August 5, 2026 • 43 mins
Brands, political parties, and businesses all want to know what you think. But what if you're not telling the truth? Cameron Fink, Ned Koh, and John Kessler, the founders of Aaru, believe the future of market research isn't polling humans—it's simulating them.
They may be onto something. The three friends founded Aaru as teenagers, and it has since grown into a billion-dollar AI startup. Cameron Fink and Ned Koh join Oz to explain how their platform uses thousands of AI-powered "digital twins" to predict everything from consumer behavior and elections to which musicians are poised to break out. They discuss why they believe traditional surveys fail and the ethical questions that come with building a machine designed to forecast the future.
See omnystudio.com/listener for privacy information.
Listen
Watch
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Speaker 1 (00:15):
Welcome to tech stuff. I'm os Vloschen. Every year we
spend billions of dollars asking real humans for their opinions
to predict how groups of people behave On a larger scale,
surveys are the foundation of political strategy, marketing, and produge innovation.
So what happens when you stop asking humans what they
think and start simulating their responses? Today I'm joined by
(00:38):
cam Think and Ned Coo, the founders of Aru. At
the ages of twenty and twenty one, they've built a
billion dollar simulation engine that uses thousands of AI powered
digital twins to deliver behavioral predictions for everyone from Ernst
and Young to Spindriff, Cam Ned, Welcome to text stuff.
Speaker 2 (00:56):
Thank you for having us having us.
Speaker 1 (00:58):
Thanks you very very excited. Tell me about the name.
How did you choose it?
Speaker 2 (01:01):
Yeah? Well, Uru is the place in Egyptian mythology where
all the worlds meet, and given we simulate human behavior,
we simulate all these worlds. It makes a lot of sense.
But it's also the oldest trick in the book. So
you know, Auru starts at two a's at the top
the top of every alphabetical list.
Speaker 3 (01:17):
The Yellow Pages trick you know, triple A plumbing, all
those sorts of things. You look at any of our investors'
websites and you see us and then Airbnb, which we
think is the proper order. I'm just kidding.
Speaker 1 (01:27):
That's very very good, Am I right in thinking, or
at least is Wikipedia right in thinking that to get
into Uru, that your heart has to be weighed? That
is true, That is true. Are you in the business
of weighing hots?
Speaker 2 (01:39):
I think by so, I think by technicality, really is
Uru is the place that describes the field of reads, right.
Speaker 1 (01:46):
So we are just kind of a heaven, shall we say,
an ancient Egypt?
Speaker 3 (01:49):
Yeah, exactlytopia almost.
Speaker 2 (01:51):
This is, by the way, this is why r dot
com our domain cost us so much, because I only
learned afterwards that anything that has anything to do with heaven,
like any domain, like our fortune, because it's really easy
to bring something that's being positive. But we are not
in the business of waite arts. I don't know if
that would be as important for the global economy.
Speaker 1 (02:10):
So, because I was thinking, in a sense, you are
definitely not judging people, but in a sense you are.
You're looking into the digital arts, or at least trying
to distill how real people might behave from people you've simulated.
Speaker 3 (02:25):
No, I think it's a it's interesting, and we like
things that are thematic and actually mean things. I think
there's a lot of tech companies nowadays that have names
that don't really don't really mean much.
Speaker 1 (02:33):
So all the ones around Tolkien were taking.
Speaker 3 (02:36):
Exactly and the ones from Dune and the ones from
Lord of the Rings, and it's, ah, well, there we
go to mythology. You know, you gotta pull one.
Speaker 1 (02:43):
I want to sort of get to how aru actually works.
But I saw the article about you two in the
Wall Street Journal or you three, actually your third co
founder is not here, and you know, basically the teenage
teenage founders of a billion dollar company. And I thought
it was a very fascinating and cool story. And I'm
a huge tennis and it's been kind of interesting for
me to observe how Djokovic, Nadal and Federer cast this long,
(03:08):
fifteen year shadow over the game of tennis where the
kind of next gen couldn't emerge. And it wasn't until
Nadal and Federer left the scene and Djokovic started to
age that Ciner and.
Speaker 3 (03:18):
Al Karaz emerged, you know, a lot of people say,
Cameron looks like by the way, Yeah, I like that.
Speaker 1 (03:27):
So, I mean, you know, it's somewhat of a parallel
with the tech industry. Right before the twenty twenty three
chat GPT moment and the diffusion of llms, basically, most
people with strong technical skills would go and work in
corporate jobs at like Facebook and Google and Amazon. Yeah,
and then something changed, and you and the censor at
(03:47):
the vanguard of that something that changed.
Speaker 2 (03:50):
I think what happened is, for like a really long time,
you had people who were very technical getting paid a
lot of money to work, largely like kushy jobs. Right
right now, we have this perception of Silicon Valley nine ninety.
Speaker 1 (04:03):
Six, like all this stuff, correct engineers.
Speaker 2 (04:05):
Yeah, it wasn't that way for the longest time. I mean,
if you go back, like even five years in history
to twenty twenty one, Silicon Valley was where people drank
beer during the work day and played ping pong, you know,
at four o'clock in the afternoon. It isn't like that anymore,
because I think what happened is people got actually attracted
to the idea of hard work. Right It's like this
backlash against what was once a work culture of like
(04:27):
go work ten hours twenty hours a week of productive
time and then not spend the rest of your time
doing as many productive things. People want to work hard,
and I think that changed.
Speaker 3 (04:36):
I mean, with a barrier entry to building a business
being so so much lower, so many things has changed, right.
I don't think it's just young people being able to
do it. I think it's also the people who are
at the top of the fields also have changed a lot.
Like something Cameron and I discuss a substantial amount is
you know, where is this generation's you know, Steve jobs
or those sorts of areas. And it's actually part of
(04:57):
our thesis around this space is that people have and
had to rally around a singular lumin area as much
because capital has been so much more accessible. Right, So
it's not just as part of language models. I think
it's a state of the you know, overall economy and
things that have changed that have allowed it to happen.
But it's very interesting. Nonetheless, Yeah, it is interesting.
Speaker 1 (05:14):
Do you think it's like a stylistic thing that your
generation want to work hard or do you think it's
an opportunity thing that like if you work hard, you
can be net and Cam and get a big Unicorn
founders before you're twenty one.
Speaker 2 (05:25):
I actually think it's I think it's a bifurcation, right,
Like I think you have a population of people who
work really hard, and that might be like the top
death style. But what just happened is like the distribution
of how hard people work has widened so much, so
like just as much as in my high school class,
there are a bunch of kids who, like maybe ten
fifteen kids who worked extremely hard orderly. You know, I
(05:47):
was not in the top performing in my high school class.
Actually proud to announce that I was thirty third rank
out of one hundred and eight.
Speaker 1 (05:53):
You were high school together, what were you?
Speaker 3 (05:55):
Probably much lower. It's very difficult for me to focus
on things I don't care about, so you know, it's.
Speaker 2 (06:00):
But you had like ten to fifteen, maybe twenty kids
who would like go study six hours and they would
be totally happy to do that, and on the contra,
like the next eighty kids just did not want to
put all the work in. And I think you see
that consistently where it's like there's this really high visibility
group of people who work a lot, but remember, there
are tons of software engineers at legacy businesses who don't
(06:21):
work as much as like the high visibility group of people.
That is like a very small small group of people
in comparison to like broader software engineering is a category.
Speaker 1 (06:31):
How do you think about your paid group? I mean,
do you know the other twenty eighth year old Unicorn founders?
Do you associate with them?
Speaker 3 (06:37):
Or?
Speaker 1 (06:37):
I mean you're in New York and most are in
San Francisco, which is a difference, right, Like, what's your
reflection on the culture that's emerging around around that? I
love the idea of Steve Jobs no longer being a
cultural reference point because the structure of how capitalist deploy
is different. That's really interesting. But how do you think
about the group you know?
Speaker 2 (06:53):
I'm curious for Ninn's thoughts as well. I really really
think the fact that more people know that is a
pathway that is open to them is brilliant. I think
that's incredible, right, like one hundred percent people should go
do what they want. I do think some of the
current environment has created people who are not in this
because they're missionaries, but are instead in this industry because
(07:16):
they're mercenaries and not to say that that is like
the founders of other unicorns or other young people. But
I just see a lot of people now who I
think go to tech because not of a mission, not
because they think it's like important philosophically important for the world,
but instead because tech is like the next highest opportunity
earning category, like finance was in the nineteen eighties. And
(07:37):
I think shifting away from that is going to be
interesting because the pendulum will always come back around.
Speaker 3 (07:44):
No, I agree entirely with that. I also think, to
the same extent Cameron mentioned, you've also seen like the
pliferation of so many different industries, not just tech, and
people saying, oh, I can actually make a business. Ow's
something I like, And it's become infinitely easier to monetize
right as like a very small business across so many
different channels, whether it's e commerce, whether that's you know,
starting a tiny little brand and CpG or whatever may be.
(08:05):
So I don't think it's a trend that's just in tech.
I do think people are a lot more public about it,
but it's been interesting to see. Nonetheless, I mean, we
have friends that also dropped out of university to start
schools and have had other companies that have been starting,
you know, starting very very young, and I think there
was this large sort of cultural shift. Cameron noted about,
you know, silkon Valian, it's perception prior, it's now in
(08:26):
some ways like sexy to be a tech founder. So
I think you've a lot of people who whose hearts
really aren't in it and these sorts of elements who
are going after it. And you know, like anything in
the world that you commit to not fully is things
have repercussions.
Speaker 1 (08:38):
There's also an accessibility thing, right, Like your third co
founder is not here. I think when you Will started
you were nineteen, you were eighteen, and he was fifteen yep,
which is kind of remarkable. But he reached out and
basically said you were an accelerator program, can I get
your advice? And then you left the call basically with
him as a third co founders That might probably have
a simplified but it.
Speaker 2 (08:57):
Was pretty close to that. He called them me on
LinkedIn and he just said, I'm applying to this accelerator,
can I have your advice on my application? I don't
even check my LinkedIn dms that often. I certainly didn't
when I was eighteen, I mean, you know what I've
been doing on LinkedIn, And so we ended up scheduling
a thirty minute zoom call. We spoke for two and
a half hours, and I called Ned immediately afterwards and
(09:19):
I said to Ned, I've just met the smartest person
I've ever met in my life. And that was the moment.
More or less immediately, I think the three of us
hopped on a call later that night and we were like,
screw it, We're starting a company together. I mean, Nut
and I were about to sell our previous business. John
literally dropped out of high school and we set everything
down and said, is, while we're passionate about.
Speaker 1 (09:39):
What was the vision you laid out to him? And
what was it that he said in response that made
you think he was the most intelligent person you ever met.
Speaker 2 (09:46):
Well, at the time, he actually was writing a paper
focused on RAG for healthcare agents, so this space I was,
Nett and I were in digital health as well, and
so I remember reading it and thinking to myself, Wow,
like just reading these benchmarks, If these benchmarks are real,
it was a material improvement in the quality of models,
specifically for retrieving like medical evidence and clinical studies. At
(10:07):
that time it was called heel actually hal and reading that,
I was just like, wow, this is insane. I spoke
to him about it, and I was kind of digging
in and I was like, well, it seems really real.
So like a fifteen year old living in a dorm
at a boarding school in like rural Massachusetts who's from
New Hampshire is able to best the labs and that
(10:29):
was incredible, And so all of us randomly having a
digital health background, but then we also all shared this
vision of population research. John was like an original population
researcher when he probably also called dm'd its way into
a lab at Tufts and a lab at MIT doing
city science and population demographic research.
Speaker 1 (10:47):
But did you already have the idea for.
Speaker 3 (10:51):
Something that was developed in its current manifestation together, right,
But in the idea of understanding and simulating populations. It's
something we thought about for a while, but our own
it's manifestation was certainly something that we developed together, which
I think is cool.
Speaker 2 (11:03):
I think it was October twenty twenty three we started
talking about predicting behavior for the first time, and I
think from there on at it was something that just
completely consumed us.
Speaker 1 (11:12):
See you mentioned missionaries versus mercenaries. What is the mission
for us?
Speaker 2 (11:16):
It's predict the future. I think it's really simple. If
you think that humans are the most impactful species on
the club, and we're able to predict human behavior, then
predicting human behaviors tend amount to predicting the future. And
if you can predict the future, you can shape it.
It's not just the power to say this is how
your product concept will perform, or this is who we
(11:37):
expect to win in an election, but instead it's actually
the power to find the product that wins and to
help the people ideas that matter win elections more consistently.
Speaker 1 (11:48):
So predict littles of influence.
Speaker 3 (11:50):
Predict and shape and help people that need to shape
behavior shape that right for the broader good. And that's
something we care a lot about. There's a lot of
predicted technology that exis for the aspect of simply measuring
behavior right. And at the end of the day, what
has to be done after that? When you get a
spreadsheet that says this is what this person said, this
is what this person said, et cetera, it's not very
(12:11):
actionable to anybody, and so it ends up being this
humongous waste of an exercise, because as much as might
have taken you months to get to that point, no
matter what measurement it gets to, it's very difficult for
someone to have to go from there and make their
own assumptions upon that data to then make a conclusion.
And then frequently that data ends up just supporting what
that initial person's decisioning has been or gets sworn that
way or another. We have no like manual waiting on
(12:33):
the output of our system. We run simulations of behavior
our system as analysis upon it, and people take action
as a result. And that's part of what makes it
so powerful.
Speaker 1 (12:41):
When I think about prediction and influence, I guess there's
like black hat predictors and influences like you know, Goebbels
or Cambridge Analytica, and then there's like would be white
hat predictors and influences like cast Sunstein and his like
famous nudge theory. Right, do you think about this spectrum
all the time?
Speaker 3 (13:01):
And if we didn't, I don't think we would believe
in our technology. But we do, and we see how
strong it is every day, and we do as much
as we can to be on the white out side
of that. Right, So, like I think about things and
you know, maybe not to announce it here, you know,
formulae leans or not. We're not without finance and all
that yet. But later this year we'll be launching a
research institute within our where we're dedicating a significant amount
(13:23):
of compute and infratructure from our side who works exclusively
on pro bono aspects. So certain these areas are you know,
women's self for example, getting people to go to preventive
screenings for you know, preventive diseases and other elements like that.
That's a behavioral problem, right. Things like getting people to
understand that nuclear energy isn't dangerous, that's a behavioral problem, right,
All those sorts of elements. There's so much good you
(13:44):
can do in the world if you understand what people's
awareness of certain issues are, how to make them aware
of certain elements and how to change their behavior on
things that are genuinely better for society. And that's something
we spend a lot of time thinking about.
Speaker 1 (13:56):
Do you have philosophical ethical debates the two of you
with your co found it do board? What is the
count sort of I was going to account some elders,
but it's not appropriate in this conversation. But how do
you debate these questions together?
Speaker 2 (14:08):
I mean, we were a company founded in like really
deep philosophical principles. We do have a lot of thoughts,
you know, about the moral and ethical use of our technology. Right, Like,
to be clear, predicting behavior is really powerful. Predicting behavior
incredibly accurately, with the capability to shape it is like
(14:29):
even more powerful. Right, this is something that not only
economically is important, but for the good of the world
it's important. And so I think part of the reason
we built these company is because we wanted to use
this technology as a force for good.
Speaker 3 (14:40):
So how does it work well within what we can share?
Of course? No, it's really interesting because you think about
the architecture or system, and you think about the proliferation
of language models and how much more accessible starting certain
elements has been and all these things. And certainly these
systems would not be capable without language models. But they're
not the only thing that makes them what they are. Right,
(15:00):
you think about what a language model is built for
and its most basic principles. It is not to capture
the irrationality of human behavior, right, It's not. It's in
fact quite the opposite. To average results exactly. It is
a beta generating machine in many ways, right, And so
you think about the things that make humans who we are,
that makes us act or buy a product from a
(15:21):
brand and a new category that they've never made before.
Of course, a language model doesn't work for that use case, right,
And so there have been many evolutions of the architecture
that we've come to today. At its core, what it
is is what we call multi agent simulation. So effectively,
we organize these large populations of AI agents, assign them
different distributions of income, age, race, gender, sex, you know,
(15:44):
credit card purchase history, all these sorts of elements that
makes them who they are. The entire principle of how
we generate those distributions of what we call audiences in
the first place, or the demographics that make that simulate
and population up or that we never actually survey people
at all, and so we're just generate these populations off
of the basic building blocks and what makes them who
they are. So what we call ground truth data, right,
(16:06):
we only look at data from areas where we don't
actually have to ask people because people lie, they have biases.
And you go to trybeca near where our office is
and you ask the average man how much money makes
He's gonna say four and a half million bucks, you know,
like whatever it may be, right, So we generate these
populations based off these ground truth distributions of traits and
then be able to sign those distribution of traits to
these thousands of models to become simulated populations, or will
(16:28):
then ask them questions to measure their behavior. Right, so
you might ask someone you know, who you vote for,
what will you buy? How much will you buy? This
for all those sorts of elements to be a measure
everything from proprency to purchase to who's going to win
an election to you know, churn risk and everything in between.
And is it falsifiable like by a validation?
Speaker 1 (16:45):
How do you like? How do you be tested exclusively
to outcomes?
Speaker 3 (16:49):
Right, So we never try to benchmarker models on surveys
or any sort of traditional primary research, so surveys, polls,
focus groups, et cetera, because again we believe this to
be biased. So exclusive look at the actual actions people
take and not what they say they do. And there's
a lot of different examples of where we've been able
to do this publicly that have been really cool. You know,
this is a fantastic example with one of our partners,
(17:10):
Ernst and Young, who you mentioned. They took a survey
that traditionally took the months to run. Thirty six hundred
people responded, thirty countries, fifty three questions. We said, don't
trust their models. Right, This is what we say to
all our customers. They gave us the questions they asked.
They gave us the people they spoke to, or the
quotas for the assets under management. It was an investor simulation.
We ran the entire questions that within our model in
(17:32):
a matter of minutes returned it to them. They did
a correlation, had a ninety percent correlation, but in the
ten percent where we were not accurate, we were significantly
closer to the real human action. So you think about
a question in that example, right. One of those in
that study was what's your likelihood to retain a wealth
manager when your parents pass away? What are the three
(17:53):
things that go through your head when you get asked
that question. The first one is I hope my parents
don't pass away. Second one is that might have been
in five years, ten years, thirty years, I don't know.
I'm going to do my money then, and the third
one is my wealth managers asked me this question. I'm
not going to tell them a lot, right, So you
have this whole scenario where at the end of the day,
eighty two percent of the people in their traditional measurement
gold standard survey said they were going to retain wealth manager.
(18:13):
Real market data from actual wealth management firms said this
between the tune of like twenty to forty percent, depending
on geography. Our simulation came back a forty percent. So
you look at a result like that, like customers do
all the time when they see the difference between our
simulations and the surveys. They run eighty two percent in
their survey forty percent in ours, and it looks totally wrong.
You look at real actions in the real world.
Speaker 1 (18:35):
We're right on the money, okay, But imagine doesn't know
even as I am. How is it possible that fake
people can get better on so those than real people.
Speaker 2 (18:42):
Because real people lie. I mean, it's that simple one.
When I go and I phone you and I say,
you know, hey, my name's Cam. I'm a researcher with
Consumer Research Incorporated. How do you feel about the coffee
you just drank from Starbucks this morning. What are you
going to tell me you're going to hang up the phone.
Are you going to say it was great and I
got off the phone. And so when you're looking at
(19:03):
that in traditional consumer research world, it just doesn't align
incentives for consumers to actually provide a real opinion, right,
a real reflection of how they care what they feel.
And so when you're looking at our data, instead, we
don't ask consumers how they care what they feel, right,
because we know that they're not going to be honest
with us. We know that nothing that is trained on
(19:24):
or built on top of surveys or traditional consumer research
is ever going to be able to reflect that accurately.
And so instead we run another direction. We train on
top of the outcomes. As Ned said, we train on
things like actual coffee mug sales and actual coffee cup sales, right,
And by looking at that data we can track far
better how humans opinions and behavior is changing over time.
Speaker 1 (19:46):
What's the most niche thing that you're proud of having
got right? What's the whole in one double egle?
Speaker 3 (19:52):
I mean, there's some really really specific ones. So there's
certainly been cases where we've done things like price elasticity
of like bundling of very specific automotive parts. That is
crazy that we get those correct. But then there are
other really interesting things, like you know, breaking new artists
in music. Right, we're on iHeart Radio right now, right,
Like looking at being able to simulate the star power
(20:14):
of a very specific emerging artist and predict that they're
going to blow up or that they have propensity to
blow up in these specific areas. That's incredibly cool.
Speaker 1 (20:22):
So this was an advanity fair piece. It was an
artist you found who had forty thousand streams when you
found them and went on to release a song with
one hundred and seventy two millions off you'd identified them.
What more can you share about this?
Speaker 2 (20:37):
Not that much, But it was a really really cool
use case. And I am proud to say that I
listened to their music on our way up there.
Speaker 1 (20:46):
Yeah, back up, yes we did. You can't say what
the artist.
Speaker 2 (20:49):
Is, No, I can't say who the artists?
Speaker 1 (20:50):
What can you share about the criteria? Yeah?
Speaker 3 (20:52):
So this is really interesting. We were testing a very
specific population. It was people simulating people listening to music
who are under the age of twenty five, right, So
we're looking at like young sort of looking at emerging
underground these sorts of areas. It was really interesting because
we ended up feeding in the Instagram feeds, Twitter feeds
some of the actual songs and prossing the audio on
(21:13):
those elements and understanding is this a res rating compared
to all these other artists, Like how often you'd listen
to this artist? Would you go to a concert for
this artist? How would you pay for a ticket to
see this artist if your friends told you about them,
like all these super specific elements that we worked with
to quantify whether or not, out of a large ranking
of these specific performers, which one would do best. And then,
(21:34):
of course that was when they were much smaller. They
ended up getting signed and sorts of things and then
had a lot of popular songs come out. So you know,
we can't take full responsibility for the fact they got
that large, but it was certainly interesting to see that
we were properly able to simulate the general populations prepensy
to listen to music, which is really cool because that
is something that's so difficult to quantify.
Speaker 1 (22:18):
Well, I actually have one that I would love to
work with you guys on. So I am the host
of this tech Stuff podcast and tex Stuff is a
co production of iHeart Where we Are and Kaleidoscope, which
is the media company I founded, and our mission is
essentially to communicate the wonder and excitement of this new
age of which you two are representative of. And that
(22:40):
idea that you articulated, which is that like a small
group of people run with everything they have towards being
like participant builders and a large group of people don't
is perhaps one of the greatest social problems we have.
And so what we think about a Kalidascope is how
can we tell stories and do media products podcast, YouTube,
(23:02):
et cetera, which are not polyannoriatee or just you know,
celebrations of the tech industry, but but rather that are
kind of engaging and inspiring storytelling about what's possible in
this new world because a lot of the narratives about it,
yes they are available, but they're also like by the industry,
for the industry.
Speaker 2 (23:18):
It's a closed and ecosystem in a lot of ways.
Speaker 1 (23:21):
I was thinking this morning on the way to talk
to you guys, how do we successfully identify the next
cohort of the best science and tech creators, influencers, et cetera.
So how would we go about doing Like, let's say
I was a paying client, which I think I might be.
(23:44):
It has to help like what would be the sales
motion as it were, if I was a real client,
Like what would you what would you guys? Where would
you go from like this problem to providing me with
the answer.
Speaker 3 (23:55):
No, I think this is really interesting. And the first
part that you're bringing up there is one of the
biggest challenges for insights, research, marketing, strategy, product pricing, whatever
you are team today because what makes research, say, it
takes so lethargic of a process today, like weeks, if
not months sometimes Yeah, part of it is like asking
the questions and finding the people to talk to and
those sorts of elements. Part of it is figuring out
(24:17):
what the numbers mean afterwards. The biggest challenge and the
place where we actually see people go the most wrong,
is figuring out what to ask in the first place, right.
And so one thing that we've built our system to
be able to do because we serve as so many
different types of operators is feed in general objectives, feed
in semantic descriptions, and our system will take that in
(24:38):
and figure out exactly what the simulations you have to
run to answer that objective are And so maybe for
that example, the first thing you want to find out is, Hey,
before we go identify what talent works with people, let's
go see who has the appetite for this in the
first place. Segmentation this is huge business for us. We
do a lot in CpG and marketing in these areas
because we don't need to get in the weeds of
your CRM to segment your audience. We can just simulate,
(25:01):
you know, half million people in downtown New York City
and see who wants to listen to your product, or
see who wants to buy it for a certain price. Right,
So it's much more first principles, like top down because
we have that scale. So maybe the first time we'd
run there would be understanding a certain population. You said
you're interested in the youth mostly right, probably as being
an audience maybe given age rage. I'm curious let's say
(25:22):
twenty one to twenty nine careert, so twenty one to
twenty nine, right, And we probably feed that in any
other descriptors that you have about this population that you're
interested in, maybe psychographically like open to sort of personal
self development. Not already checked out, but wondering how to
live a better life.
Speaker 2 (25:41):
So we would be able to take in that description,
you know, twenty one to twenty nine people who are
interested sort of open to non insiders, not insiders, people
who maybe insider rejectors, right, people don't insiders, but people
who are open to learning more about technology, and our
models are able to pull out that description and actually
say okay, based off of who even something that simple
(26:04):
is like what is their urban world breakdown? What is
their home cooking meal frequency? You know, where do they
spend their time on social media? How many hours a
day did they go to school? Yeah, what is their education?
What is their income level? You know, three hundred plus
different variables that we're storing. And then once we chart
all of that out, we're able to then go say, now,
what questions do you want to ask? And that's when
(26:24):
ND said we do a lot of work as well
in identifying kind of the right questions to ask using
our models. Then once we have that question list, we're
able to take that audience. We start with the distributions.
We generate tens of thousands of agents who are logically consistent. Right,
So in the twenty one to twenty nine year old group.
There's not going to be any seven year olds, right,
There aren't going to be any eighteen year olds. There
won't be any thirty one year olds. We make sure
(26:46):
that everything is logically consistent, and then we're able to
simulate the behavior of each individual and predict that on
an individual by individual level. Remember, these aren't real people,
but hypothetically likely people to exist. We can collect results
across all of them, and hopefully in doing that, we'll
have cracked the code and told you exactly who we
need to put on air.
Speaker 1 (27:05):
I'm sure you can't choose amongst your children, But what's
been the most fascinating project intellectually for you both so far?
Speaker 2 (27:14):
So I tell you I run a lot of simulations
all the time, just because I'm interested in it. Like
it takes all the willpower in the world not.
Speaker 1 (27:21):
To just like it.
Speaker 2 (27:22):
Sit there all day long, logged into platform dot dot
com running SIMS. So I just ran. I run a
lot of really interesting ones. I just ran one I
was really curious about. For the upcoming twenty twenty six
primaries and the twenty twenty six midterm election season, we
both kind of come from a politics background, So that's
an area where we've been interested in for a long time.
(27:45):
And then we always track glps as you may know,
you know, which we like to talk about pretty openly.
Speaker 1 (27:50):
So I interested in jlps because they're in a sense
of intervention in behavior at scale, and the implications of
that intervention are thus far not fully known.
Speaker 2 (27:58):
I think I'm I've personally I don't know about you.
I'm personally interested in GLPS because this is like the
biggest mass movement and mass change we've had in public
health care in a long long time, Like healthcare has
never been more culturally relevant. I think by the way,
a lot of things fed into that GLP moment. It
wasn't just the existence of GLPES, but like health care
and wellness has been something we talk about increasingly over
(28:19):
the last five six years. And then it is going
to create massive behavior change, and most of the largest
businesses on the globe don't know how to navigate that
behavior change with the clarity they need, and so that's
actually why we do so much work in the GLP space.
Speaker 1 (28:32):
Now you've got a little bit held your feature The
fire by Andrew Rossulkin on Squawkbox because you said, basically,
our prediction is that there may be a short term
decline alcohol consumption, but actually as people start to feel
better and more confident because they're at a healthy weight
and whatever else I feel good about themselves, they'll be
out more and so they may have won drink, won drink,
won drink five times a week versus having five or
(28:55):
ten drinks, you know, in one day. The Squawkbox team
didn't seem to quite buy it.
Speaker 2 (29:00):
Well, I tell you what, they don't have to buy
it themselves as opinion, because it's actually now starting to
become a matter of fact. I don't know if you've
been watching earnings for alcohol businesses, but Q one earnings
have been coming back, and for the first time in
a long time, we're now seeing like case volumes four
alcohol come up. That tells you a lot about the
fact that already the alcohol industry is probably starting to
(29:22):
see some of the tail ones that the market was
a little bit too negative on GLPS. I've been reading
the analyst notes. The analyst notes talk about GLPS maybe
having some long tail benefit, and so I think it
makes a lot of sense that sure has it played
out completely, No, But do I feel like we're starting
to see the early signs in an alcohol industry recovery,
mostly driven by the fact that glps aren't as bad
(29:43):
as people think one hundred percent.
Speaker 3 (29:45):
Yeah, I was just going to say, you ask again
why it is so interesting to us and why this
is so important. The most valuable people on the globe
to understand the behavior of will always be the most
difficult to access. People on GLP ones aren't even telling
their family they're on these drugs. I'm not going to
tell you to survey, right and so right now, if
you're a I'll call producer. If you are a QSR,
(30:06):
whatever it may be, you're in a really, really, really
tough spot. Same thing goes with like ultrain it with
individuals or any of those groups, Like any very difficult
to access populations are always going to be common for
us to simulate because we can offer people data that
nobody else on the planet can.
Speaker 1 (30:19):
So could you do, for example, the younger crop of
revolutionary gods in Iran who haven't been assassinated, and what
they think the deal tends to make a piece deal
would be? I mean, I'm sort of being slightly faceties,
not like do you have enough data on Iranian revolutionary
gods or ah, we do it? Do that need to
be kind of like us based, would you be ablignate
(30:41):
that population?
Speaker 2 (30:42):
So we're we're totally global, right. We cover over one
hundred and sixty countries across the globe, all of Europe,
almost all of Asia, spare a couple of countries in
the Middle East, all of the Americas and most of
africas well. I'd have to look into it, but I
can tell you we probably don't have all the data
we need to simulate that. We're very clear about where
we do and don't work, and so for us it
is important there needs to be data in order to
(31:04):
simulate an audience accurately. But I would say there needs
to be less data than you think. And so young
IRGC commanders, is there enough data on that? Probably no, right,
But is there enough data for us to understand you know,
say Filipino household purchase decision makers for you know, cleaning
products in the country, Yeah, one hundred percent. And we
(31:25):
can understand regional variations between foss food purchases in Hochiman
City versus in Hanoi and Vietnam. Right, And that's because
of really two factors, the first being less data is
required than I think most people think, and the second
being that way way more data is being created than
most people think everywhere we go nowadays.
Speaker 3 (31:45):
Yeah, there's a really big challenge when you go to
foreign markets or you go well, not even for markets.
Let's just talk about in general modeling very specific populations
and niche geographies. Of those elements, one is you want
to model this population accurately. Great, you need really up
to date, you need really fairly granular demographic data, and
all those sorts of elements. Yeah, some of that is
(32:06):
very specific, like having credit card purchase history on a
very granular level is incredibly impactful. But if you're modeling
the population of even New York City, you still need
to know like what percentage of people have running water,
and me not having credit card history on a very
specific population. With iron on understanding how certain refugee populations
in the Middle East may react to a United Nations
(32:28):
announcement about peacekeeping or delivery of food supplies, it's probably
more relevant for me to have data on whether or
not they have running water than it is for credit
card things. Right, So there are certain things that are
more or less relevant depending on geography. I also say
the next thing that is always super important to keep
in a mind when you're talking about specific populations is
understanding of the environment more than it is just understanding
(32:49):
of them. So it's great if you have a perfect
demographic representation of Singapore. What actually matters a lot if
you're trying to simulate the difference between how people feel
today and in three weeks from now, is like, was
there some crazy news that happened? Right? Is there some
like big new trend on TikTok or whatever that's like
totally altered the way people think about a certain thing.
That's more important to capture than just knowing who people
(33:11):
are alone.
Speaker 2 (33:11):
I would say even more when you look at why
people use us for totally global businesses. Right, We've some
customers were supporting in you know, dozens, if not over
one hundred different countries. They're using us in part because
no one has a consistent bar for measurement across all
these different markets today. Right. So you know here in
the United States, if a place has below four point
(33:33):
five Google rating, it's probably not great. If it has
below four, it probably is really not great. And then
it has below three I probably wouldn't go there due
to food safety issues in Japan. Most rating systems start
with three and one would be terrible, two would be bad,
three would be good, and four would be excellent and
(33:53):
five would be incredible. And so if you're trying to
take yeah, if you're trying to take consumer survey across
the United States and Japan and you get you.
Speaker 1 (34:03):
Know, there's not there's apples and orange is basic point.
One product that look the same different arts.
Speaker 2 (34:08):
One product has a top two box score of like
sixty five percent in the United States, and one product
has a top two box score of like thirty percent
in Japan. Little do you know, thirty percent top two
box score in Japan is actually like an incredible indicator.
And in the States, sixty five percent top two box
scores like not that impressive, right, So when you use us,
A lot of the reason that people are bringing us
(34:28):
in is because we offer global consistency. It's the same
level of bar for every single market we support in
every single audience because it's all based off of real,
real behavioral outcome, so you can actually trust that something
is going to be the same across different countries.
Speaker 1 (34:42):
You mentioned that the longer term mission is to predict
the future. How far out does the future that you
can predict go today, and what might lengthen the horizon.
Speaker 3 (34:52):
Yeah, so this is kind of interesting because it's not
even a time frame or the amount of years, it's
the amount of degrees behavior that leads you to that. Right,
So you take something like an election, which of course
there's going to be environmental changes that happen in between.
There's you know, hundreds of millions of people that vote
in the US. There are all those people making decisions,
(35:14):
but they're making one singular decision. So it's not that
difficult to predict. You look at a election, and then
you look at how people react, and then you look
at who protests, and then you look at how it
affects trade, and then look at It gets harder and
harder and harder to be able to model that because
even if you're ninety nine point nine percent accurate, you
multiply by itself enough times and it decreases.
Speaker 2 (35:32):
I think for us, what you see where we really
truly excel today. We're incredible at population and group level
and cohort level behavior predictions. Right, so we are the
most accurate in the world at predicting something like an
election by quite a distance, or the most accurate in
the world of predicting consumer purchase intent by quite a distance,
the most accurate in the world of predicting social media
(35:54):
shareability by quite a distance. We aren't as good at
organizational action protection. We aren't good at looking at an
individual by individual level. That's not something we offer today.
If you say, I want to take this from just
how do we impact the population to like, how do
we truly predict the future? We need to nail those
as well. And so as you see where we're headed
over the long run, we're going to take our fundamental
(36:16):
model advantage, our tech advantage, and we're going to apply
that to these new domains so that we get even
better at predicting the future.
Speaker 1 (36:24):
So you're raising eighty million Series A at a valuation
whatever billion dollars, what's the use of prosus? What are
you investing the money in? And what support is the
valuation of the investors?
Speaker 3 (36:32):
People on both sides actually in many ways, I would say,
but more importantly, what we're spending the most amount of
capital on today and really really focusing on our number
one priority is getting the best people, the best people,
the smartest people, the most driven people. The only reason
we're able to get so far is as the nature
of our team. We have aongous, humongous priority on diversity
of thought, and that goes across research, engineering, product right.
(36:54):
People who spend time looking at understanding awareness of social
media when we see put it into our models in
those elements used to do Alzheimer's and dementiary research. There
are other people on our team who come from traditional
social science, demography, applied demography backgrounds, not you know, just
applied AI in those elements, because this problem can't be
solved in one singular way.
Speaker 1 (37:14):
And I came to your office and it was notable
for the no shoe optional, no shoe policy, which I'm
a big fan of, and also for the cigarettes, which
you guys living the gen z trend.
Speaker 2 (37:27):
We predicted it. We said cigarettes were coming back in
twenty twenty four.
Speaker 3 (37:30):
Yes we did not.
Speaker 1 (37:31):
Everyone's in love with what you're doing. The New York
Times racing right OpEd with the headline this is what
we'll ruin public opinion polling for good quote, pure fictions
are on the brink of being treated as scientific and
political knowledge. If we do not pull back, our understanding
of society might become artificial too. And this was a
(37:51):
direct criticism of an Exios story that had used our
research about trust levels in doctors and nurses, which didn't
have a disclosure that it was based on synthetic respondents.
And The New York Times, these two writers basically did
a broadside against you.
Speaker 2 (38:07):
Why they wrong, I think when you actually read that story,
and by the way, I read that story, I read
many of the other ones that followed, because I think
it's important to see how people feel about technology like this.
I think a lot of people are misunderstanding what makes
our technology fundamentally different. I actually agree with them. By
the way, there are thousands of papers about this idea
(38:28):
of you know as they call it, silicon sampling as
we call it, you know, the anthology approach where you
essentually take a profile that you have like generated, or
a profile of some human being, and you give it
to a language model and you say, be this person,
that is not accurate. Right. We tried that approach, actually,
trust me, like we when we started this company, that
was our base approach, and that's when we learned that
(38:50):
it actually doesn't reflect public opinion, and we very quickly
discovered that we had to be training our own models,
We had to build our own technology. We needed frontier
research of our own in behavior simulation in order to
actually have accuracy that delivers, right. And so I think
when they talk about the issues of that technology, I agree, So.
Speaker 1 (39:09):
What do they misunderstand about what you were doing? In
that case?
Speaker 2 (39:11):
The core thing to look at us is using real
behavioral outcomes and a fundamentally different model architecture and a
fundamentally different program. Right. We're training our own models. Our
models blend LMS and traditional mL with dozens of other techniques, methods,
things that we have built internally, right, And so I
(39:31):
think if you peaked under the hood and actually saw
how this worked, it's probably a lot more similar to
the traditional models that you know, people like that have
trusted for a long time than what people have read
about this idea of silicon sampling in the papers. And
so the real thing is when you're looking at real
behavioral outcomes. We've shown time and time again that we
can be more accurate than traditional public opinion research.
Speaker 3 (39:53):
And that we are. But to be very direct on that, right,
it's like why was that piece written, right? I like
think about it from why was that piece written? Was
it written because someone has like a passion to like,
you know, break apart at Teckning methodology? Absolutely not right.
It's written because someone sees that as a flawed approach
or something that's like, you know, bad, and that's important
to recognize. And I think it's something that you know, Look,
(40:15):
we're in New York for a reason because we're outside
of the echo chamber and.
Speaker 2 (40:18):
We hear this.
Speaker 3 (40:20):
But I think, like from the perspective of how we
look at this, we didn't build this company from a
perspective of like exclusively AI efficiency. And you know, all
these ads that we see in New York that we
think are despicable around like not hiring humans in those
sorts of areas, It's not the approach that I would
ever approach that I would ever think about, right. We
think about this in perspective of We'll go to someone
(40:41):
who's been a marketer for the last two decades, last
two decades, three decades. They might have thirty different areas
and like ten different managers that tell them what they
can and what they cannot put out, and people within
organizations for the last twenty some years have been looking
for an ability, for something else to give them the
confidence to take swings. And that's the way we think
(41:02):
about it. Right, you take something that nobody's ever launches
an now before, how do you go to your CEO
and say, let's put five million dollars on this instead
of you know, paying some random athlete who you know.
Ex Soap company just put on something and it did
well for them, So why shouldn't we do the same thing.
You're going to get turned down, you'll get fired, you
won't have that right. We give people numbers that are
more accurate than anything else on the globe, that we've
(41:23):
proven to be more accurate than survey research, so that
they can take those big swings, so that they can
make those actions, so that they don't have paralysis inside
of businesses waiting for like ten other people to agree
with their decisioning. We give people quantifiable numbers and reasons
why they're right if they are, and reasons why they're
wrong if they're wrong. And that's what our goal is.
Speaker 1 (41:45):
Final question to you both. You met in high school,
you became best friends, more or less love at first sight.
Now you live together, you have a billion dollar startup together.
Speaker 3 (41:55):
Where's this gone?
Speaker 1 (41:57):
Well, do you have any fear that in enormous success,
this room maybe could come between you?
Speaker 2 (42:04):
No, not announced.
Speaker 3 (42:05):
I think too much money and effort has gone to
nurturing unremarkable talent, and we have a very heavy performance culture.
I think being able to be our age and be
able to work every single day with people that we
love and we care about on the thing that we
(42:26):
think is the most important technology in the world is
something that is more unifying that you can ever imagine.
And I could be more grateful to be able to
work alongside.
Speaker 2 (42:37):
You, much more eloquent than me. I love you. He's
a legend.
Speaker 1 (42:40):
Yeah.
Speaker 3 (42:41):
No, I mean look, I think people who say don't
do business with friends need better friends. You know.
Speaker 1 (42:44):
That's what it is. There, You go, mat Cam, thank
you so much.
Speaker 3 (42:49):
Thank you so much for having us. It's great chat.
Speaker 1 (43:09):
For tex staff iMOS Voloshin. This episode was produced by
Eliza Dennis. It was executive produced by me and Julian
Nutter for Kaleidoscope and Katria Norvel for iHeart Podcasts. Jack
Insley mixed this episode and Kyle Murdoch wrote our theme song.
Welcome to tech stuff. I'm os Vloschen. Every year we
spend billions of dollars asking real humans for their opinions
to predict how groups of people behave On a larger scale,
surveys are the foundation of political strategy, marketing, and produge innovation.
So what happens when you stop asking humans what they
think and start simulating their responses? Today I'm joined by
(00:38):
cam Think and Ned Coo, the founders of Aru. At
the ages of twenty and twenty one, they've built a
billion dollar simulation engine that uses thousands of AI powered
digital twins to deliver behavioral predictions for everyone from Ernst
and Young to Spindriff, Cam Ned, Welcome to text stuff.
Speaker 2 (00:56):
Thank you for having us having us.
Speaker 1 (00:58):
Thanks you very very excited. Tell me about the name.
How did you choose it?
Speaker 2 (01:01):
Yeah? Well, Uru is the place in Egyptian mythology where
all the worlds meet, and given we simulate human behavior,
we simulate all these worlds. It makes a lot of sense.
But it's also the oldest trick in the book. So
you know, Auru starts at two a's at the top
the top of every alphabetical list.
Speaker 3 (01:17):
The Yellow Pages trick you know, triple A plumbing, all
those sorts of things. You look at any of our investors'
websites and you see us and then Airbnb, which we
think is the proper order. I'm just kidding.
Speaker 1 (01:27):
That's very very good, Am I right in thinking, or
at least is Wikipedia right in thinking that to get
into Uru, that your heart has to be weighed? That
is true, That is true. Are you in the business
of weighing hots?
Speaker 2 (01:39):
I think by so, I think by technicality, really is
Uru is the place that describes the field of reads, right.
Speaker 1 (01:46):
So we are just kind of a heaven, shall we say,
an ancient Egypt?
Speaker 3 (01:49):
Yeah, exactlytopia almost.
Speaker 2 (01:51):
This is, by the way, this is why r dot
com our domain cost us so much, because I only
learned afterwards that anything that has anything to do with heaven,
like any domain, like our fortune, because it's really easy
to bring something that's being positive. But we are not
in the business of waite arts. I don't know if
that would be as important for the global economy.
Speaker 1 (02:10):
So, because I was thinking, in a sense, you are
definitely not judging people, but in a sense you are.
You're looking into the digital arts, or at least trying
to distill how real people might behave from people you've simulated.
Speaker 3 (02:25):
No, I think it's a it's interesting, and we like
things that are thematic and actually mean things. I think
there's a lot of tech companies nowadays that have names
that don't really don't really mean much.
Speaker 1 (02:33):
So all the ones around Tolkien were taking.
Speaker 3 (02:36):
Exactly and the ones from Dune and the ones from
Lord of the Rings, and it's, ah, well, there we
go to mythology. You know, you gotta pull one.
Speaker 1 (02:43):
I want to sort of get to how aru actually works.
But I saw the article about you two in the
Wall Street Journal or you three, actually your third co
founder is not here, and you know, basically the teenage
teenage founders of a billion dollar company. And I thought
it was a very fascinating and cool story. And I'm
a huge tennis and it's been kind of interesting for
me to observe how Djokovic, Nadal and Federer cast this long,
(03:08):
fifteen year shadow over the game of tennis where the
kind of next gen couldn't emerge. And it wasn't until
Nadal and Federer left the scene and Djokovic started to
age that Ciner and.
Speaker 3 (03:18):
Al Karaz emerged, you know, a lot of people say,
Cameron looks like by the way, Yeah, I like that.
Speaker 1 (03:27):
So, I mean, you know, it's somewhat of a parallel
with the tech industry. Right before the twenty twenty three
chat GPT moment and the diffusion of llms, basically, most
people with strong technical skills would go and work in
corporate jobs at like Facebook and Google and Amazon. Yeah,
and then something changed, and you and the censor at
(03:47):
the vanguard of that something that changed.
Speaker 2 (03:50):
I think what happened is, for like a really long time,
you had people who were very technical getting paid a
lot of money to work, largely like kushy jobs. Right
right now, we have this perception of Silicon Valley nine ninety.
Speaker 1 (04:03):
Six, like all this stuff, correct engineers.
Speaker 2 (04:05):
Yeah, it wasn't that way for the longest time. I mean,
if you go back, like even five years in history
to twenty twenty one, Silicon Valley was where people drank
beer during the work day and played ping pong, you know,
at four o'clock in the afternoon. It isn't like that anymore,
because I think what happened is people got actually attracted
to the idea of hard work. Right It's like this
backlash against what was once a work culture of like
(04:27):
go work ten hours twenty hours a week of productive
time and then not spend the rest of your time
doing as many productive things. People want to work hard,
and I think that changed.
Speaker 3 (04:36):
I mean, with a barrier entry to building a business
being so so much lower, so many things has changed, right.
I don't think it's just young people being able to
do it. I think it's also the people who are
at the top of the fields also have changed a lot.
Like something Cameron and I discuss a substantial amount is
you know, where is this generation's you know, Steve jobs
or those sorts of areas. And it's actually part of
(04:57):
our thesis around this space is that people have and
had to rally around a singular lumin area as much
because capital has been so much more accessible. Right, So
it's not just as part of language models. I think
it's a state of the you know, overall economy and
things that have changed that have allowed it to happen.
But it's very interesting. Nonetheless, Yeah, it is interesting.
Speaker 1 (05:14):
Do you think it's like a stylistic thing that your
generation want to work hard or do you think it's
an opportunity thing that like if you work hard, you
can be net and Cam and get a big Unicorn
founders before you're twenty one.
Speaker 2 (05:25):
I actually think it's I think it's a bifurcation, right,
Like I think you have a population of people who
work really hard, and that might be like the top
death style. But what just happened is like the distribution
of how hard people work has widened so much, so
like just as much as in my high school class,
there are a bunch of kids who, like maybe ten
fifteen kids who worked extremely hard orderly. You know, I
(05:47):
was not in the top performing in my high school class.
Actually proud to announce that I was thirty third rank
out of one hundred and eight.
Speaker 1 (05:53):
You were high school together, what were you?
Speaker 3 (05:55):
Probably much lower. It's very difficult for me to focus
on things I don't care about, so you know, it's.
Speaker 2 (06:00):
But you had like ten to fifteen, maybe twenty kids
who would like go study six hours and they would
be totally happy to do that, and on the contra,
like the next eighty kids just did not want to
put all the work in. And I think you see
that consistently where it's like there's this really high visibility
group of people who work a lot, but remember, there
are tons of software engineers at legacy businesses who don't
(06:21):
work as much as like the high visibility group of people.
That is like a very small small group of people
in comparison to like broader software engineering is a category.
Speaker 1 (06:31):
How do you think about your paid group? I mean,
do you know the other twenty eighth year old Unicorn founders?
Do you associate with them?
Speaker 3 (06:37):
Or?
Speaker 1 (06:37):
I mean you're in New York and most are in
San Francisco, which is a difference, right, Like, what's your
reflection on the culture that's emerging around around that? I
love the idea of Steve Jobs no longer being a
cultural reference point because the structure of how capitalist deploy
is different. That's really interesting. But how do you think
about the group you know?
Speaker 2 (06:53):
I'm curious for Ninn's thoughts as well. I really really
think the fact that more people know that is a
pathway that is open to them is brilliant. I think
that's incredible, right, like one hundred percent people should go
do what they want. I do think some of the
current environment has created people who are not in this
because they're missionaries, but are instead in this industry because
(07:16):
they're mercenaries and not to say that that is like
the founders of other unicorns or other young people. But
I just see a lot of people now who I
think go to tech because not of a mission, not
because they think it's like important philosophically important for the world,
but instead because tech is like the next highest opportunity
earning category, like finance was in the nineteen eighties. And
(07:37):
I think shifting away from that is going to be
interesting because the pendulum will always come back around.
Speaker 3 (07:44):
No, I agree entirely with that. I also think, to
the same extent Cameron mentioned, you've also seen like the
pliferation of so many different industries, not just tech, and
people saying, oh, I can actually make a business. Ow's
something I like, And it's become infinitely easier to monetize
right as like a very small business across so many
different channels, whether it's e commerce, whether that's you know,
starting a tiny little brand and CpG or whatever may be.
(08:05):
So I don't think it's a trend that's just in tech.
I do think people are a lot more public about it,
but it's been interesting to see. Nonetheless, I mean, we
have friends that also dropped out of university to start
schools and have had other companies that have been starting,
you know, starting very very young, and I think there
was this large sort of cultural shift. Cameron noted about,
you know, silkon Valian, it's perception prior, it's now in
(08:26):
some ways like sexy to be a tech founder. So
I think you've a lot of people who whose hearts
really aren't in it and these sorts of elements who
are going after it. And you know, like anything in
the world that you commit to not fully is things
have repercussions.
Speaker 1 (08:38):
There's also an accessibility thing, right, Like your third co
founder is not here. I think when you Will started
you were nineteen, you were eighteen, and he was fifteen yep,
which is kind of remarkable. But he reached out and
basically said you were an accelerator program, can I get
your advice? And then you left the call basically with
him as a third co founders That might probably have
a simplified but it.
Speaker 2 (08:57):
Was pretty close to that. He called them me on
LinkedIn and he just said, I'm applying to this accelerator,
can I have your advice on my application? I don't
even check my LinkedIn dms that often. I certainly didn't
when I was eighteen, I mean, you know what I've
been doing on LinkedIn, And so we ended up scheduling
a thirty minute zoom call. We spoke for two and
a half hours, and I called Ned immediately afterwards and
(09:19):
I said to Ned, I've just met the smartest person
I've ever met in my life. And that was the moment.
More or less immediately, I think the three of us
hopped on a call later that night and we were like,
screw it, We're starting a company together. I mean, Nut
and I were about to sell our previous business. John
literally dropped out of high school and we set everything
down and said, is, while we're passionate about.
Speaker 1 (09:39):
What was the vision you laid out to him? And
what was it that he said in response that made
you think he was the most intelligent person you ever met.
Speaker 2 (09:46):
Well, at the time, he actually was writing a paper
focused on RAG for healthcare agents, so this space I was,
Nett and I were in digital health as well, and
so I remember reading it and thinking to myself, Wow,
like just reading these benchmarks, If these benchmarks are real,
it was a material improvement in the quality of models,
specifically for retrieving like medical evidence and clinical studies. At
(10:07):
that time it was called heel actually hal and reading that,
I was just like, wow, this is insane. I spoke
to him about it, and I was kind of digging
in and I was like, well, it seems really real.
So like a fifteen year old living in a dorm
at a boarding school in like rural Massachusetts who's from
New Hampshire is able to best the labs and that
(10:29):
was incredible, And so all of us randomly having a
digital health background, but then we also all shared this
vision of population research. John was like an original population
researcher when he probably also called dm'd its way into
a lab at Tufts and a lab at MIT doing
city science and population demographic research.
Speaker 1 (10:47):
But did you already have the idea for.
Speaker 3 (10:51):
Something that was developed in its current manifestation together, right,
But in the idea of understanding and simulating populations. It's
something we thought about for a while, but our own
it's manifestation was certainly something that we developed together, which
I think is cool.
Speaker 2 (11:03):
I think it was October twenty twenty three we started
talking about predicting behavior for the first time, and I
think from there on at it was something that just
completely consumed us.
Speaker 1 (11:12):
See you mentioned missionaries versus mercenaries. What is the mission
for us?
Speaker 2 (11:16):
It's predict the future. I think it's really simple. If
you think that humans are the most impactful species on
the club, and we're able to predict human behavior, then
predicting human behaviors tend amount to predicting the future. And
if you can predict the future, you can shape it.
It's not just the power to say this is how
your product concept will perform, or this is who we
(11:37):
expect to win in an election, but instead it's actually
the power to find the product that wins and to
help the people ideas that matter win elections more consistently.
Speaker 1 (11:48):
So predict littles of influence.
Speaker 3 (11:50):
Predict and shape and help people that need to shape
behavior shape that right for the broader good. And that's
something we care a lot about. There's a lot of
predicted technology that exis for the aspect of simply measuring
behavior right. And at the end of the day, what
has to be done after that? When you get a
spreadsheet that says this is what this person said, this
is what this person said, et cetera, it's not very
(12:11):
actionable to anybody, and so it ends up being this
humongous waste of an exercise, because as much as might
have taken you months to get to that point, no
matter what measurement it gets to, it's very difficult for
someone to have to go from there and make their
own assumptions upon that data to then make a conclusion.
And then frequently that data ends up just supporting what
that initial person's decisioning has been or gets sworn that
way or another. We have no like manual waiting on
(12:33):
the output of our system. We run simulations of behavior
our system as analysis upon it, and people take action
as a result. And that's part of what makes it
so powerful.
Speaker 1 (12:41):
When I think about prediction and influence, I guess there's
like black hat predictors and influences like you know, Goebbels
or Cambridge Analytica, and then there's like would be white
hat predictors and influences like cast Sunstein and his like
famous nudge theory. Right, do you think about this spectrum
all the time?
Speaker 3 (13:01):
And if we didn't, I don't think we would believe
in our technology. But we do, and we see how
strong it is every day, and we do as much
as we can to be on the white out side
of that. Right, So, like I think about things and
you know, maybe not to announce it here, you know,
formulae leans or not. We're not without finance and all
that yet. But later this year we'll be launching a
research institute within our where we're dedicating a significant amount
(13:23):
of compute and infratructure from our side who works exclusively
on pro bono aspects. So certain these areas are you know,
women's self for example, getting people to go to preventive
screenings for you know, preventive diseases and other elements like that.
That's a behavioral problem, right. Things like getting people to
understand that nuclear energy isn't dangerous, that's a behavioral problem, right,
All those sorts of elements. There's so much good you
(13:44):
can do in the world if you understand what people's
awareness of certain issues are, how to make them aware
of certain elements and how to change their behavior on
things that are genuinely better for society. And that's something
we spend a lot of time thinking about.
Speaker 1 (13:56):
Do you have philosophical ethical debates the two of you
with your co found it do board? What is the
count sort of I was going to account some elders,
but it's not appropriate in this conversation. But how do
you debate these questions together?
Speaker 2 (14:08):
I mean, we were a company founded in like really
deep philosophical principles. We do have a lot of thoughts,
you know, about the moral and ethical use of our technology. Right, Like,
to be clear, predicting behavior is really powerful. Predicting behavior
incredibly accurately, with the capability to shape it is like
(14:29):
even more powerful. Right, this is something that not only
economically is important, but for the good of the world
it's important. And so I think part of the reason
we built these company is because we wanted to use
this technology as a force for good.
Speaker 3 (14:40):
So how does it work well within what we can share?
Of course? No, it's really interesting because you think about
the architecture or system, and you think about the proliferation
of language models and how much more accessible starting certain
elements has been and all these things. And certainly these
systems would not be capable without language models. But they're
not the only thing that makes them what they are. Right,
(15:00):
you think about what a language model is built for
and its most basic principles. It is not to capture
the irrationality of human behavior, right, It's not. It's in
fact quite the opposite. To average results exactly. It is
a beta generating machine in many ways, right, And so
you think about the things that make humans who we are,
that makes us act or buy a product from a
(15:21):
brand and a new category that they've never made before.
Of course, a language model doesn't work for that use case, right,
And so there have been many evolutions of the architecture
that we've come to today. At its core, what it
is is what we call multi agent simulation. So effectively,
we organize these large populations of AI agents, assign them
different distributions of income, age, race, gender, sex, you know,
(15:44):
credit card purchase history, all these sorts of elements that
makes them who they are. The entire principle of how
we generate those distributions of what we call audiences in
the first place, or the demographics that make that simulate
and population up or that we never actually survey people
at all, and so we're just generate these populations off
of the basic building blocks and what makes them who
they are. So what we call ground truth data, right,
(16:06):
we only look at data from areas where we don't
actually have to ask people because people lie, they have biases.
And you go to trybeca near where our office is
and you ask the average man how much money makes
He's gonna say four and a half million bucks, you know,
like whatever it may be, right, So we generate these
populations based off these ground truth distributions of traits and
then be able to sign those distribution of traits to
these thousands of models to become simulated populations, or will
(16:28):
then ask them questions to measure their behavior. Right, so
you might ask someone you know, who you vote for,
what will you buy? How much will you buy? This
for all those sorts of elements to be a measure
everything from proprency to purchase to who's going to win
an election to you know, churn risk and everything in between.
And is it falsifiable like by a validation?
Speaker 1 (16:45):
How do you like? How do you be tested exclusively
to outcomes?
Speaker 3 (16:49):
Right, So we never try to benchmarker models on surveys
or any sort of traditional primary research, so surveys, polls,
focus groups, et cetera, because again we believe this to
be biased. So exclusive look at the actual actions people
take and not what they say they do. And there's
a lot of different examples of where we've been able
to do this publicly that have been really cool. You know,
this is a fantastic example with one of our partners,
(17:10):
Ernst and Young, who you mentioned. They took a survey
that traditionally took the months to run. Thirty six hundred
people responded, thirty countries, fifty three questions. We said, don't
trust their models. Right, This is what we say to
all our customers. They gave us the questions they asked.
They gave us the people they spoke to, or the
quotas for the assets under management. It was an investor simulation.
We ran the entire questions that within our model in
(17:32):
a matter of minutes returned it to them. They did
a correlation, had a ninety percent correlation, but in the
ten percent where we were not accurate, we were significantly
closer to the real human action. So you think about
a question in that example, right. One of those in
that study was what's your likelihood to retain a wealth
manager when your parents pass away? What are the three
(17:53):
things that go through your head when you get asked
that question. The first one is I hope my parents
don't pass away. Second one is that might have been
in five years, ten years, thirty years, I don't know.
I'm going to do my money then, and the third
one is my wealth managers asked me this question. I'm
not going to tell them a lot, right, So you
have this whole scenario where at the end of the day,
eighty two percent of the people in their traditional measurement
gold standard survey said they were going to retain wealth manager.
(18:13):
Real market data from actual wealth management firms said this
between the tune of like twenty to forty percent, depending
on geography. Our simulation came back a forty percent. So
you look at a result like that, like customers do
all the time when they see the difference between our
simulations and the surveys. They run eighty two percent in
their survey forty percent in ours, and it looks totally wrong.
You look at real actions in the real world.
Speaker 1 (18:35):
We're right on the money, okay, But imagine doesn't know
even as I am. How is it possible that fake
people can get better on so those than real people.
Speaker 2 (18:42):
Because real people lie. I mean, it's that simple one.
When I go and I phone you and I say,
you know, hey, my name's Cam. I'm a researcher with
Consumer Research Incorporated. How do you feel about the coffee
you just drank from Starbucks this morning. What are you
going to tell me you're going to hang up the phone.
Are you going to say it was great and I
got off the phone. And so when you're looking at
(19:03):
that in traditional consumer research world, it just doesn't align
incentives for consumers to actually provide a real opinion, right,
a real reflection of how they care what they feel.
And so when you're looking at our data, instead, we
don't ask consumers how they care what they feel, right,
because we know that they're not going to be honest
with us. We know that nothing that is trained on
(19:24):
or built on top of surveys or traditional consumer research
is ever going to be able to reflect that accurately.
And so instead we run another direction. We train on
top of the outcomes. As Ned said, we train on
things like actual coffee mug sales and actual coffee cup sales, right,
And by looking at that data we can track far
better how humans opinions and behavior is changing over time.
Speaker 1 (19:46):
What's the most niche thing that you're proud of having
got right? What's the whole in one double egle?
Speaker 3 (19:52):
I mean, there's some really really specific ones. So there's
certainly been cases where we've done things like price elasticity
of like bundling of very specific automotive parts. That is
crazy that we get those correct. But then there are
other really interesting things, like you know, breaking new artists
in music. Right, we're on iHeart Radio right now, right,
Like looking at being able to simulate the star power
(20:14):
of a very specific emerging artist and predict that they're
going to blow up or that they have propensity to
blow up in these specific areas. That's incredibly cool.
Speaker 1 (20:22):
So this was an advanity fair piece. It was an
artist you found who had forty thousand streams when you
found them and went on to release a song with
one hundred and seventy two millions off you'd identified them.
What more can you share about this?
Speaker 2 (20:37):
Not that much, But it was a really really cool
use case. And I am proud to say that I
listened to their music on our way up there.
Speaker 1 (20:46):
Yeah, back up, yes we did. You can't say what
the artist.
Speaker 2 (20:49):
Is, No, I can't say who the artists?
Speaker 1 (20:50):
What can you share about the criteria? Yeah?
Speaker 3 (20:52):
So this is really interesting. We were testing a very
specific population. It was people simulating people listening to music
who are under the age of twenty five, right, So
we're looking at like young sort of looking at emerging
underground these sorts of areas. It was really interesting because
we ended up feeding in the Instagram feeds, Twitter feeds
some of the actual songs and prossing the audio on
(21:13):
those elements and understanding is this a res rating compared
to all these other artists, Like how often you'd listen
to this artist? Would you go to a concert for
this artist? How would you pay for a ticket to
see this artist if your friends told you about them,
like all these super specific elements that we worked with
to quantify whether or not, out of a large ranking
of these specific performers, which one would do best. And then,
(21:34):
of course that was when they were much smaller. They
ended up getting signed and sorts of things and then
had a lot of popular songs come out. So you know,
we can't take full responsibility for the fact they got
that large, but it was certainly interesting to see that
we were properly able to simulate the general populations prepensy
to listen to music, which is really cool because that
is something that's so difficult to quantify.
Speaker 1 (22:18):
Well, I actually have one that I would love to
work with you guys on. So I am the host
of this tech Stuff podcast and tex Stuff is a
co production of iHeart Where we Are and Kaleidoscope, which
is the media company I founded, and our mission is
essentially to communicate the wonder and excitement of this new
age of which you two are representative of. And that
(22:40):
idea that you articulated, which is that like a small
group of people run with everything they have towards being
like participant builders and a large group of people don't
is perhaps one of the greatest social problems we have.
And so what we think about a Kalidascope is how
can we tell stories and do media products podcast, YouTube,
(23:02):
et cetera, which are not polyannoriatee or just you know,
celebrations of the tech industry, but but rather that are
kind of engaging and inspiring storytelling about what's possible in
this new world because a lot of the narratives about it,
yes they are available, but they're also like by the industry,
for the industry.
Speaker 2 (23:18):
It's a closed and ecosystem in a lot of ways.
Speaker 1 (23:21):
I was thinking this morning on the way to talk
to you guys, how do we successfully identify the next
cohort of the best science and tech creators, influencers, et cetera.
So how would we go about doing Like, let's say
I was a paying client, which I think I might be.
(23:44):
It has to help like what would be the sales
motion as it were, if I was a real client,
Like what would you what would you guys? Where would
you go from like this problem to providing me with
the answer.
Speaker 3 (23:55):
No, I think this is really interesting. And the first
part that you're bringing up there is one of the
biggest challenges for insights, research, marketing, strategy, product pricing, whatever
you are team today because what makes research, say, it
takes so lethargic of a process today, like weeks, if
not months sometimes Yeah, part of it is like asking
the questions and finding the people to talk to and
those sorts of elements. Part of it is figuring out
(24:17):
what the numbers mean afterwards. The biggest challenge and the
place where we actually see people go the most wrong,
is figuring out what to ask in the first place, right.
And so one thing that we've built our system to
be able to do because we serve as so many
different types of operators is feed in general objectives, feed
in semantic descriptions, and our system will take that in
(24:38):
and figure out exactly what the simulations you have to
run to answer that objective are And so maybe for
that example, the first thing you want to find out is, Hey,
before we go identify what talent works with people, let's
go see who has the appetite for this in the
first place. Segmentation this is huge business for us. We
do a lot in CpG and marketing in these areas
because we don't need to get in the weeds of
your CRM to segment your audience. We can just simulate,
(25:01):
you know, half million people in downtown New York City
and see who wants to listen to your product, or
see who wants to buy it for a certain price. Right,
So it's much more first principles, like top down because
we have that scale. So maybe the first time we'd
run there would be understanding a certain population. You said
you're interested in the youth mostly right, probably as being
an audience maybe given age rage. I'm curious let's say
(25:22):
twenty one to twenty nine careert, so twenty one to
twenty nine, right, And we probably feed that in any
other descriptors that you have about this population that you're
interested in, maybe psychographically like open to sort of personal
self development. Not already checked out, but wondering how to
live a better life.
Speaker 2 (25:41):
So we would be able to take in that description,
you know, twenty one to twenty nine people who are
interested sort of open to non insiders, not insiders, people
who maybe insider rejectors, right, people don't insiders, but people
who are open to learning more about technology, and our
models are able to pull out that description and actually
say okay, based off of who even something that simple
(26:04):
is like what is their urban world breakdown? What is
their home cooking meal frequency? You know, where do they
spend their time on social media? How many hours a
day did they go to school? Yeah, what is their education?
What is their income level? You know, three hundred plus
different variables that we're storing. And then once we chart
all of that out, we're able to then go say, now,
what questions do you want to ask? And that's when
(26:24):
ND said we do a lot of work as well
in identifying kind of the right questions to ask using
our models. Then once we have that question list, we're
able to take that audience. We start with the distributions.
We generate tens of thousands of agents who are logically consistent. Right,
So in the twenty one to twenty nine year old group.
There's not going to be any seven year olds, right,
There aren't going to be any eighteen year olds. There
won't be any thirty one year olds. We make sure
(26:46):
that everything is logically consistent, and then we're able to
simulate the behavior of each individual and predict that on
an individual by individual level. Remember, these aren't real people,
but hypothetically likely people to exist. We can collect results
across all of them, and hopefully in doing that, we'll
have cracked the code and told you exactly who we
need to put on air.
Speaker 1 (27:05):
I'm sure you can't choose amongst your children, But what's
been the most fascinating project intellectually for you both so far?
Speaker 2 (27:14):
So I tell you I run a lot of simulations
all the time, just because I'm interested in it. Like
it takes all the willpower in the world not.
Speaker 1 (27:21):
To just like it.
Speaker 2 (27:22):
Sit there all day long, logged into platform dot dot
com running SIMS. So I just ran. I run a
lot of really interesting ones. I just ran one I
was really curious about. For the upcoming twenty twenty six
primaries and the twenty twenty six midterm election season, we
both kind of come from a politics background, So that's
an area where we've been interested in for a long time.
(27:45):
And then we always track glps as you may know,
you know, which we like to talk about pretty openly.
Speaker 1 (27:50):
So I interested in jlps because they're in a sense
of intervention in behavior at scale, and the implications of
that intervention are thus far not fully known.
Speaker 2 (27:58):
I think I'm I've personally I don't know about you.
I'm personally interested in GLPS because this is like the
biggest mass movement and mass change we've had in public
health care in a long long time, Like healthcare has
never been more culturally relevant. I think by the way,
a lot of things fed into that GLP moment. It
wasn't just the existence of GLPES, but like health care
and wellness has been something we talk about increasingly over
(28:19):
the last five six years. And then it is going
to create massive behavior change, and most of the largest
businesses on the globe don't know how to navigate that
behavior change with the clarity they need, and so that's
actually why we do so much work in the GLP space.
Speaker 1 (28:32):
Now you've got a little bit held your feature The
fire by Andrew Rossulkin on Squawkbox because you said, basically,
our prediction is that there may be a short term
decline alcohol consumption, but actually as people start to feel
better and more confident because they're at a healthy weight
and whatever else I feel good about themselves, they'll be
out more and so they may have won drink, won drink,
won drink five times a week versus having five or
(28:55):
ten drinks, you know, in one day. The Squawkbox team
didn't seem to quite buy it.
Speaker 2 (29:00):
Well, I tell you what, they don't have to buy
it themselves as opinion, because it's actually now starting to
become a matter of fact. I don't know if you've
been watching earnings for alcohol businesses, but Q one earnings
have been coming back, and for the first time in
a long time, we're now seeing like case volumes four
alcohol come up. That tells you a lot about the
fact that already the alcohol industry is probably starting to
(29:22):
see some of the tail ones that the market was
a little bit too negative on GLPS. I've been reading
the analyst notes. The analyst notes talk about GLPS maybe
having some long tail benefit, and so I think it
makes a lot of sense that sure has it played
out completely, No, But do I feel like we're starting
to see the early signs in an alcohol industry recovery,
mostly driven by the fact that glps aren't as bad
(29:43):
as people think one hundred percent.
Speaker 3 (29:45):
Yeah, I was just going to say, you ask again
why it is so interesting to us and why this
is so important. The most valuable people on the globe
to understand the behavior of will always be the most
difficult to access. People on GLP ones aren't even telling
their family they're on these drugs. I'm not going to
tell you to survey, right and so right now, if
you're a I'll call producer. If you are a QSR,
(30:06):
whatever it may be, you're in a really, really, really
tough spot. Same thing goes with like ultrain it with
individuals or any of those groups, Like any very difficult
to access populations are always going to be common for
us to simulate because we can offer people data that
nobody else on the planet can.
Speaker 1 (30:19):
So could you do, for example, the younger crop of
revolutionary gods in Iran who haven't been assassinated, and what
they think the deal tends to make a piece deal
would be? I mean, I'm sort of being slightly faceties,
not like do you have enough data on Iranian revolutionary
gods or ah, we do it? Do that need to
be kind of like us based, would you be ablignate
(30:41):
that population?
Speaker 2 (30:42):
So we're we're totally global, right. We cover over one
hundred and sixty countries across the globe, all of Europe,
almost all of Asia, spare a couple of countries in
the Middle East, all of the Americas and most of
africas well. I'd have to look into it, but I
can tell you we probably don't have all the data
we need to simulate that. We're very clear about where
we do and don't work, and so for us it
is important there needs to be data in order to
(31:04):
simulate an audience accurately. But I would say there needs
to be less data than you think. And so young
IRGC commanders, is there enough data on that? Probably no, right,
But is there enough data for us to understand you know,
say Filipino household purchase decision makers for you know, cleaning
products in the country, Yeah, one hundred percent. And we
(31:25):
can understand regional variations between foss food purchases in Hochiman
City versus in Hanoi and Vietnam. Right, And that's because
of really two factors, the first being less data is
required than I think most people think, and the second
being that way way more data is being created than
most people think everywhere we go nowadays.
Speaker 3 (31:45):
Yeah, there's a really big challenge when you go to
foreign markets or you go well, not even for markets.
Let's just talk about in general modeling very specific populations
and niche geographies. Of those elements, one is you want
to model this population accurately. Great, you need really up
to date, you need really fairly granular demographic data, and
all those sorts of elements. Yeah, some of that is
(32:06):
very specific, like having credit card purchase history on a
very granular level is incredibly impactful. But if you're modeling
the population of even New York City, you still need
to know like what percentage of people have running water,
and me not having credit card history on a very
specific population. With iron on understanding how certain refugee populations
in the Middle East may react to a United Nations
(32:28):
announcement about peacekeeping or delivery of food supplies, it's probably
more relevant for me to have data on whether or
not they have running water than it is for credit
card things. Right, So there are certain things that are
more or less relevant depending on geography. I also say
the next thing that is always super important to keep
in a mind when you're talking about specific populations is
understanding of the environment more than it is just understanding
(32:49):
of them. So it's great if you have a perfect
demographic representation of Singapore. What actually matters a lot if
you're trying to simulate the difference between how people feel
today and in three weeks from now, is like, was
there some crazy news that happened? Right? Is there some
like big new trend on TikTok or whatever that's like
totally altered the way people think about a certain thing.
That's more important to capture than just knowing who people
(33:11):
are alone.
Speaker 2 (33:11):
I would say even more when you look at why
people use us for totally global businesses. Right, We've some
customers were supporting in you know, dozens, if not over
one hundred different countries. They're using us in part because
no one has a consistent bar for measurement across all
these different markets today. Right. So you know here in
the United States, if a place has below four point
(33:33):
five Google rating, it's probably not great. If it has
below four, it probably is really not great. And then
it has below three I probably wouldn't go there due
to food safety issues in Japan. Most rating systems start
with three and one would be terrible, two would be bad,
three would be good, and four would be excellent and
(33:53):
five would be incredible. And so if you're trying to
take yeah, if you're trying to take consumer survey across
the United States and Japan and you get you.
Speaker 1 (34:03):
Know, there's not there's apples and orange is basic point.
One product that look the same different arts.
Speaker 2 (34:08):
One product has a top two box score of like
sixty five percent in the United States, and one product
has a top two box score of like thirty percent
in Japan. Little do you know, thirty percent top two
box score in Japan is actually like an incredible indicator.
And in the States, sixty five percent top two box
scores like not that impressive, right, So when you use us,
A lot of the reason that people are bringing us
(34:28):
in is because we offer global consistency. It's the same
level of bar for every single market we support in
every single audience because it's all based off of real,
real behavioral outcome, so you can actually trust that something
is going to be the same across different countries.
Speaker 1 (34:42):
You mentioned that the longer term mission is to predict
the future. How far out does the future that you
can predict go today, and what might lengthen the horizon.
Speaker 3 (34:52):
Yeah, so this is kind of interesting because it's not
even a time frame or the amount of years, it's
the amount of degrees behavior that leads you to that. Right,
So you take something like an election, which of course
there's going to be environmental changes that happen in between.
There's you know, hundreds of millions of people that vote
in the US. There are all those people making decisions,
(35:14):
but they're making one singular decision. So it's not that
difficult to predict. You look at a election, and then
you look at how people react, and then you look
at who protests, and then you look at how it
affects trade, and then look at It gets harder and
harder and harder to be able to model that because
even if you're ninety nine point nine percent accurate, you
multiply by itself enough times and it decreases.
Speaker 2 (35:32):
I think for us, what you see where we really
truly excel today. We're incredible at population and group level
and cohort level behavior predictions. Right, so we are the
most accurate in the world at predicting something like an
election by quite a distance, or the most accurate in
the world of predicting consumer purchase intent by quite a distance,
the most accurate in the world of predicting social media
(35:54):
shareability by quite a distance. We aren't as good at
organizational action protection. We aren't good at looking at an
individual by individual level. That's not something we offer today.
If you say, I want to take this from just
how do we impact the population to like, how do
we truly predict the future? We need to nail those
as well. And so as you see where we're headed
over the long run, we're going to take our fundamental
(36:16):
model advantage, our tech advantage, and we're going to apply
that to these new domains so that we get even
better at predicting the future.
Speaker 1 (36:24):
So you're raising eighty million Series A at a valuation
whatever billion dollars, what's the use of prosus? What are
you investing the money in? And what support is the
valuation of the investors?
Speaker 3 (36:32):
People on both sides actually in many ways, I would say,
but more importantly, what we're spending the most amount of
capital on today and really really focusing on our number
one priority is getting the best people, the best people,
the smartest people, the most driven people. The only reason
we're able to get so far is as the nature
of our team. We have aongous, humongous priority on diversity
of thought, and that goes across research, engineering, product right.
(36:54):
People who spend time looking at understanding awareness of social
media when we see put it into our models in
those elements used to do Alzheimer's and dementiary research. There
are other people on our team who come from traditional
social science, demography, applied demography backgrounds, not you know, just
applied AI in those elements, because this problem can't be
solved in one singular way.
Speaker 1 (37:14):
And I came to your office and it was notable
for the no shoe optional, no shoe policy, which I'm
a big fan of, and also for the cigarettes, which
you guys living the gen z trend.
Speaker 2 (37:27):
We predicted it. We said cigarettes were coming back in
twenty twenty four.
Speaker 3 (37:30):
Yes we did not.
Speaker 1 (37:31):
Everyone's in love with what you're doing. The New York
Times racing right OpEd with the headline this is what
we'll ruin public opinion polling for good quote, pure fictions
are on the brink of being treated as scientific and
political knowledge. If we do not pull back, our understanding
of society might become artificial too. And this was a
(37:51):
direct criticism of an Exios story that had used our
research about trust levels in doctors and nurses, which didn't
have a disclosure that it was based on synthetic respondents.
And The New York Times, these two writers basically did
a broadside against you.
Speaker 2 (38:07):
Why they wrong, I think when you actually read that story,
and by the way, I read that story, I read
many of the other ones that followed, because I think
it's important to see how people feel about technology like this.
I think a lot of people are misunderstanding what makes
our technology fundamentally different. I actually agree with them. By
the way, there are thousands of papers about this idea
(38:28):
of you know as they call it, silicon sampling as
we call it, you know, the anthology approach where you
essentually take a profile that you have like generated, or
a profile of some human being, and you give it
to a language model and you say, be this person,
that is not accurate. Right. We tried that approach, actually,
trust me, like we when we started this company, that
was our base approach, and that's when we learned that
(38:50):
it actually doesn't reflect public opinion, and we very quickly
discovered that we had to be training our own models,
We had to build our own technology. We needed frontier
research of our own in behavior simulation in order to
actually have accuracy that delivers, right. And so I think
when they talk about the issues of that technology, I agree, So.
Speaker 1 (39:09):
What do they misunderstand about what you were doing? In
that case?
Speaker 2 (39:11):
The core thing to look at us is using real
behavioral outcomes and a fundamentally different model architecture and a
fundamentally different program. Right. We're training our own models. Our
models blend LMS and traditional mL with dozens of other techniques, methods,
things that we have built internally, right, And so I
(39:31):
think if you peaked under the hood and actually saw
how this worked, it's probably a lot more similar to
the traditional models that you know, people like that have
trusted for a long time than what people have read
about this idea of silicon sampling in the papers. And
so the real thing is when you're looking at real
behavioral outcomes. We've shown time and time again that we
can be more accurate than traditional public opinion research.
Speaker 3 (39:53):
And that we are. But to be very direct on that, right,
it's like why was that piece written, right? I like
think about it from why was that piece written? Was
it written because someone has like a passion to like,
you know, break apart at Teckning methodology? Absolutely not right.
It's written because someone sees that as a flawed approach
or something that's like, you know, bad, and that's important
to recognize. And I think it's something that you know, Look,
(40:15):
we're in New York for a reason because we're outside
of the echo chamber and.
Speaker 2 (40:18):
We hear this.
Speaker 3 (40:20):
But I think, like from the perspective of how we
look at this, we didn't build this company from a
perspective of like exclusively AI efficiency. And you know, all
these ads that we see in New York that we
think are despicable around like not hiring humans in those
sorts of areas, It's not the approach that I would
ever approach that I would ever think about, right. We
think about this in perspective of We'll go to someone
(40:41):
who's been a marketer for the last two decades, last
two decades, three decades. They might have thirty different areas
and like ten different managers that tell them what they
can and what they cannot put out, and people within
organizations for the last twenty some years have been looking
for an ability, for something else to give them the
confidence to take swings. And that's the way we think
(41:02):
about it. Right, you take something that nobody's ever launches
an now before, how do you go to your CEO
and say, let's put five million dollars on this instead
of you know, paying some random athlete who you know.
Ex Soap company just put on something and it did
well for them, So why shouldn't we do the same thing.
You're going to get turned down, you'll get fired, you
won't have that right. We give people numbers that are
more accurate than anything else on the globe, that we've
(41:23):
proven to be more accurate than survey research, so that
they can take those big swings, so that they can
make those actions, so that they don't have paralysis inside
of businesses waiting for like ten other people to agree
with their decisioning. We give people quantifiable numbers and reasons
why they're right if they are, and reasons why they're
wrong if they're wrong. And that's what our goal is.
Speaker 1 (41:45):
Final question to you both. You met in high school,
you became best friends, more or less love at first sight.
Now you live together, you have a billion dollar startup together.
Speaker 3 (41:55):
Where's this gone?
Speaker 1 (41:57):
Well, do you have any fear that in enormous success,
this room maybe could come between you?
Speaker 2 (42:04):
No, not announced.
Speaker 3 (42:05):
I think too much money and effort has gone to
nurturing unremarkable talent, and we have a very heavy performance culture.
I think being able to be our age and be
able to work every single day with people that we
love and we care about on the thing that we
(42:26):
think is the most important technology in the world is
something that is more unifying that you can ever imagine.
And I could be more grateful to be able to
work alongside.
Speaker 2 (42:37):
You, much more eloquent than me. I love you. He's
a legend.
Speaker 1 (42:40):
Yeah.
Speaker 3 (42:41):
No, I mean look, I think people who say don't
do business with friends need better friends. You know.
Speaker 1 (42:44):
That's what it is. There, You go, mat Cam, thank
you so much.
Speaker 3 (42:49):
Thank you so much for having us. It's great chat.
Speaker 1 (43:09):
For tex staff iMOS Voloshin. This episode was produced by
Eliza Dennis. It was executive produced by me and Julian
Nutter for Kaleidoscope and Katria Norvel for iHeart Podcasts. Jack
Insley mixed this episode and Kyle Murdoch wrote our theme song.
TechStuff News
Hosts And Creators
Oz Woloshyn
Karah Preiss
Popular Podcasts
Stuff You Should Know
If you've ever wanted to know about champagne, satanism, the Stonewall Uprising, chaos theory, LSD, El Nino, true crime and Rosa Parks, then look no further. Josh and Chuck have you covered.
Dateline NBC
Current and classic episodes, featuring compelling true-crime mysteries, powerful documentaries and in-depth investigations. Follow now to get the latest episodes of Dateline NBC completely free, or subscribe to Dateline Premium for ad-free listening and exclusive bonus content: DatelinePremium.com
Betrayal Weekly
Betrayal Weekly is back for a new season. Every Thursday, Betrayal Weekly shares first-hand accounts of broken trust, shocking deceptions, and the trail of destruction they leave behind. Hosted by Andrea Gunning, this weekly ongoing series digs into real-life stories of betrayal and the aftermath. From stories of double lives to dark discoveries, these are cautionary tales and accounts of resilience against all odds. From the producers of the critically acclaimed Betrayal series, Betrayal Weekly drops new episodes every Thursday. If you would like to share your story, you can reach out to the Betrayal Team by emailing them at betrayalpod@gmail.com and follow us on Instagram at @betrayalpod and @glasspodcasts. Please join our Substack for additional exclusive content, curated book recommendations, and community discussions. Sign up FREE by clicking this link Beyond Betrayal Substack. Join our community dedicated to truth, resilience, and healing. Your voice matters! Be a part of our Betrayal journey on Substack.