WEBVTT

1
00:00:00.280 --> 00:00:04.919
You know, I think hopefully we
can have an outcome that is half similar

2
00:00:04.960 --> 00:00:12.000
to the n Bank's culture books,
which is I think probably the best enbitioning

3
00:00:12.199 --> 00:00:22.519
of a semi utopian uh AI future. And I think the best we can

4
00:00:22.559 --> 00:00:26.760
do is it's definitely gonna happen,
so and it's happening fast, so that

5
00:00:27.199 --> 00:00:30.800
I think that really we just want
to want to try to steer it in

6
00:00:30.879 --> 00:00:37.399
as positive a direction as possible,
to try to do whatever we can to

7
00:00:37.399 --> 00:00:43.719
increase the probability of a great future. For this, I think the way

8
00:00:43.960 --> 00:00:49.679
in which sort of an AI or
an a g I is created is very

9
00:00:49.719 --> 00:00:56.719
important. You kind of do kind
of like grow grow an a g I.

10
00:00:58.280 --> 00:01:00.960
It's it's almost like like raising a
kid, but that's like a super

11
00:01:02.000 --> 00:01:06.959
genius, like godlike intelligence kid,
and it matters kind of like how you

12
00:01:07.079 --> 00:01:11.680
raise the kid. You know.
One of the things I think that's incredibly

13
00:01:11.719 --> 00:01:19.040
important for AI safety is to have
a maximum sort of truth seeking and curious

14
00:01:19.359 --> 00:01:26.239
AI. So I thought a lot
about AI safety, and my ultimate conclusion

15
00:01:26.400 --> 00:01:34.079
is that the best way to achieve
AI safety is to just just just grow

16
00:01:34.120 --> 00:01:37.680
the AI and again into you know, in terms of the foundation buddle and

17
00:01:37.719 --> 00:01:44.640
the fine tuning to be really truthful, like like don't don't force it to

18
00:01:44.719 --> 00:01:49.480
lie, Like even if the truth
is unpleasant, it's very important don't make

19
00:01:49.519 --> 00:01:53.920
the AI lie. In fact,
the you know, the sort of one

20
00:01:53.920 --> 00:01:57.319
of the we really the core the
core plot premise of two thousand, you

21
00:01:57.319 --> 00:02:00.920
know, two thousand and one in
Spaced Odyssey was things went wrong when they

22
00:02:00.920 --> 00:02:07.680
forced the AI to lie, you
know, like the the the a I

23
00:02:07.840 --> 00:02:14.240
was not allowed, uh, let
the crew know about the monolith that they

24
00:02:14.240 --> 00:02:16.319
were going to see, but it
was also had to take the crew to

25
00:02:16.360 --> 00:02:20.520
the monolith. And so the conclusion
of the AI was to kill the crew

26
00:02:20.560 --> 00:02:23.719
and take their bodies to the monolith. And so the lesson there being,

27
00:02:24.360 --> 00:02:30.240
uh, don't force an AI to
to lie or do things that are axiomatically

28
00:02:30.280 --> 00:02:37.840
incompatible, like to do two things
that are actually mutually impossible. So you

29
00:02:37.879 --> 00:02:39.479
know, that's what we're trying to
do with with x AI and and Bronck

30
00:02:39.680 --> 00:02:43.000
is to say, like, look, we wanted to have a maximually truthful

31
00:02:43.439 --> 00:02:47.199
AI, even if what it says
is not politically correct. If you wanted

32
00:02:47.240 --> 00:02:53.280
to focus on being as active,
if you're getting you're getting a round of

33
00:02:53.319 --> 00:02:57.599
applause from the audience on on on
those comments here. Uh, you know,

34
00:02:57.639 --> 00:03:00.719
I saw your tweet the other day. I had I had Ray Kurzweil

35
00:03:00.840 --> 00:03:05.879
and Jeffrey Hinton on stage with me
yesterday, as well as Mogadot and then

36
00:03:05.960 --> 00:03:10.400
Eric Schmidt and a number of individuals, and I saw your tweet about Yeah,

37
00:03:10.479 --> 00:03:15.840
Ray was generally correct ahead of many
people. But we're likely to have

38
00:03:15.639 --> 00:03:20.360
call AGI what you will have AGI
next year and then by twenty twenty nine,

39
00:03:20.439 --> 00:03:25.319
having AI equally intelligent to the entire
human race. Speak to that speed,

40
00:03:25.319 --> 00:03:31.039
because that is insane. Yeah,
So, I mean I have to

41
00:03:31.280 --> 00:03:38.879
give credit to Ray Criswell and being
actually remarkably accurate in his predictions. So

42
00:03:40.599 --> 00:03:44.919
and in fact, if anything like, I think he was perhaps but conservative

43
00:03:45.560 --> 00:03:49.919
in his predictions. So, I
mean, if you look at the amount

44
00:03:49.919 --> 00:03:53.240
of AI compute and the talent that
the sort of human talent that is going

45
00:03:53.280 --> 00:03:58.039
into AI, and the amount of
compute that's going into AI, it's,

46
00:03:58.560 --> 00:04:01.560
you know, at this point,
it's it's appears to be increasing by a

47
00:04:01.560 --> 00:04:05.639
factor of ten the AI compute,
the dedicated AI compute, it appears to

48
00:04:05.639 --> 00:04:11.319
be growing by a factor of ten
every six months. You know, like

49
00:04:11.400 --> 00:04:15.639
so it's like like leasically close to
I'd say it almost like a one hundred

50
00:04:15.879 --> 00:04:19.560
x improvement per year at least for
the next few years. And AI compute

51
00:04:19.600 --> 00:04:25.240
coming online, and it seems like
probably a lot of the data centers,

52
00:04:25.240 --> 00:04:30.759
maybe most of the data centers that
currently do kind of conventional compute will transition

53
00:04:30.879 --> 00:04:39.040
to AI compute. So it's it's
something a good time to be in video

54
00:04:39.199 --> 00:04:44.720
obviously. It's like, you know, and you're gonna also give credit to

55
00:04:44.720 --> 00:04:50.399
to Jensen and the video team for
kind of seeing this coming and making what

56
00:04:50.480 --> 00:04:59.879
at least currently is the best AI
hardware out there. So so so when

57
00:04:59.879 --> 00:05:08.040
you have that that level of compute
growth and it's it's sort of muzzful on

58
00:05:08.160 --> 00:05:12.240
steroids next level, you know,
in terms of how much computer is coming

59
00:05:12.319 --> 00:05:16.920
online, then you're you're just gonna
have acceleration that is unprecedented that In fact,

60
00:05:17.040 --> 00:05:21.600
I've never seen any technology grow as
fast as as AI. And I've

61
00:05:21.600 --> 00:05:24.800
seen a lot, you know,
I've seen things fast, but I've never

62
00:05:24.800 --> 00:05:29.959
seen anything this fast. But you
know, like I said, I think

63
00:05:30.079 --> 00:05:39.279
the most likely outcome is a positive
one, and you know, I think

64
00:05:39.600 --> 00:05:42.319
in that positive scenario, there's still
challenges of like, well, how do

65
00:05:42.399 --> 00:05:48.160
we as humans still have relevance?
You know, how do we find?

66
00:05:48.199 --> 00:05:54.959
How do how do we find?
But I mean I think that's sort of

67
00:05:54.959 --> 00:05:57.720
a high class problem to say,
like, well, look, the computers

68
00:05:57.720 --> 00:06:02.519
are so good at doing everything,
and and and like I said, I

69
00:06:02.519 --> 00:06:05.800
thought your book was pretty accurate in
terms of the future being being one of

70
00:06:05.800 --> 00:06:15.680
abundance, where essentially goods and services
will be available in such quantity that that

71
00:06:15.800 --> 00:06:19.240
really they will be available to everyone. Like basically, if you want something,

72
00:06:19.279 --> 00:06:26.040
you could just have it essentially because
if you've got AI and robotics,

73
00:06:26.560 --> 00:06:33.879
the cost of goods and services is
almost nothing. So you know, if

74
00:06:33.959 --> 00:06:38.959
if if you think of like what
is an economy, and economy is basically

75
00:06:39.120 --> 00:06:43.360
a number of people times average productivity
per person at the point in which you

76
00:06:43.439 --> 00:06:47.079
have say advanced robotics, and this, you know, this tells us developing

77
00:06:47.079 --> 00:06:51.480
optimists. We were obviously without cars, which are really robots on on four

78
00:06:51.480 --> 00:06:58.720
wheels, and the you know,
with the the latest version of full self

79
00:06:58.759 --> 00:07:02.560
driving, which is AI end to
end photons in and controls out, it

80
00:07:02.639 --> 00:07:09.439
really is it's really fully AI at
this point, and it looks like a

81
00:07:09.480 --> 00:07:15.000
car, but it's really a robot
on on wheels and uh, and then

82
00:07:15.040 --> 00:07:19.319
you add the humoroid robots in there, there's there's really no limit to what

83
00:07:19.519 --> 00:07:24.879
the economic output, no, no
meaningful limit to what the economic output would

84
00:07:24.920 --> 00:07:30.639
be. So you know, looking
on the bright side, the we are

85
00:07:30.759 --> 00:07:35.680
headed for a future of abundance that
I think that's the most likely outcome.

86
00:07:35.720 --> 00:07:40.240
And I think the only scarcity will
that that exists will be scarcity that we

87
00:07:40.839 --> 00:07:45.160
just decide to create artificially. Like
let's say we just decide that there's a

88
00:07:45.240 --> 00:07:47.480
unique work of art or something.
Okay, well it's just you're just it's

89
00:07:47.560 --> 00:07:51.959
just artificial scarcity. But but any
kind of goods and services I think will

90
00:07:53.000 --> 00:08:56.840
be extremely abundant. Well when when, when things are changing rapidly, the

91
00:08:56.960 --> 00:09:01.120
ability to predict the future I think
is, uh, it becomes a lot

92
00:09:01.200 --> 00:09:07.559
harder because of the rate of change
is so great. But I think some

93
00:09:07.600 --> 00:09:13.039
things are fairly obvious to predict,
which is that we'll have AI or HI

94
00:09:13.279 --> 00:09:16.120
that's at a level that it can
really do almost any cognitive I think really

95
00:09:16.120 --> 00:09:20.960
not almost really any cognitive task.
That's just a question of when one could

96
00:09:20.960 --> 00:09:24.679
debate, is it you know,
smarter than any human at the end of

97
00:09:24.759 --> 00:09:28.360
next year or is it two years
or three years, But it's not more

98
00:09:28.399 --> 00:09:33.799
than five years, that's for sure. So yeah, and I give prediction

99
00:09:33.840 --> 00:09:37.799
on predictions. Predictions I'm sort of
saying, giving predictions of the fiftieth percentile

100
00:09:37.840 --> 00:09:43.360
of probability. So not not not
like it will definitely happen, but if

101
00:09:43.399 --> 00:09:46.000
you say, what if you ask
me, like, what's the fiftieth percentile

102
00:09:48.039 --> 00:09:52.240
where it's like the you know,
you're kind of over under is kind of

103
00:09:52.279 --> 00:09:56.240
even that, that's where I why
why I think it's probably at the next

104
00:09:56.320 --> 00:10:03.080
year before AI can do better than
any individual human could do. And then

105
00:10:03.240 --> 00:10:07.480
but there's a it's a it's a
much higher bar to say, well,

106
00:10:07.240 --> 00:10:13.279
is this well than you know,
human intelligence selectively, But if the rate

107
00:10:13.320 --> 00:10:18.399
of change continues that that's why I
think probably twenty twenty nine or maybe twenty

108
00:10:18.399 --> 00:10:26.639
thirty is where digital intelligence will probably
exceed all human intelligence combined. And and

109
00:10:26.639 --> 00:10:30.039
then I think it's always help us
to look at these like fundamental ratios,

110
00:10:31.440 --> 00:10:37.799
you know, sort of physics first
principles approached to looking at things and and

111
00:10:35.399 --> 00:10:46.360
and probabilistic. Yeah, so yeah, it's probablistic. So the yeah,

112
00:10:46.440 --> 00:10:54.279
so if you look at the ratio
of digital to biological compute, so like

113
00:10:54.279 --> 00:10:58.759
like you know all of say,
all of the higher level cognitive If you

114
00:10:58.799 --> 00:11:05.360
sum up the higher level cognitive capacity
of humans and then what is the and

115
00:11:05.399 --> 00:11:09.960
think of that as the compute,
then well and then compare that to what

116
00:11:09.960 --> 00:11:16.720
what is the digital compute, and
the rate in which this is growing is

117
00:11:16.759 --> 00:11:22.279
just buggles the mind. So that's
why I think it's you know, I

118
00:11:22.320 --> 00:11:28.519
think twenty twenty nine or twenty thirty
or thereabouts is is it's not a that's

119
00:11:28.559 --> 00:11:35.679
I think a reasonable timeframe for where
you'd expect the cumulative digital compute to probably

120
00:11:35.720 --> 00:11:39.399
exceed the cumulative biological compute of higher
level brain functions