WEBVTT

1
00:00:04.800 --> 00:00:07.560
With GenAI, we understand that those models

2
00:00:07.560 --> 00:00:10.200
will have a strong
command of language,

3
00:00:10.200 --> 00:00:12.680
but is that all they can do? Language?

4
00:00:12.680 --> 00:00:17.680
So far, the greatest success
we've had is with language models.

5
00:00:18.080 --> 00:00:21.160
But if we ever want to get to general

6
00:00:21.160 --> 00:00:24.920
artificial intelligence,
that might not be enough.

7
00:00:24.920 --> 00:00:28.000
In fact, we humans
learn a lot from the environment,

8
00:00:28.000 --> 00:00:30.400
we learn a lot from what we see,

9
00:00:30.400 --> 00:00:33.440
we learn a lot from interacting
with the environment.

10
00:00:33.440 --> 00:00:36.800
And there is a big research,
open question:

11
00:00:36.800 --> 00:00:41.800
Is language enough to build
our internal model of the world?

12
00:00:43.720 --> 00:00:48.720
The model that allows us to predict how
the world will respond to our actions,

13
00:00:49.960 --> 00:00:53.840
because that is the essence
for the intelligent system,

14
00:00:53.840 --> 00:00:57.920
being able to predict what's going
to happen when you act on the world.

15
00:00:57.920 --> 00:01:02.760
So there is a lot of research happening
in this field, and we're seeing a lot

16
00:01:02.760 --> 00:01:07.480
of multi-modal models appearing,
those that connect text and images,

17
00:01:07.480 --> 00:01:11.760
text and audio, text and video
and allowing it to go from text

18
00:01:11.760 --> 00:01:16.760
to generating images or taking images
and describing them with language.

19
00:01:17.840 --> 00:01:22.840
What will be the impact for business or
consumers of those multi-model models?

20
00:01:24.320 --> 00:01:26.600
I think that impact will be huge.

21
00:01:26.600 --> 00:01:31.600
We already have seen how quickly
applications like Midjourney spread

22
00:01:32.800 --> 00:01:36.760
around and how people
enjoying creating images.

23
00:01:36.760 --> 00:01:40.800
It really gives capability for those
who cannot draw to create images,

24
00:01:40.800 --> 00:01:44.920
for those who cannot design,
and so on and so forth.

25
00:01:44.920 --> 00:01:48.080
But if you think about businesses,
about creative industries,

26
00:01:48.080 --> 00:01:51.800
those technologies are clearly
transformative for those industries.

27
00:01:51.800 --> 00:01:56.800
And it will be very soon when we get
algorithms that allow us to generate

28
00:01:57.240 --> 00:02:02.240
videos, to generate audio, and that will
change how creative industry works.

29
00:02:02.600 --> 00:02:05.720
Leonid, we understand now, multi-modal models.

30
00:02:05.720 --> 00:02:08.560
Now, I've heard you
say sometimes that

31
00:02:08.560 --> 00:02:12.120
this is a first step
towards embodiment of agents.

32
00:02:12.120 --> 00:02:13.320
Can you tell us more?

33
00:02:13.320 --> 00:02:16.640
So embodiment is actually a very,
very interesting concept.

34
00:02:16.640 --> 00:02:21.640
It's really about a model having a 'body'
or acting in the physical world.

35
00:02:22.680 --> 00:02:24.120
And the way it is connected

36
00:02:24.120 --> 00:02:29.040
to multi-modality is quite simple
because in order for agent or model

37
00:02:29.040 --> 00:02:32.160
to operate, it needs
to sense the outside world.

38
00:02:32.160 --> 00:02:35.120
Sensing usually happens
through images,

39
00:02:35.120 --> 00:02:37.280
through video,
or maybe through audio.

40
00:02:37.280 --> 00:02:39.680
And then multi-modality allows us to build

41
00:02:39.680 --> 00:02:44.680
a single model that processes
all that information, and then allows

42
00:02:46.240 --> 00:02:50.440
the embodied agent,
to act on outside physical world.

43
00:02:50.440 --> 00:02:55.440
Will those embodied models change
the way we or machines learn things?

44
00:02:56.920 --> 00:03:00.000
We, as humans, we have our very individual

45
00:03:00.000 --> 00:03:02.240
experience and individual
learning pattern.

46
00:03:02.240 --> 00:03:05.480
In fact, if you, for example,
tell me how to ski,

47
00:03:05.480 --> 00:03:09.320
I will not be able to ski because I
don't have any experience in skiing.

48
00:03:09.320 --> 00:03:13.160
And just by listening to you,
I will not gain that experience.

49
00:03:13.160 --> 00:03:16.600
So we literally need to live through
our lives to gain experience.

50
00:03:16.600 --> 00:03:20.320
Now, if you think about agents,
they also need to have experience,

51
00:03:20.320 --> 00:03:22.800
experiencing the physical
world, getting feedback.

52
00:03:22.800 --> 00:03:25.520
But the difference is that
all the agents,

53
00:03:25.520 --> 00:03:28.760
multiple agents
can be connected in between themselves,

54
00:03:28.760 --> 00:03:33.000
and so they can instantly share this
experience and their learning becomes

55
00:03:33.000 --> 00:03:36.120
much, much quicker and much,
much more powerful.

56
00:03:36.120 --> 00:03:38.560
So imagine if humans could be able,

57
00:03:38.560 --> 00:03:42.320
instead of spending years in college,
just instantly share their knowledge,

58
00:03:42.320 --> 00:03:45.360
and that's what brings
the power to this

59
00:03:45.360 --> 00:03:47.520
collective intelligence
or collective learning.