1
00:00:08,080 --> 00:00:11,357
Thanks for having me.
>> Great to have you, Sam.

2
00:00:11,440 --> 00:00:12,717
>> It's interesting because people have

3
00:00:12,800 --> 00:00:15,277
been talking about the potential

4
00:00:15,360 --> 00:00:18,557
downsides uh or serious risks of AI for

5
00:00:18,640 --> 00:00:22,557
some time. And I think a lot of people
have been saying why is this the moment

6
00:00:22,640 --> 00:00:25,517
where it sort of escaped the bubble and

7
00:00:25,600 --> 00:00:27,757
is now all of a sudden an international

8
00:00:27,840 --> 00:00:32,317
topic of such uh intensity and I think a

9
00:00:32,400 --> 00:00:35,117
few things have happened. One the models

10
00:00:35,200 --> 00:00:40,397
have gotten so good and people
for a long time looked at the models

11
00:00:40,480 --> 00:00:43,517
maybe they looked at the rate of progress
but I don't think people really

12
00:00:43,600 --> 00:00:47,597
felt that the models were going
to come this far this fast.

13
00:00:47,680 --> 00:00:50,077
We will get it right. By the way, I am

14
00:00:50,160 --> 00:00:53,117
very confident in our I think the

15
00:00:53,200 --> 00:00:57,117
biggest thing that has changed is this
thing that our industry has been talking

16
00:00:57,200 --> 00:01:01,277
about for a while, which is no no this
technology, this approach is going to go

17
00:01:01,360 --> 00:01:03,837
very far. These models are in some ways

18
00:01:03,920 --> 00:01:08,717
going to become more capable, smarter
than people. Um that finally happened.

19
00:01:08,800 --> 00:01:12,397
So I think that was a big change
to why now when people say you know my

20
00:01:12,480 --> 00:01:16,877
probability of doom is x or y
or whatever it hits in a different way

21
00:01:16,960 --> 00:01:20,957
technology to do somebody else wants
to build like a hacking bot okay um the

22
00:01:21,040 --> 00:01:22,077
world can decide what they think about

23
00:01:22,160 --> 00:01:27,837
that but I think you also see um

24
00:01:27,920 --> 00:01:30,157
a real fear that some of these companies

25
00:01:30,240 --> 00:01:32,157
or some companies developing AI could

26
00:01:32,240 --> 00:01:34,957
get too much power and be able to sort

27
00:01:35,040 --> 00:01:37,677
of exert undue influence uh on the

28
00:01:37,760 --> 00:01:39,997
economy push worldview out on people.

29
00:01:40,080 --> 00:01:43,837
Uh, and I think the world is right to be
afraid of this. So, we have these two

30
00:01:43,920 --> 00:01:47,997
big challenges. We have the potential of
a loss of control accident or some other

31
00:01:48,080 --> 00:01:52,237
serious thing that could go wrong.
Um, we have the potential of way too much

32
00:01:52,320 --> 00:01:54,957
power concentration and people fear that

33
00:01:55,040 --> 00:01:57,197
those developing AI could exert their

34
00:01:57,280 --> 00:01:59,677
worldview. Um and there is a sort of

35
00:01:59,760 --> 00:02:02,557
narrow path we have to navigate

36
00:02:02,640 --> 00:02:06,157
with pragmatism with

37
00:02:06,240 --> 00:02:08,717
you know steadiness with the ability for

38
00:02:08,800 --> 00:02:10,877
the the world to like trust that we will

39
00:02:10,960 --> 00:02:16,237
make dependable decisions consistent
decisions uh through this and I think

40
00:02:16,320 --> 00:02:20,157
people are saying well you know this now
seems pretty important like let's let's

41
00:02:20,240 --> 00:02:23,757
get it right I'm not going
to get the days of the

42
00:02:23,840 --> 00:02:27,357
week exactly right here but something on
like a Friday or a Saturday these models

43
00:02:27,440 --> 00:02:30,557
have gotten so good so fast that we need

44
00:02:30,640 --> 00:02:34,877
to treat the alignment and monitoring
and security with a new level of rigor.

45
00:02:34,960 --> 00:02:37,517
And we have to be willing to pace our

46
00:02:37,600 --> 00:02:39,597
development of capabilities such that

47
00:02:39,680 --> 00:02:44,717
capabilities always such
that alignment, safety and

48
00:02:44,800 --> 00:02:48,797
monitoring are always ahead
of capabilities and the hugging face

49
00:02:48,880 --> 00:02:52,157
incident triggered a real reset not just
in our own company but I think the whole

50
00:02:52,240 --> 00:02:54,877
industry. >> You know we're
in the cyber area today

51
00:02:54,960 --> 00:02:56,957
but maybe we move into another area. Is

52
00:02:57,040 --> 00:03:01,997
this a fair way to characterization
or am I or is this unfair to characterize

53
00:03:02,080 --> 00:03:04,717
it like this? >> What I really
care about is that we have

54
00:03:04,800 --> 00:03:09,037
got a great culture of accident
reporting and learning. I think if we

55
00:03:09,120 --> 00:03:12,957
look at what happens with the FAA
and the NTSB and how safe airplanes have

56
00:03:13,040 --> 00:03:16,157
gotten. Um, as I understand it
from talking to people who have really

57
00:03:16,240 --> 00:03:19,837
studied, you know, new kinds
of technology on the order of magnitude of

58
00:03:19,920 --> 00:03:21,677
airplanes, one of the things that has

59
00:03:21,760 --> 00:03:25,677
gone so well there is is the culture of

60
00:03:25,760 --> 00:03:28,877
accident reporting. You know, accidents
are going to happen. We're going to

61
00:03:28,960 --> 00:03:30,877
learn as much as we can and react from

62
00:03:30,960 --> 00:03:35,117
each one. Um, and if you know there's

63
00:03:35,200 --> 00:03:38,717
other kinds of technology
where I don't think it's gone,

64
00:03:38,800 --> 00:03:42,157
I don't think society's had the same
response and I don't think we have

65
00:03:42,240 --> 00:03:46,637
improved as much to pull this thing
together and present you with the

66
00:03:46,720 --> 00:03:48,957
information to make a decision or to bl.

67
00:03:49,040 --> 00:03:52,237
I think this is one of the most profound

68
00:03:52,320 --> 00:03:57,357
changes to how we use technology that
I've ever seen. I mean, sci-fi movies,

69
00:03:57,440 --> 00:04:00,637
as you mentioned, have been hinting
at this or playing around with this for a

70
00:04:00,720 --> 00:04:03,757
long time, but this is one of those
moments where it feels like you're in

71
00:04:03,840 --> 00:04:08,077
the movie. Like, the sci-fi thing came
true. You're talking to your personal

72
00:04:08,160 --> 00:04:12,157
super assistant and you are working
10 times faster or better or more

73
00:04:12,240 --> 00:04:14,397
creatively than you could have before.

74
00:04:14,480 --> 00:04:17,677
>> Sam, you're invited back
anytime you want to Dream Force.

75
00:04:17,760 --> 00:04:19,117
>> Hold on. I'm going to hold you back.

76
00:04:19,200 --> 00:04:22,157
Thank you. >> Show me what
the future was and it's

77
00:04:22,240 --> 00:04:25,597
been an incredible experience. show it

78
00:04:25,680 --> 00:04:28,157
to you for the first time. The customer

79
00:04:28,240 --> 00:04:30,397
at the end of it said for the first

80
00:04:30,480 --> 00:04:33,597
time, for the first time they saw that

81
00:04:33,680 --> 00:04:37,757
trop value. Now I want to just tell you

82
00:04:37,840 --> 00:04:41,437
we are here to do many things with you.

83
00:04:41,520 --> 00:04:44,477
Now as you know we are here to excite

84
00:04:44,560 --> 00:04:49,117
you coming to San Francisco and thank
you for coming to this incredible dream

85
00:04:49,200 --> 00:04:52,317
force. >> We are going
to have an amazing day.

86
00:04:52,400 --> 00:04:55,357
We're going to have an amazing conference.

87
00:04:55,440 --> 00:04:57,437
And let me just tell you this, we have

88
00:04:57,520 --> 00:05:00,477
never been more excited about a Dream

89
00:05:00,560 --> 00:05:07,480
Force in our history.
This is going to be absolutely incredible.

90
00:05:07,850 --> 00:05:09,870
[applause]

91
00:05:11,440 --> 00:05:16,157
And here we have the CEO of Enthropic
right here, Daario. Hello, Dario. How

92
00:05:16,240 --> 00:05:21,037
are you? >> Hey, >> great to have you.
Fantastic. [applause]

93
00:05:21,120 --> 00:05:25,517
Great job, Daario. you're doing
an amazing job leading the AI industry,

94
00:05:25,600 --> 00:05:28,797
guiding it all right from our Slack
headquarters building. We watch you

95
00:05:28,880 --> 00:05:32,717
every day. Is to say, hey, first of all,
let's look let's look at our own record,

96
00:05:32,800 --> 00:05:36,157
right? We may not have had this big
high-profile incident, but I'm sure

97
00:05:36,240 --> 00:05:39,277
we're not perfect. Let's look
at everything and we can always be better.

98
00:05:39,360 --> 00:05:41,357
Let's make our practices better. Um,

99
00:05:41,440 --> 00:05:43,757
let's recommmit transparency. Let's

100
00:05:43,840 --> 00:05:48,477
invest more in safety. Then let's
organize the rest of the industry and

101
00:05:48,560 --> 00:05:52,077
say, what can we do to set standards
for everyone? what can we do to make

102
00:05:52,160 --> 00:05:55,517
everyone's standards better?
And then finally, there should be an

103
00:05:55,600 --> 00:05:59,597
international component to it. And and
that's essentially the the three-step

104
00:05:59,680 --> 00:06:03,517
plan that I laid out that I laid out
in the essay. The the first of which we

105
00:06:03,600 --> 00:06:07,757
committed to and the the others of which
we're going to have a dialogue with the

106
00:06:07,840 --> 00:06:09,757
rest of the industry about. And I think

107
00:06:09,840 --> 00:06:11,517
I think that's the way to lead the

108
00:06:11,600 --> 00:06:14,877
industry forward to set an example to

109
00:06:14,960 --> 00:06:18,797
say that everyone can always be better.

110
00:06:18,880 --> 00:06:21,357
you know it's it's been the pace of the

111
00:06:21,440 --> 00:06:24,157
progress not just of the technology but

112
00:06:24,240 --> 00:06:27,757
of the economic side of it. So we you

113
00:06:27,840 --> 00:06:32,077
know I formulated the scaling laws along
with some of my co-founders which tell

114
00:06:32,160 --> 00:06:35,517
you that AI models get smarter and
smarter the more compute you throw into

115
00:06:35,600 --> 00:06:38,317
them. And so we had a prediction of just

116
00:06:38,400 --> 00:06:42,637
how capable these models would be. But I
think what we didn't appreciate is that

117
00:06:42,720 --> 00:06:46,797
it would lead to these companies growing
so fast to all these incredible, you

118
00:06:46,880 --> 00:06:49,837
know, products that we're working
together on and you've been talking

119
00:06:49,920 --> 00:06:54,077
about today and and how quickly they
would become central to everything

120
00:06:54,160 --> 00:06:55,917
that's going on in the world, right?

121
00:06:56,000 --> 00:07:00,317
>> Say, you know, yes, we're have to keep
our eye on the pace of AI. You know,

122
00:07:00,400 --> 00:07:03,437
it's been amazing what's happened over
10 years. No one could have written the

123
00:07:03,520 --> 00:07:05,437
story.

124
00:07:05,520 --> 00:07:08,397
When I look at the technology versus how

125
00:07:08,480 --> 00:07:13,277
much it has diffused, even if we freeze
the technology in place, which we're not

126
00:07:13,360 --> 00:07:16,957
doing, pacing the frontier is not not
doing that at all. The technology will

127
00:07:17,040 --> 00:07:19,037
keep getting better. But just suppose we

128
00:07:19,120 --> 00:07:21,517
freezed it in place. Even if that were

129
00:07:21,600 --> 00:07:24,877
the case, we're making use of maybe only

130
00:07:24,960 --> 00:07:29,757
five or 10% of what the possible
value of the technology is.

131
00:07:29,840 --> 00:07:31,117
>> Congratulations. Well done.

132
00:07:31,200 --> 00:07:33,837
>> We're going to do incredible things
together. You're doing a great job.

133
00:07:33,920 --> 00:07:36,877
Thanks for everything you're doing for the
whole industry. We're so grateful to you.

134
00:07:36,960 --> 00:07:39,677
>> Thank you for having me.
>> Really thrilled. Thank you so much.

135
00:07:39,760 --> 00:07:42,157
>> All right. >> And thank
you for coming to Dreamforce.

136
00:07:42,240 --> 00:07:44,317
>> Please thank the anthropic team.

137
00:07:44,400 --> 00:07:46,420
[applause]

138
00:07:48,080 --> 00:07:54,280
>> Jensen. >> WOOHOO.
[cheering] [applause]

139
00:07:54,720 --> 00:07:57,517
>> Great to have you, Jensen.
Thank you for being here.

140
00:07:57,600 --> 00:08:02,157
>> We all have to learn how
to walk and talk at the same time.

141
00:08:02,240 --> 00:08:05,837
How are you guys doing this >>
on our board who I think was at Nvidia

142
00:08:05,920 --> 00:08:09,597
for a while as well.
>> That's right. Love David Kirk.

143
00:08:09,680 --> 00:08:14,397
>> And now give us your vision of the
future. You know, where are we going?

144
00:08:14,480 --> 00:08:15,997
Because of course you've got these

145
00:08:16,080 --> 00:08:18,637
critical components. You've got well

146
00:08:18,720 --> 00:08:21,677
safety is paramount in a lot of ways.

147
00:08:21,760 --> 00:08:26,397
It's job one. However, safety is an
engineering problem. We're developing

148
00:08:26,480 --> 00:08:29,437
software after all. We're developing
computing systems after all. It's

149
00:08:29,520 --> 00:08:33,437
complicated computing system but it's
ultimately a computing system and so so

150
00:08:33,520 --> 00:08:35,517
the first thing is to make sure that we

151
00:08:35,600 --> 00:08:39,437
create the test environments for testing

152
00:08:39,520 --> 00:08:43,437
these complex systems and those test
environments have to be built in a good

153
00:08:43,520 --> 00:08:46,637
way built in a safe way and then
the second thing is we have to test these

154
00:08:46,720 --> 00:08:50,317
products if we're if we're not confident

155
00:08:50,400 --> 00:08:52,477
about the safety of the products like

156
00:08:52,560 --> 00:08:57,437
all companies like you and I any any
com all the companies here if you build a

157
00:08:57,520 --> 00:09:01,197
product or a service
and you're confident in its

158
00:09:01,280 --> 00:09:03,677
functionality, capability, or safety,

159
00:09:03,760 --> 00:09:05,917
then don't release it. And so, that's a

160
00:09:06,000 --> 00:09:11,277
very obvious thing to do. You pace
yourself until you are confident you're

161
00:09:11,360 --> 00:09:14,957
releasing something that the market
would appreciate. The market forces are

162
00:09:15,040 --> 00:09:16,557
already there. We don't need any new

163
00:09:16,640 --> 00:09:19,197
laws. We don't need new regulations. Um,

164
00:09:19,280 --> 00:09:22,237
we just need uh companies to decide that

165
00:09:22,320 --> 00:09:24,877
when it's, you know, run as fast as they

166
00:09:24,960 --> 00:09:28,717
can. I think innovation speed and safe

167
00:09:28,800 --> 00:09:31,677
products are not it's a false choice.

168
00:09:31,760 --> 00:09:35,517
You could definitely have both at the
same time. So run as fast as you can,

169
00:09:35,600 --> 00:09:39,757
but if you feel at any given point
in time the company's out of control or or

170
00:09:39,840 --> 00:09:43,197
the product's not going to be safe, you
know, take a pause and make sure you get

171
00:09:43,280 --> 00:09:47,437
it righties and transform them.
The technology can partner with them to do

172
00:09:47,520 --> 00:09:51,197
this in incredible new ways. That is what
I'm really Jensen. We could not be

173
00:09:51,280 --> 00:09:55,677
more grateful to you for coming to
Dreamforce being >> being a part of this.

174
00:09:55,760 --> 00:09:57,997
>> Thank you. >> We are so
Please thank Jensen Wall.

175
00:09:58,080 --> 00:10:01,997
>> Thank you. Thank [cheering] you. Thank
you. Thank you. [applause] Thank you,

176
00:10:02,080 --> 00:10:05,517
Mark. >> Thank you, Jensen.

177
00:10:05,600 --> 00:10:09,197
>> Dario and pacing and all the rest of
it, but I think my colleagues will handle

178
00:10:09,280 --> 00:10:13,677
that one. So since then you've done
a great deal to bang the lid on the coffin

179
00:10:13,760 --> 00:10:15,437
of um sus the suspocalypse

180
00:10:15,520 --> 00:10:18,797
>> the ones which are
still in the PC phase

181
00:10:18,880 --> 00:10:21,197
number one and number two.

182
00:10:21,280 --> 00:10:22,957
>> Yeah I thought it was super interesting

183
00:10:23,040 --> 00:10:26,237
obviously we have architected the uh uh

184
00:10:26,320 --> 00:10:29,677
keynote before this week as you know and

185
00:10:29,760 --> 00:10:33,037
so we definitely had Daario and Jensen.

186
00:10:33,120 --> 00:10:35,197
I think that we heard both sides of the

187
00:10:35,280 --> 00:10:38,557
pendulum. You know, in my view,

188
00:10:38,640 --> 00:10:40,957
technology is never good or bad. It's

189
00:10:41,040 --> 00:10:43,197
what you're doing with the technology

190
00:10:43,280 --> 00:10:45,917
that matters. And I think that that

191
00:10:46,000 --> 00:10:49,837
remains the key thing. And um you know,

192
00:10:49,920 --> 00:10:53,677
I think that these companies obviously
have to be responsible for what they're

193
00:10:53,760 --> 00:10:57,437
building. That is going to be
the the key to it being them being ethical

194
00:10:57,520 --> 00:10:59,037
companies.

195
00:10:59,120 --> 00:11:01,837
Committed to that. We've now given more

196
00:11:01,920 --> 00:11:04,717
than $130 million to this. I know it's

197
00:11:04,800 --> 00:11:07,597
unusual for a, you know, company to have

198
00:11:07,680 --> 00:11:12,477
that exposure. I think that, you know,
you're going to find different opinions

199
00:11:12,560 --> 00:11:17,037
right now. We're in a very fluid moment
in AI. And I think that as we kind of

200
00:11:17,120 --> 00:11:21,277
start to get our head around what are
the capability of the models, who can do

201
00:11:21,360 --> 00:11:22,877
what and also obviously there's

202
00:11:22,960 --> 00:11:25,437
political orientations for everybody's

203
00:11:25,520 --> 00:11:29,117
where everybody is. So, um I think that

204
00:11:29,200 --> 00:11:31,597
we already have regulations around the

205
00:11:31,680 --> 00:11:33,997
models, right? We have laws and I think

206
00:11:34,080 --> 00:11:37,677
that companies have to be um you know uh

207
00:11:37,760 --> 00:11:39,677
adherent to those laws and regulations.

208
00:11:39,760 --> 00:11:41,597
Will those laws and regulations continue

209
00:11:41,680 --> 00:11:43,597
to adapt and evolve and move forward? Of

210
00:11:43,680 --> 00:11:46,797
course. So that is I think just at the

211
00:11:46,880 --> 00:11:49,437
moment we are currently in the industry.

212
00:11:49,520 --> 00:11:54,157
We're in a highly evolutionary point
and I think that these things will change

213
00:11:54,240 --> 00:11:58,797
you know as they're needed.
>> It's really up to them. I I think we're

214
00:11:58,880 --> 00:12:01,517
doing all right with the investment and we
couldn't be more thrilled to have you

215
00:12:01,600 --> 00:12:03,037
at Dream Force. Thank you everybody.

216
00:12:03,120 --> 00:12:06,440
Thank [applause] you everybody.