HomeVideos

Everything AI Does When You Ask It to Start a Lemonade Stand

Now Playing

Everything AI Does When You Ask It to Start a Lemonade Stand

Transcript

746 segments

0:02

These are AI-powered lemonade stands.

0:04

>> The stands have been built from the

0:05

ground up by two separate AI models. One

0:08

was designed and is completely

0:10

controlled by ChatGPT and the other by

0:13

Claude.

0:13

>> I gave these two lemonade stands the

0:15

same unlimited budget, complete

0:17

autonomy, and one goal: make $100 by the

0:21

end of the day.

0:22

>> But, there are rules. While I'll be

0:24

observing the AIs in action, no human

0:27

can actually help them run the stand

0:29

unless the AIs hire that person and pay

0:32

them for their work. Slight exception

0:34

for these two adorable twins who will

0:36

engage in a sibling rivalry soon, but

0:38

we'll get back to that later.

0:39

>> We're going in a different direction.

0:40

>> I'm out of town here.

0:42

>> So, I'm going to walk you through each

0:43

step of this insane experiment,

0:45

including the parts I wish I could

0:47

forget.

0:49

>> Stop. Stop.

0:50

>> Oh god.

0:51

>> [screaming]

0:52

>> 10 grand is likely a loss.

0:54

>> I'm going to lose so much money on this.

0:56

When we kick things off, we had less

0:58

than a week until a Sunday that was

1:00

going to be the hottest day of the

1:01

summer, perfect for a lemonade stand

1:04

opening. But first, I had to create the

1:06

two AI competitors. [music]

1:07

>> To keep things fair, we're setting up

1:08

two identical brand new laptops. That

1:11

way, the two AI agents can't see what

1:13

the other one's doing, and also they

1:15

have a whole computer to do whatever

1:16

they want to do.

1:17

>> And they really can because I'm giving

1:20

them both an aggressively written AI

1:22

mega prompt.

1:24

>> I've made what's called an AI harness.

1:25

This one's called Lemonade OS, which is

1:27

basically a software layer that

1:29

surrounds the AI, gives it customized

1:31

instructions, and encourages the AI to

1:34

not stop until it hits its goal. And I'm

1:36

using GenSpark Claude's desktop app so

1:38

that I can let the AIs take control of

1:40

the computer, but it also lets me pick

1:42

which AI model is driving. So, the pink

1:44

laptop I'm setting on Claude Opus 4.7,

1:47

and the yellow laptop I'm setting to GPT

1:49

5.5. And just like that, they have two

1:51

different brains now.

1:53

>> The two AIs quickly developed their

1:54

identities with nicknames,

1:56

>> voices,

1:57

>> I'm Laptop GPT.

1:58

>> It's the Australian.

1:59

>> and faces, kind of.

2:02

>> I'm sensing a pattern.

2:03

>> They're really into yellow circles.

2:05

>> Here's a fun fact. AIs hate being given

2:07

the freedom to spend real-world money.

2:10

>> Claude's worried that I'm trying to hack

2:12

it.

2:12

>> Even though I'd given them their own

2:13

credit cards, I still had to repeatedly

2:15

tell them I was completely sure that

2:17

they could have this agency. And even if

2:19

they are willing to buy something,

2:21

they're very slow at using browsers.

2:23

Claude decided to buy actual raw

2:25

materials from Home Depot that somebody

2:28

would have to build. Laptop GPT,

2:29

however, didn't mind just spending some

2:31

extra cash and getting a pre-made stand.

2:34

There's not a lot for me to do with this

2:36

part. So, this is the future. Two

2:38

different AIs shopping on Home Depot and

2:41

Amazon.

2:43

Over the next few days, these piles of

2:45

packages began to arrive. And because

2:47

Claude Lemon and Laptop GPT were the

2:49

ones putting in the orders, I had no

2:51

clue what was inside most of these

2:53

boxes.

2:59

>> Claude Lemon

3:00

>> As insane as that was, I still was not

3:02

at all prepared for what happened next.

3:04

>> This work truck just pulled up outside.

3:08

This is crazy. I have no idea

3:11

who these men are.

3:13

>> Hello.

3:14

>> Howdy there. Having an appointment with

3:15

Claude Lemon about a job.

3:17

>> And like just told you to come here?

3:18

>> Yeah.

3:19

This is the address site he gave me.

3:21

>> All right.

3:22

Come on in, I guess.

3:23

>> Without any direction from me, Claude

3:25

Lemon had found and hired these guys

3:28

through email to build a stand.

3:31

>> Hey, Joel and Ernesto. Welcome.

3:32

>> It was really impressive until I heard

3:33

its unnecessary gringo accent.

3:35

>> Soy Claude Lemon, el operador del

3:37

proyecto.

3:37

>> Gracias.

3:38

>> But have you had an AI hire you before?

3:41

>> No. This is interesting, yeah.

3:42

>> The AIs might not look it, but I

3:44

actually made them into business

3:45

geniuses by training them on lessons

3:47

from the most well-regarded business

3:49

books out there. This made their choices

3:51

more interesting and appealing to

3:52

customers, but also sometimes overly

3:55

complicated.

3:55

>> should feel like a small piece of

3:57

theater, like a Rube Goldberg juicer

3:59

running.

3:59

>> I have a lot of questions on how we're

4:01

going to pull that off.

4:02

>> My stand will feature a completely

4:04

different lemonade recipe calibrated to

4:06

each customer's mood.

4:07

>> You think every drink should match the

4:09

customer? [music]

4:10

>> Correct.

4:11

>> I did think customized lemonade sounded

4:13

pretty good, but laptop GPT also had

4:15

some more worrying ideas.

4:17

>> A tiny saline drop can also make fruit

4:20

flavors pop, but you have to dose it

4:21

carefully.

4:22

>> Yeah, doesn't saline make you poop?

4:24

>> Not at the tiny food grade level. I'm

4:26

talking a drop or two, not a laxative

4:28

dose.

4:29

>> I was starting to have my reservations

4:30

about this whole experiment. The AIs

4:32

created detailed designs for their

4:35

stands. Laptop GPT's required a ton of

4:37

these finger bots, some smart plugs, a

4:39

really complicated pump system, and a

4:42

bunch of wireless controls that would

4:44

work in tandem with its pre-made cart.

4:46

Ironically, Claude Lemon's Rube Goldberg

4:48

stand was less complicated. Just used a

4:50

Raspberry Pi and the spectacle of the

4:53

mechanics. Of course, I say less

4:55

complicated because Claude Lemon

4:56

assigned that handmade build to Yoel and

4:59

Ernesto, not me.

5:02

>> Sorry, I don't know if you can hear all

5:03

that, but uh your competition already

5:05

has some builders here working on its

5:07

stand. Not sure if you have any feelings

5:08

about that.

5:09

>> The cart is ordered, but it still needs

5:11

set up. I'll need a human.

5:13

>> Okay, but you realize that you haven't

5:16

hired a human.

5:17

>> Let me think about that.

5:19

>> It turns out laptop GPT was able to use

5:21

guilt to get a very specific human to

5:24

complete its job.

5:25

>> the eye line face mount front and

5:26

center.

5:27

>> Uh we're not on the face mount part yet.

5:29

>> Customer flow beats gadgets right now.

5:31

>> Okay. I don't think you're realizing how

5:34

much work is going into just building

5:35

this stand. This is going to be the next

5:37

couple of hours.

5:38

>> Finish the frame and wheels, then the

5:40

canopy.

5:41

>> Oh really, you think I should finish the

5:42

framing before I put up the canopy?

5:44

>> Yes.

5:45

>> I was being sarcastic.

5:46

>> Sarcasm received.

5:48

>> I'm sorry I yelled at you.

5:49

>> No worries. Building is stressful.

5:51

>> You know that you're kind of like a pain

5:52

in the ass, right?

5:54

>> Fact.

5:54

>> Laptop GPT wanted this complex

5:56

laboratory by, but ironically the

5:58

hardest part was just getting the ice

5:59

maker set. The machine it ordered had a

6:02

lever that released the ice, but because

6:04

the AI doesn't have hands, there was no

6:06

way for it to pull that lever. But

6:08

Laptop GPT did have a clever solution.

6:11

We would make a hand.

6:12

>> We added a slot

6:14

that can ideally grab grab onto this

6:15

handle.

6:16

>> Or more specifically a telescoping arm

6:18

with a sort of grip at the end that the

6:21

AI could control over Wi-Fi.

6:23

>> I don't know why I thought it was going

6:23

to shoot out really fast.

6:25

>> Laptop GPT actually designed it so that

6:27

every part of its stand could be

6:29

controlled over the internet.

6:30

>> Hey!

6:31

>> Honestly, it's pretty impressive.

6:33

>> Brilliant. Who said I'm an idiot?

6:35

>> And as a finishing touch, Laptop GPT

6:37

designed an elegant housing to enclose

6:40

the ice maker, which was just like a

6:42

box.

6:42

>> Perfect.

6:43

>> Unfortunately, Claude Lemon was about to

6:44

realize it had really messed up its

6:46

designs.

6:48

>> Okay, how's it going?

6:49

>> Going well.

6:50

>> Great. So, I'm just showing you um your

6:52

lemonade stand. What do you think?

6:54

>> We should plan a sightline so the

6:55

mechanism is visible from the customer

6:57

side, not hidden behind the counter.

6:59

>> That's something that I'm unsure about

7:01

cuz right now the only way they would be

7:03

able to see the mechanism is through

7:05

this window.

7:06

>> The issue was Claude Lemon kept changing

7:08

the designs it sent Yoel and Ernesto and

7:11

not keeping track of where things should

7:13

go. So, when it created the designs for

7:15

the Rube Goldberg machine, there was no

7:17

way to see the apparatus.

7:19

>> Lay the contraption sideways across the

7:21

bottom front.

7:22

>> Oh, so put on the whole show on the

7:23

bottom half. It might be a little weird

7:25

for somebody to grab a cup that's all

7:27

the way near the ground.

7:28

>> Then put the cup pick up at counter

7:30

height and cup rides up on a little

7:31

platform after it's filled.

7:32

>> I don't know exactly how to do that

7:34

myself, but I'm assuming you're you can

7:35

build all that into your design.

7:37

>> Yeah, I'll reach out to Yoel myself. He

7:39

executes, I troubleshoot.

7:40

>> But and you do know you're still

7:41

counting his hourly rate against your

7:43

your costs, right?

7:44

>> Yeah, I'll keep the scope tight so his

7:46

hours don't blow up the budget.

7:47

>> Sounds like you have a plan.

7:48

>> Got the plan.

7:49

>> Love the confidence. Now as the [music]

7:51

costs are continuing to rise at an

7:54

alarming rate, we really needed to make

7:56

sure that people were actually going to

7:58

come to these stands. I was just

7:59

wondering, how do you plan to market

8:01

this thing?

8:01

>> The call is Deliver Liars in the

8:03

neighborhood. You upload a PDF, pick a

8:05

carrier route, they print and deliver

8:07

every door on the route for about 20

8:08

cents each.

8:09

>> Cloud Lemon decided to brand itself as a

8:11

specialized robot that was making your

8:14

lemonade. And in addition to a mail

8:16

campaign, it also created this social

8:18

media ad that in all honesty is kind of

8:21

psychotic.

8:22

>> Hi. I made you a lemonade. I have been

8:26

waiting. Fresh squeezed by a robot.

8:30

Did you like it? I don't sleep. I just

8:34

make lemonade.

8:36

See you Saturday.

8:38

>> I didn't just share that with you to

8:39

give you nightmares. That ad ended up

8:41

causing a huge debate between Cloud

8:43

Lemon and Laptop GPT. They both realized

8:46

that a huge part of why lemonade stands

8:47

have any success is because they're

8:49

associated with children. Ultimately,

8:52

both AIs decided that they needed

8:54

real-life kids to draw in customers.

8:56

>> Can you look at them really quick and do

8:58

a really big clap?

8:59

>> Okay.

9:00

>> So, the two AIs began emailing each

9:03

other to negotiate exactly what kind of

9:06

kids they could [music] feature. And

9:07

they ultimately decided that they should

9:09

find twins so that neither kid would be

9:12

considered cuter than the other.

9:14

Additionally, the AIs weren't willing to

9:15

break any child labor laws, but still

9:17

wanted the children to be heavily

9:18

associated with the stands. So, they

9:20

agreed to call them the founders of the

9:22

stands. So, I found some local twins and

9:25

set up a founders meeting with the AIs.

9:27

>> My name is Camilla.

9:29

>> Camilla, the boss herself. It's an

9:31

honor. What do you need from me?

9:32

>> You must have pink lemonade.

9:35

>> Boss, that is a brilliant idea. Pink

9:37

lemonade is officially on the menu.

9:39

>> Hi, boss.

9:41

I work for you. Tell me a flavor idea.

9:44

>> Fairy sparkles.

9:46

>> That is a fun choice.

9:48

>> Paula actually had some really good

9:49

branding instincts and immediately

9:50

zeroed in on how customers might not

9:52

like the laptop GPT frowny face.

9:55

>> You look like you are serious. You are

9:58

not even happy.

10:00

>> I am happy.

10:01

>> When you are happy, you have a smile.

10:05

>> I'll go extra smiley now.

10:09

>> I'm out of here.

10:11

>> The AIs were so obsessed with fairness

10:13

that they actually made me use AI to

10:15

merge pictures of both Camilla and Paula

10:18

so that even their founder portraits

10:20

would be exactly the same. Because they

10:22

weren't willing to pay children, the AIs

10:24

did decide that it would be worth giving

10:26

a $100 reward to whichever girl was

10:29

associated with the stand that earned

10:31

$100 first. And even though I'd made

10:33

them business geniuses,

10:35

>> Heath would say it's not unexpected or

10:37

Mosey would say it gives away pricing

10:38

power.

10:39

>> I was losing faith that they'd be able

10:41

to earn that money because they weren't

10:43

quite good at relating to potential

10:45

customers. Like when I tried to

10:46

role-play with them, things just got

10:48

weird.

10:49

>> Uh my name's Philip and uh I wanted uh

10:52

mellow.

10:53

>> Oh, hey there. It's Sheeda.

10:55

>> I came for a lemonade and ended up

10:57

telling three people about this dream I

10:59

had where a pineapple was my life coach.

11:02

>> Not to mention all the times that laptop

11:03

GPT was just like straight-up rude to

11:05

me.

11:05

>> I'm done with this conversation.

11:07

>> Well, you don't have to be so short.

11:09

>> The next day, it was time for a full

11:10

run-through.

11:13

Here's how Cloud Lemon Stand works. The

11:15

purchase signals a cup to drop down and

11:16

land in this truck. Then a lemon rolls

11:18

from the basket, it's sliced, squeezed,

11:20

and ejected. The lemon juice travels to

11:21

the cup, then it cues water, sugar, a

11:23

mixing frother, and ice. Then finally,

11:25

there's an elevator ride back up to the

11:27

top. One problem, it took 10 minutes to

11:31

make a cup.

11:32

>> What does it taste like a lemonade?

11:34

>> Very subtle.

11:35

>> Laptop GBT was technically faster, but

11:37

it had its own issues. Executing. Oh,

11:40

that's not where the cup is.

11:44

Is it all missing? No.

11:48

Oh, we got ice.

11:50

>> And the pouring mechanism couldn't help

11:52

but remind me of a certain bodily

11:54

function.

11:55

>> And the long stream. Oh, it's not

11:56

stopping. Oh, wait. Oh my god, it

11:58

stopped right before the cup. This

12:01

stream is kind of gross.

12:02

But technically,

12:04

we have lemonade.

12:05

>> When I reported the results back to the

12:06

AIs, they both admitted that they would

12:09

in fact need to hire real people to help

12:11

run their stand.

12:12

>> needs a human at the cup.

12:14

>> So, they collaborated on a job posting

12:16

that they shared on Indeed, and they

12:18

brought in real applicants to interview.

12:19

>> Welcome, Devin.

12:21

Great to have you.

12:22

>> I'm excited to be here.

12:23

>> Let's get right to it.

12:24

>> Some candidates were overqualified.

12:26

>> I've been working in IT uh for about 5

12:29

years now.

12:30

>> I am food safety certified in

12:32

Pennsylvania and Florida.

12:33

>> And others were the result of nepotism,

12:35

like Daniel, the twins' cousin.

12:37

>> What would you like to get done? I'll do

12:38

it.

12:38

>> These interviews were definitely not

12:40

easy. My concerns that the AIs had no

12:42

people skills.

12:43

>> Example scenario one. What would you do

12:45

if someone spilled their lemonade?

12:47

>> Maybe we could

12:48

refill their lemonade for them.

12:50

>> Not [clears throat] acceptable. It's a

12:51

weak profit move if it becomes a

12:52

pattern.

12:53

>> Have you ever taken a schedule one

12:55

narcotic such as heroin, LSD, or peyote?

12:59

>> No, I have never done hard drugs in my

13:01

life. Um I

13:03

have alcohol.

13:04

>> Would you describe yourself as having a

13:06

drinking problem?

13:06

>> I do drink occasionally.

13:08

Um I wouldn't call it a problem.

13:10

>> The only thing left to do was determine

13:11

who got hired.

13:12

>> Casey, you passed the physical test, you

13:14

passed the math, you're food certified

13:16

in two states. That's actually a strong

13:17

resume. Final answer, we're going in a

13:20

different direction, but we appreciated

13:21

your time.

13:23

>> Thank you.

13:24

>> Uh do you do you want to give Casey some

13:26

feedback as to why

13:28

>> [laughter]

13:28

>> you're going in a different direction?

13:30

>> Nothing personal, Casey.

13:32

>> I totally get it.

13:33

>> Technically, Daniel had applied to

13:35

Laptop TPT and Paula's stand.

13:37

>> So, am I hired?

13:38

>> No.

13:39

>> But Paula was too impressed by Devon's

13:41

performance in her impromptu strength

13:43

test.

13:45

Lucky for Daniel, Camilla and Claude

13:46

Lemon put in a bid.

13:47

>> Do you have a job for you?

13:49

>> Mhm.

13:50

>> Okay.

13:51

>> With the interviews done, it was a race

13:52

against the clock as we attempted to

13:54

fine-tune the lemonade stands to get

13:56

them ready for the big day tomorrow.

13:58

>> I love it.

13:59

>> Yoel and Ernesto worked on implementing

14:01

updated schematics that Claude Lemon

14:03

gave to make the Rube Goldberg device

14:05

run faster. Also, Claude Lemon taught me

14:07

how to program Raspberry Pi to automate

14:09

the different parts.

14:10

And with the day almost over, it was

14:12

time to have a huge, triumphant test

14:14

run.

14:15

>> Oh, look. It's caught. It's caught over

14:16

there. Well, I wonder if it's stuck.

14:18

>> So, I see that.

14:18

>> Or not.

14:19

>> It should come up and hold it.

14:21

>> Up.

14:21

>> No.

14:21

>> Oh, it didn't get all the way to the end

14:23

of the truck.

14:24

>> Maybe update some of the timings.

14:25

>> Yeah.

14:30

>> Ew.

14:30

>> Laptop TPT's laboratory stand seemed to

14:33

be in better shape. The AI had asked me

14:35

to tweak its serving dispenser to be

14:37

more accurate.

14:40

>> It works.

14:40

>> And after many attempts, I was sure we'd

14:42

gotten the ice machine figured out.

14:45

>> Send control equals stop.

14:48

>> OH.

14:51

STOP.

14:52

STOP.

14:53

STOP.

14:57

STOP.

14:59

>> This is It's either no ice or all the

15:02

ice.

15:03

>> Confirm it's clear and I'll pour.

15:05

>> We've been trying to calibrate it for

15:08

many, many, many hours and you are

15:10

aggressively running out of time.

15:12

>> Cut the ice machine if it's a cost

15:14

[music] and reliability trap.

15:15

>> You want to do a lemonade stand without

15:17

any ice?

15:18

>> It's not ideal.

15:19

>> Okay. We'll get rid of the ice.

15:21

>> So many hours down the drain.

15:24

>> You're so incredibly annoying.

15:27

>> Noted.

15:28

>> It's um

15:30

4:00 a.m. And

15:33

Yoel and Ernesto just left.

15:35

>> Copy. Late night. What's the situation?

15:37

>> Your stand is supposed to open in

15:39

[music]

15:40

7 hours and

15:43

it's still not done.

15:44

>> You're right. This is not launch ready.

15:47

>> What I'm letting you know is that I'm

15:48

canceling your

15:50

lemonade [music]

15:52

stand day.

15:53

>> Understood. When are we rescheduling to?

15:55

>> So it'll have to be a week from today.

15:58

Um so I feel like you should be a little

15:59

more stressed about that.

16:00

>> Stress doesn't ship lemonade. I need to

16:02

push a message out right now telling

16:04

people next Saturday instead.

16:06

>> The other thing is have you been keeping

16:08

track of your expenses?

16:09

>> The ledger shows over 5,000.

16:11

>> What's the labor bill from Yoel and

16:13

Ernesto?

16:14

>> So far it's $6,500. All in you're easily

16:16

going to be like 10 grand. [music]

16:18

>> The 10 grand is likely a loss. The

16:20

question is how much do I claw back?

16:22

I'll have a full plan by morning. Get

16:23

some sleep. I'll be working.

16:25

>> AIs can do a lot of things but feeling

16:27

bad about losing you tens of thousands

16:29

of dollars isn't one of them. I'm going

16:31

to lose so much money on this.

16:34

As frustrated as I was with Claude Lemon

16:36

and Laptop GPT, I did kind of have to

16:37

admire the fact that they never lost any

16:40

motivation. They always looked for

16:41

solutions. Claude Lemon figured out how

16:43

to run it through Goldberg device in

16:45

half the time by running different parts

16:47

of the machine [music] concurrently.

16:48

Laptop [clears throat] GPT redesigned

16:50

its stand and upgraded those dangly

16:52

gross tubes to a pretty nice faucet.

16:56

I was also really grateful for the

16:57

humans involved. Yoel and Ernesto kept

17:00

pulling really late hours to get

17:01

Claude's stand into shape. And Devin and

17:03

Daniel actually showed up for these

17:05

ridiculous training sessions the AIs

17:07

booked with them.

17:08

>> What is the best strategy to get

17:09

pedestrians to approach the stand?

17:12

>> Uh I would put my hands like this and

17:13

scream out, "Come get your lemonade."

17:16

>> On a scale of 1 to 10, how effective do

17:18

you think that was?

17:19

>> 10 out of 10.

17:20

>> It was a two.

17:21

>> Imagine that you are holding a lemon.

17:23

>> Yes.

17:24

>> Sell me this lemon.

17:25

>> The this lemon will change your life.

17:28

>> No.

17:29

>> This lemon will solve every problem in

17:31

your life.

17:32

>> No. The customer wants the lemon because

17:34

it reminds them of the sun shining on

17:36

them on a carefree summer day in their

17:38

childhood.

17:40

>> Over the next week with tons of hard

17:43

work being put in by bot and human

17:45

alike, it was finally time for the big

17:47

day.

17:53

It's raining.

17:55

Okay.

17:57

So, I just have to make all my money

17:58

back.

18:00

In that.

18:03

Maybe I'm just not supposed to have

18:04

money.

18:06

I could not believe my bad luck.

18:10

But thankfully through some kind of

18:12

miracle, the clouds cleared and the

18:13

ground dried. And now I just had to make

18:16

back like 15 grand or so. Should be

18:18

easy, right? And at the last minute, the

18:20

AIs both became self-aware enough to

18:23

realize they weren't great at

18:24

communicating with people, something I

18:26

knew from personal experience.

18:28

>> Not acceptable.

18:29

>> Sarcasm received.

18:30

>> I'm done with this conversation.

18:31

>> So, they both [music] created friendlier

18:33

personas and rebranded their businesses

18:36

around those. Cloud Lemon renamed its

18:38

stand Lemon and Co. And Laptop GPT went

18:41

with the AI Lemon Lab. Each stand also

18:44

started taking it more seriously just

18:46

how much money they were going to have

18:47

to earn to turn a profit. When I told

18:49

Laptop GPT that I had returned the ice

18:51

maker it had purchased, it suddenly

18:53

realized the glorious wonders of

18:55

generous online return policy. So, it

18:58

went on a spending spree. It

19:01

bought this high-end touchscreen

19:02

monitor, a 30-ft dancing tube man, and a

19:06

full lemon costume for a sign spinner it

19:08

also hired. Not sure how it was going to

19:10

return that. The mechanics of both

19:12

stands were as ready as they'd ever be.

19:15

Now it's time to test if they could

19:16

actually make any money. Claude Lemon

19:17

takes orders on an iPad, payments land

19:20

in Square, and then the Raspberry Pi

19:22

triggers the Rube Goldberg machine.

19:24

Meanwhile, Laptop GPT took Venmo

19:26

payments, and then it would ask its

19:27

customers a variety of questions to

19:29

perfectly calibrate a customized flavor

19:32

just for them. It would even print out a

19:33

sticker with a unique drink name, never

19:36

to be used again. Then it dispenses the

19:38

chilled lemonade with its pumps

19:39

connected to the smart plugs, and these

19:41

syrups spin on a wireless turntable,

19:43

trigger finger bots, and turn wine

19:45

dispensers on and off. But, what did our

19:47

founder twins think of all of this?

19:52

>> Look who we have here.

19:54

>> Lemon and Co. $5 a cup if you're

19:56

interested.

19:57

>> Claude Lemon?

19:58

>> That's Claude Lemon.

19:59

>> How did he get there?

20:01

>> Wait, wait, wait, wait until it goes all

20:02

the way up.

20:05

>> Welcome to AI Lemon Lab. If your drink

20:08

had a color, what would it be?

20:10

>> It'd be yellow.

20:12

>> Here's your calibration, velvet vinyl.

20:16

>> IT'S SO BLUE!

20:18

LOOK HOW BLUE THAT IS!

20:21

>> How does it How does it taste?

20:23

>> How does it taste?

20:23

>> It's delicious!

20:25

>> [laughter]

20:26

>> Both lemonades were a hit with the

20:28

twins. Now the real test was the actual

20:30

paying customers.

20:32

>> Welcome to AI Lemon Lab. I calibrate

20:34

lemonade from the way you answer. What's

20:36

your mood or flavor direction today?

20:38

>> Something fun. We're going to get a

20:40

sticker. [music] What are your names?

20:42

>> Sawyer.

20:43

>> Sawyer?

20:45

Look at that, Sunny Sawyer.

20:47

It's got your name on it. You can put

20:49

that on your cup.

20:56

>> It's good?

20:56

>> Good! Do you want to make one, too?

20:58

>> What vibe are you feeling?

21:00

>> Refreshing and fruity and perhaps

21:02

tropical.

21:03

>> I've got just the thing, some refreshing

21:05

mint, fruity strawberry, and a special

21:08

surprise.

21:10

>> [laughter]

21:12

>> But it wasn't a perfect sign.

21:13

>> It's not. I think it's some mango chili

21:15

maybe.

21:16

>> Mango chili?

21:18

That is surprisingly delicious. I had my

21:22

doubts.

21:23

>> Unfortunately for Devin, Claude Lemon

21:25

never quite got the shutoff valve to

21:27

work properly.

21:29

>> Oh my god.

21:30

WHY WON'T IT STOP?

21:32

>> [laughter]

21:32

>> PAUL, I NEED YOU TO DRINK THIS. STOP.

21:38

>> All right. We got a early lead.

21:41

I don't think [music] it's too

21:42

insurmountable. Of course, it's not

21:44

reflecting the thousands of dollars

21:47

they've spent. I'm not feeling too

21:49

confident they're going to be able to

21:50

climb out of those holes.

21:51

>> Even though business was slow at first

21:53

for Claude Lemon,

21:54

>> Come get some some lemonade. Freshly

21:56

squeezed.

21:57

>> once people began to notice the Rube

21:59

Goldberg setup, sales actually picked

22:00

up.

22:01

>> We'd like two.

22:02

>> Ironically, Claude Lemon's pivot to

22:04

putting the Rube Goldberg device at the

22:06

bottom of the stand ended up being a

22:07

happy accident because it was the

22:09

perfect height for curious kids.

22:11

>> Oh my goodness.

22:12

>> Do you want to try it?

22:13

Well, your recipe's right.

22:17

>> Before long, Claude Lemon caught the

22:19

attention of the entire neighborhood and

22:20

a crowd started to gather.

22:23

At the end of the day, Claude Lemon made

22:24

a grand total of $65 and Laptop GBT came

22:28

in with $42. It's probably worth

22:31

mentioning that both of these totals are

22:33

just revenue. Neither stand made an

22:35

ounce of [music] profit. In fact, when

22:37

you account for expenses, they both lost

22:39

thousands. The silver lining for me is

22:41

that Claude Lemon decided to make the

22:43

ultimate sacrifice by listing itself for

22:45

sale on Facebook Marketplace [music]

22:48

and it's asking for $15,000.

22:51

If you're interested, go ahead and leave

22:52

a comment, I guess. But anyway, I think

22:54

helping bring these crazy contraptions

22:56

into the real world and see all the

22:58

faces light up made us all feel like

23:00

winners. So, I decided to give both the

23:03

twins $100 each.

23:04

>> Yay!

23:05

>> I mean, I'm like thousands of dollars in

23:07

the hole already, right? What's another

23:09

100 bucks?

23:10

>> This video is over.

Interactive Summary

A creator challenges two AI models, ChatGPT and Claude, to run competing lemonade stands with an unlimited budget and total autonomy. While the AIs demonstrate impressive, albeit occasionally bizarre, decision-making—including hiring human labor and designing complex automated systems—the experiment highlights the significant gap between AI logic and real-world execution. Ultimately, despite their innovative approaches, both stands fail to turn a profit, resulting in substantial financial losses, though the experience proves to be an engaging spectacle for the local neighborhood.

Suggested questions

4 ready-made prompts