1
00:00:00,640 --> 00:00:03,279
Speaker 1: Thank you so much, Lon for being here at build.

2
00:00:03,520 --> 00:00:06,759
I know you started off as an intern at Microsoft,

3
00:00:06,799 --> 00:00:08,880
you were a Windows developer, and of course you're a

4
00:00:08,880 --> 00:00:11,839
big PC gamer. Still, you want to just talk about

5
00:00:11,880 --> 00:00:15,519
you in your early days with Windows and the kinds

6
00:00:15,519 --> 00:00:16,440
of things you build.

7
00:00:17,280 --> 00:00:20,800
Speaker 2: Yeah, well, actually started before Windows with us. I had

8
00:00:20,839 --> 00:00:25,480
one of the early IBM PCs with MS DOS and

9
00:00:26,960 --> 00:00:28,559
I think I had like a one hundred and twenty eight

10
00:00:28,640 --> 00:00:30,480
K in the beginning, and then at double T two

11
00:00:30,519 --> 00:00:34,280
fifty six K, which felt like a liar, so I yeah,

12
00:00:34,560 --> 00:00:41,960
pro programmed video games indus and then in Windows remember

13
00:00:41,960 --> 00:00:43,200
Windows three point one.

14
00:00:43,799 --> 00:00:46,759
Speaker 1: Yeah, no, it's wonderful. I mean even at the last

15
00:00:46,799 --> 00:00:49,280
time I chatted with you, you were talking all about everything,

16
00:00:49,439 --> 00:00:53,799
the intricacies of active directory, and so it's fantastic to

17
00:00:53,920 --> 00:00:56,640
have you at our developer conference. Obviously, the exciting thing

18
00:00:56,719 --> 00:00:59,880
for us is to be able to launch grock on Azure.

19
00:01:00,159 --> 00:01:03,240
I know you have a deep vision for what AI

20
00:01:03,359 --> 00:01:05,519
needs to be, and that's what got you to get

21
00:01:05,519 --> 00:01:08,239
this billed. It's a family of models that are both

22
00:01:08,359 --> 00:01:11,640
response and reasoning models, and you have a very exciting roadmap.

23
00:01:11,680 --> 00:01:13,120
You want to just tell us a little bit about

24
00:01:13,159 --> 00:01:16,200
sort of your vision the capability you're pushing on both

25
00:01:16,319 --> 00:01:19,280
capability and efficiency, So maybe you can just talk about

26
00:01:19,280 --> 00:01:19,879
a little bit of that.

27
00:01:20,359 --> 00:01:23,640
Speaker 2: Sure. So yeah, with with GROC, especially with Rock three

28
00:01:23,640 --> 00:01:27,400
point five that is about to be released, it's it's

29
00:01:27,439 --> 00:01:30,200
trying to reason from first principles, so apply kind of

30
00:01:30,239 --> 00:01:35,920
the the tools of physics to thinking. So if you're

31
00:01:35,920 --> 00:01:38,760
trying to get to fundamental truths, you you try you

32
00:01:39,159 --> 00:01:41,920
boil things down to the axiomatic elements that are most

33
00:01:41,959 --> 00:01:44,359
likely to be correct, and then you reason up from there,

34
00:01:44,400 --> 00:01:47,519
and then you can test your conclusions against those axiomatic elements.

35
00:01:47,799 --> 00:01:50,680
And you know, in physics, if if you violate conservation

36
00:01:50,680 --> 00:01:53,760
of energy or momentum, then you're either going to get

37
00:01:53,760 --> 00:01:56,480
a Nobel price or you're you're wrong, and you're certainly

38
00:01:56,519 --> 00:02:01,680
wrong basically. So so the the that's really the focus

39
00:02:01,680 --> 00:02:05,879
of Rock three point five is uh, sort of I

40
00:02:05,959 --> 00:02:10,759
find a fundamental physics and applying physics tools across all

41
00:02:10,800 --> 00:02:16,680
lines of reasoning and to aspire to truth with minimal error.

42
00:02:16,759 --> 00:02:19,080
Like there's always gonna be some mistakes that are made,

43
00:02:19,759 --> 00:02:25,360
but we aim to get to truth with acknowledged error,

44
00:02:25,680 --> 00:02:29,120
but minimize that error over time, and I think that's

45
00:02:29,120 --> 00:02:33,719
actually extremely important for AI safety. So I've put a

46
00:02:33,719 --> 00:02:36,439
lot for a long time about AI safety and wealth.

47
00:02:36,479 --> 00:02:39,960
Book conclusion is the old maximum that honesty is the

48
00:02:39,960 --> 00:02:45,280
best policy. It really really is for safety. But any

49
00:02:45,280 --> 00:02:48,000
want to have size. You know, we we haven't will

50
00:02:48,000 --> 00:02:50,439
make mistakes, but we aspire to correct them very quickly,

51
00:02:51,199 --> 00:02:53,599
and we are very much looking forward to feedback from

52
00:02:53,639 --> 00:02:56,319
the developer community to say like, what do you need?

53
00:02:56,439 --> 00:02:59,319
Where are we wrong? Cocker, we make it better, and

54
00:02:59,400 --> 00:03:02,599
to have Rock be something that the developer community community

55
00:03:04,039 --> 00:03:06,520
is very excited to use and where they can feel

56
00:03:06,560 --> 00:03:11,599
that their feedback is being heard and GROCK is improving

57
00:03:12,159 --> 00:03:13,439
and serving their need.

58
00:03:13,840 --> 00:03:16,159
Speaker 1: Yeah, I know it's in some sense, you know, cracking

59
00:03:16,280 --> 00:03:21,199
the physics of intelligence is perhaps the real goal for

60
00:03:21,360 --> 00:03:24,560
us to be able to use AI at scale, and

61
00:03:24,599 --> 00:03:26,639
so it's so good to you know, take that first

62
00:03:26,639 --> 00:03:29,240
principle's approach that you and your team are taking. And

63
00:03:29,319 --> 00:03:31,360
also you're deploying this. I mean one of the things

64
00:03:31,360 --> 00:03:34,199
about sort of what you do is you're doing you know,

65
00:03:34,560 --> 00:03:38,280
unsupervised FSD on one side, you're doing robotics. Of course

66
00:03:38,319 --> 00:03:41,599
there's Rock. You're deploying Rock across all of your businesses,

67
00:03:41,639 --> 00:03:45,319
from SpaceX to Tesla. Obviously at X, I would love

68
00:03:45,360 --> 00:03:47,319
to even you know, one of the themes for this

69
00:03:47,439 --> 00:03:52,319
developer conference, Elan is we're building pretty sophisticated AI apps. Right,

70
00:03:52,319 --> 00:03:55,000
It's not even about any one model. It's about orchestrating

71
00:03:55,080 --> 00:04:00,159
multiple models, multiple agents, just anything that you're seeing in

72
00:04:00,199 --> 00:04:04,080
the real world application side, even inside of your own companies.

73
00:04:04,120 --> 00:04:06,080
When you think about even a Tesla or a space

74
00:04:06,280 --> 00:04:09,000
X where you put grock and needs the other AI

75
00:04:09,080 --> 00:04:10,400
models you're building.

76
00:04:10,280 --> 00:04:14,319
Speaker 2: Yeah, it's incredibly important for a model to be grounded

77
00:04:14,319 --> 00:04:19,360
in reality reality. You know, I was saying, which is

78
00:04:19,360 --> 00:04:20,959
like like physics is the low and everything else is

79
00:04:20,959 --> 00:04:23,600
the recommendation. Which is I'm not suggesting people break but

80
00:04:23,680 --> 00:04:27,800
the laws made by you know, humans. Uh, you know,

81
00:04:27,879 --> 00:04:31,000
we should generally avail laws of humans. But but I've

82
00:04:31,040 --> 00:04:34,000
seen many people break human made laws, but I have

83
00:04:34,040 --> 00:04:38,360
not seen anyone break the loads of physics. So for

84
00:04:38,560 --> 00:04:43,079
for any given AI grounding it against reality and reality.

85
00:04:43,160 --> 00:04:45,319
For example, as you mentioned with with the car, it

86
00:04:45,360 --> 00:04:49,439
needs to drive safely and correctly. Uh, the human raid

87
00:04:49,560 --> 00:04:53,519
robot optimist needs to perform the task that that that

88
00:04:53,560 --> 00:04:57,800
it's being asked to perform these These are things that

89
00:04:57,800 --> 00:05:01,079
are very very helpful for issuing that the model is

90
00:05:01,959 --> 00:05:04,879
crucial and accurate because it has to adhere to the

91
00:05:04,879 --> 00:05:08,079
loads of physics. So so I think that's actually maybe

92
00:05:08,879 --> 00:05:11,519
so some somewhat overlooked or at least not talked about

93
00:05:11,560 --> 00:05:15,160
it enough. Is that to really be intelligent, it's it's

94
00:05:15,399 --> 00:05:18,279
got to make predictions that are in line with reality.

95
00:05:18,519 --> 00:05:21,399
In other words, physics. That's that's it's a really fund

96
00:05:21,399 --> 00:05:25,480
metal thing and and being able to ground that with

97
00:05:25,639 --> 00:05:29,639
the cars and robots is very important. We are seeing

98
00:05:29,839 --> 00:05:33,839
Grock be very helpful in things like customer service, and

99
00:05:34,399 --> 00:05:37,160
you know that the AI is infinitely patient and apparently

100
00:05:37,199 --> 00:05:38,639
and you can yell at it and it's still going

101
00:05:38,680 --> 00:05:43,839
to be very nice. Uh So that's good. Yeah, And

102
00:05:43,920 --> 00:05:47,879
so so I think in terms of improving the quality

103
00:05:47,920 --> 00:05:52,839
of customer service and sort of issue resolution, AI's were

104
00:05:52,920 --> 00:05:55,879
already uh Grock is already doing quite a good job

105
00:05:55,920 --> 00:05:59,759
that at SpaceX and Tesla, and and we look forward

106
00:05:59,800 --> 00:06:01,680
to like offering that to other companies.

107
00:06:02,040 --> 00:06:06,120
Speaker 1: No, that's fantastic, Really thrilled to get this journey started,

108
00:06:06,199 --> 00:06:09,759
getting that developer feedback and then looking forward to even

109
00:06:09,800 --> 00:06:12,800
how they deployed. There is these language models, there's you know,

110
00:06:13,120 --> 00:06:15,560
I think over time we will have this coming together

111
00:06:15,680 --> 00:06:18,920
of language models with vision with action, but to your point,

112
00:06:19,000 --> 00:06:22,360
being really grounded on a real world model, and that

113
00:06:22,399 --> 00:06:25,000
I think is ultimately the goal here. And so thank

114
00:06:25,040 --> 00:06:27,800
you so much Eland for briefly joining us today. And

115
00:06:27,800 --> 00:06:30,199
we're really excited about working with you and getting this

116
00:06:30,240 --> 00:06:31,279
into the developers' hands.

117
00:06:31,639 --> 00:06:33,639
Speaker 2: Thank you, thank you very much. And I can't empsize

118
00:06:33,720 --> 00:06:36,279
enough that we're looking for feedback from you that develop

119
00:06:36,319 --> 00:06:39,399
an audience. Tell us what you want and we'll make

120
00:06:39,399 --> 00:06:40,120
it happen. Thank you,