WEBVTT
1
00:00:00.000 --> 00:00:03.120
Today I'm excited to talk to a Praveen Murugasan.
2
00:00:03.120 --> 00:00:08.600
His team at Samsara processes something like 25 trillion data points
3
00:00:08.600 --> 00:00:12.360
across 90 billion miles of road every year.
4
00:00:12.360 --> 00:00:13.840
What I think is really interesting,
5
00:00:13.840 --> 00:00:17.280
other than the scale at which he's doing this Samsara is
6
00:00:17.280 --> 00:00:21.400
a publicly traded company at Fortune ranks at number seven on the future 50,
7
00:00:21.400 --> 00:00:26.800
is how they are both approaching AI from making themselves more efficient as a company,
8
00:00:26.800 --> 00:00:30.240
how they're experimenting, how they're building products faster,
9
00:00:30.240 --> 00:00:33.520
but also how they can then roll that out to their customers
10
00:00:33.520 --> 00:00:35.520
and help us all stay safe on the roads.
11
00:00:45.760 --> 00:00:48.920
What is your most contrary thought related to AI?
12
00:00:48.920 --> 00:00:50.520
If you were just going to take a step back and say,
13
00:00:50.520 --> 00:00:52.920
here's what I believe that nobody else does.
14
00:00:52.920 --> 00:00:57.520
I'll tell you, my fascination recently has been on computer use,
15
00:00:57.520 --> 00:00:59.880
AI driven computer use.
16
00:00:59.880 --> 00:01:02.800
I've been starting to use cloud co-work.
17
00:01:02.800 --> 00:01:06.600
When you think about it, like historically enterprise software,
18
00:01:06.600 --> 00:01:11.000
even doing any workflow or task was defined by the rules of the game of structure,
19
00:01:11.000 --> 00:01:15.120
meaning you have APIs, you have, like even MCP servers,
20
00:01:15.120 --> 00:01:17.880
which try to define some structure, etc.
21
00:01:18.600 --> 00:01:20.440
I think with computer use,
22
00:01:20.440 --> 00:01:25.320
essentially enabling any sort of human tool available without structure
23
00:01:25.320 --> 00:01:28.960
for a reasoning model to make decisions,
24
00:01:28.960 --> 00:01:30.440
I thought was quite interesting.
25
00:01:30.440 --> 00:01:32.640
So one of the things I'm curious about is like,
26
00:01:32.640 --> 00:01:34.960
or at least something that I'm actively thinking of is like,
27
00:01:34.960 --> 00:01:38.200
how do we apply that into our business?
28
00:01:38.200 --> 00:01:40.320
And also from a customer centric lens,
29
00:01:40.320 --> 00:01:45.520
because like a lot of people are working on very old operational tools,
30
00:01:45.520 --> 00:01:49.000
which a lot of people will tell you it's very hard to move.
31
00:01:49.000 --> 00:01:52.400
Like the mode is actually like, I have 100 integrations.
32
00:01:52.400 --> 00:01:54.760
Need a move is going to be like a big project,
33
00:01:54.760 --> 00:02:00.280
migrations becoming like a pretty big effort for most people.
34
00:02:00.280 --> 00:02:06.600
I think there'll be a radical shift in companies that are able to leverage
35
00:02:06.600 --> 00:02:11.960
newer technology and actually like keep up to find more defensible modes.
36
00:02:11.960 --> 00:02:14.480
So computer use actually, I think, opens up the game.
37
00:02:14.520 --> 00:02:19.920
So I think a lot of the legacy players who've been like essentially
38
00:02:19.920 --> 00:02:23.840
treating the cost of integrations or like the complexity being the mode
39
00:02:23.840 --> 00:02:25.560
should be really worried at the moment, right?
40
00:02:25.560 --> 00:02:27.960
So I feel like in the next 12 months,
41
00:02:27.960 --> 00:02:30.400
that's going to be changing quite rapidly.
42
00:02:30.400 --> 00:02:32.680
And do you think that will happen on an enterprise level
43
00:02:32.680 --> 00:02:37.360
or even individuals given their new tool sets available
44
00:02:37.360 --> 00:02:41.800
will be able to begin taking part in that themselves?
45
00:02:41.800 --> 00:02:43.400
I think both actually.
46
00:02:43.440 --> 00:02:45.440
This is something I've been thinking about recently,
47
00:02:45.440 --> 00:02:51.080
where every business is right now thinking through a lens of transformation
48
00:02:51.080 --> 00:02:55.920
of their workforce in terms of how do you supercharge them with AI tools,
49
00:02:55.920 --> 00:02:56.720
right?
50
00:02:56.720 --> 00:03:01.240
But for me in engineering, this is something I think about all the time,
51
00:03:01.240 --> 00:03:04.560
meaning we have all these AI coding tools.
52
00:03:04.560 --> 00:03:07.840
How do we enable our engineers to be more productive?
53
00:03:07.840 --> 00:03:12.040
But then we still don't want to do something crazy, right?
54
00:03:12.040 --> 00:03:15.200
Like I mean, there's also lots of stories you read now about
55
00:03:15.200 --> 00:03:17.720
like production DBs, getting deleted, et cetera.
56
00:03:17.720 --> 00:03:22.280
So with some guardrails, but how do you make people supercharge?
57
00:03:22.280 --> 00:03:24.840
So that's a common problem.
58
00:03:24.840 --> 00:03:28.040
Now, what I found fascinating with computer use was like,
59
00:03:28.040 --> 00:03:31.760
I've been, I'm still more recently playing with it.
60
00:03:31.760 --> 00:03:36.000
I used it to make a deck for like one of my leadership upside
61
00:03:36.000 --> 00:03:39.960
that like, you know, what would have taken me potentially like
62
00:03:40.240 --> 00:03:42.880
about 10 hours, took me like three,
63
00:03:42.880 --> 00:03:47.120
but then I later realized a lot of people we sell to in enterprise space
64
00:03:47.120 --> 00:03:50.000
or even anyone who's actually working in front of a computer, right?
65
00:03:50.000 --> 00:03:54.240
Like think about any task you do from booking a flight ticket to anything
66
00:03:54.240 --> 00:03:57.600
that you do, actually sort of felt like it's surprisingly
67
00:03:57.600 --> 00:04:03.680
every task you do was made easier with time back to the individual, right?
68
00:04:03.680 --> 00:04:06.480
This becomes more of not only like an enterprise,
69
00:04:06.480 --> 00:04:10.480
but you play enterprise definitely because you have some of these big companies
70
00:04:10.480 --> 00:04:13.680
like which has like hundreds of thousands of employees.
71
00:04:13.680 --> 00:04:18.640
And I often wonder like, what does that look like in a couple of years, right?
72
00:04:18.640 --> 00:04:22.880
And I would think that has to radically transform quickly, right?
73
00:04:22.880 --> 00:04:25.960
So people have to go and like, you have to really rethink.
74
00:04:25.960 --> 00:04:30.960
As you begin to think about it from even your teams and how you're rolling it out
75
00:04:30.960 --> 00:04:35.480
because you've got all of that, you know, as well,
76
00:04:35.480 --> 00:04:39.360
how have you empowered the team and what guardrails have you put?
77
00:04:39.360 --> 00:04:42.920
And sort of as everyone just on like a cursor, for example,
78
00:04:42.920 --> 00:04:46.280
and doing whatever they want or, you know,
79
00:04:46.280 --> 00:04:49.040
then Claude comes out and it's like, should we move over there?
80
00:04:49.040 --> 00:04:52.040
Like how do you even think about the stack today?
81
00:04:52.040 --> 00:04:54.720
We always thought about a little bit progressive.
82
00:04:54.720 --> 00:04:59.120
I would say at some sort of like I speak to a lot of industry friends
83
00:04:59.120 --> 00:05:01.800
and I realize people are still like adopting these tools slower.
84
00:05:01.800 --> 00:05:03.800
I think we've been a little bit more progressive.
85
00:05:03.800 --> 00:05:07.040
I think the largely the way we've approached the problem was
86
00:05:07.040 --> 00:05:11.640
enable people to get the tools they want and also say that's okay to go experiment, right?
87
00:05:11.640 --> 00:05:17.440
So I think letting people try things has been a good approach.
88
00:05:17.440 --> 00:05:22.200
We've also established dedicated forums where people or engineers
89
00:05:22.200 --> 00:05:27.000
who are trying new things would come back and talk about it to the broader org.
90
00:05:27.000 --> 00:05:32.080
Now we've gotten more structured about it by collapsing all of that into like
91
00:05:32.080 --> 00:05:36.360
a little bit of a learning curriculum for new people that are starting in the organization.
92
00:05:36.360 --> 00:05:42.000
So we've always been a bit of like a lot more open versus like not necessarily like,
93
00:05:42.000 --> 00:05:46.000
you know, start with objections as the starting point, right?
94
00:05:46.000 --> 00:05:49.560
So it's culturally I think we've been a lot more open in general,
95
00:05:49.560 --> 00:05:51.080
encouraging people.
96
00:05:51.080 --> 00:05:55.560
On the other side, I think from a God rail perspective,
97
00:05:55.560 --> 00:05:59.600
we've been a bit cautious about our life cycle,
98
00:05:59.600 --> 00:06:04.560
especially with respect to programming or like shipping new models, et cetera,
99
00:06:04.560 --> 00:06:06.800
even for our customers.
100
00:06:06.800 --> 00:06:11.280
It's almost think about using some processes as a bit of a God rail,
101
00:06:11.280 --> 00:06:16.880
but then also like relax them as we learn and adapt to the new AI tool, right?
102
00:06:16.880 --> 00:06:19.720
Like a few examples would be obviously like, you know,
103
00:06:19.720 --> 00:06:25.280
we still walk away talk about this where when an AI is authoring code
104
00:06:25.280 --> 00:06:28.000
that doesn't mean the accountability is lost, right?
105
00:06:28.000 --> 00:06:32.280
Like the accountability still stays with the person who's actually like shipping the PR, right?
106
00:06:32.280 --> 00:06:34.400
And the person who's reviewing the PR, right?
107
00:06:34.400 --> 00:06:37.040
We also recognize not all code creator is equal.
108
00:06:37.040 --> 00:06:41.760
So we want to enable people to move faster in some smaller changes.
109
00:06:41.760 --> 00:06:49.280
But then we've now built capabilities to score changes to almost like identify risk, right?
110
00:06:49.280 --> 00:06:53.600
So it's the governance model or the God rail that we're putting and learning ourselves,
111
00:06:53.600 --> 00:07:00.880
where like, you know, I mean, like modeling risk correctly and enabling people to go faster, right?
112
00:07:00.880 --> 00:07:06.080
So that's been a big shift, I would say, right?
113
00:07:06.080 --> 00:07:08.880
I don't think we fully crack the code of,
114
00:07:08.880 --> 00:07:14.560
or we have the perfect governance model, but I think the attitude towards approaching this problem
115
00:07:14.560 --> 00:07:20.080
is one, let's enable people to try and experiment and then ensure people have good intent.
116
00:07:20.080 --> 00:07:24.560
Let's not relax processes to the point where we just go fully open.
117
00:07:24.560 --> 00:07:28.720
Let's make sure because we have a duty to our customers to ship good quality products
118
00:07:28.720 --> 00:07:31.200
and we need to make sure we continue to do that.
119
00:07:31.200 --> 00:07:35.840
So we have internal metrics like any business in terms of understanding or releases
120
00:07:35.840 --> 00:07:38.320
and the quality of our products.
121
00:07:38.320 --> 00:07:44.680
We've been monitoring those plus also like slowly relaxing or changing our controls to ensure
122
00:07:44.680 --> 00:07:47.960
like we also enable developer productivity, but without compromising quality.
123
00:07:47.960 --> 00:07:49.080
That's what we think about it.
124
00:07:49.080 --> 00:07:52.840
Like what are some of the wins that you've seen as part of that has been
125
00:07:52.840 --> 00:07:59.240
have you seen quality of code go up because or at least consistency of it go up and speed
126
00:07:59.240 --> 00:08:02.920
or is it not that easy to to point to?
127
00:08:02.920 --> 00:08:07.720
Speed of development has gone up, like that's like it's undoubtedly, I would say that, right?
128
00:08:07.720 --> 00:08:16.280
So the maybe a couple of things, like one is we have definitely started doing things
129
00:08:16.280 --> 00:08:21.000
with a fail fast mindset, like historically you would discuss about like whether an idea
130
00:08:21.000 --> 00:08:25.080
is worth to pursue, you'd have a lot of meetings.
131
00:08:25.080 --> 00:08:27.480
Yeah, yeah, yeah, sure, right?
132
00:08:27.480 --> 00:08:33.520
So right now the at least what culturally what we try to tell people is like actually you
133
00:08:33.520 --> 00:08:37.680
have a good idea rather than talk about the idea maybe like build a quick prototype and show
134
00:08:37.680 --> 00:08:38.680
me.
135
00:08:38.680 --> 00:08:42.680
We've seen that it enables the quick follow on conversations about okay is this worth
136
00:08:42.680 --> 00:08:47.880
pursuing should we build this as a product is there going to market for this or why should
137
00:08:47.880 --> 00:08:50.080
we pursue this versus something else or not, right?
138
00:08:50.080 --> 00:08:53.560
So going to prototypes has been super simple, right?
139
00:08:53.560 --> 00:08:55.360
So that's one.
140
00:08:55.360 --> 00:09:01.400
The other dimension I'm also seeing is we are actually addressing more product quality issues
141
00:09:01.400 --> 00:09:04.680
faster than what we would have done in the past.
142
00:09:04.680 --> 00:09:09.560
So any business would like, you know, if you're operating software you're going to have
143
00:09:09.560 --> 00:09:13.760
like some degree of support issues and bugs that come in, but usually people will always
144
00:09:13.760 --> 00:09:18.880
categorize and say I'm going to solve my P1s like P2 is well, we'll get to it, right?
145
00:09:18.880 --> 00:09:25.760
So I think the nice part now is we enabled the capability of people are able to quickly
146
00:09:25.760 --> 00:09:28.360
see P2s come in, oh, this is actually easy.
147
00:09:28.360 --> 00:09:34.480
I can just land this wire one of my agents while I work on this other thing.
148
00:09:34.480 --> 00:09:40.800
So that's become a bit of a norm, like I was seeing one of our engineers work with four
149
00:09:40.800 --> 00:09:48.640
different course sessions on four different Mac which will destops, each of them spawn
150
00:09:48.640 --> 00:09:51.440
to run like, you know, in different tasks.
151
00:09:51.440 --> 00:09:53.440
That's crazy.
152
00:09:53.440 --> 00:09:58.480
Yeah, so it's been pretty awesome actually, it's been very energizing, right?
153
00:09:58.480 --> 00:10:04.160
So that's kind of the transformation I see both along the lines of productivity, but
154
00:10:05.160 --> 00:10:11.840
can you talk about one of the customer-facing releases that you've done and how that was
155
00:10:11.840 --> 00:10:17.360
developed because you've got an engineering team which is probably really great at coding,
156
00:10:17.360 --> 00:10:20.960
you've got a product team which has a level of expertise, but by the time you get to the
157
00:10:20.960 --> 00:10:26.480
customer, I don't always imagine they know exactly what they want or the art of the possible,
158
00:10:26.480 --> 00:10:27.920
as you said.
159
00:10:27.920 --> 00:10:33.640
So, you know, what's a product that you released for them that typically wasn't something
160
00:10:33.640 --> 00:10:38.280
they might have even known that they could have or want, but because they just don't understand
161
00:10:38.280 --> 00:10:40.160
the future, the capability.
162
00:10:40.160 --> 00:10:44.440
In some markets that we operate, so in some sort of operates globally, we also operate
163
00:10:44.440 --> 00:10:49.880
in markets like Mexico, Canada, US and Europe, etc.
164
00:10:49.880 --> 00:10:57.120
In Mexico, and surprisingly in some places in Europe, like some customers of ours had told
165
00:10:57.120 --> 00:11:01.440
us that they have a bit of like fuel theft problem.
166
00:11:01.440 --> 00:11:07.640
It's actually an interesting case because, and especially now that fuel prices are super
167
00:11:07.640 --> 00:11:08.640
high, right?
168
00:11:08.640 --> 00:11:18.240
Like, you know, the problem effectively is we have a lot of like fuel theft by two different
169
00:11:18.240 --> 00:11:19.240
means.
170
00:11:19.240 --> 00:11:25.920
One is actually like people actually siphoning fuel off the vehicle and selling it happened
171
00:11:25.920 --> 00:11:31.480
surprisingly a lot, actually there were some studies done where this was in the hundreds
172
00:11:31.480 --> 00:11:34.920
of millions of dollars lost globally.
173
00:11:34.920 --> 00:11:39.840
And there's also another type of fraud that happens where people might use a fuel card that's
174
00:11:39.840 --> 00:11:42.320
issued to them, but then they fuel their personal vehicle, right?
175
00:11:42.320 --> 00:11:48.440
So, these are two different types of fuel products that actually, you know, these are problems
176
00:11:48.440 --> 00:11:49.440
that exist.
177
00:11:49.440 --> 00:11:52.480
So, from a, how do we build like psycho perspective?
178
00:11:52.480 --> 00:11:57.360
So, the way we do this is we always, some sort of culture has been more around like working
179
00:11:57.360 --> 00:12:00.440
with a customer because a lot of, when you, when you think about it and this goes all
180
00:12:00.440 --> 00:12:07.000
the way to our CEO Sanjit, like a lot of the companies, DNAs, we are all product builders.
181
00:12:07.000 --> 00:12:10.160
We are not experts in physical world operations, right?
182
00:12:10.160 --> 00:12:16.160
So you have to really lean in to your customers and relearn and understand from them who are
183
00:12:16.160 --> 00:12:18.360
the experts in their physical world problems, right?
184
00:12:18.360 --> 00:12:22.480
So we take those problems and then translate into building technology products, which is
185
00:12:22.480 --> 00:12:23.680
what we are good at.
186
00:12:23.680 --> 00:12:29.840
This actually came up as a problem that I spoke into a few customers and they had flagged
187
00:12:29.840 --> 00:12:31.320
it as a problem.
188
00:12:31.320 --> 00:12:34.560
And one of the immediate instinct was like, hey, this doesn't feel like a hard problem
189
00:12:34.560 --> 00:12:39.640
to solve because you get fuel level from every vehicle, right?
190
00:12:39.640 --> 00:12:42.160
It's, we have sensors installed on the vehicle.
191
00:12:42.160 --> 00:12:49.280
You can get the fuel level, you can over time run algorithms like to see if the fuel dropped
192
00:12:49.280 --> 00:12:51.400
and then like let's go solve it, right?
193
00:12:51.400 --> 00:12:57.080
So interestingly, like, you know, we've apparently tried doing that a while back and it just
194
00:12:57.080 --> 00:13:01.800
led to poor results, like very noisy alerts and issues, right?
195
00:13:01.800 --> 00:13:07.560
One of the things we had to do here was almost rethink the problem, friend approach it
196
00:13:07.560 --> 00:13:10.360
from like leveraging some degree of machine learning techniques.
197
00:13:10.360 --> 00:13:18.040
The problem itself was sensors are not reliable and they're not equal, right, across vehicles.
198
00:13:18.040 --> 00:13:22.720
So you would, if you start inspecting fuel level data, it's extremely noisy.
199
00:13:22.720 --> 00:13:29.400
So you can't distill real fuel levels quickly as much as what people would assume.
200
00:13:29.400 --> 00:13:36.320
And even extreme cases, because of this fuel sensor noise problem in markets like Mexico,
201
00:13:36.320 --> 00:13:41.680
they actually have a lot of companies selling fuel level sensors that you can install as
202
00:13:41.680 --> 00:13:45.200
an external thing on top of your tank, right?
203
00:13:45.200 --> 00:13:47.600
So people are trying to keep a pulse of it.
204
00:13:47.600 --> 00:13:51.880
We actually worked on building a model, but then the approach was we wanted to get quick
205
00:13:51.880 --> 00:13:52.880
feedback.
206
00:13:52.880 --> 00:13:54.080
That's just how we work, right?
207
00:13:54.080 --> 00:13:57.600
Like in terms of you, you can build the entire product and ship it to the market.
208
00:13:57.600 --> 00:14:01.440
But whereas what we did was we took like four or five customers who had come to us with
209
00:14:01.440 --> 00:14:03.960
this being a common problem.
210
00:14:03.960 --> 00:14:10.200
And we were initially just giving them algorithmic results in like spreadsheets, telling them
211
00:14:10.200 --> 00:14:14.520
to react to it, like, you know, does this feel like a useful signal for you?
212
00:14:14.520 --> 00:14:16.760
And what are the context can we add?
213
00:14:16.760 --> 00:14:22.040
Like is location interesting is like, you know, how long somebody was idling in a specific
214
00:14:22.040 --> 00:14:23.720
place before interesting?
215
00:14:23.720 --> 00:14:28.560
It's almost like a co iteration of the product with some of our customer feedback.
216
00:14:28.560 --> 00:14:35.080
And then we bring it into the market and like try to scale as we, like, and learn from
217
00:14:35.080 --> 00:14:36.080
that experience.
218
00:14:36.080 --> 00:14:40.080
So that's typically how we go about working on any sort of product.
219
00:14:40.080 --> 00:14:44.280
The other proxy signal we look for is usage or adoption.
220
00:14:44.280 --> 00:14:48.080
And this ties back to like today, the value to enter the building solutions is like so
221
00:14:48.080 --> 00:14:49.080
low.
222
00:14:49.080 --> 00:14:55.920
You could build like for you to find real value, like usage or like somebody willing
223
00:14:55.920 --> 00:14:58.680
to pay you stuff is like actually a good valuable signal.
224
00:14:58.680 --> 00:15:04.080
You're collecting all the data from all the sensors and then trying to come up with whether
225
00:15:04.080 --> 00:15:08.880
it's usage or theft based on the other metrics that's going on.
226
00:15:08.880 --> 00:15:10.440
And is that a real time?
227
00:15:10.440 --> 00:15:16.800
We have our some sort of sensor, which actually like you can hook into a vehicle's on board
228
00:15:16.800 --> 00:15:18.520
diagnostics.
229
00:15:18.520 --> 00:15:25.520
So we have access to pretty much the vehicles on board state in real time.
230
00:15:25.520 --> 00:15:32.280
Think about not only fuel level, like fuel consumption rate and like any signal you want
231
00:15:32.280 --> 00:15:36.320
engine state, like whether the engine's on or not, like whether the engine's moving,
232
00:15:36.320 --> 00:15:38.880
what's the automata state, etc.
233
00:15:38.880 --> 00:15:42.360
So you get access to all of that data in real time.
234
00:15:42.360 --> 00:15:48.840
And for this particular problem, we had to employ like a suite of algorithms.
235
00:15:48.840 --> 00:15:56.280
That was not effectively one because the problem's complexity is derived from a few different
236
00:15:56.280 --> 00:15:57.280
dimensions.
237
00:15:57.280 --> 00:16:02.360
One is the sensor gauge itself might not be precise.
238
00:16:02.360 --> 00:16:08.480
So you have to model based on the variance for distant vehicles.
239
00:16:08.480 --> 00:16:14.840
And then sometimes you have to fuse data from consumption to improve your sensor quality,
240
00:16:14.840 --> 00:16:19.000
like you also can take the sense of fuel consumption data.
241
00:16:19.000 --> 00:16:23.400
In some scenarios, not all vehicles have only one single fuel tank.
242
00:16:23.400 --> 00:16:26.960
They might have multiple tanks, so you have to accommodate for that.
243
00:16:26.960 --> 00:16:33.600
There's also scenarios where we try to predict out what is translating fuel level percentage
244
00:16:33.600 --> 00:16:39.120
into the actual volume, which actually helps you get more precise results.
245
00:16:39.120 --> 00:16:46.280
So this is just an example of a project that we worked on and a feature that's actually
246
00:16:46.280 --> 00:16:49.200
fairly successful in creating value for customers.
247
00:16:49.200 --> 00:16:52.080
All this exists in a pre-AI world.
248
00:16:52.080 --> 00:16:57.880
So when these models come out, how does it change how you fundamentally do the analysis
249
00:16:57.880 --> 00:16:59.120
and the business on that?
250
00:16:59.120 --> 00:17:04.200
How do you all think about picking a model to build on top of given the pace of change?
251
00:17:04.200 --> 00:17:10.000
So this specific algorithm, we built our own proprietary models to solve it.
252
00:17:10.000 --> 00:17:14.960
So the common question we got was like, why don't we just use an LLM to do this, give
253
00:17:14.960 --> 00:17:18.600
the samples of fuel traces.
254
00:17:18.600 --> 00:17:24.440
Language models largely were not able to accommodate for the sensor variance.
255
00:17:24.440 --> 00:17:28.200
This was almost like a noise filtering problem, so it's a very form-fitted problem.
256
00:17:28.200 --> 00:17:34.000
So it's not necessarily something that required a large language model, and it's like not
257
00:17:34.000 --> 00:17:37.720
something that's also the best suited model, I would say.
258
00:17:37.720 --> 00:17:42.560
So we didn't use a large language model in the AI model in that sense, so we actually
259
00:17:42.560 --> 00:17:49.080
built our own models to account for the noise, and how do we go about filtering.
260
00:17:49.080 --> 00:17:55.200
I would say, in this case, we've used AI in a little bit different lengths.
261
00:17:55.200 --> 00:18:00.040
One of the challenges when you think about this problem is, you need to still get samples
262
00:18:00.040 --> 00:18:07.080
from the real world, and you need to label them as like, yes, this was like actual sensor
263
00:18:07.080 --> 00:18:12.760
noise versus this was a level change that was unexpected, right?
264
00:18:12.760 --> 00:18:20.520
So we had to go do human labeling for a lot of this cases, and we significantly sped
265
00:18:20.520 --> 00:18:23.120
up that process by using LLMs, basically.
266
00:18:23.120 --> 00:18:29.560
So what we would do was use more like an LLM as a judge, and where we would have given
267
00:18:29.560 --> 00:18:37.600
the LLM original hand-labeled samples, and then leverage an LLM to go classify at scale
268
00:18:37.600 --> 00:18:38.880
like the results.
269
00:18:38.880 --> 00:18:41.400
How many data points are we talking about just in my sense?
270
00:18:41.400 --> 00:18:47.160
In general, the product runs across multiple millions of vehicles, and they're running
271
00:18:47.160 --> 00:18:54.080
across like on continuous stream of data across these vehicles when they are mobile, right?
272
00:18:54.080 --> 00:18:57.800
So we're talking easily the order of the trillions, right?
273
00:18:57.800 --> 00:19:01.840
Like when it comes down to individual data points that we work with, obviously the training
274
00:19:01.840 --> 00:19:05.280
set was a little bit more smaller.
275
00:19:05.280 --> 00:19:12.000
We would primarily pick a range and train based on different vehicles or like different
276
00:19:12.000 --> 00:19:15.360
make models, so we get a comprehensive set.
277
00:19:15.360 --> 00:19:17.080
So that's kind of how we go about it.
278
00:19:17.080 --> 00:19:21.600
Your own model is doing the analysis, but the actual data labeling of the 20 trillion
279
00:19:21.600 --> 00:19:24.200
or 25 trillion is coming in.
280
00:19:24.200 --> 00:19:30.000
Yeah, it's AI augmented when it comes down to for us to create the training set, but also
281
00:19:30.000 --> 00:19:31.800
like label the training set itself.
282
00:19:31.800 --> 00:19:36.640
So that's something that we've been successfully leveraging AI for.
283
00:19:36.640 --> 00:19:40.640
So you wrote a piece about vehicle routing, which I thought was really interesting.
284
00:19:40.720 --> 00:19:46.440
Look at a personal level, I hope, you know, I relied on ways forever to get around, but
285
00:19:46.440 --> 00:19:49.480
I think that's an individual driving one car.
286
00:19:49.480 --> 00:19:55.320
Where does AI come in from your guys perspective in terms of we can really help add massive
287
00:19:55.320 --> 00:20:00.200
value to our customers by helping them get somewhere faster and more safely?
288
00:20:00.200 --> 00:20:03.800
So one of the things we've been doing recently is we launched this product called commercial
289
00:20:03.800 --> 00:20:04.800
navigation, right?
290
00:20:05.080 --> 00:20:10.960
So it's sort of think about it for ways for enterprise users who are driving for commercial
291
00:20:10.960 --> 00:20:17.440
use, especially you might be driving trucks around and like your typical ways doesn't
292
00:20:17.440 --> 00:20:20.400
work for it directly, right?
293
00:20:20.400 --> 00:20:25.880
There are a few really cool applications that we are able to build, which like gives us
294
00:20:25.880 --> 00:20:28.600
a real leverage unlike like other players, right?
295
00:20:28.600 --> 00:20:32.920
So you said you're a user of ways, I've used ways actually and I thought some features
296
00:20:32.920 --> 00:20:34.720
were really cool in ways, right?
297
00:20:34.720 --> 00:20:41.080
So this, my understanding is fundamentally, it's primarily crowdshosed in terms of like
298
00:20:41.080 --> 00:20:43.080
intelligence that it plugs on.
299
00:20:43.080 --> 00:20:48.440
It could tell you whether there's a cop in some ways or it could tell you, I think when
300
00:20:48.440 --> 00:20:52.960
I was driving once it even told me that there's like a fallen tree on the road, right?
301
00:20:52.960 --> 00:20:54.680
Which I thought was pretty cool.
302
00:20:54.680 --> 00:21:02.880
Now what's possible for us is quite interesting because we have millions of vehicles with sensors,
303
00:21:02.880 --> 00:21:09.000
not only sensors that's connected to the cab from a getting diagnostic data perspective,
304
00:21:09.000 --> 00:21:13.080
but we have like a dual-facing camera that we also set, right?
305
00:21:13.080 --> 00:21:18.880
Now so we almost have like eyes on the road with which you can derive data like or build
306
00:21:18.880 --> 00:21:21.240
better intelligence into our products.
307
00:21:21.240 --> 00:21:28.000
So one of the things we've been doing is understanding exceptions or things that's happening
308
00:21:28.000 --> 00:21:34.520
in the real world and automatically flagging them and collecting data insights towards
309
00:21:34.520 --> 00:21:35.520
it.
310
00:21:35.520 --> 00:21:41.800
An example here would be this example that we talked about on like, let's say a tree
311
00:21:41.800 --> 00:21:44.360
fallen on the road, right?
312
00:21:44.360 --> 00:21:51.880
You do not need human input anymore to actually detect that scenario because like some, at
313
00:21:51.880 --> 00:21:59.160
some sort of like cameras like already recording it can see eyes on the road and we can program
314
00:21:59.160 --> 00:22:05.720
or like, you know, cameras or the edge to detect for certain scenarios and automatically
315
00:22:05.720 --> 00:22:12.000
capture that signal with evidence and so you set it up to the cloud, you can process
316
00:22:12.000 --> 00:22:15.920
it and you actually have many, many data points is not even like we're relying on like one
317
00:22:15.920 --> 00:22:18.640
instance because we have like millions of vehicles on the road.
318
00:22:18.640 --> 00:22:26.960
So you actually can almost run an auction for multiple data instances to agree upon and
319
00:22:26.960 --> 00:22:30.560
surface that back into the other customers, right?
320
00:22:30.560 --> 00:22:36.440
So if I might be driving towards that route, even before I drive the route, you could preemptively
321
00:22:36.440 --> 00:22:43.600
intervene and actually reroute for a person or like give people closest insights as they
322
00:22:43.600 --> 00:22:46.520
are using their navigation, hey, this is something you probably want to watch out on the
323
00:22:46.520 --> 00:22:47.520
road.
324
00:22:47.520 --> 00:22:52.840
We've really successfully done this for instances like construction detection, like for example,
325
00:22:52.840 --> 00:22:57.640
there's a road construction, we've been able to like capture some of the data around like,
326
00:22:57.640 --> 00:23:03.200
you know, automatically identifying this construction happening here and then that becomes an inside.
327
00:23:03.200 --> 00:23:10.400
We're now on the next phase working on how do we automatically like intervene when we have
328
00:23:10.400 --> 00:23:11.400
this context.
329
00:23:11.400 --> 00:23:16.120
A lot of this comes down to having the real world context and also the operational context
330
00:23:16.120 --> 00:23:21.160
for the particular user and you then get to a point of rather than just sharing it as an
331
00:23:21.160 --> 00:23:24.840
inside, start taking real world actions.
332
00:23:24.840 --> 00:23:29.960
Do you think over time you'll have enough data to get predictive on some of the stuff
333
00:23:29.960 --> 00:23:35.200
where you can say these are the patterns of this time of night, this sort of thing.
334
00:23:35.200 --> 00:23:38.160
Is it a weird place where that's possible soon?
335
00:23:38.160 --> 00:23:39.160
I think so.
336
00:23:39.160 --> 00:23:41.600
Yeah, I think we're definitely getting there.
337
00:23:41.600 --> 00:23:48.360
What I've seen from our customers is there is a degree of strategic planning that people
338
00:23:48.360 --> 00:23:53.720
try to do in terms of operations like I need to do deliveries across hundreds of my customers
339
00:23:53.720 --> 00:23:55.920
over the course of the next six months.
340
00:23:55.920 --> 00:23:57.560
How do I plan, right?
341
00:23:57.560 --> 00:24:03.560
So to define those, there's a heavy reliance on what potentially happened in history.
342
00:24:03.560 --> 00:24:09.000
If you're able to learn and derive the right insights, I think it can largely influence
343
00:24:09.000 --> 00:24:11.320
how you do strategic plans, right?
344
00:24:11.320 --> 00:24:16.120
Obviously, we talked a lot about efficiency in this context primarily from like, you know,
345
00:24:16.120 --> 00:24:21.880
hey, like we know obstructions on the road and how can you change routes, etc. from an
346
00:24:21.880 --> 00:24:22.880
efficiency angle.
347
00:24:22.880 --> 00:24:24.880
It doesn't even stop there.
348
00:24:24.880 --> 00:24:30.280
When you think about it, right, it actually extends to more ideas, like an example would
349
00:24:30.280 --> 00:24:33.640
be safety of different roads.
350
00:24:33.640 --> 00:24:36.360
Some roads are more prone to accidents.
351
00:24:36.360 --> 00:24:41.720
We have tons of data which actually can pinpoint where do accidents happen.
352
00:24:41.720 --> 00:24:47.800
We can score safety of individual street segments and even potentially take that into account
353
00:24:47.800 --> 00:24:50.200
when it comes down to route modeling, right?
354
00:24:50.200 --> 00:24:55.640
So there's an aspect of strategic planning that will heavily lean on historical data but
355
00:24:55.640 --> 00:24:58.400
also like historical modeling of data.
356
00:24:58.400 --> 00:25:04.280
So where does the sort of last five years of model development help you do that more efficiently
357
00:25:04.280 --> 00:25:10.040
or is that just, is it now something that you can present to the customer in a more accessible
358
00:25:10.040 --> 00:25:12.360
way, thanks to AI?
359
00:25:12.360 --> 00:25:16.040
A lot of the historical data always existed.
360
00:25:16.040 --> 00:25:21.600
For using it, you almost always use it in deterministic algorithms, right?
361
00:25:21.600 --> 00:25:27.400
Like imagine like even from a strategic planning perspective, a human has to sit through, let's
362
00:25:27.400 --> 00:25:31.920
say in the before the agentic revolution, right?
363
00:25:31.920 --> 00:25:37.760
The human has to take the insights of this is the safety profile of like the region you're
364
00:25:37.760 --> 00:25:39.800
trying to do routing to.
365
00:25:39.800 --> 00:25:43.080
This is probably the traffic profile.
366
00:25:43.080 --> 00:25:48.280
They have to look at all of that together and make some decisions, right?
367
00:25:48.280 --> 00:25:54.120
I think what's powerful right now with the agentic wave has been the ability to unlock
368
00:25:54.120 --> 00:25:55.760
a reason.
369
00:25:55.760 --> 00:26:01.120
And the model is actually, when you give this context, can actually give you a synthesis
370
00:26:01.120 --> 00:26:04.880
of like, you know, this is how I reason about it and this is what you need to choose.
371
00:26:04.880 --> 00:26:10.600
If you're able to give a more higher order goal or an abstraction to say, my goal is to
372
00:26:10.600 --> 00:26:17.040
still do all these deliverables, deliveries, maybe my risk profile for safety is this.
373
00:26:17.040 --> 00:26:18.960
Is it possible or not?
374
00:26:18.960 --> 00:26:22.720
Can what kind of like a pastry should I need to add, right?
375
00:26:22.720 --> 00:26:27.400
So you're operating at a much higher level and you're kind of making that person who is
376
00:26:27.400 --> 00:26:33.240
doing that work or like defining that final structure or strategic plan.
377
00:26:33.240 --> 00:26:35.520
You're making them a bit more super human, right?
378
00:26:35.520 --> 00:26:40.240
Like where you are like overall giving it a bunch of business outcomes and you're able
379
00:26:40.240 --> 00:26:41.960
to reason through the data that's available.
380
00:26:41.960 --> 00:26:46.800
I think that's what's like, you know, the models have made possible today, which we haven't
381
00:26:46.800 --> 00:26:48.600
done that in the past.
382
00:26:48.600 --> 00:26:51.200
Or it's been very expensive to do that well in the past.
383
00:26:51.440 --> 00:26:55.480
Well, so that's interesting because every company should have that ability regardless
384
00:26:55.480 --> 00:27:01.920
of their business, but that changes the role of the person doing it.
385
00:27:01.920 --> 00:27:08.520
The input question, you know, to get to the outcome is very different when it's not a
386
00:27:08.520 --> 00:27:12.440
process you're running to gather the data, to run the thing.
387
00:27:12.440 --> 00:27:15.760
It's all kind of there and now you're just trying to get it out.
388
00:27:15.760 --> 00:27:20.240
How have you guys approached that because that's something I think every company can relate
389
00:27:20.240 --> 00:27:21.240
to right now is it?
390
00:27:21.240 --> 00:27:22.240
I know it's all there.
391
00:27:22.240 --> 00:27:25.560
I just don't know how to get the insights out of it.
392
00:27:25.560 --> 00:27:31.840
It's actually people do have it, right, except they used to hire people or specialists
393
00:27:31.840 --> 00:27:32.840
to doing it, right?
394
00:27:32.840 --> 00:27:38.760
Let's say, let's think about the scenario where if I'm an operations business who needs
395
00:27:38.760 --> 00:27:43.440
to do food delivery for like hundreds of my customers, right?
396
00:27:43.440 --> 00:27:49.280
Now I then have to, the only logical way for me to go do that with respect to understanding
397
00:27:49.280 --> 00:27:55.600
my risk, et cetera, is I would delegate that command to say either I have to go pay for
398
00:27:55.600 --> 00:27:59.840
a service which promises that they will do all of that on my behalf based on the inputs
399
00:27:59.840 --> 00:28:05.160
that I control, but the inputs then mean either they are static, don't really change based
400
00:28:05.160 --> 00:28:09.000
on the context of like something that's changed, et cetera.
401
00:28:09.000 --> 00:28:13.120
So all of that's like, you know, you either stuck with a little bit of like a tool that
402
00:28:13.120 --> 00:28:17.640
you bought, which was not necessarily taking advantage of what agent capabilities available
403
00:28:17.640 --> 00:28:18.640
today.
404
00:28:18.640 --> 00:28:24.240
And option is you then have to spend to hire like a quant person or a person who actually
405
00:28:24.240 --> 00:28:30.240
understand math with respect to like, you know, taking all of the data.
406
00:28:30.240 --> 00:28:34.560
And in many cases, people also don't have the data in a structured format, so then you
407
00:28:34.560 --> 00:28:38.600
will need to go hire a data engineer who would actually go to it, right?
408
00:28:38.600 --> 00:28:43.440
So the person who like most of these companies would have like let's say a route planner
409
00:28:43.440 --> 00:28:48.560
who's ultimately responsible for planning routes and they're not always set up to succeed
410
00:28:48.560 --> 00:28:51.920
without like all of the support system in place, right?
411
00:28:51.920 --> 00:28:59.360
So I almost imagine us providing these capabilities more towards one single human who's central
412
00:28:59.360 --> 00:29:07.120
to the operations and giving them the capabilities or the box where they are able to define or
413
00:29:07.120 --> 00:29:10.680
still operate at the higher order objectives that they would probably have given to like
414
00:29:10.680 --> 00:29:15.040
a stats person in their team and still get the results.
415
00:29:15.040 --> 00:29:19.480
And also have an interface that can reason about it and speak the language in which they
416
00:29:19.480 --> 00:29:24.440
understand in terms of like why certain decisions were made in what way.
417
00:29:24.440 --> 00:29:28.800
So they can kind of co-create like the strategic plan, right?
418
00:29:28.800 --> 00:29:34.120
You see the non-technical people have greater access to that data information and do that
419
00:29:34.120 --> 00:29:36.840
and are changing how your org functions as well.
420
00:29:36.840 --> 00:29:42.320
Obviously actually this is quite interesting because we're seeing for example like one
421
00:29:42.320 --> 00:29:47.440
of the things we're starting to see a lot is like product managers trying to land code,
422
00:29:47.440 --> 00:29:48.440
right?
423
00:29:48.440 --> 00:29:53.200
Like which is a common thing that's happening and designers making changes in production
424
00:29:53.200 --> 00:29:54.280
that's happening, right?
425
00:29:54.280 --> 00:30:01.360
So there's definitely like the barrier to entry has reduced when you think about people's
426
00:30:01.360 --> 00:30:02.720
like you know, roles.
427
00:30:02.720 --> 00:30:08.400
We still obviously, again, as we talked about at the start of this conversation around
428
00:30:08.400 --> 00:30:12.000
the governance model and still having the guardrails so things don't go off, right?
429
00:30:12.000 --> 00:30:14.800
But we are definitely seeing that and we are also encouraging that, right?
430
00:30:14.800 --> 00:30:21.480
So in turn to find limits and I totally believe this is happening also with all our customers
431
00:30:21.480 --> 00:30:26.240
where people are individually testing new capabilities that are out there and trying
432
00:30:26.240 --> 00:30:29.920
to find the boundary of what's possible, which is not possible.
433
00:30:30.080 --> 00:30:35.840
Going back to the governance, the cost, like what's your stack today and is there a consistent
434
00:30:35.840 --> 00:30:41.120
one or are you just letting everybody experiment within some sort of guardrail system?
435
00:30:41.120 --> 00:30:46.800
So within models, we have a blessed weight of models anybody can use within the company.
436
00:30:46.800 --> 00:30:51.840
And as I mentioned before, like access to these things, we've tried to centralize and have
437
00:30:51.840 --> 00:30:59.440
centralized gateways so that we have a degree of security audit possible or have a report
438
00:30:59.440 --> 00:31:06.160
that I can quickly look at in terms of spend across my organization and also I would go understand
439
00:31:06.160 --> 00:31:10.800
selectively myself, like you know, what has that spend produced, which I think is important.
440
00:31:10.800 --> 00:31:14.880
We built some degree of tooling to ensure I think two things primarily.
441
00:31:14.880 --> 00:31:20.320
One is like while we want to enable people to go faster, we want to do it in a responsible
442
00:31:20.320 --> 00:31:24.320
way when it comes down to production systems, etc. So I think that's one aspect of it.
443
00:31:24.400 --> 00:31:30.720
While we enable people to try all the new tools, we need to also have a little bit of an eye on
444
00:31:31.520 --> 00:31:36.480
what it means in terms of real spend, right? So we don't want to be blind to it.
445
00:31:36.480 --> 00:31:42.160
At this point, I haven't seen a phase where like why would be super concerned like about the cost
446
00:31:42.160 --> 00:31:46.800
really is like the problem that I need to go solve, right? I think having the
447
00:31:47.440 --> 00:31:52.880
auditability and somewhat of a chargeback model is important from a governance perspective.
448
00:31:52.960 --> 00:31:55.760
I think that's something everybody should do, right? Otherwise you're going to be in for
449
00:31:55.760 --> 00:31:58.960
surprises later on, so yeah. One of the most interesting things over the last
450
00:31:58.960 --> 00:32:04.640
couple of years is companies that didn't know they had so much value but are selling their data,
451
00:32:05.200 --> 00:32:11.840
their pictures, their everything to the bigger model companies. And I would think that you have
452
00:32:11.840 --> 00:32:17.760
a tremendous amount of data. I think about this a lot from the scooter market with so many
453
00:32:17.760 --> 00:32:23.040
hundred plus thousand vehicles on the street. Is there a hidden value in that data?
454
00:32:24.080 --> 00:32:29.920
Is that something that over time that physical presence will only become more and more important
455
00:32:30.560 --> 00:32:37.760
to figuring out how the real world works because of the ability to film, take pictures,
456
00:32:37.760 --> 00:32:42.400
label? Is that like the next generation of true value unlock in these models?
457
00:32:42.480 --> 00:32:48.560
When we think about what we sell to our customers, we sell products that are like you know
458
00:32:48.560 --> 00:32:54.080
fundamentally the foundation is based on sensors that we are selling but what we eventually sell
459
00:32:54.080 --> 00:33:00.960
to customers is actually applications that help with our customer safety or efficiency goals
460
00:33:00.960 --> 00:33:05.360
that's or sustainability goals, right? So that's generally our mission. Now when you really unpack
461
00:33:05.440 --> 00:33:13.280
and transit that most of our customers spend like if you go look at their particular income statement
462
00:33:13.280 --> 00:33:21.440
they spend across safety in terms of insurance cost or like you know accidents which they want
463
00:33:21.440 --> 00:33:26.560
to reduce like you know or like operational cost when you think about like fuel cost like
464
00:33:26.560 --> 00:33:31.600
is an example maintenance is another huge example where like people spend a lot of money,
465
00:33:31.600 --> 00:33:36.720
like a lot of the physical world companies, collective intelligence that we gather from our
466
00:33:37.600 --> 00:33:43.600
install base or distribution actually enables us to build unique products which help
467
00:33:44.560 --> 00:33:51.760
the collective good of like passing down efficiency to across our customers is both from a length
468
00:33:51.760 --> 00:33:56.320
of any of these dimensions right any of the goals take safety for into account or maintenance
469
00:33:56.960 --> 00:34:02.240
cost, right? So I see a world where the world gets more predictive in terms of maintenance, right?
470
00:34:02.240 --> 00:34:07.760
The world will get more predictive in terms of fuel cost management. Today a lot of these things
471
00:34:07.760 --> 00:34:14.400
are done a lot more reactively and we are getting predictive on safety already like which is
472
00:34:15.040 --> 00:34:21.040
which is one of the biggest value propositions we've seen from some Sarah's perspective because
473
00:34:21.040 --> 00:34:25.200
we know statistically from people using our technology that like you know there's we reduced
474
00:34:25.200 --> 00:34:30.480
incidents, we've actually reduced insurance bills and generally kept people safe on the road
475
00:34:30.480 --> 00:34:35.760
right which is a good cause I think and obviously I'm proud to be a part of. I generally think we have
476
00:34:35.760 --> 00:34:42.160
an opportunity with this next wave of like you know we have fuel consumption data, fuel changes
477
00:34:42.160 --> 00:34:48.720
across different like vehicles, the other important way to add is like people's driving
478
00:34:48.720 --> 00:34:54.880
styles influence fuel consumption or what, right? We have all that data too like who's driving
479
00:34:54.880 --> 00:35:00.080
that cab at that particular point in time what did they do like did they harsh break did they
480
00:35:00.080 --> 00:35:06.320
accelerate a lot where did they do that was it on the freeway was it on the like you know city road
481
00:35:06.320 --> 00:35:11.600
etc so we have all of that context which helps us build intelligence to the point where like you know
482
00:35:11.600 --> 00:35:15.920
we know what a baseline good looks like we know what opportunity exists and we can actually like
483
00:35:15.920 --> 00:35:23.680
get a lot more efficiency back to our customers right so so I do think the data is pretty valuable
484
00:35:23.680 --> 00:35:27.680
and especially when you think about the collective intelligence in which how you can apply
485
00:35:27.680 --> 00:35:31.840
we have a unique data set which actually enables a lot of the applications and the end
486
00:35:32.720 --> 00:35:38.080
value that we can pass along to our customers. What's overhyped right now? What's overhyped huh?
487
00:35:38.640 --> 00:35:42.400
I don't know I think there's a little bit of a humorism which I think is definitely overhyped
488
00:35:43.600 --> 00:35:49.920
like an example would be you know we we have Claude Maitos the models like you know
489
00:35:50.000 --> 00:35:54.880
stalled I guess the narrative originally was like you know it's just too powerful and going to find
490
00:35:54.880 --> 00:35:58.880
like all the security issues in the world and like but later there's also an alternate narrative
491
00:35:58.880 --> 00:36:05.920
around compute limitations that they have but I generally think as these models get better like you know
492
00:36:07.040 --> 00:36:11.440
it's again like black hats versus white hats right so there's going to be models which are also like
493
00:36:11.440 --> 00:36:18.480
more smarter about patching issues before right so you just I almost see it as like a one-time
494
00:36:18.480 --> 00:36:24.640
shift in terms of like security processes and then you have the same capabilities available to
495
00:36:24.640 --> 00:36:31.040
pretty much both sides I kind of don't believe there's like a very big the negativity around like
496
00:36:31.040 --> 00:36:35.920
oh is this something that's going to like completely shift everything right. I love hearing that
497
00:36:35.920 --> 00:36:41.680
because I'm scared of that so when technical when super technical people are not as scared of it it
498
00:36:41.680 --> 00:36:46.400
makes me a little happy because I treat Claude with respect just to make sure that it doesn't kill me
499
00:36:46.400 --> 00:36:51.760
well yeah yeah no I think there's there's a little bit of like we we do tend to project out
500
00:36:51.760 --> 00:36:57.280
capabilities right and I do think there's going to be a lot of interesting things happening
501
00:36:57.280 --> 00:37:03.840
in the world right but then I wish we also focused a lot on the positivity of things that's
502
00:37:03.840 --> 00:37:09.840
possible right and I generally am a believer that these tools will be more of an enabler
503
00:37:10.720 --> 00:37:18.560
uh for like many more things or opportunities that we are not thinking about is how I generally see
504
00:37:18.560 --> 00:37:24.240
it a lot of what we talked about in terms of predictive intervention none of these things were
505
00:37:24.240 --> 00:37:29.120
possible without like a lot of the reasoning capabilities that's available so yeah so I
506
00:37:29.120 --> 00:37:33.440
generally believe that there's like going to be a lot a lot of good stuff right the other part is
507
00:37:34.080 --> 00:37:40.080
it's also going to enable a lot more access which is generally good for humanity right like where
508
00:37:41.360 --> 00:37:48.320
which I think is very interesting and I do believe that right so as cost of making things go
509
00:37:48.320 --> 00:37:52.480
down or like you know there's more people are enabled to make things so you're generally going to see
510
00:37:52.480 --> 00:37:58.720
like accessibility increase for capabilities in the real world yeah so the next few years will be
511
00:37:58.720 --> 00:38:03.200
interesting obviously it's a very exciting time like every day is we learn something new and
512
00:38:04.240 --> 00:38:07.680
I'm definitely looking at it more through a positive lens I would say
513
00:38:08.480 --> 00:38:12.960
I love it well thank you for joining us I really appreciate it and sharing your thoughts and insights
514
00:38:12.960 --> 00:38:23.760
yeah enjoy talking to you keep
1
00:00:00.000 --> 00:00:03.120
Today I'm excited to talk to a Praveen Murugasan.
2
00:00:03.120 --> 00:00:08.600
His team at Samsara processes something like 25 trillion data points
3
00:00:08.600 --> 00:00:12.360
across 90 billion miles of road every year.
4
00:00:12.360 --> 00:00:13.840
What I think is really interesting,
5
00:00:13.840 --> 00:00:17.280
other than the scale at which he's doing this Samsara is
6
00:00:17.280 --> 00:00:21.400
a publicly traded company at Fortune ranks at number seven on the future 50,
7
00:00:21.400 --> 00:00:26.800
is how they are both approaching AI from making themselves more efficient as a company,
8
00:00:26.800 --> 00:00:30.240
how they're experimenting, how they're building products faster,
9
00:00:30.240 --> 00:00:33.520
but also how they can then roll that out to their customers
10
00:00:33.520 --> 00:00:35.520
and help us all stay safe on the roads.
11
00:00:45.760 --> 00:00:48.920
What is your most contrary thought related to AI?
12
00:00:48.920 --> 00:00:50.520
If you were just going to take a step back and say,
13
00:00:50.520 --> 00:00:52.920
here's what I believe that nobody else does.
14
00:00:52.920 --> 00:00:57.520
I'll tell you, my fascination recently has been on computer use,
15
00:00:57.520 --> 00:00:59.880
AI driven computer use.
16
00:00:59.880 --> 00:01:02.800
I've been starting to use cloud co-work.
17
00:01:02.800 --> 00:01:06.600
When you think about it, like historically enterprise software,
18
00:01:06.600 --> 00:01:11.000
even doing any workflow or task was defined by the rules of the game of structure,
19
00:01:11.000 --> 00:01:15.120
meaning you have APIs, you have, like even MCP servers,
20
00:01:15.120 --> 00:01:17.880
which try to define some structure, etc.
21
00:01:18.600 --> 00:01:20.440
I think with computer use,
22
00:01:20.440 --> 00:01:25.320
essentially enabling any sort of human tool available without structure
23
00:01:25.320 --> 00:01:28.960
for a reasoning model to make decisions,
24
00:01:28.960 --> 00:01:30.440
I thought was quite interesting.
25
00:01:30.440 --> 00:01:32.640
So one of the things I'm curious about is like,
26
00:01:32.640 --> 00:01:34.960
or at least something that I'm actively thinking of is like,
27
00:01:34.960 --> 00:01:38.200
how do we apply that into our business?
28
00:01:38.200 --> 00:01:40.320
And also from a customer centric lens,
29
00:01:40.320 --> 00:01:45.520
because like a lot of people are working on very old operational tools,
30
00:01:45.520 --> 00:01:49.000
which a lot of people will tell you it's very hard to move.
31
00:01:49.000 --> 00:01:52.400
Like the mode is actually like, I have 100 integrations.
32
00:01:52.400 --> 00:01:54.760
Need a move is going to be like a big project,
33
00:01:54.760 --> 00:02:00.280
migrations becoming like a pretty big effort for most people.
34
00:02:00.280 --> 00:02:06.600
I think there'll be a radical shift in companies that are able to leverage
35
00:02:06.600 --> 00:02:11.960
newer technology and actually like keep up to find more defensible modes.
36
00:02:11.960 --> 00:02:14.480
So computer use actually, I think, opens up the game.
37
00:02:14.520 --> 00:02:19.920
So I think a lot of the legacy players who've been like essentially
38
00:02:19.920 --> 00:02:23.840
treating the cost of integrations or like the complexity being the mode
39
00:02:23.840 --> 00:02:25.560
should be really worried at the moment, right?
40
00:02:25.560 --> 00:02:27.960
So I feel like in the next 12 months,
41
00:02:27.960 --> 00:02:30.400
that's going to be changing quite rapidly.
42
00:02:30.400 --> 00:02:32.680
And do you think that will happen on an enterprise level
43
00:02:32.680 --> 00:02:37.360
or even individuals given their new tool sets available
44
00:02:37.360 --> 00:02:41.800
will be able to begin taking part in that themselves?
45
00:02:41.800 --> 00:02:43.400
I think both actually.
46
00:02:43.440 --> 00:02:45.440
This is something I've been thinking about recently,
47
00:02:45.440 --> 00:02:51.080
where every business is right now thinking through a lens of transformation
48
00:02:51.080 --> 00:02:55.920
of their workforce in terms of how do you supercharge them with AI tools,
49
00:02:55.920 --> 00:02:56.720
right?
50
00:02:56.720 --> 00:03:01.240
But for me in engineering, this is something I think about all the time,
51
00:03:01.240 --> 00:03:04.560
meaning we have all these AI coding tools.
52
00:03:04.560 --> 00:03:07.840
How do we enable our engineers to be more productive?
53
00:03:07.840 --> 00:03:12.040
But then we still don't want to do something crazy, right?
54
00:03:12.040 --> 00:03:15.200
Like I mean, there's also lots of stories you read now about
55
00:03:15.200 --> 00:03:17.720
like production DBs, getting deleted, et cetera.
56
00:03:17.720 --> 00:03:22.280
So with some guardrails, but how do you make people supercharge?
57
00:03:22.280 --> 00:03:24.840
So that's a common problem.
58
00:03:24.840 --> 00:03:28.040
Now, what I found fascinating with computer use was like,
59
00:03:28.040 --> 00:03:31.760
I've been, I'm still more recently playing with it.
60
00:03:31.760 --> 00:03:36.000
I used it to make a deck for like one of my leadership upside
61
00:03:36.000 --> 00:03:39.960
that like, you know, what would have taken me potentially like
62
00:03:40.240 --> 00:03:42.880
about 10 hours, took me like three,
63
00:03:42.880 --> 00:03:47.120
but then I later realized a lot of people we sell to in enterprise space
64
00:03:47.120 --> 00:03:50.000
or even anyone who's actually working in front of a computer, right?
65
00:03:50.000 --> 00:03:54.240
Like think about any task you do from booking a flight ticket to anything
66
00:03:54.240 --> 00:03:57.600
that you do, actually sort of felt like it's surprisingly
67
00:03:57.600 --> 00:04:03.680
every task you do was made easier with time back to the individual, right?
68
00:04:03.680 --> 00:04:06.480
This becomes more of not only like an enterprise,
69
00:04:06.480 --> 00:04:10.480
but you play enterprise definitely because you have some of these big companies
70
00:04:10.480 --> 00:04:13.680
like which has like hundreds of thousands of employees.
71
00:04:13.680 --> 00:04:18.640
And I often wonder like, what does that look like in a couple of years, right?
72
00:04:18.640 --> 00:04:22.880
And I would think that has to radically transform quickly, right?
73
00:04:22.880 --> 00:04:25.960
So people have to go and like, you have to really rethink.
74
00:04:25.960 --> 00:04:30.960
As you begin to think about it from even your teams and how you're rolling it out
75
00:04:30.960 --> 00:04:35.480
because you've got all of that, you know, as well,
76
00:04:35.480 --> 00:04:39.360
how have you empowered the team and what guardrails have you put?
77
00:04:39.360 --> 00:04:42.920
And sort of as everyone just on like a cursor, for example,
78
00:04:42.920 --> 00:04:46.280
and doing whatever they want or, you know,
79
00:04:46.280 --> 00:04:49.040
then Claude comes out and it's like, should we move over there?
80
00:04:49.040 --> 00:04:52.040
Like how do you even think about the stack today?
81
00:04:52.040 --> 00:04:54.720
We always thought about a little bit progressive.
82
00:04:54.720 --> 00:04:59.120
I would say at some sort of like I speak to a lot of industry friends
83
00:04:59.120 --> 00:05:01.800
and I realize people are still like adopting these tools slower.
84
00:05:01.800 --> 00:05:03.800
I think we've been a little bit more progressive.
85
00:05:03.800 --> 00:05:07.040
I think the largely the way we've approached the problem was
86
00:05:07.040 --> 00:05:11.640
enable people to get the tools they want and also say that's okay to go experiment, right?
87
00:05:11.640 --> 00:05:17.440
So I think letting people try things has been a good approach.
88
00:05:17.440 --> 00:05:22.200
We've also established dedicated forums where people or engineers
89
00:05:22.200 --> 00:05:27.000
who are trying new things would come back and talk about it to the broader org.
90
00:05:27.000 --> 00:05:32.080
Now we've gotten more structured about it by collapsing all of that into like
91
00:05:32.080 --> 00:05:36.360
a little bit of a learning curriculum for new people that are starting in the organization.
92
00:05:36.360 --> 00:05:42.000
So we've always been a bit of like a lot more open versus like not necessarily like,
93
00:05:42.000 --> 00:05:46.000
you know, start with objections as the starting point, right?
94
00:05:46.000 --> 00:05:49.560
So it's culturally I think we've been a lot more open in general,
95
00:05:49.560 --> 00:05:51.080
encouraging people.
96
00:05:51.080 --> 00:05:55.560
On the other side, I think from a God rail perspective,
97
00:05:55.560 --> 00:05:59.600
we've been a bit cautious about our life cycle,
98
00:05:59.600 --> 00:06:04.560
especially with respect to programming or like shipping new models, et cetera,
99
00:06:04.560 --> 00:06:06.800
even for our customers.
100
00:06:06.800 --> 00:06:11.280
It's almost think about using some processes as a bit of a God rail,
101
00:06:11.280 --> 00:06:16.880
but then also like relax them as we learn and adapt to the new AI tool, right?
102
00:06:16.880 --> 00:06:19.720
Like a few examples would be obviously like, you know,
103
00:06:19.720 --> 00:06:25.280
we still walk away talk about this where when an AI is authoring code
104
00:06:25.280 --> 00:06:28.000
that doesn't mean the accountability is lost, right?
105
00:06:28.000 --> 00:06:32.280
Like the accountability still stays with the person who's actually like shipping the PR, right?
106
00:06:32.280 --> 00:06:34.400
And the person who's reviewing the PR, right?
107
00:06:34.400 --> 00:06:37.040
We also recognize not all code creator is equal.
108
00:06:37.040 --> 00:06:41.760
So we want to enable people to move faster in some smaller changes.
109
00:06:41.760 --> 00:06:49.280
But then we've now built capabilities to score changes to almost like identify risk, right?
110
00:06:49.280 --> 00:06:53.600
So it's the governance model or the God rail that we're putting and learning ourselves,
111
00:06:53.600 --> 00:07:00.880
where like, you know, I mean, like modeling risk correctly and enabling people to go faster, right?
112
00:07:00.880 --> 00:07:06.080
So that's been a big shift, I would say, right?
113
00:07:06.080 --> 00:07:08.880
I don't think we fully crack the code of,
114
00:07:08.880 --> 00:07:14.560
or we have the perfect governance model, but I think the attitude towards approaching this problem
115
00:07:14.560 --> 00:07:20.080
is one, let's enable people to try and experiment and then ensure people have good intent.
116
00:07:20.080 --> 00:07:24.560
Let's not relax processes to the point where we just go fully open.
117
00:07:24.560 --> 00:07:28.720
Let's make sure because we have a duty to our customers to ship good quality products
118
00:07:28.720 --> 00:07:31.200
and we need to make sure we continue to do that.
119
00:07:31.200 --> 00:07:35.840
So we have internal metrics like any business in terms of understanding or releases
120
00:07:35.840 --> 00:07:38.320
and the quality of our products.
121
00:07:38.320 --> 00:07:44.680
We've been monitoring those plus also like slowly relaxing or changing our controls to ensure
122
00:07:44.680 --> 00:07:47.960
like we also enable developer productivity, but without compromising quality.
123
00:07:47.960 --> 00:07:49.080
That's what we think about it.
124
00:07:49.080 --> 00:07:52.840
Like what are some of the wins that you've seen as part of that has been
125
00:07:52.840 --> 00:07:59.240
have you seen quality of code go up because or at least consistency of it go up and speed
126
00:07:59.240 --> 00:08:02.920
or is it not that easy to to point to?
127
00:08:02.920 --> 00:08:07.720
Speed of development has gone up, like that's like it's undoubtedly, I would say that, right?
128
00:08:07.720 --> 00:08:16.280
So the maybe a couple of things, like one is we have definitely started doing things
129
00:08:16.280 --> 00:08:21.000
with a fail fast mindset, like historically you would discuss about like whether an idea
130
00:08:21.000 --> 00:08:25.080
is worth to pursue, you'd have a lot of meetings.
131
00:08:25.080 --> 00:08:27.480
Yeah, yeah, yeah, sure, right?
132
00:08:27.480 --> 00:08:33.520
So right now the at least what culturally what we try to tell people is like actually you
133
00:08:33.520 --> 00:08:37.680
have a good idea rather than talk about the idea maybe like build a quick prototype and show
134
00:08:37.680 --> 00:08:38.680
me.
135
00:08:38.680 --> 00:08:42.680
We've seen that it enables the quick follow on conversations about okay is this worth
136
00:08:42.680 --> 00:08:47.880
pursuing should we build this as a product is there going to market for this or why should
137
00:08:47.880 --> 00:08:50.080
we pursue this versus something else or not, right?
138
00:08:50.080 --> 00:08:53.560
So going to prototypes has been super simple, right?
139
00:08:53.560 --> 00:08:55.360
So that's one.
140
00:08:55.360 --> 00:09:01.400
The other dimension I'm also seeing is we are actually addressing more product quality issues
141
00:09:01.400 --> 00:09:04.680
faster than what we would have done in the past.
142
00:09:04.680 --> 00:09:09.560
So any business would like, you know, if you're operating software you're going to have
143
00:09:09.560 --> 00:09:13.760
like some degree of support issues and bugs that come in, but usually people will always
144
00:09:13.760 --> 00:09:18.880
categorize and say I'm going to solve my P1s like P2 is well, we'll get to it, right?
145
00:09:18.880 --> 00:09:25.760
So I think the nice part now is we enabled the capability of people are able to quickly
146
00:09:25.760 --> 00:09:28.360
see P2s come in, oh, this is actually easy.
147
00:09:28.360 --> 00:09:34.480
I can just land this wire one of my agents while I work on this other thing.
148
00:09:34.480 --> 00:09:40.800
So that's become a bit of a norm, like I was seeing one of our engineers work with four
149
00:09:40.800 --> 00:09:48.640
different course sessions on four different Mac which will destops, each of them spawn
150
00:09:48.640 --> 00:09:51.440
to run like, you know, in different tasks.
151
00:09:51.440 --> 00:09:53.440
That's crazy.
152
00:09:53.440 --> 00:09:58.480
Yeah, so it's been pretty awesome actually, it's been very energizing, right?
153
00:09:58.480 --> 00:10:04.160
So that's kind of the transformation I see both along the lines of productivity, but
154
00:10:05.160 --> 00:10:11.840
can you talk about one of the customer-facing releases that you've done and how that was
155
00:10:11.840 --> 00:10:17.360
developed because you've got an engineering team which is probably really great at coding,
156
00:10:17.360 --> 00:10:20.960
you've got a product team which has a level of expertise, but by the time you get to the
157
00:10:20.960 --> 00:10:26.480
customer, I don't always imagine they know exactly what they want or the art of the possible,
158
00:10:26.480 --> 00:10:27.920
as you said.
159
00:10:27.920 --> 00:10:33.640
So, you know, what's a product that you released for them that typically wasn't something
160
00:10:33.640 --> 00:10:38.280
they might have even known that they could have or want, but because they just don't understand
161
00:10:38.280 --> 00:10:40.160
the future, the capability.
162
00:10:40.160 --> 00:10:44.440
In some markets that we operate, so in some sort of operates globally, we also operate
163
00:10:44.440 --> 00:10:49.880
in markets like Mexico, Canada, US and Europe, etc.
164
00:10:49.880 --> 00:10:57.120
In Mexico, and surprisingly in some places in Europe, like some customers of ours had told
165
00:10:57.120 --> 00:11:01.440
us that they have a bit of like fuel theft problem.
166
00:11:01.440 --> 00:11:07.640
It's actually an interesting case because, and especially now that fuel prices are super
167
00:11:07.640 --> 00:11:08.640
high, right?
168
00:11:08.640 --> 00:11:18.240
Like, you know, the problem effectively is we have a lot of like fuel theft by two different
169
00:11:18.240 --> 00:11:19.240
means.
170
00:11:19.240 --> 00:11:25.920
One is actually like people actually siphoning fuel off the vehicle and selling it happened
171
00:11:25.920 --> 00:11:31.480
surprisingly a lot, actually there were some studies done where this was in the hundreds
172
00:11:31.480 --> 00:11:34.920
of millions of dollars lost globally.
173
00:11:34.920 --> 00:11:39.840
And there's also another type of fraud that happens where people might use a fuel card that's
174
00:11:39.840 --> 00:11:42.320
issued to them, but then they fuel their personal vehicle, right?
175
00:11:42.320 --> 00:11:48.440
So, these are two different types of fuel products that actually, you know, these are problems
176
00:11:48.440 --> 00:11:49.440
that exist.
177
00:11:49.440 --> 00:11:52.480
So, from a, how do we build like psycho perspective?
178
00:11:52.480 --> 00:11:57.360
So, the way we do this is we always, some sort of culture has been more around like working
179
00:11:57.360 --> 00:12:00.440
with a customer because a lot of, when you, when you think about it and this goes all
180
00:12:00.440 --> 00:12:07.000
the way to our CEO Sanjit, like a lot of the companies, DNAs, we are all product builders.
181
00:12:07.000 --> 00:12:10.160
We are not experts in physical world operations, right?
182
00:12:10.160 --> 00:12:16.160
So you have to really lean in to your customers and relearn and understand from them who are
183
00:12:16.160 --> 00:12:18.360
the experts in their physical world problems, right?
184
00:12:18.360 --> 00:12:22.480
So we take those problems and then translate into building technology products, which is
185
00:12:22.480 --> 00:12:23.680
what we are good at.
186
00:12:23.680 --> 00:12:29.840
This actually came up as a problem that I spoke into a few customers and they had flagged
187
00:12:29.840 --> 00:12:31.320
it as a problem.
188
00:12:31.320 --> 00:12:34.560
And one of the immediate instinct was like, hey, this doesn't feel like a hard problem
189
00:12:34.560 --> 00:12:39.640
to solve because you get fuel level from every vehicle, right?
190
00:12:39.640 --> 00:12:42.160
It's, we have sensors installed on the vehicle.
191
00:12:42.160 --> 00:12:49.280
You can get the fuel level, you can over time run algorithms like to see if the fuel dropped
192
00:12:49.280 --> 00:12:51.400
and then like let's go solve it, right?
193
00:12:51.400 --> 00:12:57.080
So interestingly, like, you know, we've apparently tried doing that a while back and it just
194
00:12:57.080 --> 00:13:01.800
led to poor results, like very noisy alerts and issues, right?
195
00:13:01.800 --> 00:13:07.560
One of the things we had to do here was almost rethink the problem, friend approach it
196
00:13:07.560 --> 00:13:10.360
from like leveraging some degree of machine learning techniques.
197
00:13:10.360 --> 00:13:18.040
The problem itself was sensors are not reliable and they're not equal, right, across vehicles.
198
00:13:18.040 --> 00:13:22.720
So you would, if you start inspecting fuel level data, it's extremely noisy.
199
00:13:22.720 --> 00:13:29.400
So you can't distill real fuel levels quickly as much as what people would assume.
200
00:13:29.400 --> 00:13:36.320
And even extreme cases, because of this fuel sensor noise problem in markets like Mexico,
201
00:13:36.320 --> 00:13:41.680
they actually have a lot of companies selling fuel level sensors that you can install as
202
00:13:41.680 --> 00:13:45.200
an external thing on top of your tank, right?
203
00:13:45.200 --> 00:13:47.600
So people are trying to keep a pulse of it.
204
00:13:47.600 --> 00:13:51.880
We actually worked on building a model, but then the approach was we wanted to get quick
205
00:13:51.880 --> 00:13:52.880
feedback.
206
00:13:52.880 --> 00:13:54.080
That's just how we work, right?
207
00:13:54.080 --> 00:13:57.600
Like in terms of you, you can build the entire product and ship it to the market.
208
00:13:57.600 --> 00:14:01.440
But whereas what we did was we took like four or five customers who had come to us with
209
00:14:01.440 --> 00:14:03.960
this being a common problem.
210
00:14:03.960 --> 00:14:10.200
And we were initially just giving them algorithmic results in like spreadsheets, telling them
211
00:14:10.200 --> 00:14:14.520
to react to it, like, you know, does this feel like a useful signal for you?
212
00:14:14.520 --> 00:14:16.760
And what are the context can we add?
213
00:14:16.760 --> 00:14:22.040
Like is location interesting is like, you know, how long somebody was idling in a specific
214
00:14:22.040 --> 00:14:23.720
place before interesting?
215
00:14:23.720 --> 00:14:28.560
It's almost like a co iteration of the product with some of our customer feedback.
216
00:14:28.560 --> 00:14:35.080
And then we bring it into the market and like try to scale as we, like, and learn from
217
00:14:35.080 --> 00:14:36.080
that experience.
218
00:14:36.080 --> 00:14:40.080
So that's typically how we go about working on any sort of product.
219
00:14:40.080 --> 00:14:44.280
The other proxy signal we look for is usage or adoption.
220
00:14:44.280 --> 00:14:48.080
And this ties back to like today, the value to enter the building solutions is like so
221
00:14:48.080 --> 00:14:49.080
low.
222
00:14:49.080 --> 00:14:55.920
You could build like for you to find real value, like usage or like somebody willing
223
00:14:55.920 --> 00:14:58.680
to pay you stuff is like actually a good valuable signal.
224
00:14:58.680 --> 00:15:04.080
You're collecting all the data from all the sensors and then trying to come up with whether
225
00:15:04.080 --> 00:15:08.880
it's usage or theft based on the other metrics that's going on.
226
00:15:08.880 --> 00:15:10.440
And is that a real time?
227
00:15:10.440 --> 00:15:16.800
We have our some sort of sensor, which actually like you can hook into a vehicle's on board
228
00:15:16.800 --> 00:15:18.520
diagnostics.
229
00:15:18.520 --> 00:15:25.520
So we have access to pretty much the vehicles on board state in real time.
230
00:15:25.520 --> 00:15:32.280
Think about not only fuel level, like fuel consumption rate and like any signal you want
231
00:15:32.280 --> 00:15:36.320
engine state, like whether the engine's on or not, like whether the engine's moving,
232
00:15:36.320 --> 00:15:38.880
what's the automata state, etc.
233
00:15:38.880 --> 00:15:42.360
So you get access to all of that data in real time.
234
00:15:42.360 --> 00:15:48.840
And for this particular problem, we had to employ like a suite of algorithms.
235
00:15:48.840 --> 00:15:56.280
That was not effectively one because the problem's complexity is derived from a few different
236
00:15:56.280 --> 00:15:57.280
dimensions.
237
00:15:57.280 --> 00:16:02.360
One is the sensor gauge itself might not be precise.
238
00:16:02.360 --> 00:16:08.480
So you have to model based on the variance for distant vehicles.
239
00:16:08.480 --> 00:16:14.840
And then sometimes you have to fuse data from consumption to improve your sensor quality,
240
00:16:14.840 --> 00:16:19.000
like you also can take the sense of fuel consumption data.
241
00:16:19.000 --> 00:16:23.400
In some scenarios, not all vehicles have only one single fuel tank.
242
00:16:23.400 --> 00:16:26.960
They might have multiple tanks, so you have to accommodate for that.
243
00:16:26.960 --> 00:16:33.600
There's also scenarios where we try to predict out what is translating fuel level percentage
244
00:16:33.600 --> 00:16:39.120
into the actual volume, which actually helps you get more precise results.
245
00:16:39.120 --> 00:16:46.280
So this is just an example of a project that we worked on and a feature that's actually
246
00:16:46.280 --> 00:16:49.200
fairly successful in creating value for customers.
247
00:16:49.200 --> 00:16:52.080
All this exists in a pre-AI world.
248
00:16:52.080 --> 00:16:57.880
So when these models come out, how does it change how you fundamentally do the analysis
249
00:16:57.880 --> 00:16:59.120
and the business on that?
250
00:16:59.120 --> 00:17:04.200
How do you all think about picking a model to build on top of given the pace of change?
251
00:17:04.200 --> 00:17:10.000
So this specific algorithm, we built our own proprietary models to solve it.
252
00:17:10.000 --> 00:17:14.960
So the common question we got was like, why don't we just use an LLM to do this, give
253
00:17:14.960 --> 00:17:18.600
the samples of fuel traces.
254
00:17:18.600 --> 00:17:24.440
Language models largely were not able to accommodate for the sensor variance.
255
00:17:24.440 --> 00:17:28.200
This was almost like a noise filtering problem, so it's a very form-fitted problem.
256
00:17:28.200 --> 00:17:34.000
So it's not necessarily something that required a large language model, and it's like not
257
00:17:34.000 --> 00:17:37.720
something that's also the best suited model, I would say.
258
00:17:37.720 --> 00:17:42.560
So we didn't use a large language model in the AI model in that sense, so we actually
259
00:17:42.560 --> 00:17:49.080
built our own models to account for the noise, and how do we go about filtering.
260
00:17:49.080 --> 00:17:55.200
I would say, in this case, we've used AI in a little bit different lengths.
261
00:17:55.200 --> 00:18:00.040
One of the challenges when you think about this problem is, you need to still get samples
262
00:18:00.040 --> 00:18:07.080
from the real world, and you need to label them as like, yes, this was like actual sensor
263
00:18:07.080 --> 00:18:12.760
noise versus this was a level change that was unexpected, right?
264
00:18:12.760 --> 00:18:20.520
So we had to go do human labeling for a lot of this cases, and we significantly sped
265
00:18:20.520 --> 00:18:23.120
up that process by using LLMs, basically.
266
00:18:23.120 --> 00:18:29.560
So what we would do was use more like an LLM as a judge, and where we would have given
267
00:18:29.560 --> 00:18:37.600
the LLM original hand-labeled samples, and then leverage an LLM to go classify at scale
268
00:18:37.600 --> 00:18:38.880
like the results.
269
00:18:38.880 --> 00:18:41.400
How many data points are we talking about just in my sense?
270
00:18:41.400 --> 00:18:47.160
In general, the product runs across multiple millions of vehicles, and they're running
271
00:18:47.160 --> 00:18:54.080
across like on continuous stream of data across these vehicles when they are mobile, right?
272
00:18:54.080 --> 00:18:57.800
So we're talking easily the order of the trillions, right?
273
00:18:57.800 --> 00:19:01.840
Like when it comes down to individual data points that we work with, obviously the training
274
00:19:01.840 --> 00:19:05.280
set was a little bit more smaller.
275
00:19:05.280 --> 00:19:12.000
We would primarily pick a range and train based on different vehicles or like different
276
00:19:12.000 --> 00:19:15.360
make models, so we get a comprehensive set.
277
00:19:15.360 --> 00:19:17.080
So that's kind of how we go about it.
278
00:19:17.080 --> 00:19:21.600
Your own model is doing the analysis, but the actual data labeling of the 20 trillion
279
00:19:21.600 --> 00:19:24.200
or 25 trillion is coming in.
280
00:19:24.200 --> 00:19:30.000
Yeah, it's AI augmented when it comes down to for us to create the training set, but also
281
00:19:30.000 --> 00:19:31.800
like label the training set itself.
282
00:19:31.800 --> 00:19:36.640
So that's something that we've been successfully leveraging AI for.
283
00:19:36.640 --> 00:19:40.640
So you wrote a piece about vehicle routing, which I thought was really interesting.
284
00:19:40.720 --> 00:19:46.440
Look at a personal level, I hope, you know, I relied on ways forever to get around, but
285
00:19:46.440 --> 00:19:49.480
I think that's an individual driving one car.
286
00:19:49.480 --> 00:19:55.320
Where does AI come in from your guys perspective in terms of we can really help add massive
287
00:19:55.320 --> 00:20:00.200
value to our customers by helping them get somewhere faster and more safely?
288
00:20:00.200 --> 00:20:03.800
So one of the things we've been doing recently is we launched this product called commercial
289
00:20:03.800 --> 00:20:04.800
navigation, right?
290
00:20:05.080 --> 00:20:10.960
So it's sort of think about it for ways for enterprise users who are driving for commercial
291
00:20:10.960 --> 00:20:17.440
use, especially you might be driving trucks around and like your typical ways doesn't
292
00:20:17.440 --> 00:20:20.400
work for it directly, right?
293
00:20:20.400 --> 00:20:25.880
There are a few really cool applications that we are able to build, which like gives us
294
00:20:25.880 --> 00:20:28.600
a real leverage unlike like other players, right?
295
00:20:28.600 --> 00:20:32.920
So you said you're a user of ways, I've used ways actually and I thought some features
296
00:20:32.920 --> 00:20:34.720
were really cool in ways, right?
297
00:20:34.720 --> 00:20:41.080
So this, my understanding is fundamentally, it's primarily crowdshosed in terms of like
298
00:20:41.080 --> 00:20:43.080
intelligence that it plugs on.
299
00:20:43.080 --> 00:20:48.440
It could tell you whether there's a cop in some ways or it could tell you, I think when
300
00:20:48.440 --> 00:20:52.960
I was driving once it even told me that there's like a fallen tree on the road, right?
301
00:20:52.960 --> 00:20:54.680
Which I thought was pretty cool.
302
00:20:54.680 --> 00:21:02.880
Now what's possible for us is quite interesting because we have millions of vehicles with sensors,
303
00:21:02.880 --> 00:21:09.000
not only sensors that's connected to the cab from a getting diagnostic data perspective,
304
00:21:09.000 --> 00:21:13.080
but we have like a dual-facing camera that we also set, right?
305
00:21:13.080 --> 00:21:18.880
Now so we almost have like eyes on the road with which you can derive data like or build
306
00:21:18.880 --> 00:21:21.240
better intelligence into our products.
307
00:21:21.240 --> 00:21:28.000
So one of the things we've been doing is understanding exceptions or things that's happening
308
00:21:28.000 --> 00:21:34.520
in the real world and automatically flagging them and collecting data insights towards
309
00:21:34.520 --> 00:21:35.520
it.
310
00:21:35.520 --> 00:21:41.800
An example here would be this example that we talked about on like, let's say a tree
311
00:21:41.800 --> 00:21:44.360
fallen on the road, right?
312
00:21:44.360 --> 00:21:51.880
You do not need human input anymore to actually detect that scenario because like some, at
313
00:21:51.880 --> 00:21:59.160
some sort of like cameras like already recording it can see eyes on the road and we can program
314
00:21:59.160 --> 00:22:05.720
or like, you know, cameras or the edge to detect for certain scenarios and automatically
315
00:22:05.720 --> 00:22:12.000
capture that signal with evidence and so you set it up to the cloud, you can process
316
00:22:12.000 --> 00:22:15.920
it and you actually have many, many data points is not even like we're relying on like one
317
00:22:15.920 --> 00:22:18.640
instance because we have like millions of vehicles on the road.
318
00:22:18.640 --> 00:22:26.960
So you actually can almost run an auction for multiple data instances to agree upon and
319
00:22:26.960 --> 00:22:30.560
surface that back into the other customers, right?
320
00:22:30.560 --> 00:22:36.440
So if I might be driving towards that route, even before I drive the route, you could preemptively
321
00:22:36.440 --> 00:22:43.600
intervene and actually reroute for a person or like give people closest insights as they
322
00:22:43.600 --> 00:22:46.520
are using their navigation, hey, this is something you probably want to watch out on the
323
00:22:46.520 --> 00:22:47.520
road.
324
00:22:47.520 --> 00:22:52.840
We've really successfully done this for instances like construction detection, like for example,
325
00:22:52.840 --> 00:22:57.640
there's a road construction, we've been able to like capture some of the data around like,
326
00:22:57.640 --> 00:23:03.200
you know, automatically identifying this construction happening here and then that becomes an inside.
327
00:23:03.200 --> 00:23:10.400
We're now on the next phase working on how do we automatically like intervene when we have
328
00:23:10.400 --> 00:23:11.400
this context.
329
00:23:11.400 --> 00:23:16.120
A lot of this comes down to having the real world context and also the operational context
330
00:23:16.120 --> 00:23:21.160
for the particular user and you then get to a point of rather than just sharing it as an
331
00:23:21.160 --> 00:23:24.840
inside, start taking real world actions.
332
00:23:24.840 --> 00:23:29.960
Do you think over time you'll have enough data to get predictive on some of the stuff
333
00:23:29.960 --> 00:23:35.200
where you can say these are the patterns of this time of night, this sort of thing.
334
00:23:35.200 --> 00:23:38.160
Is it a weird place where that's possible soon?
335
00:23:38.160 --> 00:23:39.160
I think so.
336
00:23:39.160 --> 00:23:41.600
Yeah, I think we're definitely getting there.
337
00:23:41.600 --> 00:23:48.360
What I've seen from our customers is there is a degree of strategic planning that people
338
00:23:48.360 --> 00:23:53.720
try to do in terms of operations like I need to do deliveries across hundreds of my customers
339
00:23:53.720 --> 00:23:55.920
over the course of the next six months.
340
00:23:55.920 --> 00:23:57.560
How do I plan, right?
341
00:23:57.560 --> 00:24:03.560
So to define those, there's a heavy reliance on what potentially happened in history.
342
00:24:03.560 --> 00:24:09.000
If you're able to learn and derive the right insights, I think it can largely influence
343
00:24:09.000 --> 00:24:11.320
how you do strategic plans, right?
344
00:24:11.320 --> 00:24:16.120
Obviously, we talked a lot about efficiency in this context primarily from like, you know,
345
00:24:16.120 --> 00:24:21.880
hey, like we know obstructions on the road and how can you change routes, etc. from an
346
00:24:21.880 --> 00:24:22.880
efficiency angle.
347
00:24:22.880 --> 00:24:24.880
It doesn't even stop there.
348
00:24:24.880 --> 00:24:30.280
When you think about it, right, it actually extends to more ideas, like an example would
349
00:24:30.280 --> 00:24:33.640
be safety of different roads.
350
00:24:33.640 --> 00:24:36.360
Some roads are more prone to accidents.
351
00:24:36.360 --> 00:24:41.720
We have tons of data which actually can pinpoint where do accidents happen.
352
00:24:41.720 --> 00:24:47.800
We can score safety of individual street segments and even potentially take that into account
353
00:24:47.800 --> 00:24:50.200
when it comes down to route modeling, right?
354
00:24:50.200 --> 00:24:55.640
So there's an aspect of strategic planning that will heavily lean on historical data but
355
00:24:55.640 --> 00:24:58.400
also like historical modeling of data.
356
00:24:58.400 --> 00:25:04.280
So where does the sort of last five years of model development help you do that more efficiently
357
00:25:04.280 --> 00:25:10.040
or is that just, is it now something that you can present to the customer in a more accessible
358
00:25:10.040 --> 00:25:12.360
way, thanks to AI?
359
00:25:12.360 --> 00:25:16.040
A lot of the historical data always existed.
360
00:25:16.040 --> 00:25:21.600
For using it, you almost always use it in deterministic algorithms, right?
361
00:25:21.600 --> 00:25:27.400
Like imagine like even from a strategic planning perspective, a human has to sit through, let's
362
00:25:27.400 --> 00:25:31.920
say in the before the agentic revolution, right?
363
00:25:31.920 --> 00:25:37.760
The human has to take the insights of this is the safety profile of like the region you're
364
00:25:37.760 --> 00:25:39.800
trying to do routing to.
365
00:25:39.800 --> 00:25:43.080
This is probably the traffic profile.
366
00:25:43.080 --> 00:25:48.280
They have to look at all of that together and make some decisions, right?
367
00:25:48.280 --> 00:25:54.120
I think what's powerful right now with the agentic wave has been the ability to unlock
368
00:25:54.120 --> 00:25:55.760
a reason.
369
00:25:55.760 --> 00:26:01.120
And the model is actually, when you give this context, can actually give you a synthesis
370
00:26:01.120 --> 00:26:04.880
of like, you know, this is how I reason about it and this is what you need to choose.
371
00:26:04.880 --> 00:26:10.600
If you're able to give a more higher order goal or an abstraction to say, my goal is to
372
00:26:10.600 --> 00:26:17.040
still do all these deliverables, deliveries, maybe my risk profile for safety is this.
373
00:26:17.040 --> 00:26:18.960
Is it possible or not?
374
00:26:18.960 --> 00:26:22.720
Can what kind of like a pastry should I need to add, right?
375
00:26:22.720 --> 00:26:27.400
So you're operating at a much higher level and you're kind of making that person who is
376
00:26:27.400 --> 00:26:33.240
doing that work or like defining that final structure or strategic plan.
377
00:26:33.240 --> 00:26:35.520
You're making them a bit more super human, right?
378
00:26:35.520 --> 00:26:40.240
Like where you are like overall giving it a bunch of business outcomes and you're able
379
00:26:40.240 --> 00:26:41.960
to reason through the data that's available.
380
00:26:41.960 --> 00:26:46.800
I think that's what's like, you know, the models have made possible today, which we haven't
381
00:26:46.800 --> 00:26:48.600
done that in the past.
382
00:26:48.600 --> 00:26:51.200
Or it's been very expensive to do that well in the past.
383
00:26:51.440 --> 00:26:55.480
Well, so that's interesting because every company should have that ability regardless
384
00:26:55.480 --> 00:27:01.920
of their business, but that changes the role of the person doing it.
385
00:27:01.920 --> 00:27:08.520
The input question, you know, to get to the outcome is very different when it's not a
386
00:27:08.520 --> 00:27:12.440
process you're running to gather the data, to run the thing.
387
00:27:12.440 --> 00:27:15.760
It's all kind of there and now you're just trying to get it out.
388
00:27:15.760 --> 00:27:20.240
How have you guys approached that because that's something I think every company can relate
389
00:27:20.240 --> 00:27:21.240
to right now is it?
390
00:27:21.240 --> 00:27:22.240
I know it's all there.
391
00:27:22.240 --> 00:27:25.560
I just don't know how to get the insights out of it.
392
00:27:25.560 --> 00:27:31.840
It's actually people do have it, right, except they used to hire people or specialists
393
00:27:31.840 --> 00:27:32.840
to doing it, right?
394
00:27:32.840 --> 00:27:38.760
Let's say, let's think about the scenario where if I'm an operations business who needs
395
00:27:38.760 --> 00:27:43.440
to do food delivery for like hundreds of my customers, right?
396
00:27:43.440 --> 00:27:49.280
Now I then have to, the only logical way for me to go do that with respect to understanding
397
00:27:49.280 --> 00:27:55.600
my risk, et cetera, is I would delegate that command to say either I have to go pay for
398
00:27:55.600 --> 00:27:59.840
a service which promises that they will do all of that on my behalf based on the inputs
399
00:27:59.840 --> 00:28:05.160
that I control, but the inputs then mean either they are static, don't really change based
400
00:28:05.160 --> 00:28:09.000
on the context of like something that's changed, et cetera.
401
00:28:09.000 --> 00:28:13.120
So all of that's like, you know, you either stuck with a little bit of like a tool that
402
00:28:13.120 --> 00:28:17.640
you bought, which was not necessarily taking advantage of what agent capabilities available
403
00:28:17.640 --> 00:28:18.640
today.
404
00:28:18.640 --> 00:28:24.240
And option is you then have to spend to hire like a quant person or a person who actually
405
00:28:24.240 --> 00:28:30.240
understand math with respect to like, you know, taking all of the data.
406
00:28:30.240 --> 00:28:34.560
And in many cases, people also don't have the data in a structured format, so then you
407
00:28:34.560 --> 00:28:38.600
will need to go hire a data engineer who would actually go to it, right?
408
00:28:38.600 --> 00:28:43.440
So the person who like most of these companies would have like let's say a route planner
409
00:28:43.440 --> 00:28:48.560
who's ultimately responsible for planning routes and they're not always set up to succeed
410
00:28:48.560 --> 00:28:51.920
without like all of the support system in place, right?
411
00:28:51.920 --> 00:28:59.360
So I almost imagine us providing these capabilities more towards one single human who's central
412
00:28:59.360 --> 00:29:07.120
to the operations and giving them the capabilities or the box where they are able to define or
413
00:29:07.120 --> 00:29:10.680
still operate at the higher order objectives that they would probably have given to like
414
00:29:10.680 --> 00:29:15.040
a stats person in their team and still get the results.
415
00:29:15.040 --> 00:29:19.480
And also have an interface that can reason about it and speak the language in which they
416
00:29:19.480 --> 00:29:24.440
understand in terms of like why certain decisions were made in what way.
417
00:29:24.440 --> 00:29:28.800
So they can kind of co-create like the strategic plan, right?
418
00:29:28.800 --> 00:29:34.120
You see the non-technical people have greater access to that data information and do that
419
00:29:34.120 --> 00:29:36.840
and are changing how your org functions as well.
420
00:29:36.840 --> 00:29:42.320
Obviously actually this is quite interesting because we're seeing for example like one
421
00:29:42.320 --> 00:29:47.440
of the things we're starting to see a lot is like product managers trying to land code,
422
00:29:47.440 --> 00:29:48.440
right?
423
00:29:48.440 --> 00:29:53.200
Like which is a common thing that's happening and designers making changes in production
424
00:29:53.200 --> 00:29:54.280
that's happening, right?
425
00:29:54.280 --> 00:30:01.360
So there's definitely like the barrier to entry has reduced when you think about people's
426
00:30:01.360 --> 00:30:02.720
like you know, roles.
427
00:30:02.720 --> 00:30:08.400
We still obviously, again, as we talked about at the start of this conversation around
428
00:30:08.400 --> 00:30:12.000
the governance model and still having the guardrails so things don't go off, right?
429
00:30:12.000 --> 00:30:14.800
But we are definitely seeing that and we are also encouraging that, right?
430
00:30:14.800 --> 00:30:21.480
So in turn to find limits and I totally believe this is happening also with all our customers
431
00:30:21.480 --> 00:30:26.240
where people are individually testing new capabilities that are out there and trying
432
00:30:26.240 --> 00:30:29.920
to find the boundary of what's possible, which is not possible.
433
00:30:30.080 --> 00:30:35.840
Going back to the governance, the cost, like what's your stack today and is there a consistent
434
00:30:35.840 --> 00:30:41.120
one or are you just letting everybody experiment within some sort of guardrail system?
435
00:30:41.120 --> 00:30:46.800
So within models, we have a blessed weight of models anybody can use within the company.
436
00:30:46.800 --> 00:30:51.840
And as I mentioned before, like access to these things, we've tried to centralize and have
437
00:30:51.840 --> 00:30:59.440
centralized gateways so that we have a degree of security audit possible or have a report
438
00:30:59.440 --> 00:31:06.160
that I can quickly look at in terms of spend across my organization and also I would go understand
439
00:31:06.160 --> 00:31:10.800
selectively myself, like you know, what has that spend produced, which I think is important.
440
00:31:10.800 --> 00:31:14.880
We built some degree of tooling to ensure I think two things primarily.
441
00:31:14.880 --> 00:31:20.320
One is like while we want to enable people to go faster, we want to do it in a responsible
442
00:31:20.320 --> 00:31:24.320
way when it comes down to production systems, etc. So I think that's one aspect of it.
443
00:31:24.400 --> 00:31:30.720
While we enable people to try all the new tools, we need to also have a little bit of an eye on
444
00:31:31.520 --> 00:31:36.480
what it means in terms of real spend, right? So we don't want to be blind to it.
445
00:31:36.480 --> 00:31:42.160
At this point, I haven't seen a phase where like why would be super concerned like about the cost
446
00:31:42.160 --> 00:31:46.800
really is like the problem that I need to go solve, right? I think having the
447
00:31:47.440 --> 00:31:52.880
auditability and somewhat of a chargeback model is important from a governance perspective.
448
00:31:52.960 --> 00:31:55.760
I think that's something everybody should do, right? Otherwise you're going to be in for
449
00:31:55.760 --> 00:31:58.960
surprises later on, so yeah. One of the most interesting things over the last
450
00:31:58.960 --> 00:32:04.640
couple of years is companies that didn't know they had so much value but are selling their data,
451
00:32:05.200 --> 00:32:11.840
their pictures, their everything to the bigger model companies. And I would think that you have
452
00:32:11.840 --> 00:32:17.760
a tremendous amount of data. I think about this a lot from the scooter market with so many
453
00:32:17.760 --> 00:32:23.040
hundred plus thousand vehicles on the street. Is there a hidden value in that data?
454
00:32:24.080 --> 00:32:29.920
Is that something that over time that physical presence will only become more and more important
455
00:32:30.560 --> 00:32:37.760
to figuring out how the real world works because of the ability to film, take pictures,
456
00:32:37.760 --> 00:32:42.400
label? Is that like the next generation of true value unlock in these models?
457
00:32:42.480 --> 00:32:48.560
When we think about what we sell to our customers, we sell products that are like you know
458
00:32:48.560 --> 00:32:54.080
fundamentally the foundation is based on sensors that we are selling but what we eventually sell
459
00:32:54.080 --> 00:33:00.960
to customers is actually applications that help with our customer safety or efficiency goals
460
00:33:00.960 --> 00:33:05.360
that's or sustainability goals, right? So that's generally our mission. Now when you really unpack
461
00:33:05.440 --> 00:33:13.280
and transit that most of our customers spend like if you go look at their particular income statement
462
00:33:13.280 --> 00:33:21.440
they spend across safety in terms of insurance cost or like you know accidents which they want
463
00:33:21.440 --> 00:33:26.560
to reduce like you know or like operational cost when you think about like fuel cost like
464
00:33:26.560 --> 00:33:31.600
is an example maintenance is another huge example where like people spend a lot of money,
465
00:33:31.600 --> 00:33:36.720
like a lot of the physical world companies, collective intelligence that we gather from our
466
00:33:37.600 --> 00:33:43.600
install base or distribution actually enables us to build unique products which help
467
00:33:44.560 --> 00:33:51.760
the collective good of like passing down efficiency to across our customers is both from a length
468
00:33:51.760 --> 00:33:56.320
of any of these dimensions right any of the goals take safety for into account or maintenance
469
00:33:56.960 --> 00:34:02.240
cost, right? So I see a world where the world gets more predictive in terms of maintenance, right?
470
00:34:02.240 --> 00:34:07.760
The world will get more predictive in terms of fuel cost management. Today a lot of these things
471
00:34:07.760 --> 00:34:14.400
are done a lot more reactively and we are getting predictive on safety already like which is
472
00:34:15.040 --> 00:34:21.040
which is one of the biggest value propositions we've seen from some Sarah's perspective because
473
00:34:21.040 --> 00:34:25.200
we know statistically from people using our technology that like you know there's we reduced
474
00:34:25.200 --> 00:34:30.480
incidents, we've actually reduced insurance bills and generally kept people safe on the road
475
00:34:30.480 --> 00:34:35.760
right which is a good cause I think and obviously I'm proud to be a part of. I generally think we have
476
00:34:35.760 --> 00:34:42.160
an opportunity with this next wave of like you know we have fuel consumption data, fuel changes
477
00:34:42.160 --> 00:34:48.720
across different like vehicles, the other important way to add is like people's driving
478
00:34:48.720 --> 00:34:54.880
styles influence fuel consumption or what, right? We have all that data too like who's driving
479
00:34:54.880 --> 00:35:00.080
that cab at that particular point in time what did they do like did they harsh break did they
480
00:35:00.080 --> 00:35:06.320
accelerate a lot where did they do that was it on the freeway was it on the like you know city road
481
00:35:06.320 --> 00:35:11.600
etc so we have all of that context which helps us build intelligence to the point where like you know
482
00:35:11.600 --> 00:35:15.920
we know what a baseline good looks like we know what opportunity exists and we can actually like
483
00:35:15.920 --> 00:35:23.680
get a lot more efficiency back to our customers right so so I do think the data is pretty valuable
484
00:35:23.680 --> 00:35:27.680
and especially when you think about the collective intelligence in which how you can apply
485
00:35:27.680 --> 00:35:31.840
we have a unique data set which actually enables a lot of the applications and the end
486
00:35:32.720 --> 00:35:38.080
value that we can pass along to our customers. What's overhyped right now? What's overhyped huh?
487
00:35:38.640 --> 00:35:42.400
I don't know I think there's a little bit of a humorism which I think is definitely overhyped
488
00:35:43.600 --> 00:35:49.920
like an example would be you know we we have Claude Maitos the models like you know
489
00:35:50.000 --> 00:35:54.880
stalled I guess the narrative originally was like you know it's just too powerful and going to find
490
00:35:54.880 --> 00:35:58.880
like all the security issues in the world and like but later there's also an alternate narrative
491
00:35:58.880 --> 00:36:05.920
around compute limitations that they have but I generally think as these models get better like you know
492
00:36:07.040 --> 00:36:11.440
it's again like black hats versus white hats right so there's going to be models which are also like
493
00:36:11.440 --> 00:36:18.480
more smarter about patching issues before right so you just I almost see it as like a one-time
494
00:36:18.480 --> 00:36:24.640
shift in terms of like security processes and then you have the same capabilities available to
495
00:36:24.640 --> 00:36:31.040
pretty much both sides I kind of don't believe there's like a very big the negativity around like
496
00:36:31.040 --> 00:36:35.920
oh is this something that's going to like completely shift everything right. I love hearing that
497
00:36:35.920 --> 00:36:41.680
because I'm scared of that so when technical when super technical people are not as scared of it it
498
00:36:41.680 --> 00:36:46.400
makes me a little happy because I treat Claude with respect just to make sure that it doesn't kill me
499
00:36:46.400 --> 00:36:51.760
well yeah yeah no I think there's there's a little bit of like we we do tend to project out
500
00:36:51.760 --> 00:36:57.280
capabilities right and I do think there's going to be a lot of interesting things happening
501
00:36:57.280 --> 00:37:03.840
in the world right but then I wish we also focused a lot on the positivity of things that's
502
00:37:03.840 --> 00:37:09.840
possible right and I generally am a believer that these tools will be more of an enabler
503
00:37:10.720 --> 00:37:18.560
uh for like many more things or opportunities that we are not thinking about is how I generally see
504
00:37:18.560 --> 00:37:24.240
it a lot of what we talked about in terms of predictive intervention none of these things were
505
00:37:24.240 --> 00:37:29.120
possible without like a lot of the reasoning capabilities that's available so yeah so I
506
00:37:29.120 --> 00:37:33.440
generally believe that there's like going to be a lot a lot of good stuff right the other part is
507
00:37:34.080 --> 00:37:40.080
it's also going to enable a lot more access which is generally good for humanity right like where
508
00:37:41.360 --> 00:37:48.320
which I think is very interesting and I do believe that right so as cost of making things go
509
00:37:48.320 --> 00:37:52.480
down or like you know there's more people are enabled to make things so you're generally going to see
510
00:37:52.480 --> 00:37:58.720
like accessibility increase for capabilities in the real world yeah so the next few years will be
511
00:37:58.720 --> 00:38:03.200
interesting obviously it's a very exciting time like every day is we learn something new and
512
00:38:04.240 --> 00:38:07.680
I'm definitely looking at it more through a positive lens I would say
513
00:38:08.480 --> 00:38:12.960
I love it well thank you for joining us I really appreciate it and sharing your thoughts and insights
514
00:38:12.960 --> 00:38:23.760
yeah enjoy talking to you keep
