1
00:00:00,360 --> 00:00:03,277
You actually think AI's
under hyped. Why is that?

2
00:00:03,360 --> 00:00:06,797
>> And I'll tell I'll tell you why. Most
of you think of AI as I'll just use the

3
00:00:06,880 --> 00:00:10,677
general term as chat GPT. For most of you,
chat GPT was the moment where you

4
00:00:10,760 --> 00:00:14,237
said "Oh my god,
this thing writes." And it

5
00:00:14,320 --> 00:00:18,597
makes mistakes, but it's so brilliantly
verbal. Right? That was certainly my

6
00:00:18,680 --> 00:00:20,437
reaction. Most people
that I knew did that.

7
00:00:20,520 --> 00:00:24,077
>> It's visceral, yeah.
>> This was 2 years ago.

8
00:00:24,160 --> 00:00:28,037
Since then, the gains in what is is
called reinforcement learning, which is

9
00:00:28,120 --> 00:00:32,667
what AlphaGo helped invent and so
forth, allow us to do planning.

10
00:00:32,750 --> 00:00:32,997
>> [music]

11
00:00:33,080 --> 00:00:37,557
>> And a good example is look at OpenAI 03

12
00:00:37,640 --> 00:00:40,597
or DeepSeek R1, and you can see how it

13
00:00:40,680 --> 00:00:43,277
goes forward and back, forward and back,

14
00:00:43,360 --> 00:00:46,237
forward and back. It's extraordinary. In

15
00:00:46,320 --> 00:00:51,357
my case, I bought a rocket company because
it was like interesting, and I know

16
00:00:51,440 --> 00:00:54,957
>> As one does. >> As one does.
And it's it's an area that

17
00:00:55,040 --> 00:00:59,277
I'm not an expert in, and I want to be
an expert, so I'm using deep research.

18
00:00:59,360 --> 00:01:03,717
And these systems are spending 15
minutes writing these deep papers. This

19
00:01:03,800 --> 00:01:08,477
is true for most of them. Do you have
any idea how much computation 15 minutes

20
00:01:08,560 --> 00:01:12,757
of these supercomputers is?
It's extraordinary. So, you're seeing the

21
00:01:12,840 --> 00:01:16,717
arrival, the shift from language
to language, then you have language to

22
00:01:16,800 --> 00:01:18,917
sequence, which is how biology is done.

23
00:01:19,000 --> 00:01:23,077
Now you're [music] doing
essentially planning and strategy.

24
00:01:23,160 --> 00:01:27,597
The eventual state of this >> [music]
>> is the computers running all business

25
00:01:27,680 --> 00:01:31,157
processes, right? So, you have an agent
to do this, an agent to do this, an

26
00:01:31,240 --> 00:01:34,837
agent to do this, an agent to do this,
and you concatenate them together, and

27
00:01:34,920 --> 00:01:38,277
they speak language >> [music] >>
among each other. They typically speak

28
00:01:38,360 --> 00:01:43,157
English language. >> I mean,
speaking of of just the sheer

29
00:01:43,240 --> 00:01:47,277
compute requirements of these systems,
[music] let's talk about scale briefly.

30
00:01:47,360 --> 00:01:50,357
You know, I kind of think of these AI
systems as Hungry Hungry Hippos. They

31
00:01:50,440 --> 00:01:54,237
seemingly soak up all the data
and compute that we throw at them. They've

32
00:01:54,320 --> 00:01:58,517
already digested all the tokens on the
public internet, and it seems we can't

33
00:01:58,600 --> 00:02:01,237
build data centers fast enough.

34
00:02:01,320 --> 00:02:05,797
What do you think the real limits are,
and how do we get ahead of them before

35
00:02:05,880 --> 00:02:08,117
they start throttling AI progress?

36
00:02:08,200 --> 00:02:09,917
>> So, there's a real
limit in energy. Give

37
00:02:10,000 --> 00:02:13,877
an example, there's one calculation,
and I testified on this this week in

38
00:02:13,960 --> 00:02:15,197
Congress,

39
00:02:15,280 --> 00:02:18,237
that we need another 90 gigawatts of

40
00:02:18,320 --> 00:02:22,277
power in America. My answer, by the way,

41
00:02:22,360 --> 00:02:27,277
is think Canada, right? Nice people,
full of hydroelectric power, but that's

42
00:02:27,360 --> 00:02:30,157
apparently not
the political mood right now.

43
00:02:30,240 --> 00:02:31,077
Sorry.

44
00:02:31,160 --> 00:02:35,157
So, 90 gigawatts is 90 nuclear power

45
00:02:35,240 --> 00:02:39,277
[music] plants in America.
Not happening. We're building zero,

46
00:02:39,360 --> 00:02:42,957
right? How are we going to get all that
power? This is a major, major national

47
00:02:43,040 --> 00:02:45,997
issue. Um you can use
the Arab world, which is

48
00:02:46,080 --> 00:02:50,997
busy building 5 to 10 gigawatts of data
centers. India is considering a 10

49
00:02:51,080 --> 00:02:52,637
gigawatt data center.

50
00:02:52,720 --> 00:02:55,237
To understand how big gigawatts are, is

51
00:02:55,320 --> 00:02:58,437
think cities per data center. That's how

52
00:02:58,520 --> 00:03:03,517
much power these things need. And the
people look at it, and they say, "Well,

53
00:03:03,600 --> 00:03:06,557
there's lots of algorithmic
improvements, and you will need less

54
00:03:06,640 --> 00:03:10,397
power." There's an old rule,
I'm old enough to remember,

55
00:03:10,480 --> 00:03:14,457
right? Grove giveth, Gates taketh away.

56
00:03:14,540 --> 00:03:17,317
>> [laughter] >> Okay?
The hardware just gets faster and

57
00:03:17,400 --> 00:03:20,157
faster. The physicists are amazing.

58
00:03:20,240 --> 00:03:23,957
Just incredible what they've been able
to do. And us software people, we just

59
00:03:24,040 --> 00:03:26,237
use it, and use it, and use it.

60
00:03:26,320 --> 00:03:30,757
And when you look at planning, to at
least in today's algorithms, it's back

61
00:03:30,840 --> 00:03:35,037
and forth, and try this, and try And just
watch it yourself. Their estimates,

62
00:03:35,120 --> 00:03:39,597
and you know this from from Industry
Horowitz reports, uh it's been well

63
00:03:39,680 --> 00:03:42,557
studied, [music] that there's
an increase in at least a

64
00:03:42,640 --> 00:03:46,837
factor of a hundred, maybe a factor of
thousand in computation required just to

65
00:03:46,920 --> 00:03:48,557
do the kind [music] of planning.

66
00:03:48,640 --> 00:03:51,997
There the technology goes
from essentially deep learning to

67
00:03:52,080 --> 00:03:55,357
reinforcement learning to something
called test time compute, where not only

68
00:03:55,440 --> 00:03:58,917
are you doing planning, but you're also
learning while you're doing planning.

69
00:03:59,000 --> 00:04:01,157
That is the, if you will, the zenith or

70
00:04:01,240 --> 00:04:03,917
what have you, of computation needs.

71
00:04:04,000 --> 00:04:09,317
That's problem number one, electricity
and hardware. Problem number two is we

72
00:04:09,400 --> 00:04:12,837
ran out of data, so we
have to start generating it,

73
00:04:12,920 --> 00:04:15,957
but we can easily do that because
that's one of the functions.

74
00:04:16,040 --> 00:04:19,757
And then the third question that I
don't understand is what's the limit of

75
00:04:19,840 --> 00:04:22,357
knowledge? Um I'll I'll
give you an example. Let's

76
00:04:22,440 --> 00:04:26,437
imagine we are collectively all of the
computers in the world, and we're all

77
00:04:26,520 --> 00:04:29,877
thinking, and we're all thinking
based on knowledge that exists that was

78
00:04:29,960 --> 00:04:32,397
previously invented. How do we invent

79
00:04:32,480 --> 00:04:35,517
something completely [music] new?

80
00:04:35,600 --> 00:04:39,317
So, Einstein. So, when you
study the way scientific

81
00:04:39,400 --> 00:04:42,637
discovery works, biology,
math, so forth and so on,

82
00:04:42,720 --> 00:04:48,157
what typically happens is a truly
brilliant human being looks at one area

83
00:04:48,240 --> 00:04:52,237
and says, "I see a pattern
that's in a completely

84
00:04:52,320 --> 00:04:55,957
different area. Has nothing to do with
the first one. It's the same pattern."

85
00:04:56,040 --> 00:04:58,197
And they take the tools from one, and

86
00:04:58,280 --> 00:05:00,437
they apply it to another. Today, our

87
00:05:00,520 --> 00:05:03,117
systems cannot do that. If we can get

88
00:05:03,200 --> 00:05:05,277
through that, um I'm working on this.

89
00:05:05,360 --> 00:05:07,677
It's a general technical term for this

90
00:05:07,760 --> 00:05:12,317
is non-stationarity of objectives.
The rules are keep changing.

91
00:05:12,400 --> 00:05:15,597
Um we will see if we can solve that
problem. If we can solve that, we're

92
00:05:15,680 --> 00:05:17,957
going to need even more data centers,

93
00:05:18,040 --> 00:05:22,717
and we'll also be able to invent
completely new schools of scientific and

94
00:05:22,800 --> 00:05:25,920
intellectual [music] thought.