8 BILION DIGITAL CLONES

A few years ago, there was this

experiment out of Stanford to see what

happens if you populate a village with

large language models. Can you simulate

what happens? Can it sort of simulate

what happens in the real world? It was

by Jun Park, Smallville, aka interactive

similacra. Absolutely loved covering

that story. Highly, highly fascinating

paper and now it's back and bigger than

ever. So, Jun Park, the the Stanford

researcher, so he's back with something

that he's calling simile. And the basic

concept is to take this idea of digital

twins and create entire societies. So

let's say you take entire demographics

based on transcripts, transaction logs,

scientific data, whatever. So you create

societies, cities, villages, whatever.

Then you run the simulation. And the

goal here is to be able to ask certain

specific questions about what happens.

What happens if we raise or lower taxes?

What happens if there's a new marketing

campaign for a specific product? How do

people react to this news breaking out?

Really anything you want, but

specifically targeting kind of like the

social interactions. How do certain news

flow through the community? How do

people respond? And this isn't like a

little project. This thing is massive

and has some pretty big heavyweights on

board. So there's a few angel investors

including Andre Karpathy. So he's the

co-founder of OpenAI. He was a former

director of AI at Tesla, a big name in

the industry. You have Faith Lee who is

the the godmother of AI. She's the

co-director of Stanford's human centered

AI institute. We also have Adam

D'Angelo. So you might remember him from

that whole OpenAI fiasco where Sam Alman

got fired. He was one of the board

members. He's still on the board of

OpenAI. He's the CEO of Quora. Then we

have GMO Raj who's the CEO of Versell.

Huge company widely used. Recently I was

using my AI agents OpenCloud to build

something with it. I suggested they use

WordPress. They were like, "No, dude.

We're not using WordPress. We're using

Versell." Is it Versel? I don't even

know how to pronounce it, but they told

me that's the way to go. And I went

along with it. Very happy. You know,

five out of five stars would let my AI

agents use it again. And it's called

Bilski, Adobe's chief strategy officer

and founder of Behance. Okay, but what's

the big deal here? So, the original

Similac paper was kind of mind-blowing

because it was very very early in sort

of like the the chat GPT story. I think

they were using chat GPT or GBT 3.5

Turbo at the time to simulate all the

agents in the village. I believe there

were 25 agents or you know little

characters and they had those cutesy

pixel art animations and each of them

had a backstory personality. They had

jobs or they went to school. They were

married to people. They had kids etc.

And they had a little schedule too. So

they kind of woke up oh 8 a.m. I got to

go to work or I got to go to school etc.

So they literally lived out their entire

lives, did their chores, interact with

their loved ones, etc. And at the time

the question was very simple. What

happens if we give one of them kind of a

inspirational idea? So imagine like

almost like it's some sort of a divine

inspiration where God shows up in your

sleep and says do this and you're like

ah I got to go do this. So they did that

to one of the characters. I think

Isabella was her name. This was years

ago. So, I apologize if I miss some

details, but they told her, "You shall

create a a Valentine's Day party for

everyone in the city." Interestingly, I

just realized that they're launching it

a day before Valentine's Day. And in the

past, that whole thing was about

organizing a Valentine's Day party. But

the point was to see how well this

community would sort of be able to

create this party because again, they

only notified one person. So this one

person had to first and foremost tell

everybody else, convince them to have

it. Then they had to plan the party,

they had to organize the party, etc.,

etc., etc. So on day one or day zero,

whatever you want to call it, it's only

Isabella knew that that she had this

idea. No one else was thinking about it.

So she went, she talked to her closest

friends and she told her, "Hey, let's do

this party." And some of them liked it

and some of them didn't. But the news

kind of percolated throughout their

little village, their their little

society. To fast forward to the end of

it, like they did have the Valentine

party. Not every single person who came

was invited by Sabella, the original

person. So some of it kind of

percolated. So she would tell person A,

person A would tell person B, and then

person B would would show up. So there

was this kind of like ways of

information traveling through people

similar to how it does in the real life

in the real world. And some people

wanted to show up, some people forgot to

show up, some people did not want to

show up and didn't show up. So it very

much simulated how, at least in that

sort of small village, small society,

how things could have went. And one

highly, highly interesting thing about

this paper, looking back at it now, is

they figured out a lot of ways and a lot

of scaffolding to create around these

large language models that were highly

highly effective. They had, for example,

a memory stream. and they would organize

their memories in terms of how relevant

they were in terms of like the recency

etc. So over time instead of just

forgetting everything they kind of had

their most core memories and that that

kind of memory pool was added to over

time so they just didn't forget

everything all of a sudden. The

interesting thing about that is now with

AI agents like OpenClaw I mean we're

seeing a lot of the same ideas being

incorporated to make these agents highly

highly highly capable. So I think a lot

of us expected more to come out of this.

We knew there was going to be more and

more projects, bigger projects coming

out of this because the whole idea

seemed number one useful for predicting

human behavior for simulating it and

just I think it captured a lot of

people's imaginations and we didn't hear

anything for quite some time and finally

today I think they're coming out of

stealth mode and talking about it and

apparently they already had a $100

million seed round. So this thing has

some serious people behind it. It's got

some serious money behind it and

apparently it already has some large

enterprise clients that are behind it

that are on board. CVS Health, a

pharmacy store here in the United States

as well as Telra. They provide mobile

coverage, mobile phones, internet, etc.

And so some of the things that these

companies might want to do is for

example have market research done to see

how a group of let's say a thousand

people might react to a brand new

marketing campaign or product testing UI

testing etc. Jun Park reported that this

has been pretty accurate in terms of

predicting the analyst questions during

simulated earning calls. So the models

correctly predict eight out of 10 of

these earning calls that are asked by

analysts. So, if you think about it,

that could be extremely extremely

useful. You're doing an earning call.

It's live. People are not going to be

asking you tough questions, but you can

run the simulation and it tells you

like, here's most likely what they're

going to ask. So, you can prepare for

it. This can also be used for social

science. How will people respond to new

health scares or economic shocks, for

example. I think they should rerun 2020

and and see like is the whole toilet

paper being gone everywhere. Was that

predictable? Like, was it was it

obvious? because I was aware of that

whole thing happening very very early on

but somehow the whole toilet paper thing

was was never on my bingo card. So why

is this such a big deal? Now if you

think about it this kind of might mark a

transition from big data to big

simulation meaning that in the past

companies the more data they had that

was kind of like the gold that was the

the oil digital oil because you can mine

that data to prepare for certain events

to forecast etc. So having that data

stored somewhere and then kind of

parsing through it and trying to extract

insights was a big big deal. If these

simulations get large enough and

accurate enough that might not be as

important like that that realworld data

stored somewhere might not be as

important. Instead it's going to

transition to running these simulations

to try to extract data like that virtual

simulation data from there. So instead

of going out there and interviewing

10,000 customers to try to see what they

think and then you know like getting

that data, you run a simulation with

100,000 customers and you kind of figure

out what they think based on that

simulation or how they would react to a

new brand new marketing campaign. The

other big shift here might be the fact

that you know a lot of startups there's

this thing that's referred to sometimes

as the innovation tax, right? You might

have heard of it as the pioneers have

arrows in their backs. Basically the

whole idea that when you're the first to

market with something, when you have

some innovative idea, there's a cost to

it because if you fail, you pay the

price because a failure is expensive.

But yet, if you figure something out,

then other people can come in and sort

of copy what you're doing, etc. So, the

first movers, you know, there's a first

mover advantage, but there's also an

innovation tax, and the two kind of

balance each other off. But imagine if

there wasn't an innovation tax. What if

a startup or a brand new company, what

if it could run a thousand simulations

about how something will unfold, a new

product or anything like that and it can

run it in those digital simulation

sandboxes, right? Gather the data, see,

okay, it seems like this is the wing

approach. This is kind of like the more

dangerous approaches, etc. So, what if

the cost of running those thousand

simulations are equivalent to the

expense of actually executing one of

those in the real world? That means you

can try a thousand times for that same

price. That really cuts down on that

innovation tax. You can fail 999

times in that simulation as long as you

find that that one sort of winning

outcome. Now, of course, a lot of this

is going to be a little bit more

statistics based, right? So maybe the

simulation is going to be accurate only

85% of the time or whatever. I say 85%

because in that statement by Jun Sun

Park, the the researcher behind this,

that's what they found in relation to

those uh earnings calls, right? it was

85% of the time it was accurate. And

again, it's only going to keep getting

better and better. But let's assume it's

just 85%. If you can run simulations

that are 85% accurate, even that would

be incredibly beneficial. But here's

kind of what I think is the big thing,

the big deal that might have a outsiz

effect. A lot of our data statistics,

they they kind of they regress to the

mean. They're kind of trying to find the

average person. What does the average

person think about this phone, right?

How do they rate it? They rate it 8 out

of 10 or whatever. But the reality is

across that average, there's a lot of

different people for a lot of different

opinions. Some of them have an outsized

effect on the market. And just looking

at the average doesn't really

necessarily predict how people will

behave. What if 99% of the people are

okay with something and then 1% of the

people that get really triggered by a

particular message that you're putting

out there or some feature that your

product has, they go, you know, declare

a war and make lots of noise about it.

The rest of the people that are, you

know, they're okay. They're look warm

with the product. They like it, but they

don't love it, but they don't hate it.

But they do see this one group of people

that are damaged by the product or

triggered. What whatever happens, right?

we're may be able to simulate those

weird idiosyncratic things that that

kind of happen out there in the real

world cuz again, you know, if you're

looking at averages, on average

everybody's fine, right? Everybody sort

of likes it. No problems. But that

minority of people might have an

outsized effect. And this would allow us

to capture those weird trends and and

weird eentricities that might happen

that that we can see in a simulation

that we can't if we're running some

statistical analysis on a bunch of data.

And of course, this could be huge for

stock market analysis. You might

simulate CEOs and market traders and a

lot of people involved like leadership

teams and companies to see how they

react to a market crash or a

competitor's moves and see if you can

predict some things before they occur.

Certainly, there were conversations in

my life that I'd love to be able to run

through something like this to to

hopefully get a glimpse into how they

how might they go. like if we have to

give this person some tough news, what

are some possible sort of outcomes of

this happening? What might they say? How

may I react? And would that trigger

anything? Anyways, very excited to see

more coming out of this company. Similey

and June, of course, I'm so happy that

he's having success. Obviously,

incredibly intelligent person, one of

the first really mind-blowing papers

utilizing Chad GPT. Absolutely loved

everything that he did there and

expecting to see a lot more from this.

They're on Twitter x simile company at

simile_ai.

So I think this company is going to

answer a lot of questions that people

have and it might even raise a new

question so to speak and that is if

running simulations to get data if it's

so incredible and I'm sure the data is

more valuable the more accurate a

simulation is then what's the chance

that this is a simulation and we're all

just here stuck reacting to some

marketing campaign. Something to think

about. If you made this far, thank you

so much for watching. My name is Wes

Roth. I'll see you in the