1
00:00:00,240 --> 00:00:03,357
I'm honored to be here and to be able
to speak with you all at this important

2
00:00:03,440 --> 00:00:07,757
time. We've been talking about
artificial intelligence for years, but

3
00:00:07,840 --> 00:00:11,917
it feels different in recent weeks
and months. Rapid model progress has made

4
00:00:12,000 --> 00:00:15,517
the timeline feel more compressed,
the upside more tangible, but also the

5
00:00:15,600 --> 00:00:18,557
stakes and the risks more immediate. We

6
00:00:18,640 --> 00:00:21,517
have a choice in front of us. AI can

7
00:00:21,600 --> 00:00:25,757
either be more like a new renaissance
of creativity and discovery or more like a

8
00:00:25,840 --> 00:00:27,757
new industrial revolution of upheaval

9
00:00:27,840 --> 00:00:30,477
and disarray. The Renaissance was a time

10
00:00:30,560 --> 00:00:33,117
when new tools, new ideas, and new

11
00:00:33,200 --> 00:00:34,877
institutions expanded what people

12
00:00:34,960 --> 00:00:37,837
believed they could do. AI can play that

13
00:00:37,920 --> 00:00:42,237
role for our time. I believe that people
are much more capable than they imagine

14
00:00:42,320 --> 00:00:46,157
themselves to be, but have been held
back by limited technology that was

15
00:00:46,240 --> 00:00:50,797
meant to empower but instead
distracts and often steals attention.

16
00:00:50,880 --> 00:00:55,117
The best version of AI is not about
making people cogs in a giant machine or

17
00:00:55,200 --> 00:00:58,797
optimizing every part of life
until the human parts disappear.

18
00:00:58,880 --> 00:01:02,717
It is about giving people more agency in
an ever more complex world. More ability

19
00:01:02,800 --> 00:01:05,677
to learn, to create, discover, build,

20
00:01:05,760 --> 00:01:07,997
participate. There are many things that

21
00:01:08,080 --> 00:01:12,717
AI cannot and should not automate.
AI should be built for people, not simply

22
00:01:12,800 --> 00:01:17,597
to turn the crank crank of the
machines faster for their own sake.

23
00:01:17,680 --> 00:01:20,717
Some of the things people working to build
AI have said are so dystopian that

24
00:01:20,800 --> 00:01:22,237
they sound like the plot of bad science

25
00:01:22,320 --> 00:01:25,037
fiction movies. Anything this power is

26
00:01:25,120 --> 00:01:29,837
also powerful is also intimidating.
This moment is complicated as we feel both

27
00:01:29,920 --> 00:01:35,197
tremendous potential and the very
understandable anxiety at the same time.

28
00:01:35,280 --> 00:01:38,477
These models can help people discover
new knowledge, new medicines, better

29
00:01:38,560 --> 00:01:42,157
healthare, stronger education, and build
more productive economies. They can be

30
00:01:42,240 --> 00:01:45,997
new tools for people to pursue dreams
and ambition that today feel out of

31
00:01:46,080 --> 00:01:49,837
reach. The complexity of life has
increased so dramatically for the

32
00:01:49,920 --> 00:01:53,437
average person. AI can be
a defensive tool against this

33
00:01:53,520 --> 00:01:57,277
complexity as much as it can be
this tool of unprecedented discovery and

34
00:01:57,360 --> 00:01:59,437
innovation.

35
00:01:59,520 --> 00:02:01,917
On the other hand, as AI systems become

36
00:02:02,000 --> 00:02:04,477
more capable and more autonomous, they

37
00:02:04,560 --> 00:02:06,637
could move faster than our institutions,

38
00:02:06,720 --> 00:02:09,357
concentrate power in two too few hands,

39
00:02:09,440 --> 00:02:14,477
or make decisions that people
no longer understand or control.

40
00:02:14,560 --> 00:02:17,917
That concern becomes especially
important as we approach systems that

41
00:02:18,000 --> 00:02:21,757
can improve themselves and future
versions of themselves, often called

42
00:02:21,840 --> 00:02:25,997
recursive self-improvement.
As the process of building AI becomes

43
00:02:26,080 --> 00:02:28,397
more automated, the pace of AI progress

44
00:02:28,480 --> 00:02:30,797
could accelerate rapidly. This moment

45
00:02:30,880 --> 00:02:35,997
calls for extreme care.
I largely agree with Professor Benjio.

46
00:02:36,080 --> 00:02:39,117
Uh although I'll say two instead of three,
I believe there are two ways that

47
00:02:39,200 --> 00:02:41,277
AI progress could go very badly that we

48
00:02:41,360 --> 00:02:44,477
have to avoid. First, we could lose

49
00:02:44,560 --> 00:02:49,197
control of the future to AI.
The risk is that it moves so fast that

50
00:02:49,280 --> 00:02:52,877
people can no longer follow what's
happening or intervene when needed. This

51
00:02:52,960 --> 00:02:56,477
would obviously be terrible.
The industry must not accept too much

52
00:02:56,560 --> 00:02:59,757
technological risk just because
the benefits are too great and that they

53
00:02:59,840 --> 00:03:03,837
feel too important to slow down. Beating
companies in competitive race is not a

54
00:03:03,920 --> 00:03:06,957
reason to make rash decisions. Nor do we
believe we are locked in a race where we

55
00:03:07,040 --> 00:03:10,157
are unable to do that. We have
unilaterally slowed down in the past. We

56
00:03:10,240 --> 00:03:13,037
will do so in the future. We do not
want to build systems. We think it's

57
00:03:13,120 --> 00:03:16,557
important that everyone avoids building
systems where we cannot make the

58
00:03:16,640 --> 00:03:20,717
alignment, monitor, and safety
guarantees that we must. We need to

59
00:03:20,800 --> 00:03:23,917
understand what these systems are doing
and have strong evidence that they will

60
00:03:24,000 --> 00:03:28,237
do what people intend even as they get
very very smart. Actually, especially as

61
00:03:28,320 --> 00:03:30,797
they get very, very smart. It doesn't

62
00:03:30,880 --> 00:03:35,757
matter whether people put the risk
of catastrophe at 10% or 1% or 12% or.1%.

63
00:03:35,840 --> 00:03:38,957
None of these levels are remotely
acceptable. And we should not train

64
00:03:39,040 --> 00:03:42,477
models that we cannot make an extremely
strong case that will be able to keep

65
00:03:42,560 --> 00:03:47,517
under human control. So that's one
way things can go wrong. In the other

66
00:03:47,600 --> 00:03:50,477
direction, these systems could
concentrate too much power in too few

67
00:03:50,560 --> 00:03:53,517
hands. No one person or company or

68
00:03:53,600 --> 00:03:55,437
country should be able to use the most

69
00:03:55,520 --> 00:04:00,717
powerful AI models to impose their
worldview on everyone else. A company or

70
00:04:00,800 --> 00:04:03,917
country that believes only it can be
trusted with this technology can use

71
00:04:04,000 --> 00:04:05,597
that belief to justify almost anything

72
00:04:05,680 --> 00:04:08,157
else. We have to reject that logic even

73
00:04:08,240 --> 00:04:11,277
when it's convenient for anyone actor. I

74
00:04:11,360 --> 00:04:15,837
believe there is a pragmatic path
through both of these. At OpenAI, we

75
00:04:15,920 --> 00:04:20,077
approach these questions with a clear
goal. AI should give people more power

76
00:04:20,160 --> 00:04:23,597
over their own lives, more freedom
to live lives the way they want, more

77
00:04:23,680 --> 00:04:28,077
ability to solve problems, big or small,
and to improve their communities. We

78
00:04:28,160 --> 00:04:31,197
don't want to fall into the trap
of blind optimism. We also don't want to

79
00:04:31,280 --> 00:04:33,517
fall into the trap of dumerism either.

80
00:04:33,600 --> 00:04:37,197
Getting pulled into either extreme will
not help us successfully figure out the

81
00:04:37,280 --> 00:04:41,117
challenges in front of us. Especially
at a time when the world and its people

82
00:04:41,200 --> 00:04:45,677
face such enormous challenges that this
technology can help solve. So we try to

83
00:04:45,760 --> 00:04:47,197
walk the middle path between these two

84
00:04:47,280 --> 00:04:49,677
poles. For us, that means three

85
00:04:49,760 --> 00:04:53,437
principles. First, as AI becomes more

86
00:04:53,520 --> 00:04:57,757
capable, people must remain
at the center of AI decision-m

87
00:04:57,840 --> 00:05:01,357
alignment is not an abstract research
question. It is the work of ensuring

88
00:05:01,440 --> 00:05:06,157
that these systems reliably remain under
human control, reflect human values, and

89
00:05:06,240 --> 00:05:10,077
help people guide the next stages
of development. We are not trying to and

90
00:05:10,160 --> 00:05:13,677
must not automate human
judgment or human values.

91
00:05:13,760 --> 00:05:17,917
Second, the benefits of scientific
progress and economic growth must be by

92
00:05:18,000 --> 00:05:21,597
people and for people. AI should help
researchers make discoveries, doctors

93
00:05:21,680 --> 00:05:25,037
treat patients, teachers teach students,
entrepreneurs start companies,

94
00:05:25,120 --> 00:05:28,877
communities solve problems that felt out
of reach, and much more. The continuous

95
00:05:28,960 --> 00:05:33,357
story of human progress is not that
tools replace human agency. It is that

96
00:05:33,440 --> 00:05:36,397
the right tools make people more capable.
And I believe this is going to

97
00:05:36,480 --> 00:05:40,957
go much further than we believe possible.
Third, this technology must

98
00:05:41,040 --> 00:05:45,037
empower people individually. A great
future will come from people realizing

99
00:05:45,120 --> 00:05:48,717
their potential and building value
for themselves, their communities, their

100
00:05:48,800 --> 00:05:52,157
countries, and the broader world.
We might be the builders, one of the

101
00:05:52,240 --> 00:05:55,837
builders of this technology, but we are
not the heroes of the story. Our role is

102
00:05:55,920 --> 00:05:59,837
to enable people all over the world
and to trust in the magic of humanity's

103
00:05:59,920 --> 00:06:04,237
skill and faculties.
We now have models capable of

104
00:06:04,320 --> 00:06:08,157
discovering knowledge that we have never
had before. Consider the progress we've

105
00:06:08,240 --> 00:06:13,357
seen in math. Three summers ago,
our models were okay at grade school math.

106
00:06:13,440 --> 00:06:17,597
Two summers ago, they were pretty good
at high school math competitions. Last

107
00:06:17,680 --> 00:06:20,877
summer, we reached gold a gold
level medal in the most prestigious

108
00:06:20,960 --> 00:06:25,357
international math competition. And just
a few weeks ago this summer, one of our

109
00:06:25,440 --> 00:06:29,837
models solved one of the Millennium Prize
problems, the Navia Stokes equations.

110
00:06:29,920 --> 00:06:33,277
These equations are used for aircraft
design, weather forecasting, and the

111
00:06:33,360 --> 00:06:36,957
study of blood flow. We are not solving
math problems for their own sake, but to

112
00:06:37,040 --> 00:06:40,957
empower people to discover new knowledge
and to use it to improve healthcare,

113
00:06:41,040 --> 00:06:44,957
raise standards of living of living,
and expand everyone's potential around the

114
00:06:45,040 --> 00:06:48,477
world. There remains so
much to be discovered,

115
00:06:48,560 --> 00:06:51,517
but it's very important that companies
not substitute for the democratic

116
00:06:51,600 --> 00:06:54,477
process. If AI is to be democratic, the

117
00:06:54,560 --> 00:06:58,717
most important decisions cannot be made
by labs in San Francisco alone. They

118
00:06:58,800 --> 00:07:02,237
must be shaped through democratic
processes and by governments accountable

119
00:07:02,320 --> 00:07:04,877
to the people that they serve. At the

120
00:07:04,960 --> 00:07:09,837
international level, this will require
cooperation. In our history, there have

121
00:07:09,920 --> 00:07:12,717
been times where countries who compete
and don't always like each other very

122
00:07:12,800 --> 00:07:16,557
much still come together for shared
interests and the collective good in the

123
00:07:16,640 --> 00:07:18,957
face of a powerful new technology. We

124
00:07:19,040 --> 00:07:22,237
believe this must be one of those times.

125
00:07:22,320 --> 00:07:25,837
We need a mechanism for complimentary
national and international frontier AI

126
00:07:25,920 --> 00:07:30,557
standards. Standards for measuring
capabilities, assessing risks,

127
00:07:30,640 --> 00:07:33,997
determining whether safeguards are
sufficient, and preserving meaningful

128
00:07:34,080 --> 00:07:37,517
human oversight as systems
become more autonomous.

129
00:07:37,600 --> 00:07:41,997
We need common standards so countries
can compare evidence, verify compliance,

130
00:07:42,080 --> 00:07:45,837
and have a shared language
and understanding about what is happening.

131
00:07:45,920 --> 00:07:49,677
We need accurate and speedy incident
reporting, classification and reporting

132
00:07:49,760 --> 00:07:54,797
protocols so the world can learn from
failures before they become catastrophes.

133
00:07:54,880 --> 00:07:58,077
And we need secure channels among
governments, critical infrastructure

134
00:07:58,160 --> 00:08:01,837
operators and technical experts
to share emerging vulnerabilities and new

135
00:08:01,920 --> 00:08:05,117
threats. Obviously,
these standards should not

136
00:08:05,200 --> 00:08:09,517
lock in incumbents or favor one business
model over another. They must support

137
00:08:09,600 --> 00:08:13,597
open and closed model developers, new
entrance and established labs, national

138
00:08:13,680 --> 00:08:17,277
authority, and international cooperation.
Each government should

139
00:08:17,360 --> 00:08:20,877
decide how to incorporate standards into
its own legal system, but we will all be

140
00:08:20,960 --> 00:08:24,797
better off if we can agree on what
good evidence, good safeguards, and good

141
00:08:24,880 --> 00:08:27,837
oversight look like on the global stage.

142
00:08:27,920 --> 00:08:31,757
We do not have all the answers,
and we're very open to any better ideas.

143
00:08:31,840 --> 00:08:34,797
We're ready to work with you all and try
our hardest to figure out the path

144
00:08:34,880 --> 00:08:38,637
forward. We are at a crossroads.

145
00:08:38,720 --> 00:08:43,117
One path leads to a world where power
is concentrated, decisions are opaque, and

146
00:08:43,200 --> 00:08:45,997
people feel that the most important
technology of their lifetime is

147
00:08:46,080 --> 00:08:48,957
something being done to them. The other

148
00:08:49,040 --> 00:08:53,437
path leads to a world where AI works for
people because people through democratic

149
00:08:53,520 --> 00:08:57,677
institutions help decide how it should
develop and guide it, where scientific

150
00:08:57,760 --> 00:09:01,597
progress improves real lives, where more
people can participate in the future,

151
00:09:01,680 --> 00:09:05,437
and where this technology continues to
give individuals and communities greater

152
00:09:05,520 --> 00:09:09,517
power to live life as they choose.
This is a future of realized potential,

153
00:09:09,600 --> 00:09:14,640
excitement, and wonder. And it is the
future we believe is worth building.