1
00:00:00,040 --> 00:00:04,280
The answer to every problem is a model. You're tempted to write a bunch of code for it? No.

2
00:00:04,280 --> 00:00:08,360
Instead, collect some data. Figure out what a good output would look like. Figure out what model

3
00:00:08,360 --> 00:00:12,280
architecture is right. Train the model and put it in a loop so that you can keep improving the

4
00:00:12,280 --> 00:00:16,840
model. That mindset just allows you to get out of that decision making process and hire people who

5
00:00:16,840 --> 00:00:21,920
are very versatile and problem solvers. It's simple. Once you sort of just internalize that

6
00:00:21,920 --> 00:00:26,960
worldview, agents and AI's will just use the web a lot more than humans ever have. New business

7
00:00:26,960 --> 00:00:32,120
models will be needed, and now there's going to be like a second birth of what the future of the open

8
00:00:32,120 --> 00:00:38,520
web is a parallel web, a parallel web, a few minutes. If we just figured out how to incentivize

9
00:00:38,520 --> 00:00:44,720
being open, how to reward high quality, unique, differentiated content that we get to keep the

10
00:00:44,720 --> 00:00:45,600
web open.

11
00:00:57,440 --> 00:01:02,009
Welcome to grit. I'm Joubin, partner at Kleiner Perkins, a show where we go beyond the highlight

12
00:01:02,049 --> 00:01:06,969
reel and explore the personal and professional challenges of building history making companies.

13
00:01:07,010 --> 00:01:13,289
Today on the show, we've got Parag Agrawal, former Twitter CEO and now founder and CEO of parallel,

14
00:01:13,330 --> 00:01:17,969
the company building web infrastructure for AI. Before starting parallel, Parag spent more than a

15
00:01:17,970 --> 00:01:23,649
decade at Twitter, where he rose from engineer to CTO to CEO. Now he's asking a different question

16
00:01:23,650 --> 00:01:29,290
what happens when AI becomes the web's second user? We're also joined by my partner, Mamoon Hamid.

17
00:01:29,290 --> 00:01:35,889
Enjoy. When we were sitting in the kitchen right now. Um, I, we were talking about

18
00:01:35,889 --> 00:01:42,249
fundraising, all sorts of stuff. And, uh, I said, can you take a, can you take a breath? And you're like,

19
00:01:42,290 --> 00:01:49,289
dude, you take a breath, this company is dead. And I was like, okay, that's a fair and

20
00:01:49,289 --> 00:01:53,809
dramatic point of view. I like, but okay, like, I understand. And he's like, do you not feel that way?

21
00:01:53,849 --> 00:02:00,649
And yeah, Parag asked me like, do you feel that way? And I said, well, I definitely feel that

22
00:02:00,649 --> 00:02:05,449
way, except I just want to make sure that whatever I'm running at is the right thing, because there's

23
00:02:05,449 --> 00:02:11,849
so many things that are feel like they're on fire and it's full chaos. And let me ask you, can you

24
00:02:11,850 --> 00:02:16,289
take a breath? No, but I just want to make sure. He's money in company. Right? I know, I know, he just

25
00:02:16,289 --> 00:02:20,769
told me, I didn't know I didn't like I didn't. Know it until. That company has a whole team I know

26
00:02:20,770 --> 00:02:26,409
elsewhere. It's half your team size. Yeah. At this point, I did not know that. Yeah. Walking into it. I

27
00:02:26,410 --> 00:02:31,049
found out. In the kitchen and I'm on his board. Yeah. That's right, that's right. I just wanted I

28
00:02:31,089 --> 00:02:36,209
you know, your point resonated. The reason I took a beat was because I was like, the temptation when

29
00:02:36,210 --> 00:02:41,649
everything feels like it's so insane, where if you take a breath, you die, makes you feel like you

30
00:02:41,649 --> 00:02:47,169
have to do everything all at once. And for me, I was like, okay, what am I actually doing that moves

31
00:02:47,169 --> 00:02:51,809
the needle? Because otherwise I could be doing stuff, making myself feel like the most productive

32
00:02:51,850 --> 00:02:55,689
human in the world. We can. We can talk about all. These things like. Yeah, we're live, we. Can just die.

33
00:02:55,689 --> 00:03:02,419
We're already live. We just go. Just go. No no, no. So I am there. So

34
00:03:02,460 --> 00:03:07,259
I'm the kind that is extraordinarily comfortable dropping balls. And I don't know if I should say

35
00:03:07,259 --> 00:03:12,579
this in front of my board, but, like, I drop balls all the time, and so there'll be, like, a bunch of

36
00:03:12,580 --> 00:03:17,979
things. I'll come at me. But every week, every day I have like a few things that I care about or that

37
00:03:17,979 --> 00:03:22,419
matter. And I'm just going to do those things and there'll be a bunch of other balls that are less

38
00:03:22,419 --> 00:03:29,059
important that I will drop. And I think ultimately catch them, though, right? No. Someone will. Someone

39
00:03:29,059 --> 00:03:35,779
will. I am betting that my team, because I drop balls, is just built in the structure of

40
00:03:36,579 --> 00:03:43,460
evolving to catch the balls I drop. And so now you build this sort of culture, this team where.

41
00:03:43,740 --> 00:03:47,820
And if it's important enough, it's going to come back to me. It's going to bounce back to me. What

42
00:03:47,820 --> 00:03:54,179
if the ball that is dropped is only known to you? You know that is a problem and you don't drop

43
00:03:54,179 --> 00:04:00,749
those ones. So you again. You don't get to drop random balls. Yeah, you drop the balls. You can

44
00:04:00,750 --> 00:04:05,350
drop partly by design. So you have to prioritize the ones that you don't drop. How many tiers of

45
00:04:05,350 --> 00:04:11,790
memory do you have. So you have cash. And then the second tier I have a hard disk drive. You know, you

46
00:04:11,790 --> 00:04:18,589
have a flash memory in between. Uh, I think it's like. A list of three things every week. Every day.

47
00:04:18,590 --> 00:04:25,309
It's like three things. Can you give an example of what I would be surprised

48
00:04:25,549 --> 00:04:29,189
that you're dropping, that you're willing to drop? Like, what are some balls that you're like, yep. No

49
00:04:29,189 --> 00:04:34,989
problem. Let's see if it bounces back there hard, like in the moment that you're like, oh, I really

50
00:04:34,989 --> 00:04:39,229
don't. I should respond to this, but I can't because it's not the most important thing. So it's

51
00:04:39,230 --> 00:04:46,070
it's you think it would be that deliberate and intentional a process? It's not

52
00:04:46,109 --> 00:04:52,390
right. It is more that like for example, I separate out reading and responding to emails,

53
00:04:53,029 --> 00:04:59,710
the two different work cycles for me. But instead of sitting down and trying to respond to

54
00:04:59,749 --> 00:05:06,590
emails, I read all emails. Then I sit down in the morning and I'm like, okay, what are the

55
00:05:06,590 --> 00:05:12,789
three things this week? Or what are the three things today that I need to proactively do? And a

56
00:05:12,790 --> 00:05:18,590
bunch of things you do reactively, but you can't let the reactive stuff swarm the proactive stuff.

57
00:05:19,070 --> 00:05:24,189
And so I'll write down these proactive things, and they're informed by the things that are in the

58
00:05:24,190 --> 00:05:27,390
back of my mind because I've been reading emails, because I've been thinking, because I've been

59
00:05:27,390 --> 00:05:32,350
doing stuff, because I've been hacking. Right. But I'll write three things down and I will make sure

60
00:05:32,350 --> 00:05:37,269
those three things get done. Now, other things, almost by definition, have a risk of getting

61
00:05:37,269 --> 00:05:44,190
dropped. Now I'll try to go skim at my. I reread my email all the time and if it's

62
00:05:44,190 --> 00:05:50,149
important enough, it is bubble into the top three and make its cut. So I'm not making the decision

63
00:05:50,149 --> 00:05:55,358
that I'm going to drop this ball. I am making the decision that it like, which is like. And I'm not

64
00:05:55,399 --> 00:05:59,599
even making the decision. This is not in the top three. But then when you do the top three, you're

65
00:05:59,600 --> 00:06:04,679
not looking at your inbox. If something was important enough, it better stick in your head as

66
00:06:04,720 --> 00:06:08,440
being an important enough upside opportunity, downside protection opportunity. Something should

67
00:06:08,440 --> 00:06:15,239
do. And so those are the three things that I must do. Now I have time beyond doing those three

68
00:06:15,239 --> 00:06:20,359
things. And so I will go respond to a lot more emails. But then I'm not like some machine trying

69
00:06:20,359 --> 00:06:25,199
to prioritize every decision because that's exhausting too. Okay. Can you when you were when

70
00:06:25,200 --> 00:06:31,969
you were in the kitchen, you were describing how, uh, Friday afternoons you have like demos 430 to

71
00:06:31,969 --> 00:06:38,920
6 and it's your way as a founder to kind of internalize the rhythm of the business on

72
00:06:38,920 --> 00:06:42,760
the on the product side. And then going into Friday night, you feel like, all right, things are

73
00:06:42,760 --> 00:06:48,199
happening like, I know what's happening. And then Saturday morning you'll wake up with some

74
00:06:48,200 --> 00:06:55,049
existential panic that something is breaking. Can you describe that? Yeah. So this is

75
00:06:55,049 --> 00:07:01,969
my rhythm weekly rhythm. So there is like a team is amazing. We're doing a bunch of

76
00:07:01,969 --> 00:07:08,890
really cool things. And I am often not even fully aware of really good things that are cooking

77
00:07:08,890 --> 00:07:13,889
every week, or some cool result that someone came up with, or some cool insight someone came up with

78
00:07:13,889 --> 00:07:20,729
and they shipped something. And the best way I found is I like to see things which are like

79
00:07:20,730 --> 00:07:26,649
raw and in progress. So our demos are designed to be not like, oh, here's a package product that's

80
00:07:26,650 --> 00:07:32,729
going out to customers on Monday. That's not what Friday evening demos are for. Friday evening demos

81
00:07:32,730 --> 00:07:39,089
are for. I was doing this this week. It's not done. It's raw. Here's why I'm excited about it. And you

82
00:07:39,089 --> 00:07:45,449
should all know about it. And so that is the stuff that like if I'm prioritizing the three things

83
00:07:45,450 --> 00:07:52,289
that matter, that is the stuff that never Bubbles up and then I just wouldn't

84
00:07:52,329 --> 00:07:57,369
know. And I would find out when. It's like a packaged up thing. And that's what sort of gets

85
00:07:57,370 --> 00:08:04,048
the both the job satisfaction and the ideas and the rolling for me. And so

86
00:08:04,530 --> 00:08:10,129
getting that dose of infectious excitement about the work we are doing every Friday is really

87
00:08:10,130 --> 00:08:17,009
rewarding because you start thinking about it. And then some of those boil into, oh, why did

88
00:08:17,010 --> 00:08:20,769
I? There's so much more we could do with this thing. And so now you're paranoid about like, a

89
00:08:20,769 --> 00:08:27,449
missed opportunity. Some of those are about, wait, this is so cool. If someone else figured this

90
00:08:27,450 --> 00:08:32,809
out and then like, your head, like your mind goes in many, many different directions, and then you

91
00:08:32,809 --> 00:08:39,249
have to boil all of those things into the one that matters. To be truly

92
00:08:39,969 --> 00:08:45,690
either paranoid about or excited about, or is actionable. Right? And by Monday, you have to make

93
00:08:45,690 --> 00:08:52,459
it actionable. Can we go back to the top three things? Yeah. So a few weeks ago, you showed

94
00:08:52,460 --> 00:08:57,099
me this vibe coded thing that you built for yourself. Can you talk about that? Of course. Yeah.

95
00:08:57,140 --> 00:09:01,379
Why not? So you showed it to me. And it's incredible, actually. And I think it was we were

96
00:09:01,379 --> 00:09:07,380
having lunch together at our CEO summit, and you showed me the app, and I'm like, this is mind

97
00:09:07,380 --> 00:09:13,459
blowing. What was even more mind blowing was that by 1 p.m. when we were having lunch together, you

98
00:09:13,460 --> 00:09:20,379
had spent, I believe, $378 in on tokens that day. And I asked

99
00:09:20,380 --> 00:09:25,619
you, like, what the heck are you doing in the background? And you're you're clearly token maxing.

100
00:09:25,620 --> 00:09:30,859
And I think at the time you're saying, hey, like, I'm just trying to do as much as possible and

101
00:09:30,859 --> 00:09:34,659
actually don't care right now. Two questions embedded. One is like, how does that help you with

102
00:09:34,659 --> 00:09:40,658
the top three things for that day? And then how does it help you, those drop balls that come back

103
00:09:40,659 --> 00:09:45,539
up. How does that help you with that? And then thirdly anxiety. How does it help you on Saturday

104
00:09:45,539 --> 00:09:49,909
morning Oh, you know, the anxiety is less because I just built this amazing thing that helps me be

105
00:09:50,229 --> 00:09:56,629
less anxious. So, number one, me building that wasn't one of those top three things. Just as a

106
00:09:56,630 --> 00:10:03,549
clarification, it is just because I was just excited about it. I wanted to play with it and I'm

107
00:10:03,550 --> 00:10:10,109
going, I was I built it with the rationale of making a two, understanding that if I built a

108
00:10:10,110 --> 00:10:15,229
thing that was that could see everything I can see, and it was doing all the work proactively

109
00:10:15,230 --> 00:10:22,029
that I should be doing with the most limited access that I could give it. Um, what would that

110
00:10:22,030 --> 00:10:28,788
thing do? How much would it rely on all of my internal data, like slack and

111
00:10:28,789 --> 00:10:35,669
granola and emails versus the open web? Like, we obsess about the open web at parallel. And I

112
00:10:35,669 --> 00:10:42,629
really like I tell investors, I tell customers that when you do work, you're combining across all

113
00:10:42,629 --> 00:10:46,270
the information available to you, whether it is your personal information, whether it is your

114
00:10:46,270 --> 00:10:52,869
company information, or whether it is the open web. And so this project's rationale was to see what

115
00:10:52,869 --> 00:10:59,549
the machine likes to spend on parallel versus everything else in terms of tokens in a in the

116
00:10:59,550 --> 00:11:06,349
purest form I could while being productive on my behalf. Uh, and so this was a journey to

117
00:11:06,390 --> 00:11:12,990
one be useful and discover that for my role, uh, it has ended up

118
00:11:12,990 --> 00:11:18,709
being pretty useful for me, for my top three things. Like it makes me feel more secure that I

119
00:11:18,710 --> 00:11:25,710
am not dropping the important balls because it is also paranoid on my behalf. It

120
00:11:25,710 --> 00:11:32,710
is inferring my priorities. Uh, I write down my top three priorities so it knows what those are. Uh, it

121
00:11:32,710 --> 00:11:38,749
will push back on other things. I'm balls, I'm dropping, and I'm slightly more deliberate about

122
00:11:38,789 --> 00:11:44,799
dropping those balls than I was before I had this thing. It's token uses outrageously inefficient

123
00:11:44,800 --> 00:11:51,639
and I wouldn't read too much into the goal. Isn't token maxing the the. Yeah, I just it was not

124
00:11:51,680 --> 00:11:57,399
worth it to optimize. But I think it can be materially materially optimized. Just not. The

125
00:11:57,400 --> 00:12:02,960
second part of the question was would you let everyone on the parallel team go build that

126
00:12:02,960 --> 00:12:09,760
version of themselves and spend, let's say, $500 a day on things to make

127
00:12:09,760 --> 00:12:16,399
people less anxious? Probably, yeah. I haven't thought it through. I haven't done the full math,

128
00:12:16,400 --> 00:12:23,200
but yeah, if once useful. Totally. I think it's like the utility. It takes some real work and

129
00:12:23,200 --> 00:12:29,599
effort to make it actually useful to who you are, at least today. For me. Like it like out of the box.

130
00:12:29,600 --> 00:12:36,559
Nothing is useful unless you mold yourself and mold it around you. And I don't know if you can

131
00:12:36,840 --> 00:12:42,809
if I yet know how to productize it for everyone and all the different roles at parallel and but

132
00:12:42,849 --> 00:12:49,330
yeah, anyone who does the work and can drive through value like $50 a day's fine if you get

133
00:12:49,330 --> 00:12:54,168
really maximize yourself. Yeah, I remember you had an offsite coming up the week thereafter. So you

134
00:12:54,169 --> 00:12:58,729
were really token maxing because you had some stuff running in the background. And I imagine

135
00:12:58,729 --> 00:13:05,689
today less tokens have been used. I have since optimized it. It uses far less tokens, and it

136
00:13:05,690 --> 00:13:10,529
works in certain periods more intensely and other periods less intensely. And how much do you think

137
00:13:10,530 --> 00:13:16,368
you'll spend today? Like 100. Okay, a. Couple hundred bucks. Do you care? Like, if you look at your

138
00:13:16,369 --> 00:13:23,249
portfolio now, like, obviously token spend is going up. Like, we're spending a lot of money

139
00:13:23,250 --> 00:13:29,649
on cloud code, basically everybody. Are you starting to look at that like, uh, as you dig right

140
00:13:29,650 --> 00:13:33,369
now, do you actually care or are you more curious, like, is this do you think this is a line item

141
00:13:33,370 --> 00:13:38,089
that's going to show up at the board level where you're like, this is getting insane. Yeah, I ask

142
00:13:38,090 --> 00:13:44,249
most of my CEOs at this point and we talk about it because I'm trying to understand the curve of

143
00:13:44,250 --> 00:13:50,450
token use, like use and cost coming down. And at what point does it sort of stabilize or does it

144
00:13:50,450 --> 00:13:55,449
keep going through the roof where if it's right now, maybe like one tenth of a person's salary,

145
00:13:55,489 --> 00:14:01,889
does it go to 50% of someone's salary to one times a person's salary? And I don't know where it

146
00:14:01,929 --> 00:14:05,889
sort of lands, but I'm trying to make a sense. And that's part of our job is to understand, like,

147
00:14:05,889 --> 00:14:12,889
where do these, uh, macro trends land for, you know, for knowledge work for other types of work in

148
00:14:12,890 --> 00:14:19,329
the future? I will say my I have two takes on this. Number one, I think

149
00:14:19,369 --> 00:14:25,569
most token maxing a lot of token maxing that is happening isn't that valuable,

150
00:14:26,570 --> 00:14:32,649
including a lot that I have done over time in the last few months. So at the same time,

151
00:14:33,449 --> 00:14:40,219
if you have to. So clearly there is a failure mode to just token max. Not because it's

152
00:14:41,020 --> 00:14:47,499
expensive in terms of tokens, but it's expensive in terms of like AI,

153
00:14:47,540 --> 00:14:54,218
psychosis or time you spend token maxing, right? It takes some effort and

154
00:14:54,219 --> 00:15:00,499
energy to token max. And sometimes it's actually not super productive. So I think it's there are

155
00:15:00,499 --> 00:15:05,139
some people who are really, really good at it and actually drive themselves and accelerate

156
00:15:05,140 --> 00:15:11,819
themselves forward. And not everyone is. So I think it's really important to not measure by token

157
00:15:11,819 --> 00:15:18,059
maxing. The second framework that I have is like, listen, none of us know what the

158
00:15:18,260 --> 00:15:24,700
perfect line is at any time on this. Which way would you rather be wrong on? Right? Would you

159
00:15:24,700 --> 00:15:29,699
rather be wrong because you're too conservative and just not spending tokens? Because you're too

160
00:15:29,700 --> 00:15:35,899
afraid of either spending tokens the wrong way, or spending your time on tokens the wrong way? Or

161
00:15:35,900 --> 00:15:41,149
would you rather go on the other side and spend a little bit too much time with the models? Aren't

162
00:15:41,150 --> 00:15:46,909
good enough. Spend a little bit too much tokens, and at least my personal decision is I'd rather

163
00:15:46,950 --> 00:15:53,069
be wrong on the token maxing side than the non right, and you're never going to be right. But if

164
00:15:53,110 --> 00:15:59,389
you know which way you'd rather be wrong, that tells you what you should do. You track it. And

165
00:15:59,390 --> 00:16:05,949
Roadrunner. Uh, my co-founder Eugene tracks it. We first it was like,

166
00:16:06,469 --> 00:16:12,069
let it rip. That was like, that's been the last six months. Was let it rip and everyone let it rip. And

167
00:16:12,069 --> 00:16:17,909
then you're like, okay, this is getting expensive. And so then we started to make sure that it's all

168
00:16:17,909 --> 00:16:23,510
in one credit card. So you can track like, all right, where is it actually being spent. And then

169
00:16:23,549 --> 00:16:29,229
you started to ask yourself, okay, well what model is it being spent on. Okay. So then we started

170
00:16:29,229 --> 00:16:33,429
figuring out like, all right, should we just let everybody use whatever model. And so we've kind of

171
00:16:33,469 --> 00:16:40,229
like kept pushing in a little bit. And the short answer is we track it, but we

172
00:16:40,229 --> 00:16:47,229
don't enforce anything today. Um, we're also watching to

173
00:16:47,269 --> 00:16:53,469
make sure that at some point, if an engineer is spending their salary on tokens,

174
00:16:54,189 --> 00:17:00,869
they better be pretty good. You know, they better be pretty good. And so we I would say we,

175
00:17:00,870 --> 00:17:07,469
um, we watch it, but our best engineers are doing such unique techniques and how these agents are

176
00:17:07,469 --> 00:17:12,069
working that I think limiting the ways that they're using these tokens feels like a mistake

177
00:17:12,069 --> 00:17:17,550
to me, because they're on the bleeding edge of pushing the boundaries of like, sequencing five

178
00:17:17,550 --> 00:17:22,109
agents together to do all these interesting things. And I would rather spend more now to let

179
00:17:22,109 --> 00:17:27,029
them be on the bleeding edge. And look, they happen to also be the best engineers that we have. Do you

180
00:17:27,030 --> 00:17:33,709
agree with that? Yeah, I think there are. We definitely see, I think one way

181
00:17:33,760 --> 00:17:39,319
people perhaps undersell the use of tokens is like, oh, you're doing the same work that you would

182
00:17:39,319 --> 00:17:45,399
do faster. I actually think the what excites me is people

183
00:17:46,959 --> 00:17:53,959
like broadening what they could have done, or taking more end to end ownership of things by

184
00:17:54,400 --> 00:18:01,159
using agents and models. So the example is like a a back end engineer will now do more front end

185
00:18:01,159 --> 00:18:08,119
work, or a front end engineer will change more of your API ship or. Someone will

186
00:18:08,120 --> 00:18:13,479
use agents to do a bunch of security reviews. Someone will use agents to push a bunch of

187
00:18:13,479 --> 00:18:19,959
optimizations which they would otherwise not have been doing, and design something and do it. So it's

188
00:18:20,000 --> 00:18:26,999
actually people expanding their span of influence and ownership on the product. That's

189
00:18:27,000 --> 00:18:33,529
where I think there is true business value being generated in my mind. And so

190
00:18:33,769 --> 00:18:39,688
people who end up sort of feeling comfortable about that and like really pushing their own sort

191
00:18:39,689 --> 00:18:46,010
of learning curves really fast. That's where I think the real value is, at least for me, that I

192
00:18:46,010 --> 00:18:52,889
see. Mhm. Yeah. The, the, the real life use case that came up this morning for me was every time

193
00:18:52,930 --> 00:18:59,329
a customer wants to see the proof in the pudding of Roadrunner, they want to see it with their SKUs.

194
00:18:59,329 --> 00:19:05,369
They want us to do like hairy SKUs and show a demo of what it looks like. And so I asked our

195
00:19:05,370 --> 00:19:09,649
head of solutions architecture, how long does it take you right now to like, set up this

196
00:19:09,650 --> 00:19:15,448
environment? It's like hand-to-hand combat. And she's like a couple days and I'm like, what would

197
00:19:15,449 --> 00:19:21,369
you need to make it a couple of hours? And so just like I started like peeling back like a few whys

198
00:19:21,370 --> 00:19:27,129
and eventually it was like, well, I need an engineer and I need these skills that they have

199
00:19:27,130 --> 00:19:33,929
to do these things. And then I'm like, okay. But like, can't you just embed those skills into

200
00:19:33,930 --> 00:19:39,529
Claude and then start to try? By the way, you might fail, but at least start to try to automate this

201
00:19:39,530 --> 00:19:46,489
away, because if we have 15 customers at the same time, this isn't going to work. So like

202
00:19:46,529 --> 00:19:52,609
either we go hire another ten people or we figure out how to automate this process and we can hire

203
00:19:52,650 --> 00:19:59,569
three. And I think like that to me is where I'm like, yeah, let it rip. Like go, go. Because the the

204
00:19:59,569 --> 00:20:06,289
trade off is I have to go hire more hands and I'd prefer not to. That is interesting. Can I at least

205
00:20:07,209 --> 00:20:14,088
I struggle with this. I think there's this. We are hiring engineers as fast as

206
00:20:14,569 --> 00:20:21,290
we can. Despite all of this today. And I find that counterintuitive,

207
00:20:22,170 --> 00:20:27,929
but also not because if you think about it, like if you have an extremely

208
00:20:28,569 --> 00:20:35,379
interesting set of things that you can build, which are extremely valuable, and your customers

209
00:20:35,420 --> 00:20:41,379
need it. Even if each engineer is more productive, buy a lot

210
00:20:42,379 --> 00:20:47,859
more is better until you start saturating your opportunity space, right? So I do think what's

211
00:20:47,860 --> 00:20:51,579
interesting is, and I think there is an explanation for why there is a class of

212
00:20:51,580 --> 00:20:57,259
businesses growing at crazy rates. It is because you get to multiply the two effects together,

213
00:20:57,420 --> 00:21:04,300
which is like how fast and well can you hire and how quickly can you channel your

214
00:21:04,339 --> 00:21:11,300
capacity to build into really valuable things, and how fast your customers can adopt

215
00:21:11,300 --> 00:21:16,099
these new things that you're making available? And all of these things are multiplying together into

216
00:21:16,099 --> 00:21:23,059
these outrageous growth rates that we're starting to see now. And so I have

217
00:21:23,060 --> 00:21:29,349
never once sat down and been like, oh, Everyone's going to be so much more

218
00:21:29,349 --> 00:21:35,829
productive. So maybe I need fewer people. Yeah. And it's the exact opposite that you feel. Yeah, I

219
00:21:35,829 --> 00:21:42,309
think I agree with you, except, like, uh. So we have insatiable appetite to hire as many engineers as

220
00:21:42,310 --> 00:21:46,989
we can, which is. Yeah, counterintuitive. Maybe. I don't even know if it's that kind of. It's not

221
00:21:46,989 --> 00:21:51,069
counterintuitive. Yeah. It's just like, you know, you guys have the tools, you know how to use them. You

222
00:21:51,069 --> 00:21:57,589
know how to build more product. You have large, wide, blue oceans to go after. Uh, and yeah, like go

223
00:21:57,630 --> 00:22:04,429
build more. And, uh, if you know, and you've got capital, you've got all the ingredients. Yeah, yeah.

224
00:22:04,430 --> 00:22:10,709
You're welcome. Uh, to go build and sell and generate revenue. And which is why it goes back to

225
00:22:10,709 --> 00:22:15,949
the point, the, the multiplicative effects of all this are that companies are growing at rates

226
00:22:15,949 --> 00:22:22,468
we've never, ever seen before. And that's great. That's amazing. And so if you're not

227
00:22:23,390 --> 00:22:29,709
if you're not growing up those rates uh, you're a good company. Not a great company. Mhm. But don't

228
00:22:29,709 --> 00:22:34,309
you think there is certain types of work that, like the truly great engineers, just don't want to

229
00:22:34,310 --> 00:22:39,468
do. Like this example of like seeding a customer environment and having to do that 20 times like

230
00:22:39,469 --> 00:22:46,388
that seems to me a little bit adjacent to core product development. Or do you also

231
00:22:46,389 --> 00:22:53,389
have infinite appetite for hiring hands and that type of work as well? No, but I think instead of

232
00:22:53,390 --> 00:22:59,669
defining what kind of work, I think great engineers want to do

233
00:23:01,349 --> 00:23:08,349
high impact work, which happens to be hard. So if you have great

234
00:23:08,349 --> 00:23:13,269
engineers, you have to kind of trust their judgment and their taste that if they can

235
00:23:13,870 --> 00:23:19,948
automate something away with the existing tools, they will. And so

236
00:23:20,910 --> 00:23:26,319
we just we have a few simple things we talk about, which is like beyond the meta layer, if you can

237
00:23:26,359 --> 00:23:33,119
like use an AI to automate something like there are big parts of our system which are loops

238
00:23:33,120 --> 00:23:38,759
with models in them, which in an old world you could have thought that this was a.

239
00:23:40,879 --> 00:23:45,279
Like an engineer would have done that, right. Like if you're doing quality work, like if I think back

240
00:23:45,280 --> 00:23:49,759
to our time at Twitter when we were building, like these sort of recommendation systems, and you were

241
00:23:49,760 --> 00:23:56,759
doing a bunch of work to improve its quality and run, build better models, improve them.

242
00:23:56,839 --> 00:24:03,799
And there were engineers, a bunch of them, iterating on models by using intuition,

243
00:24:03,800 --> 00:24:09,399
running a bit as finding new features, building data pipelines. You can take some of those things

244
00:24:09,400 --> 00:24:16,160
and now have them be in a more automated loop, because wherever there is a nice, clean way

245
00:24:16,160 --> 00:24:21,559
of measuring things, right, you can automate a lot of those things, and not a lot breaks because

246
00:24:21,560 --> 00:24:27,569
they're going to go into opaque models anyways, right? So those things are in our system and we

247
00:24:27,569 --> 00:24:34,568
don't even think about them as, oh, this was work that an engineer did earlier which

248
00:24:34,569 --> 00:24:41,449
are being done by models now. Does that make sense? And so to me, it's like what an engineer

249
00:24:41,490 --> 00:24:48,170
does become more and more meta over time as you find feedback loops that you

250
00:24:48,170 --> 00:24:53,129
trust and can automate and make measurable, and you can trust on your models and then you don't,

251
00:24:53,170 --> 00:24:59,888
you have to look less under the hood of how it's working. Right. And our goal is

252
00:24:59,889 --> 00:25:06,170
to push more things inside learned models rather than in code.

253
00:25:07,050 --> 00:25:13,809
Right. So if there is a I've been saying one sentence internally quite a lot like the answer

254
00:25:13,810 --> 00:25:20,770
to every problem is a model. So and which is sort of how we work. So you have a problem.

255
00:25:20,770 --> 00:25:26,129
You're tempted to write a bunch of code for it? No. Instead, collect some data. Figure out what a good

256
00:25:26,129 --> 00:25:32,809
output would look like. Figure out what model architecture is right. Train the model and put it

257
00:25:32,810 --> 00:25:38,688
in a loop so that you can keep improving the model. The more data you can collect and the where

258
00:25:38,689 --> 00:25:45,369
do you want to point it? So it's simple. Once you sort of just internalize that worldview

259
00:25:45,610 --> 00:25:51,649
and the whole debate of like, is this worth like? I can't sit there and curate what work is

260
00:25:52,249 --> 00:25:59,209
worth doing versus not. And so that mindset just allows you to get out of that decision making

261
00:25:59,209 --> 00:26:04,529
process and hire people who are very versatile and problem solvers. How many engineers do you

262
00:26:04,530 --> 00:26:10,890
have? 30 odd. And how many did you have six months ago? 15

263
00:26:12,329 --> 00:26:18,529
I want to say. Yeah. So about two months ish. Yeah, yeah. In about 6 or 7 months, we've doubled. I want

264
00:26:18,529 --> 00:26:23,019
to say we. Haven't gotten the parallel. Yet. I know I was about to. Yeah, well, I guess we're kind of

265
00:26:23,020 --> 00:26:29,459
talking about it, but can I tell you a story about parallel story? So I, like I haven't gotten to, um,

266
00:26:30,099 --> 00:26:36,780
know you that well yet until recently. And I got a call from Liam on my

267
00:26:36,780 --> 00:26:43,539
team. I was on a work trip, and this was like, gosh, eight weeks ago, not even six weeks

268
00:26:43,540 --> 00:26:49,099
ago. And he was like, hey, I got to tell you something. And I'm like, what? And he's like, you

269
00:26:49,099 --> 00:26:54,899
know how sometimes founders, when we work really well together, you know, they kind of want to see

270
00:26:54,900 --> 00:26:58,539
if there's ways for us to work more formally together. And Liam, for those listening as their

271
00:26:58,540 --> 00:27:03,899
sales operating partner, I'm like, yeah. He's like, and you know how I always like I'm, you know, I'm

272
00:27:03,900 --> 00:27:10,779
not interested. I'm like, yeah. He's like, I think I'm interested in one. And I'm like, huh? Why?

273
00:27:11,220 --> 00:27:18,139
And he's talking about parallel. And he was like, well, number one, they've sold like tens

274
00:27:18,140 --> 00:27:24,509
of millions of dollars of revenue with like for people and like it's in spite of themselves.

275
00:27:24,510 --> 00:27:30,709
And he and he has spent a lot of time in your company at this point and he's talked to all the

276
00:27:30,709 --> 00:27:35,630
reps and so he can like see it. Like, you know, sales has a really good nose for like, can you

277
00:27:35,670 --> 00:27:41,748
stick it to this product to sell. And so that was one. And then he was like um to

278
00:27:42,469 --> 00:27:48,509
Parag and, and your co-founder Shivers, they're the real deal. Like they're the real deal. And

279
00:27:49,910 --> 00:27:56,348
I said, um, handicapped for me. Like, how serious are you? And he was like

280
00:27:57,150 --> 00:28:03,989
at least 50%. And within a week we went from 50%. Mamoon and I

281
00:28:04,030 --> 00:28:09,550
talked, we talked to Liam. It was very obvious, like this was like his calling was materializing in

282
00:28:09,550 --> 00:28:16,068
front of us. I think it's a fair way to describe it. And and this happens all the time. Like this I,

283
00:28:16,069 --> 00:28:21,149
I tell the team, your measurement of if you're doing a good job is if a founder tries to hire

284
00:28:21,150 --> 00:28:24,909
you like that is truly the measure of greatness. That's like the highest compliment you give

285
00:28:24,910 --> 00:28:31,549
someone. And, you know, eventually, uh, he it became obvious that this was his thing.

286
00:28:31,829 --> 00:28:38,509
And, um, and so that was the first time I had like, that was the first time you really caught my

287
00:28:38,509 --> 00:28:45,309
attention. And Mamoon and I talked and, um, were genuinely both two things happened at

288
00:28:45,310 --> 00:28:49,909
once, in my opinion. Tell me if you think I'm wrong. One, we were sad to see Liam go, but really excited

289
00:28:49,909 --> 00:28:56,429
and two, way more bullish on parallel. So I think, you know, to be fair, we've

290
00:28:56,429 --> 00:29:01,669
been fortunately, uh, our job is invest in great companies. And the things that Liam saw, we'd seen

291
00:29:01,669 --> 00:29:06,189
six months before, that's that's how I put it. Yeah. That's fair. But then when you see Liam and Graham

292
00:29:06,189 --> 00:29:12,149
and like, this group of people also then see it and then Parag's ability to recruit them, that's

293
00:29:12,150 --> 00:29:18,680
like the, the vision coming to life. It is. But let me tell you the the I'll double down on the in

294
00:29:18,680 --> 00:29:25,599
spite of ourselves comment. So the way this whole thing happened, I think there is a real story

295
00:29:25,600 --> 00:29:31,639
to tell here. So we were at this board meeting at the beginning of the year, and

296
00:29:32,679 --> 00:29:38,998
Mamoon, in his sort of simple way, asked a question, uh, which was like,

297
00:29:40,160 --> 00:29:46,879
you guys like you're in some when a customer is looking at you, like when you show up, like,

298
00:29:47,039 --> 00:29:53,280
how often do you win? How often do you close? And we said we tend to always close

299
00:29:54,400 --> 00:30:00,719
and once like so you just not enough deals. It's a very simple comment. And the one that really hurts

300
00:30:00,719 --> 00:30:06,118
when you hear it and think about it. Right. And so it really hurt me that you're like you just not

301
00:30:06,119 --> 00:30:12,639
enough fights. You're just not showing up. Mhm. Uh, so I called Mamoon back the same day. I was like,

302
00:30:12,680 --> 00:30:19,370
okay, I have a problem. Help me fix it. And Mamoon was gracious

303
00:30:19,370 --> 00:30:24,089
enough to get Liam to help us solve this problem. We'd been looking and recruiting for a head of

304
00:30:24,089 --> 00:30:30,129
sales for some time, but these things take time, and I don't think we fully internalized

305
00:30:31,369 --> 00:30:38,369
how well our product was being pulled by the market. Until

306
00:30:38,369 --> 00:30:44,649
I can. Liam just showed up as an outsider on week one and looked at things

307
00:30:44,649 --> 00:30:51,569
objectively, and it was him telling us about how

308
00:30:51,569 --> 00:30:58,129
unusual it is, what we were seeing on the inside, because we had no calibration points. He

309
00:30:58,129 --> 00:31:05,129
did. That actually gave us more confidence in what we were doing and more direction in what we were

310
00:31:05,129 --> 00:31:12,129
doing. So it was really helpful in so many ways to get to work with him. Uh, and yeah, I think

311
00:31:12,169 --> 00:31:18,969
we. It's amazing to. To get to work with Liam and I'm like forever grateful. Like

312
00:31:18,969 --> 00:31:25,290
keep telling. It was like a hard conversation to call Mamoon and ask about this. And I knew it

313
00:31:25,290 --> 00:31:30,290
when you texted me and said, hey, can we go for a walk tomorrow? As I was out of the country and I

314
00:31:30,290 --> 00:31:36,970
said, I land at 3 p.m. on Friday, how about 4 p.m.? And we met at 4 p.m. for a walk. And

315
00:31:36,970 --> 00:31:42,529
I, I sort of knew what was coming. Yeah. It is. Listen, like it was on the top three for the week.

316
00:31:42,529 --> 00:31:48,850
Right. So so then you do not waive time. You're like, yeah, you land at 315. We can leave at four.

317
00:31:48,890 --> 00:31:55,169
Yeah. Like, why would you. I did take a shower and then I met you after a 12 hour flight. But yeah. I

318
00:31:55,209 --> 00:31:59,088
generally also think like when you want something from someone, then being jet lagged is usually in

319
00:31:59,089 --> 00:32:06,009
your favor. That's right. Crafty. Crafty. And tell me. Tell you. How was the

320
00:32:06,010 --> 00:32:11,299
conversation? No, like I was it was it was great. So one I'm glad you knew what it was because the

321
00:32:11,300 --> 00:32:18,259
Mamoon came. Mamoon's first reaction was amazing. He's like number one. We're

322
00:32:18,259 --> 00:32:24,899
all about founders. And so if it's right for you and parallel, it's right for me. Second, don't do

323
00:32:24,899 --> 00:32:31,739
it. And I kind of needed to hear both of those things. Uh, because if you had not said the

324
00:32:31,739 --> 00:32:38,618
second thing, then I'd be like, wait, why is he not trying to save Liam? And so then he went and

325
00:32:39,379 --> 00:32:46,059
went to bat to save Liam and negotiated compromises. And it was

326
00:32:46,579 --> 00:32:52,979
clear that from that conversation that I had made the right bet,

327
00:32:53,020 --> 00:32:59,540
both on trying to hire Liam, but more importantly in working with Mamoon. So

328
00:32:59,780 --> 00:33:06,659
I felt validated on two of my big decisions that I had made over the last few months

329
00:33:06,740 --> 00:33:13,150
in that one walk. And then you also got Graham. Which happened right after. Yeah. So

330
00:33:13,510 --> 00:33:19,869
also Mamoon first introduced me with Graham. Who ran sales at Windsurf

331
00:33:19,910 --> 00:33:26,189
previously. Yes. And Mamoon introduced me with Graham. Well, before

332
00:33:26,870 --> 00:33:33,549
you invested in the company. Graham believed in what we were doing early. And I think

333
00:33:33,589 --> 00:33:40,549
ultimately you're building a company of believers and missionaries, and no matter what you roll,

334
00:33:40,550 --> 00:33:46,829
what function you're in. And so he, him and I like spoke

335
00:33:47,909 --> 00:33:54,709
quite a lot for several months. And it was when we truly got to the

336
00:33:54,709 --> 00:34:01,108
point of being able to fully benefit from someone like Graham. Like really just like

337
00:34:02,310 --> 00:34:07,389
scaling up the GTM motion and crushing the market.

338
00:34:08,710 --> 00:34:14,269
But that early introduction with Graham getting to know each other for several months. Him being

339
00:34:14,269 --> 00:34:20,229
enough of a believer in the company to want to invest in it. Those are all small things that all

340
00:34:20,230 --> 00:34:26,549
compound into being able to build a really powerful, really deep partnership with extremely

341
00:34:26,549 --> 00:34:32,749
special people. And to be clear, he was at Codeium before it was even called Windsurf. And my intent

342
00:34:32,750 --> 00:34:38,709
was purely for you to get to know someone who could be a really good sounding board for GTM way

343
00:34:38,709 --> 00:34:44,590
before you had any salespeople. And obviously I didn't want you to hire him from Windsurf because

344
00:34:44,669 --> 00:34:50,549
we were on the board there too. Yeah. Uh, yeah. No, I think but I think once I

345
00:34:51,749 --> 00:34:57,629
every special person you end up talking to as a company builder, you think about

346
00:34:58,189 --> 00:35:04,989
hiring if they are the right person. Like, it's very hard to not think about. Like, it's

347
00:35:04,989 --> 00:35:11,438
simultaneously very hard to believe that someone is exceptional at what they do, and at some point

348
00:35:11,439 --> 00:35:17,039
I will need that kind of exceptional in my business and also not visualize trying to make

349
00:35:17,040 --> 00:35:23,479
that happen, right? Like, I just don't know how to do that. So I walked the world trying to figure

350
00:35:23,480 --> 00:35:30,039
out how parallel can be better, because there are some amazing people that can who

351
00:35:30,479 --> 00:35:35,519
are great at what they do and actually believe in the mission to join us. Can you talk about the

352
00:35:35,520 --> 00:35:42,479
mission? Why did you start the company? What are you trying to do? Yeah, we started because

353
00:35:42,479 --> 00:35:48,959
I got obsessed with this notion that I wrote down one day, which is like, agents and AI's will just

354
00:35:48,960 --> 00:35:54,720
use the web a lot more than humans ever have. And I wrote down a number which is, like 1000 x, and

355
00:35:54,720 --> 00:35:58,279
this is like two and a half years ago, we wrote down this random number 1000 x more than humans

356
00:35:58,280 --> 00:36:03,839
ever have. And I said, like, I have built systems before at 1000 x the scale, none of no system

357
00:36:03,840 --> 00:36:08,009
works and no system is designed for it from an architecture perspective. From infrastructure

358
00:36:08,009 --> 00:36:14,129
perspective. I also wrote down that new business models will be needed

359
00:36:14,730 --> 00:36:20,529
for how agents use the web. The all the business models that have existed on the web for and have

360
00:36:20,529 --> 00:36:25,969
been built and evolved for like last three decades, let's call it, are about to evolve and

361
00:36:25,969 --> 00:36:31,090
change completely. And that seems like an opportunity. And it seems like an opportunity that

362
00:36:31,090 --> 00:36:37,689
is like really close to me because like the open web is a like for

363
00:36:38,010 --> 00:36:42,648
where I grew up and how I grew up. Open web is like this generational change in the world that I

364
00:36:42,850 --> 00:36:49,610
have experienced and lived through, and now there's going to be like a second birth of what

365
00:36:49,610 --> 00:36:56,089
the future of the open web is. A parallel web. A parallel web, if you may, that came to us later, by

366
00:36:56,090 --> 00:37:02,489
the way, not in the founding moment. Uh, the so, so this notion of

367
00:37:02,529 --> 00:37:07,849
like this observation that the business model will change. And like there is this existential

368
00:37:07,850 --> 00:37:14,490
threat, right? That the whole contract, the whole bargain of the open web, is that

369
00:37:15,250 --> 00:37:22,129
people put content out in the open because there are amazing ways

370
00:37:22,129 --> 00:37:28,689
of getting distribution, driving, monetization, selling goods. There are many incentive systems

371
00:37:28,689 --> 00:37:35,489
built in which keep things open. Now, the risk is, if we don't innovate

372
00:37:35,490 --> 00:37:42,289
enough that the web starts closing up and we're starting to see that already, right. And so there

373
00:37:42,290 --> 00:37:49,089
is some missing incentive alignment today in terms of agents using the web

374
00:37:49,290 --> 00:37:55,649
and people putting content out on the open web that needs solving. Now, the good

375
00:37:55,649 --> 00:38:02,260
news is, if agents are going to use the web 1000 x more, and I actually think 1000 x, did we under

376
00:38:02,260 --> 00:38:08,899
shoot? I'm already using the web with this token maxing. Uh, 1000 x more.

377
00:38:09,419 --> 00:38:14,619
I think it's going to be way more than 1000 x. That's not work. That is just a pure waste. I

378
00:38:14,620 --> 00:38:21,418
actually think we will generate real value, satisfaction, all kinds of things as a result of

379
00:38:21,620 --> 00:38:28,539
these tokens over time. And if we do that, there is enough value generated so

380
00:38:28,540 --> 00:38:35,379
that if we just figure out how to incentivize being open, how to reward high

381
00:38:35,379 --> 00:38:42,099
quality, unique, differentiated content that we get to keep the web open and prospering and becoming

382
00:38:42,139 --> 00:38:47,819
like a whole new thing, like we can't imagine. And so that's the core mission of the business, to

383
00:38:47,860 --> 00:38:53,779
keep the web open, to create incentive alignment

384
00:38:53,980 --> 00:39:00,419
between all of the plenty we will create with agents having access to more high quality

385
00:39:00,469 --> 00:39:07,309
information, and the people who curate create, produce that information.

386
00:39:07,389 --> 00:39:13,389
So when we, I think, first spoke about this two and a half years ago. I remember it was like

387
00:39:13,389 --> 00:39:19,949
Thanksgiving of 20, 23, 23, 23, 23. Right.

388
00:39:19,990 --> 00:39:26,869
Yeah. And the notion of a web for AI agents completely, 100% made sense. Um,

389
00:39:26,870 --> 00:39:32,749
and a lot has happened since then, uh, in terms of models that are

390
00:39:33,669 --> 00:39:40,589
actually utilizing the web in ways that at the time, you maybe not had foreseen.

391
00:39:40,909 --> 00:39:46,509
I certainly hadn't foreseen, but I knew that AI agents would exist, and they would need a web or a

392
00:39:46,510 --> 00:39:51,548
parallel version of the web that they would use a lot more than humans use the web presently today.

393
00:39:51,549 --> 00:39:58,429
And so maybe take us through what that inflection point was in the significant

394
00:39:58,550 --> 00:40:04,790
tailwinds that parallel has gotten as a business. So 2 or 3 things. And some of this is

395
00:40:05,149 --> 00:40:11,149
like fitting to what we've observed over the last couple of years, right? If we all go back in time

396
00:40:11,149 --> 00:40:17,869
and think about like the initial version of ChatGPT, it did not use the web and we still

397
00:40:17,909 --> 00:40:24,709
thought it was really cool, right? For people who've now been using any of these products

398
00:40:24,709 --> 00:40:31,029
ChatGPT cloud for a period, if you go back and just use a model without giving it access to the

399
00:40:31,030 --> 00:40:37,989
web for 75% of what you do with them, you'll find it

400
00:40:37,989 --> 00:40:44,389
unusable. Now, why is that? We should observe. Why did we suddenly reorient

401
00:40:44,389 --> 00:40:51,349
ourselves? It's because these models present interfaces

402
00:40:51,350 --> 00:40:58,278
and a mental model, which is that they're both smart and all knowing Once you build

403
00:40:58,279 --> 00:41:04,678
that mental model into a product, you must use the web because otherwise you're not following. It

404
00:41:04,679 --> 00:41:11,319
feels entirely stupid that the thing doesn't know the date of the last Super Bowl. It just it just

405
00:41:11,360 --> 00:41:16,399
it's so confusing to people. The same thing happened with coding agents, right? Coding agents. I

406
00:41:16,399 --> 00:41:22,919
remember every coding agent. Like I would be pasting links to docs

407
00:41:22,919 --> 00:41:29,638
into a UX to have it fetch and index it

408
00:41:29,639 --> 00:41:35,438
locally so that it could build in an API that was not in the pre-training data set.

409
00:41:36,279 --> 00:41:42,320
Now we'll feel like cavemen doing it. And so really what's happening is

410
00:41:42,360 --> 00:41:49,279
every almost every application we build with models creates the

411
00:41:49,280 --> 00:41:56,128
customer expectation of that application that this is all knowing. And the that

412
00:41:56,129 --> 00:42:03,129
intuition is gone from oh, a model is smart to a model plus the web. And now over time it's

413
00:42:03,129 --> 00:42:08,449
evolving into an it has access to all of my data. Right. But the web is going to remain in that

414
00:42:08,449 --> 00:42:14,968
expectation. So we really think of parallel as a adjacency

415
00:42:15,010 --> 00:42:21,610
infrastructure, adjacency to every model to inference for all kinds of work.

416
00:42:21,649 --> 00:42:28,649
Right. So if you think of how we evolve, it's when different categories of work,

417
00:42:28,649 --> 00:42:35,128
when people derive value from models in the work context, we take off. And that really started

418
00:42:35,129 --> 00:42:42,009
happening in my observation last summer in some like nine months ago

419
00:42:42,129 --> 00:42:47,689
in some areas, and has been exploding since then. And we

420
00:42:48,810 --> 00:42:55,810
show up with like the highest quality way, which is extremely agent native to help

421
00:42:55,810 --> 00:43:02,809
an agent do more, do it faster, and do it cheaply. When you're building an agent for

422
00:43:02,810 --> 00:43:08,089
any kind of knowledge, work end to end. And so since last summer, we're now seeing like across

423
00:43:08,090 --> 00:43:15,009
categories, the people who are at the, the tip of the spear, the frontier of using agents

424
00:43:15,009 --> 00:43:21,570
to do really cool things, incorporate our APIs and make their own products

425
00:43:21,689 --> 00:43:28,570
even better than they already were. And we write that and then, um, as sort of

426
00:43:28,809 --> 00:43:35,329
so we've seen these models being really useful in finance, really useful. Can you give some examples.

427
00:43:35,370 --> 00:43:39,049
Maybe like the Harvey use case. So you guys just announced a big a big partnership with Harvey.

428
00:43:39,569 --> 00:43:44,809
Like what are they like what are they using it for. We have a pretty big product suite at that

429
00:43:44,809 --> 00:43:50,610
point. So they use it for multiple things. Uh, one of the things they do is just use our search

430
00:43:50,610 --> 00:43:57,259
products so that All engagements with Harvey. The the product always knows what's out there

431
00:43:57,259 --> 00:44:04,179
and it knows it without burning too many of its tokens. It knows it with fresh data. It knows

432
00:44:04,179 --> 00:44:10,898
it with hard to reach data. Now there is a second element to this, which is there is a lot

433
00:44:10,940 --> 00:44:17,860
of data that over time, let's call it like I think in

434
00:44:17,860 --> 00:44:24,659
the in 2000 this was called the deep web, which is content that is

435
00:44:25,299 --> 00:44:30,500
extraordinarily hard to get to on the web. Let me give you an example. There are many portals where

436
00:44:32,219 --> 00:44:39,179
they're not there's content not in the a typical search index which is optimized on

437
00:44:39,179 --> 00:44:45,059
ten blue links. Why isn't it there? Because a lot of this content isn't reachable by just clicking

438
00:44:45,060 --> 00:44:49,860
on a link. You kind of have to end up at a portal. You have to either navigate somewhere or search

439
00:44:49,860 --> 00:44:56,029
somewhere and it doesn't have a a URL for that content. Now,

440
00:44:56,030 --> 00:45:02,269
historically people, when it's hard to crawl two it's less valuable to crawl because you can't

441
00:45:02,270 --> 00:45:08,749
actually land a person on the content. In a world of eyes. Our job is to bring the content into the

442
00:45:08,749 --> 00:45:15,269
context window of an AI. So we now go crawl. We built crawlers for

443
00:45:15,629 --> 00:45:22,388
extra like things that historically search engines haven't been crawling. And

444
00:45:22,709 --> 00:45:29,349
when you're building Harvey, when you're optimizing for having access

445
00:45:29,350 --> 00:45:35,549
to authoritative source documents across the world

446
00:45:35,950 --> 00:45:42,269
so that every bit of output in Harvey is well grounded with an authoritative document you care

447
00:45:42,269 --> 00:45:47,829
about one having completion on these hard to crawl information that is in the public domain.

448
00:45:48,350 --> 00:45:54,429
And two, you care about really good search ranking in search quality on that. And these are two

449
00:45:54,429 --> 00:46:01,149
things we do really, really well right. And so through this partnership we are able to push

450
00:46:01,149 --> 00:46:08,110
and stretch our capabilities because they're a very demanding customer. And they are able to, uh,

451
00:46:08,149 --> 00:46:14,509
have access to extremely hard to reach high quality public data with great ranking on top. Do

452
00:46:14,509 --> 00:46:20,949
you want to talk about just having run one of the largest web scale systems at

453
00:46:20,990 --> 00:46:27,908
Twitter and, uh, the the parallels that you draw to what you have to ultimately

454
00:46:27,909 --> 00:46:33,789
build at parallel or what you may have already built, uh, in terms of the supporting the scale

455
00:46:34,270 --> 00:46:40,789
latency. Obviously, latency being so critical and important for, you know, a real time social network.

456
00:46:42,389 --> 00:46:48,919
Yeah. No, I think so. Parallel is fundamentally a technical Product. It's

457
00:46:48,960 --> 00:46:55,519
an extraordinarily large scale system. I would say over time,

458
00:46:55,519 --> 00:47:02,439
its scale is meaningfully larger than Twitter, mostly because Twitter is,

459
00:47:02,439 --> 00:47:08,279
like amazing and influential. And I'm really proud of what we built it. It's a small slice of the web.

460
00:47:08,679 --> 00:47:14,959
It's a small slice of public content out in the world. The opportunity with parallel feels like

461
00:47:14,999 --> 00:47:20,158
outrageously larger, uh, in terms of keeping the web open. The mission is the same, right? Twitter

462
00:47:20,159 --> 00:47:27,080
was trying to keep more high quality, unique content out

463
00:47:27,080 --> 00:47:33,999
in the open versus a disappearing into closed worlds. Right. Giving

464
00:47:34,000 --> 00:47:40,799
direct access to the best minds to everyone in the world at the

465
00:47:40,800 --> 00:47:47,768
same time, instead of via intermediaries or via closed off groups. That is the same mission

466
00:47:47,769 --> 00:47:54,529
at parallel. To keep the web open and having more information out in the public at all times,

467
00:47:55,449 --> 00:48:02,049
but it feels materially larger. Now going to the infrastructure, though, I think having

468
00:48:02,610 --> 00:48:09,529
great having known great engineers, having known how to

469
00:48:09,570 --> 00:48:16,129
build and design systems pragmatically for a level of scale, knowing when

470
00:48:16,169 --> 00:48:20,850
to rebuild them for the next level of scale. Having done that like several times over at

471
00:48:20,850 --> 00:48:25,849
Twitter. At Twitter. Like we went from this ride of trying to grow systems and like every year we

472
00:48:25,850 --> 00:48:32,009
would have to rebuild entire systems because they weren't built for the next order of magnitude of

473
00:48:32,010 --> 00:48:38,689
scale. But if you build for like four years out, you don't ship anything. And so you have to

474
00:48:38,729 --> 00:48:43,849
constantly keep evolving for scale. And we're running the same ish

475
00:48:45,570 --> 00:48:51,649
Set of lessons at parallel. The other, I think, really interesting thing that I carry from the

476
00:48:51,689 --> 00:48:58,569
the ML infrastructure part of Twitter is, you know, how at

477
00:48:58,570 --> 00:49:02,528
Twitter, when you're building these recommendations systems or these models, they

478
00:49:02,529 --> 00:49:08,168
initially start as these sort of cobbled up things which are like you train a bunch of models

479
00:49:08,169 --> 00:49:14,169
and then you have a bunch of heuristics on top. And over time, as you learn about your problem

480
00:49:14,169 --> 00:49:18,529
space, more things become more end to end trained. This is the same story that's played out with

481
00:49:18,570 --> 00:49:24,409
like self-driving cars in Waymo. Started with like a few perception model in this model. And over

482
00:49:24,409 --> 00:49:30,208
time, things become more end to end as you know how to collect more data, right? I think today

483
00:49:31,129 --> 00:49:37,129
you don't necessarily have to go down that journey that slowly, the way we used to do

484
00:49:37,290 --> 00:49:42,529
five, seven, ten years ago in building these systems. And you can shoot more end to end from

485
00:49:42,530 --> 00:49:49,219
the very beginning at scale, and I think doing so puts you on this path of rapid

486
00:49:49,459 --> 00:49:55,579
self-improvement. And so in some sense, we saw the journey at Twitter and at parallel, we're sort of

487
00:49:55,580 --> 00:50:01,419
taking a leap to not have to go through the three steps and jumping to the fourth step parallel

488
00:50:01,419 --> 00:50:07,979
started. You had an amazing run at Twitter CEO start parallel. Was everybody,

489
00:50:07,979 --> 00:50:14,579
like clamoring to come work with you like you had this this vision in this mission? Did

490
00:50:14,580 --> 00:50:21,458
you come out of the gates? Like, tell me about the very, very early days. Was

491
00:50:21,459 --> 00:50:28,059
it obvious to was it obvious because you are like pretty bona fide to go work with as a

492
00:50:28,060 --> 00:50:34,819
founder. So we were very quiet and stealthy. So we literally

493
00:50:35,260 --> 00:50:41,668
no one knew that I was starting a company. I had worked with a bunch of people. I It

494
00:50:42,629 --> 00:50:49,389
called The People. I wanted to join me at parallel. And then there was this mission

495
00:50:49,389 --> 00:50:56,029
alignment problem. So I didn't get an inbound call because no one knew what I was going to do. The

496
00:50:56,030 --> 00:51:02,469
one thing I did early on was decided that we would keep more than half the team new, and not

497
00:51:02,470 --> 00:51:08,590
people I had already worked with, and I had sourced really talented people through the almost

498
00:51:08,590 --> 00:51:14,789
a year before I truly started hiring for parallel, where I'd met a lot of people exploring lots of

499
00:51:14,789 --> 00:51:19,509
things and asked them a standard question which is like, what are the best three people you've

500
00:51:19,510 --> 00:51:26,189
ever worked with? And I would go meet some of them over time. And so I ended up hiring a

501
00:51:26,629 --> 00:51:33,549
and working with the initial founding team of me, plus six other people, and half of them

502
00:51:33,550 --> 00:51:40,309
were people I had not previously known, but met via this sort of multi

503
00:51:40,350 --> 00:51:47,269
hop dance off like them being some among the top three of someone I thought

504
00:51:47,270 --> 00:51:53,870
was already exceptional. And so this way I think we got to build a team that was

505
00:51:54,229 --> 00:52:01,109
beyond my in starting network from Twitter and would ultimately we've done a lot of

506
00:52:01,110 --> 00:52:07,309
hiring via this kind of a channel of like extraordinarily unique and talented people,

507
00:52:07,990 --> 00:52:13,749
uh, but built via this sort of network hiring approach of exceptional people to tap outside of

508
00:52:13,750 --> 00:52:20,629
my personal network. And I think that's been really, really good for us. 40 people

509
00:52:20,629 --> 00:52:27,350
today. 50. Yeah. What do you think you'll be at the end of the year? Are you hiring?

510
00:52:27,350 --> 00:52:34,349
I am hiding as fast as we can. Across every function. We would hire across every function,

511
00:52:34,350 --> 00:52:41,039
but we proactively put our energy towards a few. Which are. Always

512
00:52:41,040 --> 00:52:47,800
engineering. Now GTM and the new ones

513
00:52:47,879 --> 00:52:54,759
are marketing. Developer relations are the top new ones on my list

514
00:52:54,760 --> 00:53:01,439
where we just need to do more. Yeah, it's like time to build the company around the product.

515
00:53:01,999 --> 00:53:07,199
Exactly. So we started with building the technology. When we launched the product last year,

516
00:53:07,200 --> 00:53:13,959
we had we're building the product. Now we're building the company around it with includes GTM,

517
00:53:13,960 --> 00:53:20,879
but includes all the other things you need, uh, to actually show up for, uh, customers.

518
00:53:21,280 --> 00:53:28,159
So maybe just, uh, jump on that. Uh, so some of my favorite

519
00:53:28,159 --> 00:53:34,159
products that I use, AI native products use parallel, and I'm not sure they're all fully

520
00:53:34,159 --> 00:53:40,049
announced in terms of as customers or you have some public statement with them, but, uh, maybe just

521
00:53:40,090 --> 00:53:46,570
help us. One of these products. Uh, Rogo Harvey. Profound

522
00:53:46,889 --> 00:53:53,850
granola. Uh, Clay. There's more. It's good list.

523
00:53:53,889 --> 00:54:00,329
Yeah. We love. Listen, I could list. I love so one of

524
00:54:00,330 --> 00:54:07,249
our core beliefs is that you have to experience your product in the context

525
00:54:07,250 --> 00:54:13,929
of every customer. So we are particularly interested in being a part of

526
00:54:14,729 --> 00:54:19,489
the journey of products that we personally love. Right. So.

527
00:54:22,129 --> 00:54:26,168
It gives me deep, deep, deep personal satisfaction, like I think at Twitter, like I was a heavy duty

528
00:54:26,169 --> 00:54:30,729
Twitter user before I even started working there as an engineer. I was an addict, uh, before I went

529
00:54:30,729 --> 00:54:34,729
to work at Twitter. And the reason I ended up there was because I was in love with the product.

530
00:54:34,770 --> 00:54:41,569
Right? And it's the same thing. Like, my team derives extraordinary satisfaction to have

531
00:54:41,610 --> 00:54:48,009
built something that is useful for products that we love, using ourselves for our work.

532
00:54:48,330 --> 00:54:55,209
And it is truly the most rewarding thing to be able to solve problems that

533
00:54:56,169 --> 00:55:02,929
these have. The other side effect is. I believe we have good taste in the

534
00:55:02,930 --> 00:55:09,209
products we like to use and to be able to. Those products are often built by people with really

535
00:55:09,210 --> 00:55:15,529
good taste, and so to have them as customers, we put ourselves in a place where we are

536
00:55:16,529 --> 00:55:23,488
the demanding, extraordinarily demanding in terms of what products they would use. They push

537
00:55:23,489 --> 00:55:30,129
us and our technology and our product forward. And my core belief is that if we can be

538
00:55:31,249 --> 00:55:37,219
a product that they love putting into their products to power them.

539
00:55:38,459 --> 00:55:44,579
There's a lot more that follows from that. And so it's really nice that over the last several

540
00:55:44,579 --> 00:55:51,219
months, we've sort of built our product up to the place where the best

541
00:55:51,220 --> 00:55:57,819
products we like are starting to use us. So with that, uh, I imagine

542
00:55:57,820 --> 00:56:04,219
more AI native products will use parallel. You know, I think you'll blanket sort of that world.

543
00:56:04,219 --> 00:56:11,059
What else can we dream about where parallel is deeply embedded. And so, in other words,

544
00:56:11,060 --> 00:56:18,019
how does your market expand and how does this company really rise to the level of one

545
00:56:18,019 --> 00:56:23,259
of the most exciting companies that are is being built right now? So one, I don't think we have to

546
00:56:23,259 --> 00:56:28,739
expand the market like we are. Like if you use an A model for building

547
00:56:30,459 --> 00:56:37,388
anything, it's a product, a Workflow automation for anything related to work that you

548
00:56:37,389 --> 00:56:44,389
do. I almost believe you must give it the web and if you must, give it the

549
00:56:44,389 --> 00:56:51,349
web. The question is through which of our products right? And we

550
00:56:52,270 --> 00:56:59,030
have and must have the best way, no matter what your circumstances for the product or the

551
00:56:59,030 --> 00:57:05,188
workflow that you're doing to use us. Now, this need is broad based. If you're a large enterprise

552
00:57:05,389 --> 00:57:12,189
that is now trying to make your operations more streamlined and you're starting to use a model

553
00:57:12,189 --> 00:57:19,149
for it, you could. There are many ways you could use parallel. You could have an internal team by

554
00:57:19,149 --> 00:57:25,750
an API and parallel APIs, and hook it up to your internal data via a bunch of

555
00:57:25,789 --> 00:57:32,350
MCPs or clients, and build workflow automation, for example, like I've built my own app for my

556
00:57:32,350 --> 00:57:39,069
productivity. Right. Or you could go and you can say, no, no, I don't want to do it myself. I actually

557
00:57:39,069 --> 00:57:46,028
want someone who's prioritizing it for others like me for this function. And then

558
00:57:46,029 --> 00:57:52,628
we must be working with whoever it is that solves the end problem. All we care

559
00:57:52,629 --> 00:57:58,668
about is when agents access the web. We have the best products for them, no matter who they are.

560
00:57:58,990 --> 00:58:05,310
What are you going to see over the next year is there'll be a class of the

561
00:58:05,310 --> 00:58:10,869
best AI native agent products that will incorporate parallel. If we do our jobs right,

562
00:58:11,509 --> 00:58:16,709
there will be a class of large enterprises which will take all kinds of

563
00:58:17,589 --> 00:58:24,548
differentiated work that they do themselves and supercharge them with parallel. And the biggest

564
00:58:24,549 --> 00:58:29,710
thing that excites me is people often frame this as like, oh, we were already doing this work, now we

565
00:58:29,710 --> 00:58:35,079
are going to make it more efficient and we can do the same work, but it's going to be faster and

566
00:58:35,079 --> 00:58:41,959
it's going to be cheaper and more automated. Now people do work beyond what was

567
00:58:41,959 --> 00:58:48,878
previously happening. I'll give you a very simple example. Let's say you are a semi hypothetical,

568
00:58:49,399 --> 00:58:56,279
a PE firm thinking about buyouts. And in the previous world you'd say, okay,

569
00:58:57,000 --> 00:59:00,999
I could buy out these categories of businesses in these kinds of neighborhoods and these kinds of

570
00:59:01,000 --> 00:59:06,479
cities, and you'd use some gut and human judgment to narrow down to like 4 or 5 interesting

571
00:59:06,479 --> 00:59:13,119
opportunities. And then you'd go do extra work on them. Right. Their biggest limited

572
00:59:13,120 --> 00:59:20,039
resources, how many of these they can do through capital and human capital in a year? And

573
00:59:20,040 --> 00:59:24,720
the biggest job is to figure out which ones to do and prioritize that well. Now they can use

574
00:59:24,720 --> 00:59:30,929
parallel and models to really run what feel like almost simulations on various criteria. Did they

575
00:59:30,969 --> 00:59:37,928
set and do a lot of compute to exhaustively come up with

576
00:59:37,929 --> 00:59:43,649
a rational list of opportunities to actually explore, and then pick the ones they do better?

577
00:59:43,689 --> 00:59:49,289
They're previously pruning the space of what they would explore to this tiny, top down, intuitive

578
00:59:49,289 --> 00:59:56,170
thing. And now they get to see real information, real data about what it could be and if

579
00:59:56,170 --> 01:00:03,089
it makes them pick one better. There's so much value created that it's worth token maxing in

580
01:00:03,129 --> 01:00:10,129
that context. Right. And so when you see these use cases about like just entirely new ways

581
01:00:10,569 --> 01:00:17,449
of operating businesses that have existed forever, that's really

582
01:00:17,449 --> 01:00:22,489
exciting to me. And I think that's the next year of growth, which isn't just like replace what

583
01:00:22,490 --> 01:00:28,570
we're doing today. It's going back to what does it take to make it 1000 x, right? It is work that was

584
01:00:28,570 --> 01:00:35,208
not happening yesterday. Well, sir, thank you for doing this. Well said. Uh, when you hear the word

585
01:00:35,209 --> 01:00:41,330
grit, I ask everybody the same question. What do you think of. The conversation we were having

586
01:00:41,330 --> 01:00:47,609
earlier when you asked me in the kitchen about taking a breath. Uh,

587
01:00:48,810 --> 01:00:55,409
I think grit is just doing hard things. And the only way you can do it is if

588
01:00:55,410 --> 01:01:01,169
it's for something meaningful to you. And it's fun.

589
01:01:02,649 --> 01:01:09,648
Thank you. Thank you. That's it for now. If you liked the episode, please leave us a

590
01:01:09,649 --> 01:01:14,049
review or go back into the archives where we've done more than 200 episodes with some fantastic

591
01:01:14,050 --> 01:01:19,050
folks. This podcast is a Kleiner Perkins production, and I'm Joubin. Thanks for listening.