1
00:00:00,000 --> 00:00:02,837
at unprecedented scale.

2
00:00:02,920 --> 00:00:04,917
Next speaker is, ladies and gentlemen,

3
00:00:05,000 --> 00:00:08,197
Mr. Alexander Wong, Chief AI Officer at

4
00:00:08,280 --> 00:00:11,317
Meta, the founder of Scale AI. Alexander

5
00:00:11,400 --> 00:00:13,517
Wong built the data infrastructure that

6
00:00:13,600 --> 00:00:16,517
powers much of the modern AI industry

7
00:00:16,600 --> 00:00:20,197
before joining Meta as Chief AI Officer.

8
00:00:20,280 --> 00:00:26,200
So, with a round of applause,
please welcome Mr. Alexander Wong.

9
00:00:31,640 --> 00:00:33,117
Thank you so much for having me.

10
00:00:33,200 --> 00:00:38,037
Namaste. Uh it's fair
to say my upbringing wasn't

11
00:00:38,120 --> 00:00:41,197
typical. My parents
were physicists in a town

12
00:00:41,280 --> 00:00:43,677
called Los Alamos in New Mexico.

13
00:00:43,760 --> 00:00:48,397
Los Alamos is a government lab town
where for decades scientists have come

14
00:00:48,480 --> 00:00:51,557
to push the boundaries of what's
possible in mathematics, in

15
00:00:51,640 --> 00:00:56,677
supercomputing, in human genome studies,
vaccine research, space explorations,

16
00:00:56,760 --> 00:01:00,237
and material science.
My mother studies how plasma behaves

17
00:01:00,320 --> 00:01:03,957
inside stars. At the dinner
dinner table, we'd talk

18
00:01:04,040 --> 00:01:08,277
about physics problems, scientific
tradeoffs, the reasoning behind how

19
00:01:08,360 --> 00:01:11,717
systems work.
One kid in my town made huge balls of

20
00:01:11,800 --> 00:01:15,877
plasma in their garage for a science
fair project. You know, normal high

21
00:01:15,960 --> 00:01:19,397
school stuff. Growing up
in a place like Los Alamos

22
00:01:19,480 --> 00:01:22,597
leaves two things deeply ingrained in you.

23
00:01:22,680 --> 00:01:25,357
A belief that anything is possible

24
00:01:25,440 --> 00:01:28,517
and that science should serve society.

25
00:01:28,600 --> 00:01:31,077
Those ideas are what led me to study AI

26
00:01:31,160 --> 00:01:33,997
while I was at I was at college at MIT.

27
00:01:34,080 --> 00:01:36,997
They led me to start
my own company, Scale AI.

28
00:01:37,080 --> 00:01:42,597
And last year, they led me to Meta
where I'm now the Chief AI Officer.

29
00:01:42,680 --> 00:01:45,077
If you believe that anything is possible,

30
00:01:45,160 --> 00:01:47,077
Meta is one of the few companies with

31
00:01:47,160 --> 00:01:52,157
the resources, talent, and ambition
to push the science of AI forward at

32
00:01:52,240 --> 00:01:55,237
scale. If you want to make technology that

33
00:01:55,320 --> 00:01:59,837
serves society, Meta has an incredible
opportunity to get this technology in

34
00:01:59,920 --> 00:02:04,357
into people's lives.
Three and a half billion people use at

35
00:02:04,440 --> 00:02:06,797
least one of our apps every day.

36
00:02:06,880 --> 00:02:09,437
That blows my mind. It's more than half

37
00:02:09,520 --> 00:02:12,197
a billion people in India alone.

38
00:02:12,280 --> 00:02:16,117
People are already using
our AI to do amazing things.

39
00:02:16,200 --> 00:02:18,797
Across India, creators use our AI to

40
00:02:18,880 --> 00:02:23,197
automatically translate Reels into
the language of the person watching.

41
00:02:23,280 --> 00:02:26,917
Small businesses talk to customers
through WhatsApp business agents that

42
00:02:27,000 --> 00:02:31,397
they create in 10 minutes on their phones,
and they use our GenAI tools to

43
00:02:31,480 --> 00:02:35,917
create ads and reach customers way more
efficiently than they ever could before.

44
00:02:36,000 --> 00:02:40,357
And India has world-class developers
building genius things to solve societal

45
00:02:40,440 --> 00:02:44,637
challenges. For example,
there are more than 20

46
00:02:44,720 --> 00:02:48,597
million people with disabilities
in India who are locked out of education,

47
00:02:48,680 --> 00:02:50,597
jobs, and digital services because the

48
00:02:50,680 --> 00:02:53,077
digital world wasn't designed for them.

49
00:02:53,160 --> 00:02:56,357
So, I-Stem built voice-first, AI-powered

50
00:02:56,440 --> 00:03:01,477
infrastructure that helps people with
disabilities to learn, discover careers,

51
00:03:01,560 --> 00:03:05,397
and complete digital tasks independently,
like converting textbooks

52
00:03:05,480 --> 00:03:08,957
into usable formats or giving
personalized career guidance that takes

53
00:03:09,040 --> 00:03:11,117
into account their disability.

54
00:03:11,200 --> 00:03:15,157
In healthcare, researchers
at Ashoka University used

55
00:03:15,240 --> 00:03:18,597
our SAM 3 model, which
is trained on billions of natural

56
00:03:18,680 --> 00:03:23,357
images to speed up the identification
and segmentation of cancer tumors and

57
00:03:23,440 --> 00:03:27,597
at-risk organs.
Their model, OncoSeg, can help

58
00:03:27,680 --> 00:03:32,077
radiologists and radiology oncology
teams do in seconds what it takes hours

59
00:03:32,160 --> 00:03:36,877
to do manually. The beauty
of general-purpose models is

60
00:03:36,960 --> 00:03:41,597
the same technology that can segment
tumors in the brain can segment leaves

61
00:03:41,680 --> 00:03:42,877
to help farmers assess the health of

62
00:03:42,960 --> 00:03:46,557
their crops, as AgriPoint has done.

63
00:03:46,640 --> 00:03:48,757
We recently open-sourced our omnilingual

64
00:03:48,840 --> 00:03:53,637
models, which recognize speech across
more than 1,600 languages and can

65
00:03:53,720 --> 00:03:58,157
rapidly adapt to new languages
with just a few audio samples.

66
00:03:58,240 --> 00:04:02,277
It's not a fantasy that in a few years
we'll have real-time voice-to-voice

67
00:04:02,360 --> 00:04:06,397
translation for every
spoken language on Earth.

68
00:04:06,480 --> 00:04:10,997
Now build that into your glasses,
real-time translation in any language

69
00:04:11,080 --> 00:04:14,677
just for you.
That's transformative. Perhaps most

70
00:04:14,760 --> 00:04:17,357
especially in countries like India where

71
00:04:17,440 --> 00:04:22,437
so many languages are spoken.
In fact, language is an area where we're

72
00:04:22,520 --> 00:04:25,077
collaborating with the
Indian government on.

73
00:04:25,160 --> 00:04:29,197
Through its AI Kosh platform, we're
providing data sets in 10 major Indian

74
00:04:29,280 --> 00:04:33,597
languages so people can build AI models
that deeply understand Indian languages

75
00:04:33,680 --> 00:04:37,637
and context. I'm sure you're
used to people from big

76
00:04:37,720 --> 00:04:42,157
tech world making lots of grand
but vague assertions about what AI will be

77
00:04:42,240 --> 00:04:45,477
able to do. But we don't
have to be vague. People

78
00:04:45,560 --> 00:04:48,077
use our AI right here, right now.

79
00:04:48,160 --> 00:04:51,597
They're getting value from it and they're
building amazing things with it.

80
00:04:51,680 --> 00:04:55,117
And that gives us confidence
about what we're building towards.

81
00:04:55,200 --> 00:04:58,757
We're releasing new models this year
with the first coming in the next couple

82
00:04:58,840 --> 00:05:01,477
of months. These will be
deeply integrated with our

83
00:05:01,560 --> 00:05:04,317
products in a way we're
really excited about.

84
00:05:04,400 --> 00:05:06,837
We're optimistic about
the trajectory we're on.

85
00:05:06,920 --> 00:05:10,277
The first models will be good and as the
year goes on, I think we're going to be

86
00:05:10,360 --> 00:05:14,157
pushing the frontier.
Our vision is personal super

87
00:05:14,240 --> 00:05:18,957
intelligence.
AI that knows you, your goals, your

88
00:05:19,040 --> 00:05:22,597
interests, and helps you with
whatever you're focused on doing.

89
00:05:22,680 --> 00:05:28,080
It serves you, whoever
you are, wherever you are.

90
00:05:29,200 --> 00:05:32,877
We all lead busy lives.
I'm sure you'd want to do more

91
00:05:32,960 --> 00:05:35,397
if only you had the time and head space.

92
00:05:35,480 --> 00:05:37,957
That's how I think about
personal super intelligence.

93
00:05:38,040 --> 00:05:41,677
Say you want to be healthier.
Your personal AI could help you see

94
00:05:41,760 --> 00:05:47,077
through a personal health plan covering
diet, exercise, and sleep in your daily

95
00:05:47,160 --> 00:05:50,757
routine. Or you have
a project you'd like to get

96
00:05:50,840 --> 00:05:54,757
done like putting on an event.
You can track your progress, reach out to

97
00:05:54,840 --> 00:05:57,477
venues, arrange invites, remind you of

98
00:05:57,560 --> 00:05:59,917
things you haven't considered and more.

99
00:06:00,000 --> 00:06:04,197
If you love to go fishing or paint
or want to travel more, it can help free

100
00:06:04,280 --> 00:06:07,917
you up so you can do more of these things.
And it can give you advice when

101
00:06:08,000 --> 00:06:10,037
you need it and or help you show up as a

102
00:06:10,120 --> 00:06:13,237
better friend or in your community.

103
00:06:13,320 --> 00:06:14,997
It won't just do your admin, it'll be an

104
00:06:15,080 --> 00:06:19,757
extension of you so you can be you more.

105
00:06:19,840 --> 00:06:24,637
I get that some people will worry
uh what companies that what companies

106
00:06:24,720 --> 00:06:27,877
like Meta really want is to get you
hooked and leave you passively staring

107
00:06:27,960 --> 00:06:30,597
at screens.
But the whole point of personal

108
00:06:30,680 --> 00:06:33,117
superintelligence is the opposite.

109
00:06:33,200 --> 00:06:37,517
It's about helping you be more active
in your life, in pursuing your goals, and

110
00:06:37,600 --> 00:06:41,397
deepening your relationships.
I know people are going to be skeptical

111
00:06:41,480 --> 00:06:45,637
when I say we're doing we're
going to do this work responsibly.

112
00:06:45,720 --> 00:06:49,357
But you don't have to take us at our word,
take us at our incentives.

113
00:06:49,440 --> 00:06:53,717
This is a competitive space, which
is why we're seeing so much innovation.

114
00:06:53,800 --> 00:06:57,357
Given how intimately your
personal AI will know you,

115
00:06:57,440 --> 00:06:58,997
people aren't going to hire us for the

116
00:06:59,080 --> 00:07:02,437
job if we're not doing it responsibly.

117
00:07:02,520 --> 00:07:04,917
Our AI needs to work the way we say it

118
00:07:05,000 --> 00:07:09,677
does, as well as we'd say it does, and
as safely and as securely as you need it

119
00:07:09,760 --> 00:07:12,797
to. It needs to help
you in your life and if

120
00:07:12,880 --> 00:07:15,917
it doesn't, people simply won't use it.

121
00:07:16,000 --> 00:07:19,357
We'll lose customers, we'll lose
public trust, and we'll lose out to our

122
00:07:19,440 --> 00:07:22,717
competitors. That's why
we're transparent about our

123
00:07:22,800 --> 00:07:27,117
models. We publish model cards
and evaluation benchmarks and data so you

124
00:07:27,200 --> 00:07:31,597
can see how they work, their intended use,
and how we assess their performance.

125
00:07:31,680 --> 00:07:36,197
And as they get more advanced,
we're looking at ways to share even more.

126
00:07:36,280 --> 00:07:39,637
It's why we invest in the
science of model evaluation,

127
00:07:39,720 --> 00:07:43,837
both improving the existing tests
and building new ones for risks we haven't

128
00:07:43,920 --> 00:07:48,237
yet confronted.
And it's why over many years

129
00:07:48,320 --> 00:07:52,477
we've developed ways to identify
and mitigate potential risks. Before we

130
00:07:52,560 --> 00:07:55,997
release a model through risk
assessments, scaled evaluations, red

131
00:07:56,080 --> 00:07:58,157
teaming, and fine-tuning.

132
00:07:58,240 --> 00:08:03,117
And we can monitor aggregate trends
in how people use AI in our apps. So, we

133
00:08:03,200 --> 00:08:06,917
have a feedback loop that can flag
potential risks and help us improve our

134
00:08:07,000 --> 00:08:09,520
models.

135
00:08:09,880 --> 00:08:11,837
As the models improve, the governance

136
00:08:11,920 --> 00:08:16,837
around them has to keep pace. So, we're
innovating with how they learn and apply

137
00:08:16,920 --> 00:08:21,357
principles and how they're tested and
evaluated using AI to strengthen checks

138
00:08:21,440 --> 00:08:25,477
and balances. Realizing
the full promise of AI is as

139
00:08:25,560 --> 00:08:29,557
much a matter of getting policy
right as it is investment.

140
00:08:29,640 --> 00:08:32,277
There are four building blocks for AI.

141
00:08:32,360 --> 00:08:36,917
Talent, energy, data, and compute.

142
00:08:37,000 --> 00:08:40,957
Governments and industry
need to work together

143
00:08:41,040 --> 00:08:45,277
to make sure there's access to each,
so we can realize AI's potential and do it

144
00:08:45,360 --> 00:08:50,077
in a way that means you can
build for your needs, not ours.

145
00:08:50,160 --> 00:08:51,957
That's in part about having bold

146
00:08:52,040 --> 00:08:57,117
national AI strategies and policies that
encourage innovation, not patchworks of

147
00:08:57,200 --> 00:09:00,517
inconsistent regulations
that make it harder.

148
00:09:00,600 --> 00:09:03,517
But above all, it's about
collaboration between public

149
00:09:03,600 --> 00:09:07,677
and private sectors to deliver these
four building blocks. And to design and

150
00:09:07,760 --> 00:09:12,517
deploy AI that works for your
citizens and your economies.

151
00:09:12,600 --> 00:09:16,557
I don't want these amazing
technologies to be one-size-fits-all.

152
00:09:16,640 --> 00:09:19,837
I want them to serve your needs,
designed for the challenges and

153
00:09:19,920 --> 00:09:24,437
opportunities that are unique to India,
to societies across the global south,

154
00:09:24,520 --> 00:09:27,997
and all over the world.
I want them to serve you as an

155
00:09:28,080 --> 00:09:31,717
individual, no matter who
you are, where you live,

156
00:09:31,800 --> 00:09:35,117
what language you speak,
or what culture you're a part of.

157
00:09:35,200 --> 00:09:38,797
That's only going to be possible if
the public and private sector are on the

158
00:09:38,880 --> 00:09:43,677
same side. We need to be partners
working together in a spirit of openness

159
00:09:43,760 --> 00:09:45,317
and collaboration.

160
00:09:45,400 --> 00:09:48,517
And with a sense of shared ambition.

161
00:09:48,600 --> 00:09:52,957
I truly believe we're on the cusp
of a moment where really anything is

162
00:09:53,040 --> 00:09:56,477
possible. We want to work
with you to build AI

163
00:09:56,560 --> 00:10:02,080
that serves our societies.
I hope you'll work with us. Thank you.

164
00:10:07,800 --> 00:10:09,877
Thank you, Mr. Alexander Wang. Our

165
00:10:09,960 --> 00:10:12,037
heartfelt gratitude to you for your

166
00:10:12,120 --> 00:10:14,997
remarks, for sharing your findings as

167
00:10:15,080 --> 00:10:17,677
well. And well, your vision for AI and

168
00:10:17,760 --> 00:10:22,240
innovation. It does inspire many of us.