have the same thing.

s23
00:02:36.879 --> 00:02:43.200
So that drastically improved their expressiveness, their reasoning ability and adaptivity.

s24
00:02:43.200 --> 00:02:50.879
So that's why I think we have really entered a new evolutionary stage of machine intelligence.

s25
00:02:52.080 --> 00:03:00.160
And uh it didn't take long for these language agents to find their first mass markets which is coding.

s26
00:03:00.160 --> 00:03:11.440
And the best way to illustrate this is probably through the uh revenue graph of anthropics right in just under two years their revenue has grown 400 times uh to

s27
00:03:11.440 --> 00:03:27.680
uh 40 billion I think the newest number is maybe 60 billion uh annualizer runway and it's largely driven by coding and coding related productivity uh capabilities but if we think about it right coding is the

s28
00:03:27.680 --> 00:03:33.840
really the ideal market for this language anguage agents because code is already a language native world.

s29
00:03:33.840 --> 00:03:48.959
Everything is already represented symbolically and like uh recorded uh in a very structured way and you get your rewards, you get your uh like tests all in place in symbolic ways.

s30
00:03:48.959 --> 00:03:54.799
So then what happens when we leave the privileged world of code?

s31
00:03:54.799 --> 00:03:56.640
Well, not so well.

s32
00:03:56.640 --> 00:04:17.040
Um we are running into a lot of challenges deploying these agents in enterprise settings and the uh also in personal settings like the uh like open clause uh constantly make this uh like uh quite brittle and silly uh errors

s33
00:04:17.040 --> 00:04:28.720
and then to the extent that Andrew Capathy uh said that it's not going to be the year of agents it's going to be the decade of agents because they cannot do computer use they don't have continuous learning

s34
00:04:28.720 --> 00:04:42.639
um I don't know how much uh thought has changed uh since like uh last time because of the coding agent everything but I think the difficulties with computer use with container learning still largely the same right now

s35
00:04:42.639 --> 00:04:57.520
so how can something be so small but also so brittle at the same time here's my thesis around it I think we are actually witnessing a modern version of the Maravx paradox,

s36
00:04:57.520 --> 00:04:57.840
right?

s37
00:04:57.840 --> 00:05:06.720
So the par morave paradox essentially says that for AI uh hard things are easy, easy things are hard.

s38
00:05:06.720 --> 00:05:20.240
So the modern version here is that we are very good at this symbolic reasoning task like coding and math which were considered crown jewel of uh of intelligence uh earlier.

s39
00:05:20.240 --> 00:05:40.160
But then we still struggle with this everyday digital work because they really require quite different set of cognitive competencies to excel as them and more specifically I think modern society is really not just one unified world it's millions of these micro worlds like every

s40
00:05:40.160 --> 00:05:53.919
domain every every profession is different every company is different even if you're using the same software every company company configure it differently so is extremely idiosyncratic especially in the digital world.

s41
00:05:53.919 --> 00:06:02.160
It has its unique local physics like different structures, constraints, affordances and dynamics that you have to learn.

s42
00:06:02.160 --> 00:06:12.160
It's just like too hoggenous and dynamic for any uh monolithic model to try to compress it into one static representation.

s43
00:06:12.160 --> 00:06:23.199
So agents must continually learn on the job to acquire what I call specialized expertise for each specific micro world.

s44
00:06:24.560 --> 00:06:31.360
The second part of the talk I will try to establish the differences between intelligence and expertise.

s45
00:06:31.360 --> 00:06:33.840
Here are the working definitions.

s46
00:06:33.840 --> 00:06:41.520
For intelligence, it's the capacity to reason through unfamiliar problems from available context.

s47
00:06:41.520 --> 00:06:41.840
Right?

s48
00:06:41.840 --> 00:06:45.199
This is what the frontier models are increasingly good at.

s49
00:06:45.199 --> 00:06:53.680
Um, you give it the provin statements, the context, the tools, and they can reason through this even if it's seeing them for the first time.

s50
00:06:53.680 --> 00:06:56.800
Uh, and they can do a great job.

s51
00:06:56.800 --> 00:07:01.440
Every episode is more or less like independent from each other here.

s52
00:07:01.440 --> 00:07:03.199
But expertise is different.

s53
00:07:03.199 --> 00:07:08.160
Expertise is really accumulated and situated competence.

s54
00:07:08.160 --> 00:07:19.599
It's the ability to act reliably, efficiently and with judgment to achieve reproducibly super real uh performance in a particular domain.

s55
00:07:19.599 --> 00:07:20.000
Right?

s56
00:07:20.000 --> 00:07:24.319
So this is in stark contrast with intelligence.

s57
00:07:25.360 --> 00:07:31.360
And to uh show what does expertise actually contain.

s58
00:07:31.360 --> 00:07:42.880
I think uh the key idea from cognitive science is that experts don't just know more facts, we actually see the world differently, right?

s59
00:07:42.880 --> 00:07:49.120
So um the uh expertise allows you to do different pattern recognition.

s60
00:07:49.120 --> 00:07:59.840
So you see through the surface patterns like if you're looking an expert is looking at like a gigantic bug reports they can immediately locate like the most plausible places

s61
00:07:59.840 --> 00:08:01.599
where things could go wrong.

s62
00:08:01.599 --> 00:08:07.039
Um and you think about the problem with like a very deep structure, right?

s63
00:08:07.039 --> 00:08:20.560
When you are scheduling a meeting, you know that it's not just like finding the shared slots on everyone's calendar is actually a constraint optimization problem over everyone's authority, the priorities,

s64
00:08:20.560 --> 00:08:22.560
the urgency and everything.

s65
00:08:22.560 --> 00:08:29.039
Um and we don't expert don't just operate with a set of rules, a set of facts.

s66
00:08:29.039 --> 00:08:32.640
We know that every single thing is conditional, right?

s67
00:08:32.640 --> 00:08:36.240
Every rule has like the preconditions where it applies.

s68
00:08:36.240 --> 00:08:44.000
But we also know when we can bend the reality, we can bend the rules when exceptions happen, right?

s69
00:08:44.000 --> 00:08:53.600
And finally that also give us judgment and taste is importantly what's high quality and uh very importantly when to stop when is good enough.

s70
00:08:53.600 --> 00:09:15.120
Um so all this together I think experts effectively has have built a world model of their environments right that 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

s71
00:09:15.120 --> 00:09:19.760
uh reasoning decision- making uh and judgment.

s72
00:09:21.519 --> 00:09:25.920
So intelligence and expertise are really quite different across many dimensions.

s73
00:09:25.920 --> 00:09:36.560
Uh but some of the import uh interesting ones here are like intelligence is about hey when we have the context uh how to solve the problems through the context

s74
00:09:36.560 --> 00:09:48.880
but expertise actually will bring you 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

s75
00:09:48.880 --> 00:09:56.560
and because of that uh intelligence tend to expand our search like every problem solving is a search problem.

s76
00:09:56.560 --> 00:10:06.720
So, intelligence tend to brute force it, try to uh try to spin up like 100 different uh like parallel ways to to try to solve the problem.

s77
00:10:06.720 --> 00:10:22.640
Well, expertise will actually try to compress the search space because expertise has constructed this has learned this essential shortcuts for the problem space so that whenever you have a problem, you know the most plausible

s78
00:10:22.640 --> 00:10:24.160
ways to solve it.

s79
00:10:24.160 --> 00:10:38.000
Um and then I also think the final part here is that I think continuous learning is the important bridge from intelligence to expertise.

s80
00:10:38.000 --> 00:10:48.560
But first let me try to define continuous learning because it's such a confusing term um and and Jack just uh gives some definition earlier uh with like 10 different names.

s81
00:10:48.560 --> 00:10:53.279
Um but here's my the definition I work with.

s82
00:10:53.279 --> 00:11:02.079
I think continue learning is adaptive compression of the experience into reusable structures for future behavior.

s83
00:11:02.079 --> 00:11:06.000
So all of these four elements here are very important.

s84
00:11:06.000 --> 00:11:20.480
For experience, we need to uh answer the question like what kind of experience we're talking about is more like episodes of experience or it's like these semantic facts or procedures or feedback from human or environments.

s85
00:11:20.480 --> 00:11:22.480
And how do we compress that?

s86
00:11:22.480 --> 00:11:28.160
Uh so in we embed them into vectors or we index them into some symbolic structure.

s87
00:11:28.160 --> 00:11:33.839
Uh we uh distill them into model parameters or do some kind of reinforcement learning.

s88
00:11:33.839 --> 00:11:37.279
And it's not just like one time compression.

s89
00:11:37.279 --> 00:11:47.680
It needs to be adaptive compression like what you have learned what you have compressed so far should largely uh influence how you compress further.

s90
00:11:48.320 --> 00:11:50.399
And what kind of structure we're looking at?

s91
00:11:50.399 --> 00:11:58.800
It's just like parameters like adapters of your language models or it's vectors, graphs, skills or even word models.

s92
00:11:58.800 --> 00:12:01.600
And then how do you use these reusable structures?

s93
00:12:01.600 --> 00:12:08.320
It's like you use it just to recall these facts or use it for prediction of like future states.

s94
00:12:08.320 --> 00:12:21.200
is use for uh for better planning for or even for the control like actuation layer of the agent or as a value function for potential states right so it's because of this

s95
00:12:21.200 --> 00:12:33.279
uh the continent learning problem is so rich like it has this four different aspects and if different aspects can be instantiated in different ways that makes this field so confusing

s96
00:12:33.279 --> 00:12:40.959
but hopefully this is a definition that uh encompasses uh most of the versions of continuous learning.

s97
00:12:42.000 --> 00:12:47.839
Then I think that uh this is maybe the most important figure in this talk.

s98
00:12:47.839 --> 00:13:02.079
Um if we put raw intelligence as the x-axis and uh expertise as the yaxis I think we'll find that they are largely orthogonal to each other.

s99
00:13:02.079 --> 00:13:13.760
If you don't have continuous learning uh 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

s100
00:13:13.760 --> 00:13:27.040
right super smart it can try to uh try to attack at any problem uh provide given to it but it doesn't accumulate expertise so it end up just like brute forcing its way at every problem

s101
00:13:27.040 --> 00:13:37.839
then if you have continued learning uh like Different continuous learning algorithms will essentially set the slope of your learning uh curve here, right?

s102
00:13:37.839 --> 00:13:52.160
If you have a sloppy CL algorithm, maybe some kind of simple incontext learning then uh with like uh increasing intelligence then your expertise will increase like a little bit but you have a really strong

s103
00:13:52.160 --> 00:13:57.600
uh contin um increase like rapidly.

s104
00:13:57.600 --> 00:14:04.399
Of course, this is assuming like a given time horizon and experience horizon.

s105
00:14:06.639 --> 00:14:13.040
And then among all of these potential futures that good continued learning will bring us.

s106
00:14:13.040 --> 00:14:23.920
I think this is pro 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.

s107
00:14:23.920 --> 00:14:24.240
Right?

s108
00:14:24.240 --> 00:14:38.560
What if we can uh come up with a computer learning algorithm such that um given up once the raw intelligence has crossed a certain threshold we don't need stronger intelligence anymore

s109
00:14:38.560 --> 00:14:57.279
like learning will bring us like unbounded expertise once we have like a reasonable level of intelligence right then we can call this the escape intelligence and if this is indeed true then it will have a a lot of imp implications for the whole ecosystem,

s110
00:14:57.279 --> 00:14:57.680
right?

s111
00:14:57.680 --> 00:15:07.440
Do we need to continually train to train these larger and larger models or like this models like uh methods uh maybe they're already good enough?

s112
00:15:07.440 --> 00:15:11.920
What we're missing is just like better continuous learning algorithms.

s113
00:15:14.000 --> 00:15:25.120
So to 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 the overarching

s114
00:15:25.120 --> 00:15:40.959
question is like given any domain or environments 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 like how do you even measure

s115
00:15:40.959 --> 00:15:52.560
uh define and measure expertise and this is probably environment specific and how to handle the trade-off between reliability and plasticity, right?

s116
00:15:52.560 --> 00:16:02.240
Um we want these agents to be both reliable and plastic, but they are inherently conflicting with each other, right?

s117
00:16:02.240 --> 00:16:09.680
Reliable systems or stable systems, they resist the change, but the plastic systems likes change.

s118
00:16:09.680 --> 00:16:11.839
So how do we reconcile that?

s119
00:16:11.839 --> 00:16:25.360
Um but fortunately we do have a living existence proof which is us ourselves humans uh that we are incredibly plastic but also manage to be dependable most of the time.

s120
00:16:25.360 --> 00:16:37.199
Um then uh from a technical perspective like when we talk about learning largely there are like two forms of learning parametric or nonparametric.

s121
00:16:37.199 --> 00:16:46.480
So how and my uh belief here is that both are really needed for uh this type of machine learning to to actually work.

s122
00:16:46.480 --> 00:17:01.759
But how do we synergize the two 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.

s123
00:17:01.759 --> 00:17:15.679
You know we are we have exhausted the public data for training LLMs but the next stage of training the next internet scale data opportunity is actually in all of these different uh like private worlds.

s124
00:17:15.679 --> 00:17:29.760
If we can make this specialized agent work they can learn in sichu and uh channel back the learning to the general model I think that may be the next uh internet scale uh data opportunity.

s125
00:17:30.080 --> 00:17:30.640
Okay.

s126
00:17:30.640 --> 00:17:48.480
So finally a call to action um think let's start scaling expertise this will be a new dimension for us to scale because intelligence is already becoming abundance the frontier models they are probably smarter than average humans

s127
00:17:48.480 --> 00:18:04.880
um but expertise is still scarce and we want to build a world where expertise becomes abundant where everyone can gets expert uh support because in an ideal world everyone can

s128
00:18:04.880 --> 00:18:18.000
can have their personal health care, personal financial advisor and uh personal tutors and so on so forth and then every company can build their their own learning loop, right?

s129
00:18:18.000 --> 00:18:30.559
They can, as Satia said 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

s130
00:18:30.559 --> 00:18:38.480
uh own modes and to uh to still be in charge of their means of production.

s131
00:18:38.720 --> 00:18:53.039
And finally um I think with abundance 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

s132
00:18:53.039 --> 00:19:00.799
that uh that are locked up because the friction is just so high to make them uh econ economically viable.

s133
00:19:00.799 --> 00:19:11.520
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.

s134
00:19:11.520 --> 00:19:21.039
So this is the future we're building uh towards new cognition and happy uh to share this with you and uh thanks for the attention.

s135
00:19:21.328 --> 00:19:23.328
[applause]