AI Agents: Mirage Or Real Revolution? — With Dmitry Shevelenko

Channel: Alex Kantrowitz

Published at: 2026-05-07

YouTube video id: g1gAY09O0Oc

Source: https://www.youtube.com/watch?v=g1gAY09O0Oc

Is the near uniform move of AI companies
to agent super apps going to pay off?
Let's ask Perplexity's chief business
officer right after this. Welcome to Big
Technology podcast, a show for
cool-headed and nuanced conversation of
the tech world and beyond. [music]
We have the chief business officer of
Perplexity here with us. Dmitri
Shevelenko is here with [music] us in
studio, and Perplexity, as you may know,
is one of the many companies moving
towards this agentic super app-style
product with Perplexity computer. Now,
they are joining OpenAI with Codex and
Anthropic with Claude Code as one of the
[music] many companies moving towards
this agent that can control your
computer and get stuff done for you. And
today we'll talk about where that's
going [music] and whether it's going to
be a real business. Dmitri, great to see
you. Welcome to the show. Thanks for
having me. Looking forward to the
conversation. So, we're here in
mid-2026, and I got to be honest, I
thought at this point you would be a
subsidiary of Apple. Hasn't happened
yet.
Well, sorry, uh, your Polymarket, uh,
bet there, uh, ho- hopefully, you know,
didn't pan out, but
>> Just to be clear, they were No
Polymarket bet. I just thought it was a
good idea, but it's it hasn't happened.
Um, you know, we're We have a great
blossoming partnership with Apple. Uh,
they actually are really excited about
what we're doing with personal computer
and how it uses Mac minis. Um, so, uh,
>> growth area for them. Yeah, so so that
is, uh, you know, we we we found a way
to to work together there. Uh, but we're
having too much fun being independent,
and um, you know, a lot of the world is
realizing that
the power of multimodal orchestration,
mass multimodal orchestration, you know,
what what was first a wrapper is now a
harness.
Um, so, uh, so we're really excited
about the future ahead. Yeah, and
that's, uh, to me, the main criticism I
was obviously very vocal saying Apple
should buy Perplexity. I think they
actually gave you a call. I'm not taking
credit, but maybe I contributed. Um,
the reason why I thought it would be a
good tie-up is because, you know, all
the criticism was, "Oh, Perplexity is
just a wrapper company." and I was like,
these guys actually know how to build AI
products. Obviously, the search engine,
the browser comment looked pretty cool.
And then this new computer
application where Perplexity will take
over your computer on your request and
do things for you is really where AI is
heading. And as you mentioned, it
accesses multiple models as opposed to
just being tied to one.
So, I thought that would be a good
acquisition for Apple, which has clearly
struggled to take these models and
translate them into working products, at
least so far. Maybe they'll figure it
out with with Gemini.
Um
what do you think about their CEO, John
Ternus? Or their incoming CEO, John
Ternus?
Well, Apple's always been an incredible
hardware company and I think
this is, you know, an era where hardware
will matter even more because software
is going to face waves of
commoditization pressure.
So, I I actually think it's, you know, a
really smart pick and we're excited to
see what they build and, you know, we
want to
build really powerful solutions that
work well with Apple hardware. Okay,
we're going to get you have a
partnership with Samsung. So, we'll get
to that in a bit. Um let's not bury the
lead here, though, which is that
you know, Perplexity gained, I would
say, mass awareness,
at least in the tech industry, because
of the search engine that you built.
Arvind, the Perplexity CEO, was very
vocal in saying, we're going to take on
Google. We have this new way of doing
search and look out. Um
and
when we look at the usage of consumer
AI,
something very interesting has happened
over the past, I would say, 6 months,
which is that usage pretty much
flatlined. If you look at the DAUs of
of generative AI apps,
from Apptopia, for instance, there is
sort of a a flattening that starts late
2025.
Uh, even looking at Perplexity's market
share
um of of AI of AI search,
it was close to 20%. I think this is
again according to Apptopia,
mid-2025 and and it really has
decreased.
Uh,
kind of uh flat over the past
uh month or so according to SimilarWeb,
your traffic about 5.2 million average
daily visits, up 2% over the past month
compared to 182 million for ChatGPT,
which also isn't growing
too significantly. That's up 5%.
Um
the question for you is everyone is now
pivoting to this super app, this app
that can control your computer.
You guys, OpenAI, Anthropic,
I'm wondering, is this happening from a
position of strength, um which is that,
okay, if we're just going to move move
here because the technology is so
strong, um or is it potentially a
reaction to the fact that
consumer
AI hit a ceiling and you need something
else?
So, um well, I'll tell you that that I
don't know those metrics that you
shared, but the stats I look at every
morning is our revenue.
Um and we started the year at under 250
million ARR and Aravind recently shared
that, you know, as of a month ago,
uh we crossed 500 million ARR. Um and so
clearly uh
we're creating value for our users. And
when we actually go back and understand
who is using Perplexity even when it
was, you know, more focused on, let's
say, consumer AI as you define it,
people were actually using Perplexity
for work and knowledge-related tasks.
Uh so, they were coming to us, you know,
as much as we were talking up, you know,
this is the Google search killer, um
people were were using Perplexity to to
get ahead at work
even when they weren't using the
enterprise version. This was their
secret weapon to be more productive,
have greater leverage as they build
businesses, create businesses. And so in
some ways, you know,
we're not we we haven't shifted our
focus. We're really going to meeting our
users where they were always were.
And what's possible now is and this
really started
you know you couldn't have built
something like computer before November,
December of last year
because model capabilities advance where
you can have
longer time horizons for running tasks,
right? Where you're not just answering a
question but you're actually doing work
as an agent on behalf of of the user.
And
one thing Perplexity has always prided
ourselves on is being the best at
understanding what the new emergent
capabilities are and finding ways to
make that accessible and useful for a
broader population. And that's that's
that's that's where we focus but I think
revenue is a much more honest metric
than than kind of top line MAUs which
which I think you know can include in it
a lot of hype and exploratory activity
but aren't as tightly coupled with
value.
Okay, but I I'm going to give you the
off the alternative perspective here
which is that the MAUs matter. Like
typically MAU of course monthly active
user. When you're typically in a growth
surge,
you start talking. I mean every company
every tech company they grow users and
then
they have this big user base and then
when the growth slows you start hearing
about average revenue per user. You need
more users to have a bigger user to be
have a bigger revenue base, don't you?
So Well, we're We're about average
revenue. We're talking about total
revenue.
>> Right. Right. So, so pay we're talking
about Yeah, I guess maybe that's the
next step, but go ahead.
I mean, I I would say historically
that's been true for consumer internet
companies because MAU is a proxy for ad
revenue, right?
Um and as been as been reported, like
we're not focused on advertising-based
monetization. Um
we realize that there is um
when when a core value prop of
Perplexity is accuracy,
uh it's really hard to reinforce that to
users when you also have ads running
alongside the answer.
Um and so I I I think some of why MAU
matters less is at least for us
is we're not trying to go to advertisers
and say, "Look look at all these users
that that you can show ads to across all
these different demographics." Uh so
that that may be part of the the shift
in focus as well. Yeah, I mean, you
could you could the the to support your
argument Anthropic does not have deleted
users whatsoever and doing crazy amounts
of revenue. So, if you fi- figure out
this enterprise use case,
you could be a massive company. I mean,
we're looking at they're both Anthropic
and OpenAI both going to have
trillion-dollar IPOs and will have many
large companies I think that will follow
them in the generative AI world. But let
me get your take on
on if you
Well, let me just get your take on the
consumer side of things and then we'll
move move more on the enterprise side.
Uh
I mean, even if people are using these
products for work,
they're such powerful tools and um you
know, they were like ChatGPT was the
fastest-growing consumer product ever. I
guess it still is. Uh but that growth
has tailed off. Um what do you think is
behind this
um
flattening of consumer AI product growth
overall? Let's just take it with the
whole industry because it's certainly
happening. Um is it just that, like,
they kind of hit saturation? Or is it,
you know, we know there are fears about
AI. Is it people are just too afraid of
AI? What's your best diagnosis there?
I I think um
there's
we we
so [snorts] some of the use cases got
ahead of
uh where where people were curious to
explore like what is this AI thing, but
their behaviors didn't change. Um
but but I also think
there there's a fusion of consumer and
prosumer that we find very interesting.
Um a lot of people are now empowered to
uh
explore launching a side business.
Uh or, you know, explore like doing
that, you know, you know, that project
that they never had the activation
energy for. And now because you have
these super powerful tools at your
disposal, uh you're more than happy to
to spend, you know, you know, money
behind that because you feel like you
get leverage there. So, I I think
consumer to us is not just people using
uh Perplexity to look up the weather,
right? You don't need AI for that. Um
and so, I think part part of
what the broader industry needs to do is
educate users on what is possible now.
Like people refer to this as the
capabilities overhang, right? Where uh
the models got a lot more powerful,
especially in the last 6 months, and
people are still using them in a very,
you know, web 1.0 way. Um and that's
just going to take time for for for that
that discovery to catch up. Uh but we're
we're, you know, I I'd say this is less
relevant for Perplexity, but I I'm
confident that uh
everyone will prefer to have a more
intelligent uh set of software they use
to to help run their life. Web point 1.0
meaning like information retrieval?
Yeah, just like the most basic, yeah.
Yeah, like okay, like sports scores, you
know, like weather, you know, basic
news. Like that that's you know, that's
where still a lot of people are. You
don't necessarily need, you know, these
new agentic capabilities for that.
There's all kinds of, you know, other
things people can be doing and the thing
that we're going to realize is the
constraint on making the most of AI is
our own curiosity. Right? Like, you
know, that that's that's the bottleneck.
And that's why you know, perplexity's,
you know, we design our products to
spark curiosity, to activate it, to, you
know, that's a big part of our brand is
curiosity because like when we
when we kind of zero out like, you know,
what gets commoditized, what doesn't,
the uniquely human
ingredient to taking advantage of all
this will be curiosity and agency.
Let me let me give you my
uh
belief on why we're seeing this
slowdown. And it we can sort of cuz this
does lead right into the agentic use
cases.
Uh when we've seen the biggest spikes,
they've been around
some of these multimodal use cases. So
not text. I mean, ChatGPT got to 200
million users because of text. People
were interested to see what AI could do.
So I think that novelty and that
interest, you know, built the
foundation. But we're we we're I'll just
use OpenAI for an example. Where OpenAI
saw the biggest surges was after voice
hit. Remember that demo that where it
sounded so much like Scarlett Johansson,
she threatened to sue OpenAI.
You you see an inflection point in
growth there. And then images. The
Studio Ghibli moment still was just one
of the like I need I mean,
I know somebody that created like seven
OpenAI accounts just to cuz they kept
getting rate limited on the usage. And
so of course you'll probably see a user
spike there even if it's not, you know,
individual users.
So that to me is like as that as
companies have shifted away from those
things, we know that Sora is going away
at at OpenAI. Obviously, they're still
doing images. They just released a great
second generation of their latest image
product OpenAI did.
But but there is going to be this
sort of moment of adjustment among
people from going from what the AI
companies were initially telling them,
you know, chat and images and voice to
this new use case, which is like
we think that the model should take your
computer over or whatever the model
through a harness should take your
computer over and let you do stuff and
that will naturally lead to a divot.
Yeah, I mean, I think I agree with the
thesis, right? A lot of those spikes in
usage were novelty-driven, right? Like I
mean, your friend that created the seven
OpenAI accounts, you know,
I bet they haven't created any Studio
Ghibli images in the last 30 days,
right? Like I don't see those around
anymore. It's probably gone from the
family chat. Yeah, yeah, [laughter] it
is though you still see some people's
profile pictures are like Studio Ghibli.
And so that that is a warm reminder of
of that era of AI.
I think the novelty spikes are great
because it raises, you know, broad
awareness and it brings, you know, it
brings people in
and then people have to, you know,
discover their own kind of habitual use
cases. But you can't Yeah,
you know, novelty is what it is. I mean,
Nano Banana had a similar, you know,
moment for Gemini and I think you could
see now it's kind of,
you know, that there's been
a reduction there, too.
Ultimately, like we we see value in the
most economically productive aspects of
AI, right? And that's why, you know, for
us a a core foundational investment has
been accuracy and you almost think of
search and accuracy as, you know, two
sides of the same coin, right? You need
to have best-in-class search so that
whatever you're doing with AI is
grounded in the most up-to-date, you
know, highest quality sources, best
snippets of that information working for
you.
Um and and so I I do think the um
Yeah.
I I don't think it's fair to call uh us
what we're doing a pivot, uh but I think
we're mapping our investments towards
what are the most economically
productive uses of AI
um that have the most enduring value.
And Right. Effectively, what what's
happened now and I mean you're you're
probably a great example of this, you
know, you're running, you know, an
independent business, right? Uh that
previously, if you were not using AI,
which I'm sure you're using in many big
and small ways, you'd probably need to
hire, you know, a lot of people. Yeah.
Maybe a software developer. It's It is
crazy at being so heavily invested in
learning the tools, what you can do.
Yeah, so so like I mean you're you're
like the you know, you know, we should
do a case study on you
uh because you're you're exactly like
what we see as the future of the
economy, right? Like someone with high
agency, right? You had a vision of, you
know, running your own media business
that, you know, hopefully one day
becomes a media empire,
uh and you're able to make very quick,
rapid progress on it because you have a
a team, you know, I I think of it like
we all just got 100 employees, right?
And um the shift we're seeing in both
prosumers and in in the workforce is
everyone now gets to operate as an
executive because your job is to wake up
in the morning and think about, okay,
what are the useful tasks that I can,
you know, deploy the 100 agents that
that are on standby
uh to grow this thing.
Um and so that that's a you know, that's
very again, very different than like,
you know, you know, casual chat and
generating images. Like I think those
things feed into each other because
sometimes, you know, the spark of
curiosity requires kind of the quick
question and answer. And so, you want to
make that minimally, you know,
you want to make that delightful, easy,
low-friction. So, then people are
inspired to go after the longer horizon
tasks. And so, we see them working well
together.
But, you know, the future of AI is is
what you're doing. Yeah, and it is
interesting because I do use these and
you know, I just decided that the the
groups I wouldn't need to hire because
I'm using this stuff well. But, um
by having access to the tools, I'm
actually able to do a lot more, I would
say, economically productive activity
than I would have been if I wasn't
constrained by them.
Uh so, for instance, because like I'll
have like a little extra margin because
I don't have that marketing agency,
well, maybe I can use that to host the
host an event. Which, by the way, folks,
we're going to be doing on June 18th.
Arvind Srinivas, CEO of Perplexity, is
going to come speak with us. I'll I'll
link it in the show notes. Uh if there
are still tickets, you should definitely
join. Uh but that's something that
exists because, you know, there's a
little bit higher margin and we can
invest in doing an event because of
that. Uh so, I think there's like we'll
see a very interesting transformation of
the economy if this stuff works the way
that many anticipate that it will. And
I'm I'm not I've never really been
bought into the gloom and doom
uh hypothesis around it. But, I guess
that's that's a different discussion.
Let me just sort of ask the natural
follow-up to what you just said, though,
which is
if if chat,
images, voice
were part novelty to cause this
explosion of interest in generative AI,
why why are you sure that this
computer-style use or super-agent use
case is not going to be similar? For
instance, just to make the bear case,
maybe it is also a lot of people trying
out this,
um, these these apps and saying, "Oh,
that might be useful." But then there
could be a pullback,
uh, from it. I'll just give one example
and then I'll turn over to you. Um,
I'm sticking my teeth into Perplexity
computer, which is Perplexity's agent,
uh, or super agent, I guess, is the best
way to describe it. Um, and I I at its
suggestion created a daily digest email
for myself. So, it connected to my
Gmail, it's connected to my calendar,
uh, it tells me which emails I need to
respond to, what's going on today, what
I should be thinking of, the headlines.
It's pretty cool.
Um,
but
is there also a chance that that could
just potentially be like, "Oh, that was
a kind of a cool new use case."
Um, but not like a revolutionary use
case, cuz you could have said the same
thing about chat, images, voice, that
they were cool use cases, potentially
revolutionary. Maybe they're not, maybe
they have potential to be that way. So,
why is this not, you know, another one
of those novelty use cases? Yeah, so
what we're seeing with computer is
people are are generally using it the
way you're describing,
um, the way you're running your
business, where it's like you now don't
need to hire, you know, dedicated staff
or a dedicated, you know, agency to do
your marketing, to do event production.
You're gaining leverage from these
tools, right? And and what we're seeing
is the longer people have had access to
computer, I mean, the stuff is still
brand new,
um, but they're using it
consuming more computer credits every
week, uh, than the previous week, right?
So, we're we're actually just in in the
extreme upward part of the ramp. That's
a big part of why revenue's ramping as
well. Um, so we're we're certainly not
seeing that and
I think the fact that, uh,
we're people are now meant the mental
model is not this is like, I'm spending
on software.
Uh, people are thinking about this as,
you know,
this is actually
part of my payroll budget, right? I have
a team of digital agents, digital
workers, and, you know,
sure, like with the workers have to like
show up and do a good job to to earn
their paycheck, uh, just like, you know,
people do. Um, but their capabilities
are, you know, increasing, um, and we're
we're getting better every day of
connecting the models, uh, to different
tools, uh, you know, improving, you
know, the virtual machine that it runs
on.
Um, so I I think the uh,
nothing none of the usage of computer
right now that we're seeing has a
novelty effect. It's all kind of, you
know, being tied in where where people
are willing to pay for it, um, is tied
into those economically productive
scenarios. Uh, so we're we're incredibly
bullish on it. And as people in AI like
to say, like, the models are only going
to get better from here, right? So the
the capabilities will increase. Um,
I think consumer is really hard to get
right, uh, if you don't have network
effects.
Um, and so again, I think some of, you
know, the Studio Ghibli, like the the
voice, uh, those early video gen
examples, I think that's very different
than what we're seeing with computer
now.
So what should I mean, you mentioned
that people as they use it, they use
more credits. Yeah. What are some of the
use cases that you're seeing? I mean, I
might my email I thought I think is
pretty fun. I I let that go. Um,
but I also see taxes. Yeah. I mean, it's
it's any, uh, so we actually are
launching this week 36 different
workflows that go on top of computer.
Um, so this is everything from building
a financial model uh, of a company,
uh, to filing your taxes, uh, if you're
a wealth manager,
uh, prepping for a meeting with a
client. Um, and again, this takes
advantage of connecting to, you know,
your internal data systems, your your,
you know, your snowflake, your data
bricks.
Uh just last night
uh I ran a
analysis of, you know, what are the
models that are being used inside of
Perplexity right now. Like, what's the
distribution of between, you know, Opus
47 and uh GPT and Gemini and got a very
elaborate uh result back and I know zero
SQL.
I don't I can't code if my life depended
on it and I didn't bug a single data
scientist at Perplexity.
Um and I was able to do this because we
connected Perplexity computer uh to our
snowflake and I was able to, you know,
pull in that that analysis, you know,
within a few minutes that in a previous
world, you know, that would have been 10
emails and I certainly would not have
been able to get it at at midnight um as
as I wanted to kind of dive into that,
right? So, um
what we're seeing people do is is be
able to operate with much greater
velocity uh whether they're
accomplishing marketing objectives,
analytical objectives like building
product. Um you know, we're we're now
able to prototype new features
uh instantly.
Uh we have people on our content team
that
submit pull requests, basically ship,
you know, code uh that goes live into
production without engineers being in
the loop.
Um and that's all being run through
Perplexity computer. How much can you
trust this stuff? Uh you know, when I
again going back to this taxes example,
I don't trust it to do my taxes. Am I Am
I just a Luddite or is there legitimacy
to the worry that if it gets something
wrong, I could get a letter from the
IRS? Well, well, actually I would I
would flip it the other way. The way
people are using computer is to
double-check the work done by their
accountant and finding significant
errors done there. Wow. Right? So so
actually one of the workflows that we're
most excited about is called Final Pass.
And you submit PDF or presentation or
spreadsheet and it basically does a
detailed fact check on every assertion
and claim in that document and and both
in terms of fact checking against the
outside world and then for internal
consistency.
And we we actually did you know, ran
through a Gartner
press release about their earnings and
found like four glaring you know, like
mistakes in it where they like misstated
the earnings.
And you know,
we're going to have a fun marketing
exercise where basically go through
public companies press releases and and
run Final Pass through them
and show just how much you know, error
lives in the world right now.
And so I I think you know, there's
But to get to the heart of your
question, I think there's always going
to be three fundamentally like human
activities when it comes to using AI.
One is we talked about curiosity, right?
You have to give it the spark. Like you
have to define, you know, we we say, you
know, we're shifting from an era of
instructions to objectives, right? So
you have to define where the objectives
for you know, what what what what is the
marketing success that you want to see
and then the AI will accomplish it for
you. So you need the agency. The the
second part is just like you need to,
you know, error correct and double check
the work of a human, we need to get
really good at understanding where AI
might go sideways and you know, do
validation testing. And and that's going
to mean different things in different
use cases.
And then the third piece is good taste,
right? Only humans are going to deeply
know what other humans will find
interesting and cool.
Um, and I don't think AI is going to AI
can be a great brainstorming partner,
but ultimately that that's going to
require discretion. Um,
and so
yeah, I I think, you know,
fact-checking, error correction,
uh, those are going to be essential
skills.
Um, but it goes both ways. Like, you
know, as I said, you know, with taxes,
uh, there's plenty of errors that that
humans are making right now and let's
use AI to catch those. The question is
if people will stop uh,
stop at people will use these tools the
way that you intend or whether they will
just say, all right, screw it. I'm going
to replace my account entirely. But I
guess you're responsible for that if if
you do that.
>> Yeah, I mean, just like you're
responsible if you hire if you hire a
cheap accountant, you know, and they
mess up, like ultimately, you know,
that's that's going to create a headache
for you. If you use a bad AI or not
using it properly,
um, you know, that's also on you. Uh,
so, you know, Careful
>> accountability accountability doesn't
doesn't doesn't go, you know, go away
with AI.
Um, and yeah, we we need to develop a
good sense of how do we, you know,
like I I have a good way of spot
testing, you know, when I get an output
from AI, like what are the things I'm
going to like double click on to make
sure there was no silly mistakes. Yeah,
and I love the final pass idea. I mean,
I've been doing that for all my stories.
I like we'll upload the interviews and
then upload my draft and be like, what
did I miss? What outside context is
there that I should be considering? And
so it's just natural that that type of
approach would be applied to other
things like taxes, financial
projections,
um, even I don't know, marketing
presentations could be thrown in and be
like, just triple check the numbers,
which I've been doing and it's quite
good at that. Yeah, I mean, the the
really fun one was I I presented to uh,
uh, the senior leadership of of Bain and
management consultant, you know,
management consultancies they publish
uh,
uh, all kinds of, you know, you know,
you know, reports and and like we had a
lot of fun, you know, showing them some
some errors and and some of the public
reports they they found and like the
people that worked on it were in the
room and so they were they were giving
each other, you know,
some some trouble for it. But yeah,
there's
I think there's still a lot of value to
unlock in using AI to to fact-check
humans. Okay.
But to get this to work right, you have
to trust a company like yours
tremendously, actually.
Let me just read you some of the
permissions I had to enable for my just
my daily email. See and download I don't
I can't believe I actually went through
with this, by the way.
See and download contact info
automatically saved in your other
contacts. See and download your
contacts. See the list of Google
calendars you're subscribed to. See,
add, and remove Google calendars you're
subscribed to. View and edit events on
all your calendars. View availability in
your calendars. See and download any
calendar you can access on your Google
calendar. Read, compose, and send emails
from your Gmail account. See and
download your organization's Google
workspace directory.
I guess I see now why people are working
on the Mac mini
because, you know, and and this is
enabled for me right now as we speak
that Perplexity has all this access to
like,
you know, all of my mission-critical,
you know,
technological infrastructure. I mean,
maybe computer right now is like writing
up client emails and sending them.
I don't know.
Well, you do know, right? Because you're
ultimately, you know, you're you're
choosing to initiate the task. Like that
nothing is happening kind of
autonomously, right? Like again, the the
agency
is still, you know, human-triggered.
Like you're you're ultimately still
directing and you know,
you don't need to give all those
permissions to get a lot of value out of
Perplexity computer. I mean, this is a
conversation I have with with many
businesses is, you know, start with zero
connectors,
and just, you know, see the value there
because there's a lot you can do with,
you know, just interfacing with with all
the outside world's data and making more
sense of it.
Um
But, you're you're ultimately, you know,
to unlock the full value, if you think
about this as a digital worker, you
know, you know, if you hire people, you
also give them access to even greater
permissions, right? And uh and people
make mistakes, too, right?
>> Wait, it works slower than the AI does.
>> Yeah, and and, you know,
again, another like, you know, crawl,
walk, run that I would suggest is we we
have the capability for businesses
uh to allow for read access, but not
giving write access, meaning they can,
you know,
you know, it can
create the daily digest, but it won't
send the emails on your behalf, right?
Which is like the that's the part where
people are like, well, what if it like
goes [snorts] and, you know, spams a
thousand folks
uh with, you know, with the wrong the
wrong
>> confidential information.
>> Yeah, so so again, so that's like the
read rate. I think that's like a way,
you know, and again, we, you know, with
our business versions, we offer very
granular controls. Um and um I think
that that's the path forward there.
Um but, we spend a lot of time getting
the engineering on this right. Um you
know, one of our advantages in the space
is
the only thing we do is build the
product. Uh we don't train pre-trained
foundation models,
uh which means all our locus of effort
is exactly on, you know, making those
interactions
um you know, first of all, transparent
to the user, right? You know, you were
able to know exactly what you're giving
us permissions for,
uh and then make sure that, you know, it
is error-proof in terms of adhering to
those uh those permissions.
So, do you think that the technology
today is trustable enough that what I
did is not crazy? And if so, why do you
think so many people are running this on
a Mac mini. I mean, there was a Mac mini
in your ad for perplexity computer. Oh,
so the Mac mini is it's actually the
other way where it lets you get even
more, right? Because with the Mac mini,
you can then get access to your
iMessages, which you can't with the
permissions you got there. With the Mac
mini, also the agent can run 24/7,
right? Even when your laptop is closed,
it can, you know, run those long horizon
tasks. So, I wouldn't necessarily
interpret the Mac mini as like a
uh I want cuz the the inference is not
yet happening locally, right? It's still
um happening
>> think it will?
Well, I certainly think that um
as models get more powerful, you will
certainly be and as, you know, local
CPUs get more powerful as well, you're
going to be able to distill powerful
reasoning models to a size where they
can run on a Mac mini.
Um
Now,
I'm not going to offer you like a
timeline on, you know, when that's going
to, you know, when when you're going to
get the 80/20 where
uh some of these workflows can shift
towards local inference,
um but I think hybrid compute where
uh certain tasks will run in cloud and
certain will run locally,
um I think that's a pretty safe bet to
assume that that will be like the, you
know, the
the the right way to anticipate how
these systems will work in the near
future.
>> Yeah, that's the bare case to the um
data center build out is that eventually
like you do all the training in these
massive data centers and then you sort
of distill it and run locally on a Mac
mini. Well, again, I I didn't say 100% I
said hybrid.
Well, it but like if the if the work
that you're doing the cloud is so
computationally intensive, you might
still need all that data center build
out, right? So, I I don't I don't, you
know, there's kind of um
I think we're under anticipating all of
the broad types of computation that more
powerful models will, you know,
bring to bear.
Uh and so I, you know, I
from the Perplexity point of view, like
we don't have strong opinions on the
data center build-out, but there's
nothing I see that indicates that that
is, you know, a bubble or anything like
that. Yeah. Okay, so just to sort of
wrap this part of our discussion,
the Mac mini is not a way to ensconce
the agent away. It's to give it access
to more and let it work harder.
Yeah, and and again, with with kind of
even more granular control, right, and
more access to to your local files, um
obviously you're you're giving those
granular permissions, but yeah, where
currently those systems don't support
local inference.
Uh
obviously, you're doing this.
Uh we've just heard
at length from OpenAI on this show about
their ambitions to build this super app
with Codex at the at the heart of it
that obviously will take your computer
over. They call it a new way of using a
computer.
Uh and then of course, Anthropic has
done this with Claude Code and Claude
Co-work, um
which I I can't believe how I'm still
like stunned at how much permission I've
given these things, but the payoff is
pretty intense when in a good way when
you when you do.
Guess you got to take risks in life. Uh
why is Perplexity going to be able to
compete with these two giant companies
in the same product arena?
Yeah. So,
when we first set about building
Perplexity,
uh we made a very intentional decision
to be model agnostic.
Um and that was that was kind of very
contrarian at the time because the
easiest way is easiest way to raise
capital in 2022 was to say you're
training a model. Yeah, especially with
with our founders' background, that that
could have been a very easy story for
them. Um
they believed back then, and it's proven
to be the case, that models would end up
specializing.
Um, and that is that is actually one of
the most powerful things about computer
is on a single given task, it will use
different models for different parts of
that task, right? So, uh I have little
kids and I love like when whenever I'm
trying to get them to learn about
things,
uh I'll create like mini podcasts for
them. They're they're very personalized.
Um,
and when I do that, computer will use
uh
this is kind of and this changes week to
week, but it'll it'll it'll like to use
Opus for planning the task. It'll use
GPT models for writing the script cuz
GPT is a good writer. It'll then use
Gemini models for
uh generating the audio.
It will then um
sometimes actually use Grok for fast
research uh cuz Grok is a very fast
model. Uh
it will use uh Sonnet for writing the
Python code to stitch together all the
audio clips. And that's just in one, you
know, uh
single, you know, deliverable task, it
used four different models.
Uh so, the one thing that Codex is never
going to be able to support is running
Gemini models. Uh
you know, they're they will always be in
the GPT family. Same thing for, you
know, Claude, like they're not going to,
you know, have GPT models. Gemini is not
going to have Grok models. So, our value
as a multimodal orchestrator and being
an aggregator is we can tell a user
whatever is the best intelligence that
exists in the world today that can help
you accomplish your task, we're going to
be using it. And we're not going to be
discriminating because of the models we
happen to train or the ones we we have a
special relationship with. And that is a
very powerful value prop. Um
and and that's something that over time.
I I think the second piece that is uh
foundational that that, you know, I
spoke to briefly earlier is accuracy.
Um you know, when we were focused on,
you know, the the V1 of Perplexity,
which was, you know,
ushering in this transition from links
to answers,
uh the core technology investment we
made in our own tool was search. Uh you
need the most accurate grounding so that
whatever the intelligence is processing,
uh the the source input is as high
quality as it can be. Um and and so
that's something where we have a very
powerful data flywheel
um that's been running for over 3 years
uh of of compounding, you know, as
people use the product, we see which
snippets the models use, which ones they
don't. That reinforces the intelligence
of the index um and what we do on
search. And so accuracy is another thing
that is very differentiated in
Perplexity compared to some of those
other products. And so,
um you know, and I say the third
uh structural differentiator,
this one you're going to say might be
like soft and fuzzy, but uh
I I think it matters
um is usability.
Um you know, when when I talk to
businesses, something I I, you know,
comes up often is the alpha for a
company that is not an AI company is is
not in them building their own internal
tools with AI necessarily.
Um it is in the depth of their adoption,
right? Like how do they culturally
uh how do they through training,
uh you know, through the right type of
management actually get everyone to use
these superpowers uh the way you're
using them, right? Where And you you're
doing it cuz you have to, right? Cuz
like you you wouldn't, you know, like
you're seeing the necessity. And cuz I'm
yeah, I'm a psycho who likes to
pressure test these things. No, but
you're but you're seeing but you're
seeing
Yeah, like you wouldn't be I mean, I
don't think your type of business model
would work necessarily with that with I
mean, it'd be much harder. Yeah.
Yeah, it wouldn't be you wouldn't be
able to grow this fast, right? And so,
if you're, you know, part of a 5,000
person organization, you don't
necessarily feel that same
pressure that you feel, right? And so, I
think the
organizations need to figure out how do
you actually, you know, how do you
create that that pressure for for that
middle line, you know, worker?
So so that so they feel that and
uh we need to do our part in that in
making Perplexity computer super easy to
use. That's why we're launching
workflows because the, you know, the
example you had of you know how to
prompt AI to do the fact-checking on
your articles, right? And you probably
have a certain, you know, process that
that you use there that you repeat.
Um
for a lot of folks, they look at the
open prompt and it's terrifying. Yeah.
They don't know like they don't they
don't they don't
>> page for a writer. It's a new writer's
block. It's the scariest thing you could
ever look at.
>> right?
Yeah, and it's like and you hear about,
you know, I mean, all your reporting is
like, "Oh my god, AI is changing
everything. I need to you know, you need
to be ahead. You're going to get
disrupted." And
you know, that that's again why we need
something like workflows
which, you know, takes all these
complicated
uh you know, scenarios and and use case
of AI and just breaks it down into a
simple UI
uh where you don't need to provide
open-ended instructions, right? Or
objectives.
Um and so So yeah, so so summing it up,
the reason we're going to continue
thriving uh in a very competitive space
is we're the best orchestrator and
aggregator of all the intelligence.
We're the only AI company
fundamentally committed to accuracy as
like a core principle and that's where
we've made our big technology
investments along with orchestration and
usability which is is really a design
problem as much as engineering problem.
It matters and it's something that you
know, we've always had an edge in and
we're going to keep innovating on. Yeah,
well the question is if these AI
providers allow you to continue to use
the models cuz they have shut down
competing companies. So I want to take a
break and I want to go over that with
you and then talk a little bit about the
variety of models you do orchestrate
including the Chinese models. You have
Kimi K2 in there. So let's do that right
after this.
And we're back here on Big Technology
podcast with Dmitri Shevelenko. He's the
chief business officer of Perplexity.
Dmitri, this is really great rich
conversation. I appreciate it. Um
The I've written about this. One of the
big problems um
with all these AI use cases converging
is that
it used to be for the um
for the these big AI model providers,
they'll build the the um
they have the demo products like the
chat GPT. This is the previous way of
operating and they'll offer their model
that you can you know, pay for
intelligence and build whatever you want
on top of it. Um
but as we get to this style of agentic
use case where everybody wants to build
this stuff, you know, some will not be
competing but there there's interest to
have their own products like Claude
co-work, like Codex be the sort of
system or agent of record so to speak um
that handles all this stuff. And I think
they they might even prefer world where
you know, that would just be the single
app to rule them all. Um
you're orchestrating their models. So
long term, aren't you sort of at least
dependent on their benevolence to allow
you to use these models
um
even as you compete with their core
products now.
Yeah, I I think ultimately all these
companies are platform businesses in
addition to product businesses, and they
you know, they aggressively
petition us to use their models.
They want They They give us early
access. They want us to run evals.
And so we we have, you know, the exact
opposite dynamic right now where they're
they're, you know, more than happy to
take revenue from us.
And you know, they're the beneficiary of
of, you know, more consumption of
computer credits as well.
And and I think they, you know,
because they are all competing with each
other on their platform businesses as
well.
And, you know, there's open source,
which which is, you know, you know,
continuing to push at the frontier, not
necessarily at the frontier, but pushing
at it.
All those competitive dynamics are very
healthy for us. Now, I I agree with you
if we lived in a world where there was
just one
frontier model that was twice as good as
the next best model,
that wouldn't that would be a bad
scenario for perplexity.
I
I wouldn't deny that. But, you know,
since this industry has kicked off,
there's never been a moment where the
delta between the, you know, the best
model and the second best model was like
more than maybe like a 10-15% gap. And
again, like best model is is probably I
shouldn't even be using that phrase
because
it's best model at what, right? There
there's Yeah, there there's there's you
know, that it's the the sub
specialization, right? And so,
the the specialization is also a hedge
against
you know, the those sort of competitive
dynamics.
So, I don't uh
I I lose more sleep about
us preserving our execution velocity
and, you know, continue to build our our
you know, our culture in our
um
through the intensity of the space,
rather than, you know, a
uh us getting cut off scenario, um
because I'm not seeing indicators of
that. If the models
your your example of the models sort of
competitiveness is is very interesting.
I mean, we're at this point where the
models are very smart, right? We have
Anthropic, for instance, won't release
Mythos because it believes it's too
intense for cybersecurity. Great
marketing, by the way. You think it's
marketing? No, I'm saying it regardless
of whether it is or isn't, it is great
marketing.
>> it's mostly marketing or or truth about
the product?
I I I I think
>> everybody this, so I'm curious
>> I don't I think everyone it will have
their own Yeah, I I I don't think
um
we we don't have access to Mythos, so I
I I can't speak to it out of um, you
know, first-hand exposure. Yeah. But the
people you speak with in the industry,
believers or mostly
I I I think there is a I I think what is
a real concern is that models will be
better at exploiting
cyber vulnerabilities than they are at
fixing them, right?
So, so problems in the consultant
presentations.
>> Yeah, so so I think that arbitrage, I
think that's a real
concern. I think that has
already you know, but I don't I don't
know if there's been some new capability
that like didn't already exist.
Um I mean, you've been noticing like
there's been more hacks and things over
the last few years, you know, before
Mythos, so like I I think this has been
building up for a while. I guess like
the the that was a a long windup on my
question to say, um
isn't there going to come a point where
these models are just all kind of smart
enough and compute becomes a commodity
that like right now we're in this
build-up and eventually we just see
parity among models, even though they're
unbelievably smart and just like a lot
of compute infrastructure, and then sort
of a price war that brings the price of
all this stuff way down.
Well,
if if a um Be good for you. Yeah, that'd
be good. I mean, that's
like in in that scenario because again,
open source would catch up, too, right?
Uh But but again, like you start um
if if we reach some kind of plateau,
then you'll actually see even, you know,
the local inference becomes more
relevant because there'll be more
investment there.
Um
I think it's really hard to make
long-term predictions in this space.
Um
I'm I'm fond of saying that the thing
I'm most confident in is that 6 months
from now, I'm going to personally have a
perplexity a top three priority that
today I don't know what it is.
Um
and and the model companies themselves,
you know, when they're when they're
baking the cake of a new model, like
they don't know what it's going to taste
like until it comes out, right? Meaning
the capabilities,
like when you train a model, you're not
necessarily training it, you know,
you're you're making improvements, but
you don't know exactly what the new
capabilities are until it's out there
and people start using it, and that is,
you know,
in some ways it's, you know, that that's
a core skill we've developed at
perplexity is like zeroing in on when a
new model becomes available, where is
the, you know, actionable value for a
user? Yeah. Um you I mentioned this
before the break, but you use the
Chinese models.
Uh Kimikate 2 is in perplexity.
Um I don't see Deep Seek getting in
there anymore.
So, uh to to clarify, we never
uh
integrate into perplexity any product
that or API that is hosted in China.
Uh we have ourselves post-trained
we have post-trained open-source models
that that come
that they're developed by by Chinese
labs. We run those in US data centers.
We post-train them for accuracy and
removing, you know, things that that are
not accurate from them. Well, like, you
know, different countries might have,
you know, certain, you know, political
agendas that they try to
integrate into models. And again,
>> those in the models? I mean, we we've
we've published some some research on
that when with with Deep Seek.
If if you go back to it,
>> on Tiananmen Square?
Yeah, there there's the those sorts of
Now now you have like that's Yeah, we
also solve for that with grounding with
with accurate search, right? And that
that ends up, you know, if you're using
the model fundamentally for for
reasoning, that becomes less of an
issue.
But it's really impressive what the
Chinese labs are doing
and the progress they're able to make.
I think open source is good overall for
users.
Um,
it's ensuring that, you know, pricing
remains competitive.
Um, and
obviously, there's more
we can do
in the post-training space on an open
model than a closed model.
And so, that lets us kind of, you know,
accelerate our work around accuracy,
conciseness,
you know, adhering to certain task
workflows. When Jensen says it's
important for the entire world to
uh
have their AI built on a Western or US
uh AI infrastructure stack.
If you could do what you just did, what
you just told me with Kimmy K2, which is
download the weights, post-train it the
way that you want, um
why why does it matter where the models
are developed? What does it matter if,
let's say, China has the lead in open
source?
What would What would be a bad scenario
is say that the best open source models,
their architecture is done in such a way
where they don't run on Nvidia chips.
They only run on Huawei chips.
Right? So, that the the the kind of I
think the scenario Jensen is concerned
about, rightfully so, is where, you
know,
software drives the the hardware cycle,
right? And and where, you know, imagine
that the flip of the scenario where
right now Chinese companies are trying
to get access to Nvidia chips because
that's where the model architecture is,
right? And that they they need Nvidia
chips to be able to run them in an
efficient way.
What if it was flipped the other way
around where, you know, it's the Huawei
chips are the ones that US companies
would need to to get, right? And
>> makes a lot of sense. So, then China can
export control the US and control AI.
Yeah. So, I think that's that's that's
the I think that is, you know, when you
have this like
>> just say that in the DWARKISH interview?
It's like it's very straightforward
answer, anyway.
The uh well, no, it's
Jen- Jensen is very good at coms, so I
wouldn't uh yeah, I
I think there's a new I mean I mean
there's certain things he can't say,
probably, too, that, you know, can't say
certain names, but yeah.
>> Yeah.
That's fast, but the model Chinese
models are good.
Um They are, you know,
they're pushing the frontier. Uh they're
not at the frontier, but they're pushing
it. Yep. All right. I I want to end
here. There is this interesting
argument, and I think you're you're you
have a
perspective on it at Perplexity. Um
that this is a great article on from
CNBC that Deirdre Bosa wrote.
Um AI demand is inflated and only
Anthropic is being realistic.
Um I think that the crux of the argument
is that like people have been running
massive amounts of work of um
of of workflows on these like $20 or
$200 a month plans, and you know, they
are there's like a a lack of ability to
serve them and so therefore these AI
companies are showing immense demand and
going and raising money based off of it.
Um where like the everything's going to
change once you have to actually charge
per token as opposed to unlimited. Like
you wouldn't do an unlimited electricity
plan or an unlimited fuel plan, but for
some reason a lot of these companies
have been doing this. Um
do you think that this is like a
legitimate issue that she's pointing out
that basically like we don't really know
what AI demand is because it's been
subsidized so heavily for so long
uh and if so, what's the answer here?
So we uh at Perplexity, we've never
subsidized paying users. So if you're on
a you know, pro or max plan,
uh you know, we're thank you. You're
you're contributing to our success.
You're welcome.
The um and and you know, we see great
retention, so clearly folks are finding
value there.
Uh and that's actually why computer
credits are so important, right? So that
as you have uh cuz you can have a
certain computer task
uh cost you $50. For you know, say it's
like video generation and it's like long
horizon running, you know, you you can
one task can cost up to that much and
then you have certain tasks that cost
uh you know, 5 cents.
Uh and so there's no way to encapsulate
all of that in a you know, subscription
product, right? So I I think
the mental model I would have is AI is
going to become a lot like Costco
uh where you pay for the membership,
right? And that gets you in the store.
Uh and that's actually the part of
Costco's business that that is you know,
the highest margin. And then you have
you know, everything you're buying in
the Costco, you know, you have
confidence that there's like a max
margin, right? And those are kind of
like computer credits, right? And it's
you know, some people go to Costco and
they just buy the hot dog, Uh
and then you know, you know, there's
people who go and spend, you know,
thousands of dollars every trip and that
depends on their needs.
You know, but I don't
I I think
I think she's reacting to some to I
think it was Cursor kind of advanced
this this data point that like Claude
Code was subsidizing,
you know, a subscription tier.
Um
I think that will normalize over time,
but the the behavior we're seeing with
computer credits where like people are
paying for usage, right? Like there's no
there's no subsidization, there's no
there's no kind of breakage that that's
driving it
and finding value and paying more every
month
as they as they use it more. I I think
we're, you know, I think it's a safe
investment in all the computing data
centers. Okay, really the final
question. I mean, how do you how do you
keep up? Like Perplexity has been
I would say early on three trends,
right? AI search, AI browsers, and now
this computer
use. Um
Must be tough to set strategy as a
company with things changing as quickly
as they do. So, what is the process that
Perplexity uses to make decisions about,
you know, strategic direction and
product plans, you know, with all these
capabilities just like kind of blasting
all the time? Yeah. I think part of it
is
uh
keeping a very lean team. You know, as
we've increased our ARR by 5x, you know,
from 100 million to 500 million, we only
grew head count 34%.
>> You only have 300 people? Yeah. Uh
That's crazy. So, you know, that is uh
and and I mean, this is what I try to
share with with companies outside our
walls is, you know, you're going to be,
you know, the world is will keep
changing faster and so your your only
way to adapt to that is to be
quick at making decisions
Um and not like you know tying yourself
to one path. That's also a lot of the
you know not to bring it back to why
perplexity computer is great but you
don't want to be you know tied into one
model
if another model is going to be better
three weeks from now right the world is
very unpredictable
and so you want to have agility and and
you want to make quick decisions and be
willing to revisit your decisions right
and you know I think you know I I think
having the humility of not knowing
what the world's going to look like two
years from now is a big part of being
successful in that world.
Yeah I mean
it's all I mean I wrote a book with this
title but it is always day one
really really sort of felt that way
beforehand but in this world
you can't be tied to any legacy you have
to just
basically see what the new is today and
how it works and
and take charge and you guys have been
good at doing that so Thank you. Dmitry
it's great to see you again and and
thank you again for coming on the show
hopefully we can do this again soon. My
pleasure thank you. All right folks
definitely check out the link in the
show notes for the 618 event would love
to see you there and until then we'll
see you next time on big technology
podcast. [music]