[00:00] (0.16s)
YC's next batch is now taking
[00:02] (2.48s)
applications. Got a startup in you?
[00:04] (4.72s)
Apply at y combinator.com/apply.
[00:07] (7.68s)
It's never too early and filling out the
[00:10] (10.08s)
app will level up your idea. Now, on to
[00:12] (12.72s)
the video.
[00:13] (13.28s)
You have to innovate. You have to move
[00:14] (14.48s)
faster than everybody else. And it's
[00:16] (16.72s)
like running a marathon but at an
[00:19] (19.28s)
extremely high velocity. Right. The only
[00:21] (21.52s)
uh mode you have is speed. I read all
[00:23] (23.44s)
the Twitter comments every time. Google
[00:25] (25.04s)
IO last year was AI overview and
[00:27] (27.20s)
perplexity is dead. This year was AI
[00:29] (29.28s)
mode and perplexity is dead and I read
[00:31] (31.20s)
all of that too and it's it's always
[00:32] (32.96s)
fun. I love it actually.
[00:39] (39.76s)
Aravant, I see you every I don't know 2
[00:42] (42.08s)
or 3 months and you give me an update on
[00:44] (44.08s)
the latest on Perplexity. Why don't you
[00:46] (46.24s)
just tell these folks where you're at?
[00:48] (48.48s)
How are things going? Do people use
[00:50] (50.32s)
Perplexity?
[00:52] (52.64s)
Do you guys use Perplexity?
[00:55] (55.52s)
Well, whether you believe it or not,
[00:56] (56.88s)
like I have infra issues every day. So
[00:59] (59.92s)
there are a lot of people using it and
[01:02] (62.08s)
um this usage is actually growing to the
[01:04] (64.72s)
extent that we don't actually know how
[01:06] (66.72s)
to deal with it. We have to rebuild the
[01:09] (69.28s)
infra to scale the next 10x. So
[01:12] (72.08s)
definitely a lot of people in the world
[01:13] (73.60s)
using it. Thanks to all of you as well.
[01:16] (76.24s)
What is next for us? The browser. That's
[01:19] (79.12s)
the big bet we're making as as far as
[01:21] (81.68s)
the future of the company goes.
[01:23] (83.68s)
Everyone's here is like why should I use
[01:25] (85.44s)
perplexity when there's search and other
[01:27] (87.68s)
AI apps of course chat GPT has a bigger
[01:30] (90.24s)
distribution than us every other AI app
[01:32] (92.88s)
is trying to put search as a layer in it
[01:35] (95.84s)
all of them support citations a lot of
[01:38] (98.48s)
them support some of the verticals we
[01:40] (100.16s)
put work into yes like we're always
[01:41] (101.92s)
going to continue to remain better than
[01:43] (103.60s)
others in that category but I think the
[01:46] (106.24s)
browser and agents are truly the next
[01:49] (109.44s)
bet that we want to make we think about
[01:51] (111.52s)
it as an assistant rather than a
[01:53] (113.36s)
complete autonomous agent but one omni
[01:56] (116.72s)
box where you can navigate you can
[01:58] (118.64s)
askformational queries and you can give
[02:01] (121.12s)
agentic tasks and your AI with you on
[02:03] (123.68s)
your new tab page on your site car as an
[02:06] (126.00s)
assistant on any web page you are makes
[02:08] (128.48s)
the browser feel like more like a
[02:10] (130.56s)
cognitive operating system rather than
[02:12] (132.48s)
just yet another browser and we hope to
[02:16] (136.40s)
make it like a cloud where you launch
[02:18] (138.40s)
several tasks in parallel that are
[02:20] (140.88s)
running asynchronously Okay. And pulling
[02:23] (143.60s)
all your personal contacts, your email,
[02:26] (146.40s)
your calendar, your Amazon, your, you
[02:29] (149.28s)
know, all all sorts of social media
[02:30] (150.88s)
accounts that you have and you go and do
[02:33] (153.84s)
research on real estate, the markets,
[02:36] (156.40s)
and these are all like just processes
[02:38] (158.24s)
running on your browser. That's never
[02:40] (160.56s)
been possible before. And Chrome was
[02:43] (163.52s)
exciting when each tab was its own
[02:45] (165.36s)
process. You think about each query or
[02:47] (167.68s)
each prompt could be that
[02:49] (169.36s)
and that will be our new browser comet.
[02:51] (171.36s)
So, we're putting all our energy into
[02:53] (173.28s)
This was going to be the hard question I
[02:54] (174.72s)
saved for the end, but since you queued
[02:56] (176.32s)
it up, I'll do it right now. Um, I think
[02:58] (178.56s)
if Sam Alman were still on the stage
[03:00] (180.40s)
today, he would say, "Oh, yeah, that's
[03:02] (182.96s)
what we're doing." Um, and I think
[03:05] (185.68s)
Sundar at Google probably would say
[03:08] (188.56s)
that's the direction we're headed as
[03:10] (190.16s)
well. So, it feels like there are a
[03:12] (192.80s)
bunch of players now, many of them very
[03:15] (195.36s)
wellunded, going in generally the same
[03:17] (197.84s)
direction. How do you see the world?
[03:19] (199.68s)
Like do you think that there's going to
[03:21] (201.36s)
it's going to play out where there's
[03:22] (202.64s)
actually like a bunch of different use
[03:24] (204.24s)
cases and you can own a very important
[03:26] (206.24s)
one that others won't want to own or are
[03:29] (209.52s)
we in for like a major competitive
[03:31] (211.44s)
battle?
[03:32] (212.00s)
Look, if something is really worth
[03:33] (213.60s)
doing, it's it's only natural that
[03:36] (216.56s)
people with a lot of funding will go and
[03:38] (218.24s)
do it. Um people said perplexity is a
[03:41] (221.68s)
great product. Now everyone is trying to
[03:44] (224.00s)
do something that can answer any
[03:45] (225.92s)
question with sources. Cursor was a
[03:48] (228.40s)
great product.
[03:49] (229.60s)
Now, OpenAI is trying to buy cursor's
[03:52] (232.32s)
competitor anthropic launch codeex uh
[03:55] (235.04s)
like clot code. Google has its own like
[03:57] (237.76s)
rival tool. So, it's only natural that
[03:59] (239.52s)
when there's a lot of money to be made
[04:00] (240.80s)
in a certain sector, people are going to
[04:03] (243.36s)
try to copy it. And there's only a
[04:06] (246.40s)
limited amount of things you can be
[04:08] (248.00s)
world class at, whether it's being
[04:10] (250.32s)
building great models or building one or
[04:12] (252.32s)
two really good products. So, you're
[04:14] (254.08s)
obviously not going to win on
[04:15] (255.12s)
everything. For us, this is the only
[04:17] (257.12s)
thing we care about. accuracy at the
[04:19] (259.52s)
level of answers, accuracy at the level
[04:21] (261.52s)
of tasks, orchestrating all these
[04:23] (263.60s)
different tools. The browser is much
[04:25] (265.92s)
harder to copy than like uh yet another
[04:28] (268.16s)
chat tool. That said, I'm I'm fully
[04:31] (271.68s)
working with the assumption that uh
[04:33] (273.68s)
OpenAI will also build its own browser.
[04:35] (275.92s)
Anthropic will also try to build its own
[04:37] (277.68s)
browser. Google already has one called
[04:39] (279.68s)
Chrome. So, it's completely reasonable
[04:41] (281.52s)
to expect them to do it. And the only uh
[04:44] (284.00s)
mode you have is speed. You have to
[04:45] (285.36s)
innovate. You have to move faster than
[04:46] (286.80s)
everybody else. And it's like running a
[04:49] (289.52s)
marathon but at an extremely high
[04:52] (292.00s)
velocity, right?
[04:52] (292.80s)
Yes. Yeah. I I really agree with your
[04:55] (295.84s)
statement that like you can only focus
[04:58] (298.08s)
on one thing and be world class at one
[04:59] (299.84s)
thing. And just to give you guys like a
[05:01] (301.44s)
little glimpse into it, we were
[05:03] (303.28s)
backstage before this talk and he was
[05:04] (304.88s)
showing me some of the new stuff that
[05:06] (306.16s)
they're working on and there was like a
[05:07] (307.68s)
bug, right? And he stopped everything he
[05:10] (310.48s)
was doing to like figure out what was
[05:12] (312.08s)
wrong with this bug. Why was it not
[05:13] (313.52s)
doing the right thing? And if you think
[05:15] (315.12s)
about like what would the CEO of a large
[05:17] (317.60s)
company do in that situation? Probably
[05:19] (319.76s)
they would like hand it off to somebody
[05:21] (321.36s)
else on their team. So that's like a
[05:23] (323.52s)
good piece of evidence that you actually
[05:25] (325.04s)
mean what you say.
[05:26] (326.32s)
Yeah. I I I love I love triaging and
[05:28] (328.56s)
fixing bugs. I know it sounds trivial.
[05:31] (331.04s)
Like is that the best use of the time of
[05:33] (333.52s)
a CEO? There are a lot of people who
[05:36] (336.16s)
would think otherwise. Recently people
[05:38] (338.16s)
are like uh oh like like there I hope
[05:40] (340.96s)
this behavior is rubbing off on others.
[05:42] (342.48s)
Like I've noticed even Sundar is doing
[05:44] (344.08s)
bug support on X right now. So
[05:46] (346.24s)
I'm happy that like you know that that's
[05:48] (348.24s)
setting a good example.
[05:49] (349.36s)
Great. Okay, let's go back to the
[05:51] (351.20s)
beginning. Like most of the folks in the
[05:52] (352.48s)
audience here are either students or
[05:54] (354.88s)
recent grads or grad students. Um and I
[05:57] (357.92s)
think hearing your story of like how you
[06:00] (360.32s)
started Perplexity uh would be really
[06:02] (362.24s)
interesting to them because it's
[06:03] (363.20s)
probably exactly the world that they're
[06:05] (365.12s)
in now.
[06:06] (366.16s)
Tell us how you got started. We started
[06:07] (367.84s)
the company without actually having
[06:09] (369.36s)
clear idea of what to build which is the
[06:11] (371.60s)
opposite of what YC advises which is
[06:14] (374.48s)
start from a project and turn it into a
[06:16] (376.56s)
company. I really think at this point in
[06:19] (379.44s)
time when AI is improving so fast you
[06:23] (383.28s)
don't have to rigidly stick to any one
[06:25] (385.12s)
idea when you're getting started but the
[06:28] (388.00s)
most important thing is you don't change
[06:29] (389.36s)
the idea every week like that you
[06:31] (391.44s)
shouldn't do either. So start with
[06:34] (394.72s)
something like brainstorm, think about
[06:36] (396.64s)
it and then try to immediately build it
[06:39] (399.44s)
and get it in the hands of people. Uh
[06:41] (401.84s)
one tool that we were building was
[06:43] (403.28s)
natural language SQL which we actually
[06:46] (406.16s)
thought about it as a search tool
[06:47] (407.68s)
searching over relational databases. I
[06:50] (410.80s)
allowed Twitter search but it never like
[06:52] (412.96s)
like the original version of Facebook
[06:54] (414.72s)
graph search. I allowed that when I was
[06:56] (416.88s)
much younger. So I wanted to like
[06:59] (419.12s)
rebuild that but using language models.
[07:02] (422.16s)
Um, and I love Twitter as a platform.
[07:04] (424.72s)
So, it was there's no good way to search
[07:06] (426.80s)
over Twitter. There still is no good way
[07:08] (428.56s)
to search over Twitter. But at least at
[07:10] (430.80s)
the time, we organized the entire
[07:13] (433.04s)
Twitter's data in the form of like
[07:14] (434.96s)
relational tables and just converted
[07:17] (437.84s)
every user's query into a SQL query and
[07:20] (440.88s)
ran it as against the database and it
[07:22] (442.72s)
was really really good and um that's
[07:25] (445.36s)
what got us started. But at some point
[07:28] (448.48s)
we figured it's better to like scale
[07:30] (450.48s)
this across the web and we cannot make
[07:33] (453.60s)
every website in the form of tables and
[07:37] (457.04s)
neither is it actually easy to answer
[07:38] (458.88s)
all sorts of questions. So we bet on the
[07:41] (461.28s)
fact that language models can do all the
[07:43] (463.20s)
reasoning and parsing and like
[07:45] (465.28s)
structuring later but the more important
[07:47] (467.44s)
thing is to start with something more
[07:48] (468.80s)
unstructured and that ended up becoming
[07:51] (471.36s)
perplexity.
[07:52] (472.48s)
Got it. Um, and maybe one step before
[07:55] (475.68s)
you actually left to go start the
[07:57] (477.76s)
company, like how did you find your
[07:59] (479.68s)
co-founders? How did you decide that
[08:02] (482.00s)
machine learning and AI was like the
[08:04] (484.32s)
area you wanted to focus on?
[08:05] (485.84s)
Cuz that was the only thing I was good
[08:07] (487.12s)
at. I was not good at anything else.
[08:09] (489.04s)
So, what's the point in starting a
[08:10] (490.32s)
company? I cannot start a delivery
[08:11] (491.84s)
company or a social media company. Like,
[08:14] (494.48s)
I'm not I'm not the right fit, right?
[08:16] (496.24s)
Uh, the only thing I knew was AI and
[08:18] (498.24s)
machine learning. In fact, it's funny. I
[08:20] (500.40s)
we started an AI company. uh but we're
[08:23] (503.20s)
not made fun of like not even training
[08:25] (505.04s)
our own models like mo but only the
[08:27] (507.36s)
foundation models are stuff we don't
[08:28] (508.88s)
train we train so many different models
[08:31] (511.36s)
but uh that's the extent to which you
[08:33] (513.36s)
need to have the intellectual humility
[08:35] (515.04s)
to know like what you're good at what is
[08:37] (517.44s)
actually doable for you within with the
[08:39] (519.44s)
resources that you have access to and
[08:42] (522.24s)
the co-founders are like people I like
[08:44] (524.40s)
like I knew from grad school so we had
[08:47] (527.04s)
been talking and discussing ideas for a
[08:48] (528.80s)
long time and um I think grad school is
[08:51] (531.20s)
a great way to like uh you know like
[08:53] (533.84s)
like identify your co-founders. You
[08:55] (535.68s)
don't talk to them with the you know
[08:57] (537.68s)
with the long-term calculation of like
[08:59] (539.92s)
oh this could be my co-founder of my
[09:01] (541.68s)
future company. You talk to them because
[09:03] (543.68s)
they're interesting people. And I think
[09:05] (545.76s)
that's essentially the value of the Y
[09:07] (547.36s)
cominator network. So even if your first
[09:09] (549.76s)
startup year fails, you get access to a
[09:12] (552.48s)
lot of amazing people and maybe they
[09:13] (553.84s)
could be your future co-founders. So
[09:15] (555.28s)
that that's essentially what grad school
[09:16] (556.72s)
was for me.
[09:17] (557.44s)
Yeah, that's awesome. Um, okay. So, you
[09:19] (559.84s)
launched this first version of
[09:21] (561.28s)
Perplexity, which is largely to like do
[09:24] (564.00s)
Twitter search effectively. Um,
[09:26] (566.64s)
at what point did it like start to work
[09:29] (569.52s)
and you maybe internally felt like, oh,
[09:31] (571.76s)
we should keep working on this. This is
[09:33] (573.36s)
going to be something to explore.
[09:35] (575.60s)
Yeah. So, whoever we gave early access
[09:37] (577.68s)
to, they were all very excited about it.
[09:39] (579.60s)
They kept using it repeatedly. I think
[09:42] (582.08s)
there's a phenomenon in products where
[09:44] (584.16s)
there's an initial wow factor. Yep.
[09:46] (586.48s)
And then mostly either drops completely
[09:48] (588.56s)
that that means you never had real
[09:50] (590.00s)
retention or it it definitely drops but
[09:52] (592.88s)
there sustained usage. So when we saw
[09:55] (595.36s)
that for the relational database
[09:57] (597.60s)
searches like Twitter, LinkedIn, GitHub,
[10:00] (600.56s)
we knew that we have like there was
[10:02] (602.00s)
something magical about combining large
[10:04] (604.24s)
language models in search. But then what
[10:06] (606.40s)
we did is like we dreamt bigger and said
[10:08] (608.88s)
what if we just give answers and site
[10:11] (611.92s)
the relevant sources. We launched that
[10:14] (614.00s)
as a discord bot and that was also
[10:15] (615.68s)
continually being used. It was not like
[10:17] (617.44s)
a one-day usage and people started
[10:19] (619.12s)
ignoring it. So that's when we decided
[10:21] (621.52s)
we had the courage to launch it. We
[10:23] (623.04s)
launched it 7 days after the chat GPT
[10:25] (625.44s)
launch especially at a time when chat
[10:27] (627.76s)
GPT did not have web right web search.
[10:29] (629.92s)
So that was a good moment. And I think
[10:31] (631.44s)
like many of the successful AI products
[10:33] (633.60s)
that people speak about today cursor
[10:35] (635.68s)
included all were like 2022 launches or
[10:38] (638.56s)
like like early 2023 or late 2022
[10:41] (641.36s)
launches. So they're all like old
[10:43] (643.12s)
people, you know, in the in the in the
[10:44] (644.72s)
in this AI time scale.
[10:46] (646.88s)
For me, the aha moment was like the New
[10:49] (649.60s)
Year Eve, there was like close to
[10:51] (651.84s)
700,000 queries. And I was like, okay,
[10:54] (654.64s)
this has the crappiest name for a
[10:56] (656.40s)
consumer product. It's called
[10:57] (657.44s)
Perplexity. Very hard. Nobody even knows
[11:00] (660.48s)
how to share it. And then it was so
[11:02] (662.48s)
slow. Took seven seconds to answer for a
[11:04] (664.72s)
query at the time. And um it was making
[11:08] (668.32s)
a lot of mistakes, hallucinations
[11:10] (670.88s)
and like a no-name company, no-name
[11:13] (673.68s)
founder, very one or two million dollars
[11:16] (676.08s)
in seed funding. Despite that, people
[11:18] (678.24s)
were caring enough to sharing
[11:19] (679.60s)
screenshots and like and and a new year
[11:21] (681.92s)
eve and you could be, you know, watching
[11:23] (683.60s)
Netflix, right? So, uh that's when I
[11:25] (685.92s)
knew there was something real here and I
[11:27] (687.52s)
started like optimizing for like, you
[11:29] (689.36s)
know, committing to this this vision.
[11:30] (690.88s)
Okay. And at that point, like on that
[11:33] (693.44s)
New Year's Eve, did you in your head
[11:35] (695.84s)
think I'm building a thing that could
[11:38] (698.00s)
really compete with Google and like take
[11:39] (699.76s)
over a market as big as what Google
[11:41] (701.76s)
offers or was it just a toy for you?
[11:44] (704.40s)
The first time the thought occurred to
[11:45] (705.76s)
me was when um Google wrote a blog post.
[11:48] (708.48s)
Sundar wrote a blog post about Bard.
[11:51] (711.92s)
Like that was around the time when we
[11:53] (713.60s)
were raising series A funding and
[11:56] (716.40s)
everybody said, "Okay, Bard is going to
[11:57] (717.92s)
do whatever you're doing." And it's like
[12:00] (720.16s)
why do I have to build bar like why not
[12:02] (722.24s)
just do it on Google right where you you
[12:05] (725.04s)
have all the distribution in the world
[12:06] (726.88s)
so why do you have to build a separate
[12:08] (728.32s)
product like just just update your core
[12:12] (732.16s)
the best possible asset to do this
[12:13] (733.68s)
exactly and I kept thinking it was
[12:16] (736.56s)
pretty obvious you cannot like if if if
[12:19] (739.60s)
people can get answers to best hotels to
[12:22] (742.48s)
stay in San Francisco with a view of the
[12:24] (744.08s)
Golden Gate Bridge or like Bay Bridge or
[12:26] (746.40s)
like where can I stay in New York like
[12:28] (748.32s)
next to the Central park with good
[12:30] (750.08s)
amenities or like which flight is the
[12:32] (752.48s)
best thing for me to take to fly from SF
[12:34] (754.32s)
to London. If you get direct answers to
[12:37] (757.20s)
these questions with booking links right
[12:39] (759.28s)
there, how are you going to mint money
[12:41] (761.68s)
from booking and Expedia and Kayak and
[12:44] (764.24s)
like like you know or like same same
[12:47] (767.20s)
thing for shopping. How are you going to
[12:49] (769.28s)
take money from um Amazon and like
[12:52] (772.24s)
Walmart for the same ad where they're
[12:54] (774.00s)
all bidding against each other? It's not
[12:56] (776.16s)
in their incentive to give you good
[12:57] (777.52s)
answers at all. So that's when I
[12:59] (779.60s)
realized that they have to build a
[13:00] (780.96s)
separate product but they can never
[13:03] (783.12s)
capitalize on their core distribution
[13:06] (786.40s)
and 2024 2023 especially and a large
[13:10] (790.24s)
part of 2024 too Google had like maybe
[13:14] (794.16s)
the fourth or fifth best models at any
[13:16] (796.08s)
moment. So as a startup outside Google,
[13:18] (798.96s)
you had access to AI that was better
[13:21] (801.52s)
than what Google internally had, which
[13:24] (804.00s)
was unprecedented, right? Until then, if
[13:27] (807.28s)
you had to compete with Google and you
[13:29] (809.12s)
had to build something that needed a lot
[13:30] (810.80s)
of AI in it, good luck, right? Like cuz
[13:33] (813.52s)
you never have an AI outside Google
[13:35] (815.60s)
that's even equal, leave alone being
[13:37] (817.28s)
better. But now it's a completely
[13:39] (819.20s)
reversal of the situation thanks to open
[13:41] (821.20s)
AI or anthropic or open source models.
[13:44] (824.08s)
So that plus innovator dilemma plus the
[13:47] (827.60s)
fact that we could make a lot of
[13:48] (828.88s)
mistakes and it's fine. Whereas for
[13:50] (830.64s)
Google, one mistake tanks their stock.
[13:52] (832.80s)
Like you remember the live demo of Bard
[13:54] (834.56s)
where it failed and the stock went down
[13:57] (837.84s)
So we knew that there was a lot of
[13:59] (839.28s)
advantages for us.
[14:00] (840.80s)
Yeah. And I heard you talk about this
[14:02] (842.48s)
recently but you know Google
[14:04] (844.16s)
specifically has been trying to build
[14:06] (846.88s)
perplexity like experiences and you know
[14:10] (850.56s)
AI mode.
[14:11] (851.20s)
Yeah. They just like change the name of
[14:12] (852.80s)
it each Google IO and then not really
[14:15] (855.36s)
true. I'm not I'm not I'm not like
[14:16] (856.80s)
saying something wrong.
[14:18] (858.40s)
Right. So it's like I I
[14:20] (860.72s)
look it might sound a little um cocky to
[14:24] (864.16s)
say that but it's true. Um the same
[14:26] (866.56s)
feature is being launched year after
[14:28] (868.40s)
year after year
[14:29] (869.60s)
with a different name with a different
[14:31] (871.28s)
VP with a different group of people but
[14:34] (874.40s)
it's the same thing except maybe it's
[14:36] (876.88s)
getting better but it's never getting
[14:38] (878.88s)
launched to everybody. One of the things
[14:40] (880.48s)
I I've come to admire about you is you
[14:43] (883.68s)
really have a focus on the user
[14:45] (885.44s)
experience and and you told me how you
[14:47] (887.76s)
kind of learned that from Larry Page.
[14:50] (890.00s)
Yeah. By reading the book about uh
[14:52] (892.24s)
Google. Um why do you think Google has
[14:55] (895.28s)
lost that ability?
[14:58] (898.08s)
Well, it's a much bigger business,
[14:59] (899.52s)
right? And it's not founder anymore. Uh
[15:02] (902.64s)
it's hard to take risks. Um I think they
[15:05] (905.28s)
have great people. Nobody like no one in
[15:07] (907.68s)
this audience would think Google has
[15:09] (909.04s)
incompetent people. I think they're like
[15:10] (910.48s)
really great engineers. It's largely the
[15:12] (912.72s)
incentive structure. It's hard to like
[15:14] (914.88s)
you know take a hit on your own stock
[15:17] (917.52s)
and do the thing that's long-term
[15:19] (919.36s)
correct. So you know honestly I'm happy
[15:22] (922.72s)
that that sort of dilemma exists because
[15:25] (925.12s)
otherwise where is the opening for
[15:27] (927.04s)
startups right and and then if startups
[15:29] (929.44s)
can succeed then it's going to be
[15:31] (931.60s)
monopolies getting bigger and bigger and
[15:33] (933.04s)
that's not great for the world. I
[15:34] (934.80s)
actually am very happy that we are able
[15:36] (936.80s)
to win and they're also like able to
[15:38] (938.80s)
like ship new products and people are
[15:41] (941.60s)
like first time comparing right earlier
[15:44] (944.48s)
for access to information you would
[15:46] (946.32s)
never even bother to compare an
[15:48] (948.64s)
alternative to Google.
[15:49] (949.68s)
It's true.
[15:50] (950.16s)
Like that was like considered a waste of
[15:51] (951.68s)
time a joke. Now at least you're like oh
[15:54] (954.80s)
I first go ask this app like I I'll ask
[15:56] (956.88s)
Google or I'll ask Chad GPT or I'll ask
[15:59] (959.12s)
Perplexity or ask Gemini and then maybe
[16:01] (961.68s)
you don't even ask Google anymore. You
[16:03] (963.68s)
just ask the AI apps and there are a
[16:05] (965.84s)
bunch of AI assistants and the phone
[16:08] (968.24s)
makers will start offering all of them
[16:10] (970.24s)
as alternatives. It's not going to be
[16:11] (971.60s)
like a locked in default search option.
[16:14] (974.56s)
So, I'm really happy that they're
[16:16] (976.24s)
competing in a world where a monopoly
[16:18] (978.64s)
hopefully doesn't exist and that creates
[16:21] (981.76s)
a more fair ground for everybody.
[16:23] (983.44s)
Yeah. Yeah. You were also telling me
[16:25] (985.28s)
backstage about um you know you are
[16:28] (988.08s)
facing this increased competition from a
[16:30] (990.48s)
variety of folks but if you look at your
[16:32] (992.16s)
numbers you haven't really seen effect
[16:34] (994.80s)
you know I read all the Twitter comments
[16:36] (996.48s)
every time the Google IO exactly the
[16:39] (999.44s)
same set of comments repeated this year
[16:41] (1001.68s)
u Google IO last year was AI overview
[16:44] (1004.32s)
and perplexity is dead this year was AI
[16:47] (1007.36s)
mode and perplexity is dead and I read
[16:49] (1009.84s)
all of that too and it's it's always fun
[16:52] (1012.16s)
I love it actually um because like like
[16:55] (1015.12s)
they know that they're all thinking like
[16:56] (1016.88s)
I don't I don't even expect these things
[16:58] (1018.40s)
or the people in the company are like
[16:59] (1019.76s)
thinking Google wouldn't build this or
[17:01] (1021.52s)
something like that but it's the reality
[17:03] (1023.52s)
is like nobody actually gets exposed to
[17:05] (1025.76s)
those features but competition is real
[17:08] (1028.48s)
okay let's assume let's let's
[17:11] (1031.04s)
accept that open AI is extremely
[17:12] (1032.88s)
wellunded doesn't have all these
[17:14] (1034.88s)
innovative DMA problems wants to
[17:17] (1037.28s)
actually ship search on chat GPT chat
[17:19] (1039.84s)
GPT is the most successful consumer AI
[17:22] (1042.32s)
product out there And so competing
[17:24] (1044.48s)
against it is very difficult which is
[17:26] (1046.72s)
why I I really want to like um push the
[17:29] (1049.36s)
company more on the browser side and I
[17:33] (1053.04s)
think comet the browser will be an
[17:35] (1055.36s)
abstraction layer above chat bots. I you
[17:37] (1057.84s)
could even imagine like I you know if
[17:39] (1059.84s)
you permit comet all your chat GPT chats
[17:42] (1062.56s)
can you know be fed into that AI and
[17:45] (1065.12s)
like you don't even have to worry about
[17:46] (1066.96s)
memory or personalization or like you
[17:48] (1068.88s)
know any of these things and it'll do a
[17:50] (1070.72s)
thing a lot of new things that a chatbot
[17:53] (1073.36s)
cannot do like accessing other tabs
[17:55] (1075.92s)
accessing your browsing history going
[17:58] (1078.00s)
and completing forms for you like paying
[18:00] (1080.08s)
your credit cards buying stuff for you
[18:03] (1083.44s)
um and being your scout you know going
[18:05] (1085.60s)
and doing all the research for to that
[18:07] (1087.68s)
sort of thing like periodic recurring
[18:09] (1089.68s)
tasks. I think that's the magic that the
[18:11] (1091.84s)
browser enables for you. And putting it
[18:13] (1093.92s)
into like mobile like like building
[18:15] (1095.84s)
mobile versions of this browser is going
[18:18] (1098.32s)
to be very hard like just engineering
[18:20] (1100.56s)
wise it's going to take many months. So
[18:22] (1102.48s)
I'm not really worried about like
[18:23] (1103.68s)
someone else trying to copy this. It's
[18:25] (1105.04s)
going to take time for anybody.
[18:26] (1106.64s)
Switching to a different browser is like
[18:28] (1108.32s)
a pretty big decision for a user. What
[18:30] (1110.88s)
what do you think will be the very
[18:32] (1112.88s)
shortterm things that your browser will
[18:35] (1115.76s)
do? so much better than what I can get
[18:38] (1118.08s)
today in Chrome that will make me want
[18:39] (1119.84s)
to switch.
[18:40] (1120.48s)
The perfect blend of AI, navigation, and
[18:43] (1123.44s)
agents is is is what we're going to
[18:46] (1126.00s)
offer. And um might sound like a boring
[18:49] (1129.20s)
answer, but no one's done that. And
[18:51] (1131.12s)
there are like hundreds of millions,
[18:52] (1132.64s)
probably close to a billion people using
[18:54] (1134.32s)
AI these days. So, the market's already
[18:57] (1137.04s)
pretty big.
[18:57] (1137.60s)
What's like a specific example of how I
[18:59] (1139.44s)
would do that, you know, if I had access
[19:00] (1140.96s)
to it tomorrow?
[19:02] (1142.32s)
You can schedule your meetings. uh you
[19:04] (1144.80s)
can reply to some of your emails that
[19:06] (1146.88s)
you don't even want to read. You can
[19:08] (1148.56s)
like for example let's say you're
[19:09] (1149.76s)
hosting a Y cominator event and you say
[19:12] (1152.08s)
I only want to accept Stanford dropouts
[19:14] (1154.40s)
and it can go through the entire list of
[19:16] (1156.24s)
people who applied and just filter based
[19:19] (1159.68s)
on who's you know took scrape their
[19:21] (1161.76s)
LinkedIn URLs
[19:23] (1163.28s)
filter based on whether they were
[19:24] (1164.80s)
Stanford and whether they dropped out or
[19:26] (1166.64s)
not and then accepted like that level of
[19:28] (1168.96s)
multi-step reasoning
[19:30] (1170.64s)
is something you can uniquely do. By the
[19:32] (1172.64s)
way, I'm not saying that's a good
[19:33] (1173.60s)
filter. Uh I wouldn't get in otherwise.
[19:36] (1176.72s)
And so hopefully you don't you're more
[19:38] (1178.80s)
open. Yeah.
[19:39] (1179.60s)
We look for deep mind researchers also.
[19:41] (1181.68s)
Yeah. Yeah. Don't worry. Um Okay, cool.
[19:44] (1184.88s)
Let's talk a little bit about how you
[19:46] (1186.56s)
run the company now, right? I don't know
[19:49] (1189.12s)
if you wanted to say how many employees
[19:50] (1190.56s)
you have.
[19:51] (1191.12s)
Yeah, we have about 200.
[19:52] (1192.48s)
Okay, so the company's getting bigger.
[19:54] (1194.32s)
Um you now have access to code writing
[19:58] (1198.16s)
AI tools. Um, are you guys just like
[20:00] (1200.80s)
full in on that stuff? Are you vibe
[20:02] (1202.40s)
coding everything? How what what's it
[20:04] (1204.32s)
look like?
[20:04] (1204.72s)
I mean, you you don't want to wipe code
[20:06] (1206.32s)
everything, right? Like like like we
[20:08] (1208.40s)
frequently run into infra issues and you
[20:10] (1210.72s)
don't want a wipe coder right there
[20:12] (1212.88s)
fixing it on live things on production
[20:15] (1215.12s)
like I do want like people well trained
[20:16] (1216.88s)
in regular software engineering,
[20:19] (1219.20s)
infrastructure, distributed systems like
[20:21] (1221.52s)
you don't want to like replace these
[20:22] (1222.88s)
skills. But yeah front-end design that's
[20:27] (1227.04s)
where we are seeing tremendous adoption
[20:28] (1228.96s)
like cursor is being used by everybody.
[20:31] (1231.28s)
uh we made it mandatory to use at least
[20:33] (1233.84s)
one AI coding tool and internally at
[20:36] (1236.96s)
perplexity it happens to be cursor and
[20:38] (1238.64s)
like a mix mix of cursor and github
[20:40] (1240.48s)
copilot but yeah we definitely made it
[20:42] (1242.96s)
compulsory and so the way machine
[20:45] (1245.04s)
learning people are do using it AI
[20:47] (1247.04s)
people are like sometimes they read a
[20:49] (1249.20s)
paper and they can just upload a
[20:51] (1251.68s)
screenshot of the pseudo code and uh ask
[20:55] (1255.04s)
cursor to like just edit the files to
[20:57] (1257.04s)
implement this new algorithm and then
[20:59] (1259.20s)
it's able to like uh write it on unit
[21:01] (1261.44s)
tests and then uh run an experiment
[21:03] (1263.92s)
pretty quickly that uh is reducing the
[21:06] (1266.40s)
experimentation time from like 3 4 days
[21:09] (1269.12s)
to like literally 1 hour or like there
[21:11] (1271.52s)
are people who don't know design and so
[21:13] (1273.44s)
sometimes I'm I just give them feedback
[21:15] (1275.84s)
where I take a screenshot of my iOS app
[21:18] (1278.72s)
and I say this button needs to move here
[21:20] (1280.40s)
with an arrow and they upload my
[21:22] (1282.08s)
screenshot to cursor and then ask it to
[21:24] (1284.64s)
like write write a change to the swift
[21:26] (1286.72s)
UI file. So that level of change is
[21:29] (1289.52s)
incredible. Like like the speed at which
[21:31] (1291.44s)
you can fix bugs and ship to production
[21:34] (1294.48s)
is is crazy.
[21:36] (1296.08s)
The more bugs there are as long as you
[21:37] (1297.84s)
can fix them fast.
[21:38] (1298.80s)
Yeah, bugs are always ahead of like how
[21:40] (1300.48s)
fast people can write code though.
[21:41] (1301.92s)
But just just to be clear, I'm a big fan
[21:43] (1303.92s)
of all these tools, but it is also
[21:46] (1306.16s)
introducing new bugs and many people
[21:49] (1309.52s)
don't know how to fix them
[21:51] (1311.28s)
and they don't even know how the bug got
[21:53] (1313.20s)
introduced and they have to go find it
[21:54] (1314.72s)
again. So it's not perfect and I
[21:57] (1317.84s)
actually like the more newer tools like
[22:00] (1320.00s)
clawed code seems to be far smarter than
[22:03] (1323.28s)
like what cursor is able to do. So I'm
[22:06] (1326.40s)
actually like like really positive that
[22:08] (1328.72s)
this is the right future but there are
[22:10] (1330.16s)
there are issues right now.
[22:12] (1332.16s)
Yeah. um in talking to a lot of the
[22:13] (1333.84s)
folks here, one of the major questions
[22:16] (1336.16s)
that I've heard is um as these coding
[22:19] (1339.20s)
tools get better and better and better,
[22:21] (1341.60s)
what is the like actual
[22:24] (1344.24s)
enduring value of a company like yours
[22:26] (1346.40s)
if if increasingly it's easy to
[22:28] (1348.80s)
replicate what you have done using these
[22:30] (1350.96s)
tools? What's your take on that general
[22:33] (1353.04s)
type of question?
[22:35] (1355.04s)
I mean brand definitely has a big value,
[22:37] (1357.68s)
right? Like there are cursor
[22:40] (1360.16s)
competitors, perplexity competitors like
[22:43] (1363.44s)
OpenAI will have like their own cursor.
[22:46] (1366.00s)
OpenAI has perplexity within chat GPT
[22:48] (1368.40s)
that did not kill any of these
[22:50] (1370.24s)
companies. So there is a certain brand
[22:52] (1372.80s)
value that once you acquire at the scale
[22:54] (1374.64s)
of like several millions of users,
[22:56] (1376.56s)
paying users, you don't actually die
[22:59] (1379.28s)
that fast. You earn the right to survive
[23:01] (1381.12s)
and keep building. Uh so brand is
[23:03] (1383.68s)
important. uh narrative is very
[23:06] (1386.00s)
important to the brand like you have to
[23:08] (1388.24s)
communicate to people why do you even
[23:10] (1390.08s)
need to exist for us it's the focus on
[23:12] (1392.56s)
accuracy okay let there exist 100 chat
[23:15] (1395.36s)
bots but we are the most focused on
[23:17] (1397.52s)
getting as many answers right as
[23:19] (1399.84s)
possible we focus on speed time to first
[23:22] (1402.88s)
token on app or web like we're still the
[23:25] (1405.68s)
fastest despite doing search uh we focus
[23:28] (1408.32s)
a lot on like how we present the answer
[23:30] (1410.64s)
so there are some things you are
[23:32] (1412.32s)
obsessed about because you care about it
[23:34] (1414.48s)
and that becomes your narrative and your
[23:36] (1416.00s)
brand identity. And if you manage to get
[23:40] (1420.00s)
reasonable amount of distribution, not
[23:41] (1421.92s)
saying 100 million users, but tens of
[23:44] (1424.40s)
millions, then you earn the right to
[23:46] (1426.64s)
keep playing the game no matter what
[23:48] (1428.16s)
other people ship. Until then, it's
[23:50] (1430.08s)
definitely a challenge. You have to
[23:51] (1431.44s)
worry about it. Even now, we worry about
[23:53] (1433.36s)
it and the only solution is to move fast
[23:55] (1435.04s)
and keep shipping. be beyond brand like
[23:57] (1437.84s)
do you think about any network effect
[24:00] (1440.40s)
types of things emerging with perplexity
[24:04] (1444.00s)
I mean brand has network effects right
[24:05] (1445.84s)
like people people tell each other about
[24:07] (1447.52s)
the brand
[24:08] (1448.16s)
but no AI product has within app network
[24:12] (1452.16s)
effect like like it's not like WhatsApp
[24:14] (1454.08s)
where if you build a WhatsApp rival meta
[24:17] (1457.12s)
has a definitely like a questionable
[24:19] (1459.04s)
brand right like people don't
[24:21] (1461.36s)
necessarily trust Meta's products they
[24:23] (1463.60s)
think like these are ad products Despite
[24:26] (1466.32s)
that, nobody's able to switch off
[24:27] (1467.92s)
WhatsApp that easily because all your
[24:30] (1470.40s)
contacts, your groups, everything is
[24:32] (1472.32s)
there. AI doesn't quite have that yet.
[24:35] (1475.20s)
Uh mainly because you can easily export
[24:38] (1478.24s)
your chat GPT history, upload it
[24:40] (1480.64s)
somewhere else or uh things like that. I
[24:44] (1484.08s)
think the browser will definitely be one
[24:45] (1485.68s)
play to like you know figure this out
[24:48] (1488.40s)
because as your browsing history and
[24:50] (1490.24s)
like which again you can still export
[24:51] (1491.84s)
but not the same as just getting a dump
[24:54] (1494.16s)
a CSV dump
[24:55] (1495.84s)
uh and your passwords your wallet your
[24:58] (1498.48s)
agent remembers you there's a lot of
[25:00] (1500.00s)
tasks that are running on the browser
[25:01] (1501.84s)
that you rely on your day-to-day life
[25:03] (1503.76s)
and work that's one way to like keep get
[25:06] (1506.40s)
the product a lot more sticky and like
[25:08] (1508.32s)
create more network effects especially
[25:10] (1510.00s)
if multiple people rely on the same set
[25:11] (1511.76s)
of tasks you're sharing it with them.
[25:14] (1514.08s)
That's one way to get all this into like
[25:15] (1515.76s)
the next level.
[25:16] (1516.88s)
It also sounds like a lot of the stuff
[25:18] (1518.40s)
that you aspire to solve for users
[25:21] (1521.12s)
requires integrations or partnerships or
[25:23] (1523.92s)
something with a bunch of other
[25:25] (1525.20s)
companies in the world. Yeah. And if you
[25:27] (1527.60s)
can get those to be good, then there is
[25:30] (1530.40s)
somewhat of a network effect in the
[25:31] (1531.92s)
sense that your product will be good and
[25:33] (1533.76s)
some competitor would have to build the
[25:35] (1535.52s)
same integration or same deal with with
[25:38] (1538.08s)
these providers.
[25:39] (1539.76s)
What does that look like do you think in
[25:40] (1540.96s)
the future? like does perplexity do
[25:43] (1543.04s)
deals with all the airlines in the world
[25:46] (1546.16s)
and all the hotels and all the
[25:48] (1548.16s)
e-commerce providers?
[25:49] (1549.68s)
So we we already work with self book uh
[25:52] (1552.48s)
they power all the hotel bookings
[25:54] (1554.32s)
natively done on perplexity. We work
[25:56] (1556.56s)
with trip advisor to surface all the
[25:59] (1559.12s)
reviews of hotels and different you know
[26:02] (1562.40s)
places. We have like collaborations for
[26:04] (1564.40s)
the maps. We work with Yelp. uh we also
[26:08] (1568.16s)
like you know for shopping we have a lot
[26:10] (1570.80s)
of merchants who are directly selling on
[26:13] (1573.68s)
uh and then we work with firmly to like
[26:15] (1575.60s)
support the bookings like like native
[26:17] (1577.60s)
purchases so there's already a lot of
[26:19] (1579.84s)
partnerships Shopify is one of our
[26:21] (1581.52s)
partners on finance we work with FMP
[26:24] (1584.72s)
sports we work with like uh stats
[26:26] (1586.64s)
perform so there's a lot of data
[26:28] (1588.48s)
providers already working with us to on
[26:31] (1591.12s)
these verticals and we just think it's
[26:32] (1592.96s)
going to expand further as agents start
[26:35] (1595.52s)
to do things. Different people are okay
[26:38] (1598.16s)
with like becoming MCP servers. Some
[26:40] (1600.80s)
people are not. Some people just want to
[26:42] (1602.56s)
like like preserve their websites. The
[26:44] (1604.88s)
browser agent will be generic enough
[26:46] (1606.48s)
that it'll respect whatever the third
[26:48] (1608.88s)
party wants because at the end of the
[26:50] (1610.88s)
day, the agent is the one that's being
[26:53] (1613.36s)
permitted by the user to act on their
[26:55] (1615.04s)
behalf. And um if there is no MCP
[26:58] (1618.40s)
server, it's still fine. You can just
[27:00] (1620.48s)
use these tabs as if the user would have
[27:03] (1623.20s)
done it. And uh that's the key advantage
[27:06] (1626.08s)
of the browser that you do not have if
[27:08] (1628.72s)
you commit entirely to just the MCP
[27:10] (1630.80s)
vision. If you commit entirely to MCP
[27:13] (1633.36s)
vision, you require these third party
[27:15] (1635.28s)
MCP servers to work reliably. Uh the
[27:18] (1638.08s)
data that they send you uh on with the
[27:20] (1640.40s)
MCP protocol has to be perfect. Your
[27:22] (1642.64s)
chatbot has to like like you know deal
[27:24] (1644.24s)
with all these issues that exist. On the
[27:27] (1647.76s)
other hand, if you just ground up design
[27:29] (1649.52s)
it as the way a human would use that
[27:31] (1651.20s)
website, you have full control over like
[27:33] (1653.84s)
how to how to do it, you don't have to
[27:35] (1655.76s)
rely on someone else doing the
[27:37] (1657.12s)
engineering well on their end.
[27:39] (1659.52s)
Um, let's talk next about business
[27:42] (1662.16s)
model. Your main competitor Google,
[27:44] (1664.24s)
their business model, as you've talked
[27:45] (1665.52s)
about, is selling ads. Um, and you think
[27:48] (1668.48s)
that prevents them from being really
[27:50] (1670.24s)
good at what you're doing. So what what
[27:52] (1672.08s)
will your business model be and how will
[27:53] (1673.60s)
you get it to be on the order of
[27:55] (1675.36s)
magnitude of of Google's?
[27:57] (1677.52s)
I don't know if you'll ever get order of
[27:59] (1679.52s)
magnitude profits as Google. Uh just to
[28:02] (1682.48s)
be clear and I don't think that's
[28:04] (1684.08s)
needed. No one in the history even
[28:06] (1686.16s)
Google themselves never has had another
[28:08] (1688.80s)
business that had the margins that
[28:10] (1690.80s)
Google has. So it's completely
[28:12] (1692.80s)
reasonable to get something far far
[28:15] (1695.36s)
better than any public company out there
[28:17] (1697.12s)
right now and and way be still way below
[28:19] (1699.92s)
Google. Number two, I think the
[28:21] (1701.60s)
subscription revenue is like really
[28:24] (1704.08s)
encouraging. We never expected to get
[28:26] (1706.32s)
this far and then we think like we can
[28:28] (1708.32s)
grow at least, you know, a few billions
[28:30] (1710.72s)
a year in just subs, which is a great
[28:32] (1712.88s)
business. usage based pricing where
[28:36] (1716.80s)
people are paying an agent for
[28:38] (1718.40s)
completing a task or people have
[28:40] (1720.64s)
recurring tasks and they pay based on
[28:43] (1723.04s)
you know every single use of the task
[28:45] (1725.60s)
and they normalize this system based on
[28:48] (1728.24s)
like how much it would take to hire a
[28:49] (1729.84s)
person to do that for them uh is going
[28:52] (1732.32s)
to be a thing I don't exactly know how
[28:54] (1734.24s)
it's going to play out what are the
[28:55] (1735.84s)
margins going to be on that potentially
[28:57] (1737.84s)
it's going to be way better than um
[28:59] (1739.92s)
subscriptions in terms of volume of
[29:02] (1742.08s)
people who would pay for it but it might
[29:03] (1743.60s)
be lower margins because it's usage
[29:05] (1745.84s)
based. So you're still going to be
[29:07] (1747.04s)
spending on all those queries. Someone
[29:08] (1748.96s)
might be paying a subscription to like
[29:10] (1750.56s)
one of these AI apps and might not have
[29:12] (1752.24s)
used it for an entire month. And so
[29:14] (1754.08s)
that's good margins on that user, right?
[29:16] (1756.48s)
So I don't actually have a clear sense
[29:18] (1758.16s)
of how this is all going to like evolve.
[29:20] (1760.32s)
But all I know is subscriptions and
[29:22] (1762.48s)
usage based pricing are going to be a
[29:24] (1764.24s)
thing. transactions, you know, uh if
[29:27] (1767.28s)
people start buying more through AIS,
[29:30] (1770.64s)
uh taking a cut out of the transactions
[29:32] (1772.56s)
is good. It's going to be CPAs have
[29:34] (1774.80s)
historically been way lower margins than
[29:36] (1776.48s)
CPCs, which is why Google never became a
[29:39] (1779.04s)
transaction platform, which is why I
[29:41] (1781.28s)
said like you're going to make a lot of
[29:42] (1782.56s)
money here. You may never make as much
[29:44] (1784.32s)
money as Google.
[29:45] (1785.20s)
Yeah. I Google Google's business model
[29:47] (1787.60s)
is potentially the best business model
[29:49] (1789.28s)
ever. Ever. So, yeah, it's fair to not.
[29:51] (1791.68s)
Maybe it was so good that you needed AI
[29:53] (1793.76s)
to kill it basically.
[29:55] (1795.12s)
All right, let's we're going to do some
[29:56] (1796.72s)
audience Q&A in a little bit, but um
[29:58] (1798.88s)
before we get there, I kind of wanted to
[30:00] (1800.48s)
understand your advice for the folks in
[30:03] (1803.12s)
this room, right? Like if you were in
[30:04] (1804.96s)
their position back whatever it was four
[30:07] (1807.36s)
years ago, what would you advise they
[30:10] (1810.72s)
I would say uh work incredibly hard.
[30:13] (1813.84s)
There is no substitute for it. Don't
[30:16] (1816.72s)
think like you're very smart like like
[30:19] (1819.68s)
strategizing the right way to build a
[30:21] (1821.52s)
company despite all like what big model
[30:24] (1824.32s)
labs are doing. You should assume that
[30:26] (1826.96s)
if you have a big hit, if your company
[30:30] (1830.40s)
is something that can make revenue on
[30:32] (1832.08s)
the scale of hundreds of millions of
[30:33] (1833.44s)
dollars or potentially billions of
[30:35] (1835.04s)
dollars, you should always assume that a
[30:37] (1837.60s)
model company will copy it. Mainly
[30:40] (1840.16s)
because they are really looking for
[30:41] (1841.52s)
revenue. They raise like tens of
[30:43] (1843.76s)
billions or close to 50 billion and they
[30:46] (1846.48s)
need to justify all that capex spend and
[30:48] (1848.96s)
they need to keep searching for new ways
[30:50] (1850.64s)
to make money. So they will copy
[30:52] (1852.80s)
anything that's good. I think you got to
[30:54] (1854.56s)
live with that fear. You have to embrace
[30:56] (1856.16s)
it and realize that like your mode comes
[30:59] (1859.28s)
from moving fast and building your own
[31:01] (1861.12s)
identity around what you're doing
[31:02] (1862.96s)
because users at the end care. Like when
[31:04] (1864.88s)
you're trying to get like a specific
[31:06] (1866.72s)
person for your house help, you are
[31:08] (1868.88s)
searching for that specific person. and
[31:10] (1870.96s)
you're not like going for a general
[31:13] (1873.28s)
agency that handles all of it. So, um I
[31:16] (1876.48s)
think there's like like real benefit
[31:18] (1878.24s)
from embracing that fear and like
[31:20] (1880.32s)
sleeping with that fear and waking up
[31:22] (1882.72s)
every day and like feeling excited about
[31:24] (1884.40s)
what you're going to build because
[31:25] (1885.44s)
that's the only thing that'll keep you
[31:26] (1886.96s)
going.
[31:27] (1887.84s)
Well, you guys are the perfect example
[31:29] (1889.52s)
of how it's possible to go up against
[31:31] (1891.52s)
somebody as big as Google. That's great.
[31:33] (1893.44s)
All right, let's do some Q&A. We'll
[31:35] (1895.76s)
start on the left side here. Go for it.
[31:38] (1898.72s)
Uh, hi, my name is Sammy and I just want
[31:41] (1901.68s)
to personally thank you for helping me
[31:43] (1903.68s)
get a 100 in my theory of knowledge
[31:45] (1905.60s)
course. Uh, would not have been able to
[31:47] (1907.52s)
do it without you. No shame. Um, quick
[31:50] (1910.08s)
question for you. Uh, you know, with
[31:51] (1911.60s)
your recent partnership with Nvidia to
[31:54] (1914.08s)
ship AI models across Europe, um,
[31:56] (1916.40s)
there's been talks about perplexity
[31:57] (1917.92s)
being installed on um, all Samsung
[32:00] (1920.32s)
phones or pre-installed. Um, and that
[32:02] (1922.64s)
could lift your valuation towards 14
[32:04] (1924.80s)
billion uh, according to sources like
[32:06] (1926.40s)
Bloomberg. Um it's a heavy
[32:08] (1928.40s)
responsibility being the default search
[32:10] (1930.24s)
engine for you know the mainstream
[32:12] (1932.00s)
population. What do you think are the
[32:13] (1933.92s)
most important factors at Perplexity to
[32:16] (1936.32s)
prevent hallucinations or incorrect uh
[32:19] (1939.20s)
data from you know being given to the
[32:21] (1941.12s)
masses? Thank you so much.
[32:22] (1942.88s)
Thank you. Uh hallucinations is
[32:25] (1945.44s)
something we we really care about. We
[32:27] (1947.68s)
we're building benchmarks internally uh
[32:30] (1950.32s)
to keep up to date with that. The only
[32:33] (1953.04s)
way there is to keep building a better
[32:34] (1954.96s)
search index. uh keep capturing better
[32:37] (1957.92s)
snippets of all the web pages and then
[32:39] (1959.92s)
like these models like like you know are
[32:41] (1961.92s)
getting fast enough that you can have
[32:44] (1964.32s)
them reason multistep for every query
[32:47] (1967.04s)
without incurring too much cost and so
[32:49] (1969.04s)
that's another way to reduce
[32:50] (1970.48s)
hallucinations.
[32:51] (1971.68s)
I I want to ask you about like the
[32:54] (1974.08s)
innovator's dilemma. So, if you were in
[32:56] (1976.88s)
Sund's shoes or in um in like the Google
[33:00] (1980.48s)
co-founder's shoes, like what would you
[33:01] (1981.92s)
do and how would you kind of come up
[33:03] (1983.84s)
with maybe changing their business model
[33:05] (1985.28s)
even if it's a worse model?
[33:07] (1987.12s)
So, if you were running Google competing
[33:09] (1989.28s)
against yourself, what would you do?
[33:11] (1991.52s)
I I think I don't envy that job at all.
[33:14] (1994.32s)
Um I I nobody in the world wants that
[33:16] (1996.72s)
job. It's it's it's a very difficult
[33:19] (1999.92s)
job. Would you sacrifice the business
[33:22] (2002.72s)
model in order to get like a new
[33:25] (2005.44s)
get the next product
[33:26] (2006.80s)
or would you ship it as a separate
[33:28] (2008.16s)
product? Like if you're Google, would
[33:29] (2009.60s)
you just build a separate thing that is
[33:31] (2011.52s)
the perplexity competitor and and
[33:33] (2013.68s)
sacrifice the distribution advantage
[33:35] (2015.20s)
that you have in the short run?
[33:39] (2019.76s)
Yeah, I I don't like genuinely I don't
[33:41] (2021.92s)
know. I think uh I can say all what I
[33:44] (2024.56s)
want but they have more data on like
[33:46] (2026.08s)
what their users are doing and um there
[33:49] (2029.20s)
are a lot of people in the world who
[33:50] (2030.24s)
hate AI by the way so I think just
[33:53] (2033.28s)
throwing AI down people's throats on
[33:55] (2035.60s)
such a you know massive distribution
[33:57] (2037.44s)
area is not easy what I would do I would
[34:00] (2040.08s)
I definitely don't know and I I don't
[34:02] (2042.16s)
want to be in that position also by the
[34:03] (2043.76s)
way if if if ads are part of every AI
[34:06] (2046.40s)
answer you're going to hate it too um
[34:09] (2049.12s)
and and so um it's good that There
[34:12] (2052.08s)
there's alternatives like us.
[34:14] (2054.80s)
All right.
[34:17] (2057.20s)
Hey Arvin, my name is Akshad. Uh so in a
[34:20] (2060.56s)
recent interview with Nikl Kamat, he
[34:22] (2062.72s)
asked you for an internship at
[34:24] (2064.48s)
Perplexity. So I was just wondering how
[34:26] (2066.80s)
that um arrangement is going.
[34:28] (2068.80s)
He he came he came to the office. He
[34:31] (2071.04s)
spent a couple of days. I mean he hasn't
[34:33] (2073.04s)
posted about it. So I'll let him post
[34:34] (2074.96s)
about it. But I we did spend time with
[34:36] (2076.88s)
him. Uh it was not a proper internship
[34:39] (2079.12s)
but we did speak to him for a while.
[34:41] (2081.92s)
I want to start by saying thank you very
[34:43] (2083.76s)
much for your very candid answers. I
[34:45] (2085.76s)
really appreciated that. So um a lot of
[34:48] (2088.24s)
startups they find some like cool
[34:49] (2089.60s)
application of foundation models and
[34:52] (2092.48s)
then they'll like build something off of
[34:54] (2094.00s)
that but then if it does gain traction
[34:57] (2097.28s)
then the foundation models will
[34:59] (2099.28s)
consolidate that into their own
[35:00] (2100.72s)
infrastructure. And Perplexity sort of
[35:03] (2103.04s)
has that issue too with like a lot of
[35:04] (2104.64s)
LMS adding search like TAGBT,
[35:08] (2108.16s)
Gemini, companies like Coher. So I was
[35:12] (2112.16s)
just wondering like how would you
[35:13] (2113.92s)
approach something like that? Would you
[35:15] (2115.44s)
try to pivot just get better at what you
[35:17] (2117.44s)
do or I think I would say uh pick
[35:20] (2120.64s)
something you want to like be known for?
[35:22] (2122.88s)
Uh yes there are other people
[35:24] (2124.64s)
integrating search but we still want to
[35:27] (2127.12s)
be the fastest and most accurate and
[35:29] (2129.44s)
obviously I cannot just say that and and
[35:31] (2131.44s)
then stop like we need to figure out a
[35:34] (2134.16s)
new strategy too and and build new
[35:36] (2136.32s)
products that don't exist yet. Uh so our
[35:39] (2139.12s)
browser will be that bet for us and
[35:41] (2141.28s)
browser and search are not two
[35:43] (2143.12s)
distinctive products. They're actually
[35:45] (2145.44s)
like the browser is a natural graduation
[35:47] (2147.36s)
step from search just like how Google
[35:49] (2149.76s)
graduated from Google search to Chrome
[35:51] (2151.84s)
and Chrome is the main reason they got
[35:53] (2153.76s)
billions of daily queries from hundreds
[35:55] (2155.68s)
of millions. So when Google IPOed they
[35:57] (2157.92s)
had no browser and they had like maybe a
[36:00] (2160.00s)
100 million queries now you know like
[36:02] (2162.32s)
it's like 10 billion or something. So
[36:04] (2164.80s)
the browser is an important part of that
[36:06] (2166.40s)
and then so that's why we we are making
[36:08] (2168.08s)
a massive bet on that and agents can
[36:10] (2170.32s)
only be built with a browser. I'm very
[36:12] (2172.16s)
like convinced about that vision that if
[36:14] (2174.24s)
you want to have a mobile agent that you
[36:17] (2177.76s)
can actually build and implement without
[36:21] (2181.36s)
being restricted by whatever OS rules
[36:23] (2183.68s)
that Apple or Google sets in terms of
[36:25] (2185.84s)
not being able to call third party apps.
[36:28] (2188.80s)
Expecting every mobile app to have MCP
[36:31] (2191.28s)
servers and then like connecting all
[36:33] (2193.20s)
their data to your thing is not going to
[36:35] (2195.84s)
be that straightforward. Like nobody
[36:37] (2197.60s)
wants to be disintermediated by an AI
[36:39] (2199.36s)
that quickly. So the browser will be a
[36:41] (2201.68s)
great way to build all these things.
[36:42] (2202.88s)
Thank you.
[36:44] (2204.00s)
So as a lot of us here have done um
[36:46] (2206.96s)
we've tried we've failed at our startups
[36:49] (2209.68s)
um you know some of us have been more
[36:51] (2211.12s)
successful than others some like me have
[36:53] (2213.04s)
failed when you're in that moment
[36:55] (2215.28s)
failing over and over again. What do you
[36:58] (2218.00s)
tell yourself as CEO or as an
[37:01] (2221.20s)
entrepreneur to to win to teach yourself
[37:03] (2223.60s)
to win?
[37:04] (2224.40s)
What what do I tell myself when I feel
[37:06] (2226.00s)
like I might fail?
[37:07] (2227.84s)
Yeah. or when you're in that very
[37:09] (2229.60s)
specific moment of failing where you
[37:11] (2231.28s)
feel like everything's crashing down on
[37:12] (2232.80s)
you or this feature isn't working or
[37:14] (2234.40s)
this bug has popped up, how do you get
[37:16] (2236.32s)
through that and what do you think your
[37:17] (2237.60s)
biggest motivational factor is in that
[37:19] (2239.60s)
realm?
[37:20] (2240.56s)
Or maybe like at the beginning before it
[37:22] (2242.48s)
started to take off, what gave you the
[37:25] (2245.12s)
hope to keep working on it versus just
[37:27] (2247.44s)
go back to OpenAI and get your job?
[37:30] (2250.24s)
I just watched uh the Elon Musk videos
[37:33] (2253.04s)
on YouTube.
[37:34] (2254.88s)
No, I'm serious.
[37:36] (2256.57s)
[Applause]
[37:42] (2262.96s)
I I can tell you which video. There's a
[37:44] (2264.72s)
video where there's like a third failure
[37:47] (2267.12s)
in a row and like what do you think? And
[37:49] (2269.44s)
he's like I don't ever give up. I would
[37:51] (2271.12s)
have to be dead or incapacitated.
[37:53] (2273.60s)
So you'd say you're also never going to
[37:55] (2275.04s)
give up?
[37:56] (2276.08s)
Yeah, I I hope to I hope to like stay
[37:58] (2278.24s)
that way. It's not easy. Uh I think he's
[38:00] (2280.80s)
done it for way longer and that's why
[38:02] (2282.48s)
you all like respect him. But that's you
[38:05] (2285.28s)
know there are examples of great
[38:07] (2287.36s)
entrepreneurs who have done this despite
[38:10] (2290.00s)
all the odds stacked against them. So
[38:12] (2292.72s)
what do you have to lose? Just keep
[38:14] (2294.80s)
going.
[38:15] (2295.92s)
Thank you.
[38:18] (2298.43s)
[Applause]
[38:23] (2303.92s)
Uh yeah my question is about uh kind of
[38:26] (2306.48s)
the sustainability of perpetuity not in
[38:28] (2308.40s)
terms of the business model but just in
[38:29] (2309.84s)
terms of the web in general. Um, you
[38:31] (2311.52s)
know, a lot of studies have come out
[38:32] (2312.64s)
recently showing that AI search engines
[38:34] (2314.40s)
like Perplexity drive a lot less traffic
[38:36] (2316.48s)
to websites. So, I'm curious, what do
[38:38] (2318.08s)
you think like the web will look like in
[38:39] (2319.44s)
5 to 10 years when a lot of these
[38:41] (2321.20s)
websites, you know, they're not getting
[38:42] (2322.64s)
as much traffic and so they have to kind
[38:44] (2324.24s)
of cease their operations and like the
[38:45] (2325.84s)
web will just be a lot quieter of a
[38:47] (2327.84s)
place for content creation. How do you
[38:49] (2329.44s)
think perplexity fits into that? And
[38:50] (2330.88s)
what do you think the web will look like
[38:52] (2332.00s)
in that era?
[38:53] (2333.60s)
I think that there are going to be uh,
[38:55] (2335.52s)
you know, the web is already pretty long
[38:57] (2337.28s)
tail uh, and there's a massive power
[38:59] (2339.60s)
loss. So I I feel like the parallel is
[39:01] (2341.52s)
going to get even more skewed. That is
[39:04] (2344.00s)
very obvious. There are going to be
[39:06] (2346.24s)
certain brands that are wellnown and
[39:08] (2348.08s)
they're going to preserve direct organic
[39:10] (2350.16s)
visits, but those who are trying to game
[39:12] (2352.72s)
the SEO system and trying to get
[39:14] (2354.40s)
traffic, I think they're definitely
[39:15] (2355.76s)
going to have a harder time.
[39:17] (2357.52s)
Okay. Yeah.
[39:19] (2359.12s)
Hi Rean, good afternoon. Uh firstly,
[39:21] (2361.60s)
where do you place the line between
[39:23] (2363.44s)
summarization and plagiarism in report
[39:25] (2365.76s)
generation? And how do you avoid IP
[39:28] (2368.16s)
violations in your product? And
[39:30] (2370.48s)
secondly, how do you deal with political
[39:32] (2372.48s)
bias? Bias and political sorry,
[39:35] (2375.28s)
political bias and personal interest in
[39:37] (2377.76s)
news articles and other human written
[39:39] (2379.52s)
sources.
[39:40] (2380.64s)
Yeah, I think there are cases where you
[39:42] (2382.48s)
actually have objective truth, right?
[39:45] (2385.20s)
Like what was the score in the NBA game?
[39:47] (2387.84s)
what is the live weather right now in
[39:50] (2390.00s)
San Francisco where you don't want to
[39:52] (2392.08s)
you don't want to be wrong ever on those
[39:53] (2393.84s)
queries and and people know what is true
[39:56] (2396.72s)
but you even there you're trusting right
[39:58] (2398.72s)
like you're trusting some data provider
[40:00] (2400.40s)
who's tracking the live game the TV
[40:02] (2402.40s)
that's showing you the number or Apple
[40:04] (2404.64s)
or Google's like acute weather all these
[40:07] (2407.04s)
things so at some point it all relies on
[40:09] (2409.92s)
trust and trust is built over time based
[40:13] (2413.12s)
on being being accurate reliably and so
[40:16] (2416.32s)
trying to surface the right data from
[40:17] (2417.92s)
the right people who have earned the
[40:19] (2419.92s)
right to like like be surface in AI is
[40:22] (2422.80s)
is how we think about it for accuracy.
[40:25] (2425.28s)
Now there are things that don't have one
[40:27] (2427.60s)
clear accurate answer. I think there
[40:31] (2431.92s)
the best thing we can do is offer all
[40:34] (2434.00s)
the perspectives
[40:35] (2435.68s)
and not really take a clear opinion on
[40:38] (2438.72s)
like what is right and wrong when
[40:40] (2440.48s)
there's no clear uh answer to that
[40:42] (2442.64s)
question. Do you measure how accurate
[40:45] (2445.68s)
you are at that job by user feedback in
[40:48] (2448.32s)
some way?
[40:49] (2449.36s)
We don't actually measure it today. We
[40:51] (2451.60s)
we should an eval set uh should be built
[40:55] (2455.12s)
for that like like questions where there
[40:57] (2457.36s)
is no one objective answer. The problem
[40:59] (2459.68s)
with building an automated eval for for
[41:01] (2461.84s)
that type of thing is what what what is
[41:04] (2464.48s)
the right answer? It's subjective,
[41:06] (2466.24s)
right? like like if if there are
[41:08] (2468.16s)
questions about the origins of CO and
[41:10] (2470.00s)
there's so many different opinions of
[41:11] (2471.92s)
that relying a lot on Wikipedia as a
[41:14] (2474.72s)
source and and you know can say maybe
[41:16] (2476.80s)
for a human raider you're like okay
[41:18] (2478.88s)
saying all the things Wikipedia said
[41:20] (2480.24s)
it's a good answer but maybe what you
[41:22] (2482.32s)
want is to say stuff that is not there
[41:24] (2484.00s)
in Wikipedia and that relies on like
[41:26] (2486.56s)
having a much better human evaluators
[41:29] (2489.36s)
like pool much smarter people who who
[41:32] (2492.80s)
are supposed to rate these things and
[41:34] (2494.56s)
they're not like available for like you
[41:36] (2496.16s)
scale AI style evaluations. Right.
[41:38] (2498.24s)
Right.
[41:39] (2499.84s)
Okay. I think we have time for one final
[41:41] (2501.68s)
question. You get it?
[41:44] (2504.00s)
Awesome. Hi, my name is Angela. Thank
[41:46] (2506.00s)
you so much for talking to us. I have a
[41:47] (2507.92s)
question about your go to market
[41:49] (2509.68s)
strategy. You had a great campaign for
[41:51] (2511.76s)
students. That's how I and assume many
[41:53] (2513.52s)
college students learn about you guys.
[41:55] (2515.28s)
But then also you had a collaboration
[41:56] (2516.88s)
with Koshi which is a little bit
[41:59] (2519.04s)
different audience. So I'm just trying
[42:00] (2520.56s)
to understand how do you decide which
[42:02] (2522.64s)
audience you're trying to get? I think
[42:04] (2524.80s)
like what one one perspective here is
[42:08] (2528.16s)
trying to get into distributions of
[42:10] (2530.08s)
users that you don't typically have
[42:11] (2531.76s)
access to on your traditional marketing
[42:13] (2533.76s)
channels. You know, there are a lot of
[42:15] (2535.04s)
people who don't use Twitter or LinkedIn
[42:18] (2538.48s)
and and and and they're all like they
[42:20] (2540.16s)
all exist in the world. We just are
[42:21] (2541.68s)
living in a bubble here. U and and there
[42:24] (2544.40s)
are some other businesses that have good
[42:26] (2546.08s)
access to them. uh like you know
[42:28] (2548.64s)
traditional businesses like like you
[42:30] (2550.72s)
could imagine the the kind of people who
[42:32] (2552.48s)
use Costco regularly may not even be
[42:35] (2555.20s)
using AI on a regular basis and so if
[42:37] (2557.76s)
that's the kind of people you're going
[42:38] (2558.88s)
for then it makes sense to change your
[42:41] (2561.68s)
strategy to reach them but also remember
[42:44] (2564.08s)
that uh it's good to grow with
[42:46] (2566.72s)
adjacencies like you do want to have
[42:48] (2568.80s)
some overlapping sets of people who
[42:50] (2570.72s)
would be the word of mouth carriers as
[42:53] (2573.68s)
they help you expand to more you know
[42:56] (2576.32s)
non-over overlappinging circles. So I
[42:58] (2578.24s)
think that that's how I think about it.
[42:59] (2579.28s)
Like there should be some overlap, but
[43:01] (2581.20s)
your distribution should keep evolving
[43:03] (2583.92s)
over time. Thank you.
[43:06] (2586.00s)
All right, Arvin, thanks for joining us.
[43:07] (2587.84s)
Thank you everybody.