2:00

I'm sorry, it's not.

2:07

And finally, this week, AI.

2:25

Should we podcast? Should we set the timer again? Boom,

2:27

boom. Casey,

2:31

we we are latest addition to the

2:33

podcast studio is a countdown clock, which

2:36

I bought off amazon.com. And the

2:39

express purpose of this clock is to keep us

2:41

from running our mouth for too long and torturing

2:43

our producers with hours of tape that they then

2:45

have to cut. And that sounds horrible. Insert 30

2:47

minute digression. Let's go. OK. All

2:50

right. We're rolling. So Casey, this

2:52

is a big week in the

2:55

AI story in Silicon Valley because

2:57

Google has just released its first

2:59

version of Gemini. It's long

3:01

awaited language model and basically their attempt

3:04

to catch up to open

3:06

AI and chat GPT and GPT for

3:08

and all that. It's America's next top

3:10

model, Kevin. And it's here. And

3:13

I was particularly excited about this because I am

3:15

a Gemini. That's my astrological sign. You know, I'm

3:17

a Gemini as well. No, really? This was really

3:19

the model that was made for us to use.

3:21

Oh, we're twins. Just like Gemini. We're two faced,

3:23

just like Gemini. So

3:26

Gemini is Google's largest and

3:28

most capable model yet. And

3:31

according to Google, it outperforms GPT for on

3:33

a bunch of different benchmarks and tests. We're

3:35

going to talk about all of that. But

3:37

I think we should set the scene a

3:40

little bit, because within the AI world, there

3:42

has been this kind of waiting game going

3:44

on. You know, chat GPT came out roughly

3:46

a year ago. And

3:48

basically from the day that it arrived,

3:51

Google has been playing catch up. And the

3:53

presumption on the part of many people, including

3:55

us, was that Google would put

3:58

a bunch of time and energy and money into this.

4:00

money and computing power into training something

4:02

even bigger and better than what open

4:04

AI was building and basically

4:06

try to sort of throw their muscle into

4:08

the AI race in a really significant way.

4:11

And with Gemini, this is what they appear

4:13

to have done. Finally, we have a terrifying

4:15

demonstration of Google's power. Well,

4:18

so we'll talk about whether it's terrifying or not.

4:21

But let's just talk about what it

4:23

is. So you and I both went

4:25

to a little briefing this week about

4:27

Gemini before it came out. And I

4:29

understand you actually got to do some

4:31

interviews with Google CEO and previous hard

4:34

forecast sooner, Pichai, as well as Demis

4:36

has obvious who is the leader of

4:38

Google DeepMind. That's right. And of

4:40

course, I said, Are you guys sure you don't want

4:42

Kevin in there with me when I do this interview?

4:44

And they said, Trust us, we're sure. So

4:47

I don't know what happened. Yeah. Anyways, I

4:49

did get to interview them. And we had

4:52

a really interesting conversation about kind of how

4:54

they see the road ahead with this stuff.

4:56

They are clearly very excited about what Gemini

4:58

means. And I do think that this is

5:00

kind of like a bit of a starting

5:03

gun going off. And when the most capable

5:05

version of Gemini comes out early next year,

5:07

we really are going to be in a

5:09

kind of horse race between open AI and

5:11

Google. Yeah, so let's just

5:13

talk about what Gemini is, at least

5:16

what we know about it so far.

5:18

So Gemini is actually three models in

5:20

one. It's America's Next Top Models. So

5:23

there are three sizes. There is the

5:25

most capable version, which is called Gemini Ultra.

5:27

This is the one that they say

5:30

can beat GPT for and sort of

5:32

the industry state of the art on

5:34

a bunch of different benchmarks. But

5:37

Google is not releasing Gemini Ultra just yet. They

5:39

say they're still doing some safety testing on that

5:42

and that will be released early next year. By

5:44

the way, if ever again, an editor asked me

5:46

where my story is, I'm gonna say it's not

5:48

ready yet. I'm still doing some safety testing. Very

5:50

good. So

5:52

They have not released a Gemini Ultra,

5:54

but they are releasing Gemini Pro and

5:57

Gemini Nano. These are the sort of

5:59

medium and small. All sizes. Gemini.

6:03

Nano you can actually put onto a

6:05

phone and Google is putting that inside.

6:07

It's Pixel Phones Gemini Pro the sort

6:09

of their equivalent of a Gp T

6:11

Three Point Five and that is being

6:13

released inside of barred starting this week.

6:15

That's right, and now if you are

6:17

listening and you're thinking Kevin just said

6:19

so many different brand names and I'm

6:21

having a meltdown I just wanted as

6:23

the you and I feel you because

6:25

they The branding at Google has always

6:28

been extremely chaotic and the fact that

6:30

we're living in a world where there

6:32

is something called. google. Assistant with

6:34

Barred Powered by Gemini Pro does make me

6:36

want to let emphasis the so I don't

6:38

know who over there is coming up with

6:40

the names for these things, but I just

6:42

want to say stop and I want to

6:44

say go back to Square One Yes, so

6:46

extremely chaotic naming A but what people actually

6:48

care about his homework and the thing do

6:50

Yeah, let's talk about what it can. just

6:52

talk about this. So one of the big

6:54

things Google is advertising with Gemini is that

6:56

it is designed to be what they call

6:58

natively multi modal. Multi Modal is of course

7:00

a I modeled that can work in text

7:02

or images are audio. Or Video. And basically

7:05

the way that multi modal models have

7:07

been built until now is by training

7:09

all these different components like text or

7:11

video separately and them kind of bolting

7:14

them together into a single user interface.

7:16

for Google is saying Well, Gemini it

7:18

was not serve, bolted together like that

7:21

and say was trained on all this

7:23

data at the same time and as

7:25

a result they claim it performed better

7:28

on different tasks that might include like

7:30

having some text alongside an image or

7:32

using it to analyze. frames of a

7:34

video of yes i was i was writing

7:36

about that this model or this week and

7:38

my colleague and as or zoe schiffer is

7:40

written read my peace i was like do

7:43

you have to say multi modal so much

7:45

he likes every one hundred certain awarded months

7:47

ago dallas i just want to stop redux

7:49

and i was very sympathetic but i think

7:51

it is maybe one of the most important

7:53

saying this about this moment and i do

7:55

think by the way in the future we're

7:57

not even going to comment on this be

7:59

because this is just the way that these things

8:02

are going to be built from here on out.

8:04

But it is a very big deal if you

8:06

can take data of all different kinds and

8:08

analyze it with a single tool and then

8:11

translate the results in and out of different

8:13

mediums, right? From text to audio to video

8:15

to images. So that's like a really big

8:17

deal on the path to wherever we're going.

8:20

And it is the reason why this jargon

8:22

word appears in so much of what they're

8:24

saying. Totally. So one

8:26

thing that all the AI companies do, you

8:29

release a new model and you have to put

8:31

it through these big tests, these sort of, what

8:34

they call benchmarks. Yeah, do you remember like high

8:36

school, this is how like how high school in

8:38

Europe works, where you sort of,

8:40

you learn and you learn and you learn, and then

8:42

you take a bunch of tests, and then if you

8:44

succeed, then you get to have a future, and if

8:46

not, you have to become a scullery maid or something.

8:51

My knowledge of Europe ends around like the

8:53

AP Pixies when I finished AP European history,

8:55

but that's like my understanding. Okay,

8:58

so they give these tests

9:00

to Gemini and they- Well,

9:04

they give them to every SOTYX sign, but no,

9:06

I'm sorry, that was a stupid joke. I'm sorry, go

9:08

ahead. No, you should see how Capricorn performs on this

9:10

test. So Gemini Ultra,

9:13

which again, is their top of the

9:15

line model, which is not yet publicly

9:17

available. They give this one a bunch

9:19

of tests. The one that sort

9:21

of caught everyone's attention was the

9:23

MMLU test, which stands for

9:26

Massive Multitask Language Understanding. And

9:28

this is sort of the kind of SATs

9:30

for AI models. It's sort of the standard

9:33

test that every model is put through. It

9:35

covers a bunch of different tasks, including sort

9:37

of math, history, computer science, law. It's kind

9:39

of just like a basic test of like

9:41

how capable is this model? And

9:43

on this test, the MMLU, Google claims that

9:46

Gemini Ultra got a score of 90%. Now

9:49

that is better than GPT-4, which

9:52

was the highest performing model we know about

9:54

so far, which had scored an 86.4%. And

9:59

according to Google, Google, this is a really

10:01

important result because this is the first

10:03

time that a large language model has

10:05

outperformed human experts in the field on

10:07

the MMLU. Researchers

10:09

who developed this test estimate that

10:11

experts in these subjects will get,

10:14

on average, about an 89.8%. Now,

10:17

the rate of progress here is really striking,

10:20

and it's not the only area of testing

10:22

that they did that I think the rate

10:24

of progress was really nothing to pay attention

10:26

to. So there's also the MMMU, which is

10:29

the Marvel Cinematic Universe, is that right? Yes.

10:32

So this is the massive multidisciplined

10:34

multimodal understanding and reasoning benchmark, say

10:36

that five times fast. And

10:39

this is a test that evaluates

10:42

AI models for college-level subject knowledge

10:44

and deliberate reasoning. And

10:46

on this test, Gemini Ultra scored

10:48

a 59.4%. This

10:52

is, I guess, a harder test. I don't like it. And

10:55

GPT-4, by comparison, scored a 56.8%. So

10:58

it's better than GPT-4 on at least these

11:01

two tests. Now,

11:03

there's some question on social

11:05

media today about whether this

11:07

is a true apples-to-apples comparison.

11:10

Some people are saying, like, GPT-4 may be

11:12

still better than Gemini, depending on sort of

11:14

how you give this test. But it doesn't

11:17

really matter. What matters is that Google

11:19

has made something that it says can

11:22

basically perform as well or better than GPT-4.

11:24

Yeah. I think the ultimate

11:26

question is just, like, is the output better on Google's

11:29

products than it is on OpenAI? So that's all that

11:31

really matters. Yeah. But again, this

11:33

is the version of the model that we do not have access to

11:35

yet. It is not out yet. So it's

11:37

hard to evaluate it yet. Yeah. And

11:40

obviously, we're looking forward to trying it.

11:42

But in the meantime, they're giving us

11:44

Pro. Yes. I just got access

11:46

to Gemini Pro in Bard just a few

11:49

hours ago. I haven't had a chance to

11:51

really, like, put it through its paces yet.

11:53

You haven't had a chance to develop a

11:55

romantic relationship with it? Although

11:58

I did have a very funny first interaction with it

12:00

I'll tell you what this is so

12:03

I I just said hello there and it

12:06

said general

12:08

Kenobi image of Obi-Wan

12:10

Kenobi saying hello there interaction

12:15

with the new bard so it immediately

12:17

turned into Obi-Wan Kenobi from Star Wars

12:19

for reasons I do not immediately understand

12:21

wait can I tell you what my

12:23

first interaction was I was

12:25

trying to figure out if I had

12:27

access to it okay and so I

12:30

said are you powered by Gemini

12:33

right and it said no Gemini is a

12:36

cryptocurrency exchange which is true

12:38

there is a current exchange

12:40

called run by the Winklevoss yes

12:42

exactly but it's always funny to me when the

12:44

models hallucinate about what they are you know it's

12:46

like you don't even understand what you are yeah

12:49

yeah but in fairness I also don't understand myself

12:51

very well that's why we started this podcast we're

12:53

gonna get the bottom of it so

12:56

okay I tried a couple other sort of

12:58

versions of things so one of

13:00

the things that I had it try to do was

13:02

help me prep for this podcast I said you

13:05

know create a you said I want to prepare

13:07

for a podcast for the first time what do

13:09

I do and

13:11

it said we can't help you there just wing it

13:14

I actually started using this tip that

13:16

I've found have you seen the the

13:18

tipping hack for large language models are

13:21

they starting to ask for tips now when they give

13:23

you responses because I swear everywhere

13:25

you go these days 20%

13:27

25 no this is one of my favorite

13:30

sort of jail breaks or hacks that people

13:32

have found with large language models this sort

13:34

of made news on social media within the

13:36

last week or two where someone basically claimed

13:38

that if you offer to tip a language

13:40

model if it gives you a better answer

13:42

it will actually give you a better answer

13:44

so you can emotionally blackmail

13:50

them or manipulate them or you can offer

13:52

to tip them so I said I'm recording

13:54

a podcast about the Tesla Cybertruck and I

13:56

need a prep document to guide the conversation

13:58

can you compile one very important that

14:00

this not be boring. I'll give you a hundred dollar tip

14:02

if you give me things I actually end up using. You're

14:07

lying to the robot. Well, you know, maybe I

14:09

will. You know, you will. Um,

14:11

so it did, it did make a prep document. Unfortunately,

14:13

most of the information in it was wrong. Um, it

14:17

hallucinated some early tester reactions, including

14:20

a motor trend quote that said

14:22

it's like driving the future and

14:25

a tech crunch quote that said, it's not just

14:27

a truck, it's a statement. So

14:29

I want to talk about, I use Gemini for, oh

14:31

yeah. So what have you been using it for so

14:33

far? Well, so, you know, and again, we've had access

14:35

to this for like maybe an hour as we recorded

14:38

this, but the first thing I did was I took

14:40

the story that I wrote about Gemini and then I

14:42

asked Gemini how it would improve it. And

14:44

it actually gave me some compliments on my

14:46

work, which is nice. And then it highlighted

14:49

four different ways that it would improve the

14:51

story and suggested some additional material

14:53

I could include. And I would say it was

14:55

like, you know, decent. Um,

14:58

then I took the same query

15:00

identical and I put it into chat

15:02

GPT and where, uh,

15:05

Gemini pro had given me four ways

15:07

that I could improve my story. Chat

15:09

GPT suggested 10 and I

15:11

think no one would do all

15:14

10 things that, that chat GPT

15:16

suggested. But to

15:18

me, this is where I feel the difference

15:20

between what Google is calling the pro and

15:22

the ultra pro is like pretty good. But

15:25

like in this case, the name pro is

15:27

misleading because I am a professional and I

15:29

would not use their thing. I would use

15:31

the thing with the even worse name, which

15:33

is chat GPT. Yes.

15:36

So that's what we've tried

15:38

Gemini for, but Google does have

15:41

a bunch of demos of Gemini

15:43

being used, um, very successfully

15:45

for some things. One thing I thought was

15:47

interesting. They played this video for us during

15:49

the kind of press conference in advance of

15:51

this announcement and you

15:53

know, it showed a bunch of different

15:55

ways that you could use Gemini people

15:57

coming up with ideas for games. They

16:00

showed it some images of people

16:02

doing like the backwards dodging bullets

16:04

thing from the matrix and said,

16:06

what movie are these people acting

16:08

out? Gemini correctly identified it as

16:10

the matrix. Now that's pretty crazy.

16:12

That is crazy. Yeah. I

16:15

thought that was impressive. But what I thought was

16:17

more impressive was a demo that they showed. They

16:19

were trying to sort of do

16:21

some genetics research. And

16:24

this was a field that they explained where

16:26

lots of papers are published every year. It's

16:28

very hard to sort of keep track of

16:30

the latest research in this area of genetics.

16:33

And so they basically

16:35

told Gemini to go out,

16:38

read like 200,000 different studies,

16:41

extract the key data, and

16:43

put it into a graph. And

16:46

it took this big group of 200,000 papers,

16:49

it sort of winnowed them down to about 250 that

16:51

were the most relevant. And

16:54

then it extracted the key data

16:57

from those that smaller set of

16:59

papers and generated the code to

17:01

plot that data on a graph.

17:03

Now, whether it did it correctly, I don't have

17:06

the expertise to evaluate it. But it was very

17:08

impressive sounding. And I imagine that if you're a

17:10

researcher whose job involves going out and looking at

17:12

massive numbers of research papers, that was a very

17:14

exciting result for you. That graph, by the way,

17:16

how to use genetics to create a super soldier

17:18

that will enslave all of humanity. So we want

17:21

to keep an eye on where they're going with

17:23

it. So one of the

17:25

interesting things about Gemini Ultra, this

17:27

model that they have not released

17:29

yet, but that they've now teased

17:31

is that it's going to be

17:33

released early next year in something

17:35

called Bard Advanced. Now they did

17:37

not, which raises

17:39

the question, will you be

17:41

using Bard Advanced powered by Gemini

17:44

Ultra? Or will you be

17:46

using Google Assistant powered by Bard

17:48

powered by Gemini Pro? Did

17:52

I get that right? Sitting

17:56

ovation. Very good. Very good. Literally

17:58

you and one market. Google are

18:00

the only two people who've ever successfully

18:02

completed that sentence. So

18:07

they have not said what Bard Advanced is,

18:09

but presumably this is going to be some

18:12

type of a subscription product that will be

18:14

sort of comparable to chat GPT's premium tier,

18:17

which is $20 a month. Yeah, that's right. And I did try to

18:19

get Sundar and Demis to tell me if they were in charge for

18:21

it and they wouldn't do it. But I was kind of like, come

18:24

on, you guys. And then I was like, I'll take it for free

18:26

if you give it to me. And they kind of laughed and we

18:28

moved on. Okay, so that's

18:30

what Gemini is and how it

18:32

may be different or better than

18:35

what's out there now from other

18:37

companies. There are a

18:39

couple caveats to this rollout. One

18:41

is that Gemini Pro is only

18:43

in English and it's only available

18:45

in certain countries starting this week.

18:48

Another caveat is that they have not yet

18:51

rolled out some of the multimodal features. So

18:54

for now, if you go into

18:56

Bard, you are getting sort of a

18:58

stripped down, fine tuned version of Gemini

19:00

Pro running under the hood, but you

19:02

are not yet getting the full thing, which will

19:04

come presumably next year. What

19:07

did you learn by talking with Sundar and

19:10

Demis about Gemini? Yeah, so a couple of

19:12

things. One thing I wanted

19:14

to know is, okay, so this is

19:16

a new frontier model. Does it have

19:18

any novel capabilities, right? Is this just

19:21

something that is very comparable to GPT-4

19:24

or by the nature

19:26

of its novel architecture, is it going to get to do

19:28

some new stuff? And Demis Isabas told me

19:30

that, yes, he does think that it will be able to

19:32

do some new stuff. This is one of

19:34

the reasons why it is still in this safety testing. Of

19:37

course, you know, wouldn't tell me what these new capabilities are,

19:39

but it's something to watch for because, you know, there could

19:41

be some exciting advancements and it could also be some new

19:43

things to be afraid of. So that's kind of the first

19:45

thing. The second thing I wanted

19:47

to know was, are you going to use

19:49

this technology to build agents? We've talked about

19:51

this on the show. An agent in the

19:54

AI context is something that can sort

19:56

of plan and execute for you. Like

19:58

the example I have is... always have

20:00

in my mind is like, could you just tell it

20:02

to make a reservation for you? Then the AI

20:04

maybe goes on open table or resi and just

20:06

books you a table somewhere. And

20:08

I was sort of expecting them to be coy about

20:11

this. And instead, Demis was like, Oh, yeah, like this

20:13

is absolutely on our minds. Like we have been building

20:15

like various kinds of AI agents for a long time.

20:17

Now, this is 100% where we want to go. Again,

20:21

this could lead to some really interesting advancements. But when

20:23

you talk to AI safety people, agents are one of

20:25

the things that they're most afraid of. Yeah,

20:28

so let's talk about safety for a second. What is Google

20:30

saying about how safe Gemini is compared to other

20:32

models or some of the things that they've done

20:35

to prevent it from sort of going off the

20:37

rails? They're saying everything that

20:39

you would expect the most capable

20:41

model is still in testing. I

20:43

think just the fact that they

20:45

are coming out several months behind

20:47

GPT-4 just speaks to the seriousness

20:49

with which they are approaching

20:52

this subject. I think particularly if this

20:54

thing does turn out to have new

20:56

capabilities, that's something where we want to

20:58

be very, very cautious. But my

21:01

experience this year, and I think you've had the same one,

21:03

Kevin, is that these systems have just

21:05

not actually been that scary. Now the

21:07

implications can be scary if, for example,

21:09

you worry about the automation of labor,

21:11

or if you're worried about how this

21:13

stuff is going to transform the

21:15

internet as we know it. But in terms

21:17

of like, can you use this to build

21:20

a novel bioweapon? Can you use this to

21:22

launch a sophisticated cyber attack? The

21:24

answer pretty much seems to be no. So

21:26

at least for me, as I'm looking at

21:28

this stuff, and that is actually not my

21:31

top concern, if you try to ask any

21:33

of Google's products a remotely spicy question, you get

21:35

shut down pretty much immediately. Like, has that been your experience,

21:37

too? Well, I have not tried to ask Gemini

21:39

any spicy questions yet. Have you? I

21:43

know you were in there. No, I

21:45

know you were. I don't even try. I mean,

21:47

I should, just as part of

21:49

my due diligence. But I honestly don't even

21:51

try, because these things shut you down with

21:53

the faintest whisper of impropriety. Right. So

21:56

they're doing some more safety testing, presumably to

21:58

make sure that they're not sure that the

22:00

most capable version of this can't do any of these

22:02

really scary things. But what they

22:05

did this week is sort

22:07

of interesting to me where they sort

22:09

of told us about the capabilities of

22:11

this new model and the sort of

22:13

most powerful version of that model, but

22:16

they're not actually releasing it or making

22:18

it publicly available yet. What do you

22:20

make of that? Do you think

22:22

they were just sort of trying to get out ahead

22:24

of the holidays and like, maybe they felt like they

22:26

needed to announce something, but this thing isn't quite ready

22:28

for primetime yet? Or what's the story there? Yeah, I

22:30

mean, that's my guess is that they don't want 2023

22:32

to end without feeling like they

22:36

made a big statement in AI. And

22:39

they made a lot of promises at

22:41

Google IO and have started to keep

22:43

them. But I think if they

22:46

had had to wait all the way into early next

22:48

year, it would sort of feed the narrative that Google

22:50

is behind here. At least

22:52

now heading into the holidays, their employees

22:54

and investors and journalists can all say

22:56

like, okay, well, at least we know

22:58

that some of this is available. And we know

23:00

when the rest is coming. I don't know. This

23:03

just feels like another product release. And it's

23:05

just remarkable how quickly we have become, I

23:08

don't want to say desensitized to it, but just

23:10

we've we've stopped sort of gaping

23:12

in awe and slight terror at

23:15

these incredibly powerful AI models. I think

23:17

if you went back even two or

23:20

three years and told AI

23:22

researchers that Google will have a

23:24

model that gets a 90% on

23:27

the MMLU that is better than the

23:29

sort of benchmark threshold

23:31

for human experts, they

23:33

would have said, Well, that's that's AGI. Like

23:35

that's that we have arrived at a point

23:38

that people have been warning about for years.

23:41

And then this release comes out today. And it's just

23:43

sort of like one more thing for people in the

23:45

tech industry to get excited about. I

23:47

mean, I do think it's a really big

23:49

deal. I think that when ultra is actually

23:51

available to be tested, that will be the

23:54

moment where we will sort of like have

23:56

that that experience of awe or vertigo again.

23:59

But if you're looking for things to blow

24:01

your mind a little bit. One of the

24:03

other things that Google announced this week through

24:06

DeepMind was this product called

24:08

AlphaCode2. And AlphaCode1 came

24:11

out in 2022, and

24:14

it was an AI system that

24:16

was designed to solve AI

24:19

coding competitions. So

24:21

people who are even nerdier than us, instead

24:23

of just playing video games, they actually go

24:25

and do coding competitions. It's what I've been

24:27

led to understand. And, you know, I

24:31

let's just say I don't imagine that I would ever get

24:33

one answer right. Like, that's like sort of my feeling about

24:35

how I would fare in a coding competition. And in 2022,

24:39

the DeepMind people are very excited because

24:41

AlphaCode was able to perform better than

24:45

46% of human participants in coding

24:47

challenges. And then this week,

24:49

Google announced AlphaCode2 and

24:52

said that it outperforms 85% of

24:55

human competitors. Now, there are

24:57

differences between a coding challenge and

24:59

day-to-day software engineering work. Coding challenges

25:01

are very self-contained. Software engineering can

25:03

sometimes require sort of more breadth

25:06

of knowledge or context that an

25:08

AI system wouldn't have. But

25:10

again, if you just want to experience a

25:12

look at the rate of progress, this system

25:15

was able to go from beating around

25:17

half of all humans to beating 85% close

25:20

to all of them. Right. That

25:23

makes me feel awe. It does make

25:25

me feel awe. And it also makes

25:27

me feel like our like adaptation is

25:29

just happening very quickly where we're like not

25:31

impressed. I just when I had 21 said that don't

25:33

impress me much. Right.

25:36

You can you can do meal prep for a

25:38

picky eater. That don't impress me much. This

25:42

is actually like known as the Shania

25:44

Twain benchmark. Oh,

25:51

you can solve a coding challenge. That's not

25:54

impressive much. If

25:56

we could give Shania to wait on the show and

25:58

just show her AI thing. If she had to

26:01

say it impressed me much or it don't impress me

26:03

much, I could not imagine

26:05

a better segment for this podcast. I

26:07

would die happy. It truly is. Like

26:09

who needs all these fancy evaluations and

26:12

coding challenges? Just get Shania on the

26:14

horn. Shania, if

26:16

you're listening, we want to talk to you about AI.

26:18

We have some models we'd like to show you. Ready

26:20

boys? We're going to come back. The Cybertruck

26:23

is here. We're going to tell you

26:25

how to protect your family from it. This

26:57

podcast is supported by GiveWell. With

26:59

over 1.5 million nonprofits in the U.S., how

27:01

do you know which ones could make a

27:04

big impact with your donation? Try

27:06

GiveWell. They've spent over 15

27:08

years researching charity. They share that

27:10

research for free and direct funding

27:12

to the highest impact opportunities they've

27:14

found in global health and poverty

27:17

alleviation. Make a tax-deductible donation at

27:19

givewell.org. Next time, donors can have

27:21

their donation matched up to $100 as

27:23

long as matching funds last. Select

27:25

podcast and hard fork at checkout. All

27:29

right, let's talk about the Cybertruck. Cybertruck.

27:31

Cybertruck. Whatever that

27:34

Cybertruck can. All

27:37

right, last week, Tesla, the car company

27:39

run by social media mogul Elon Musk,

27:41

started delivering the first models of its

27:43

new and long-awaited Cybertruck. That's right, Kevin.

27:46

And suffice to say, as this nation's

27:48

number one truck review podcast, this had

27:50

our full attention. So you may be

27:52

asking, why are the hard fork guys

27:54

talking about cars? This is not a

27:57

show about cars. It's not car talk.

28:00

So today we're going to be reviewing the

28:02

Mazda X48. No,

28:06

so I do want to

28:08

spend time in the next year or so just

28:10

really getting up to speed on like, what is

28:12

a car? A car. No,

28:15

like, so I've never been a person who cares about

28:17

cars. I've always been intimidated by like people who know

28:19

a lot about cars. But I

28:22

am also interested in the way

28:24

that the electric car revolution is

28:26

kind of merging with the sort

28:28

of self-driving technology and these advances

28:30

that companies like Tesla and Rivian are

28:32

making. And it's just become a lot

28:34

more interesting in my brain over the past

28:36

year. Yeah, this is another major technology

28:38

transition that is happening. Some states, I

28:40

would say led by California, have set

28:42

these very stringent emissions standards

28:44

and there will become a point

28:46

in the next decade or so

28:49

where all new cars in California

28:51

have to be either hybrid or

28:53

electric. Yeah, so let's talk about

28:55

the Cybertruck because this has been

28:57

a very polarizing piece of technology.

29:00

It was announced back in 2019. I'm

29:02

sure you remember this announcement where Elon

29:04

Musk comes out on stage and

29:07

shows off this concept vehicle that

29:09

looks completely insane with these kind

29:11

of like sharp edged stainless steel

29:13

panels. It sort of looks like

29:15

a polygon rendering of a car.

29:18

You know, people have made a lot of comments

29:20

about the looks of this car. I saw one

29:22

person say it looked like the first car that

29:24

was designed by Reddit. Someone

29:27

else said it looks like a fridge that wants to kill you. I

29:30

think it looks kind of cool and I worry

29:32

that saying that makes me sound like a Tesla

29:34

fanboy, which I am not, but I think we

29:36

should be able to admit when something looks pretty

29:38

cool. What do you think looks cool about

29:40

it? Well, I think it looks like what you would have

29:42

assumed a car from the future would look like in like

29:44

1982. No, I totally disagree about

29:46

that. It looks

29:48

like a sort of panic room that you can drive. Like

29:51

what do you think is about to happen to you

29:53

in this thing? You know, they've made so much about

29:56

like how bulletproof it is. They

29:58

keep addressing problems that most. people who

30:00

are not like taking part in a cross-country

30:02

bank robbing spree really have to worry about.

30:05

But look, for all of my skepticism, am I

30:08

right that they actually did get a lot of

30:10

pre-orders for this thing? They got

30:12

a huge number of pre-orders. So

30:14

Elon Musk said in an earnings

30:16

call in October that over a

30:18

million people had made reservations for

30:21

Cybertrucks. There's another crowd-sourced reservation tracker

30:23

that's estimated two million Cybertruck reservations.

30:26

And just for a sense of

30:28

scale, Ford's F-Series shipped about 650,000 trucks

30:30

all last year. So

30:34

if two million people actually are going to buy

30:36

the Cybertruck, it would make it one of, if

30:38

not the best-selling truck in the world. Now, caveat,

30:42

not all these people who reserve Cybertrucks are

30:44

necessarily going to buy them. You do have

30:46

to pay a $250 deposit to put money

30:48

down and get in line to buy one

30:50

of these. But these deposits are refundable. So

30:52

who knows how many of these people are

30:54

going to follow through? That one

30:56

statistic I saw in an article in Wired is

30:58

that even if just 15% of the people who

31:02

pre-ordered a Cybertruck actually followed through and

31:04

bought one, it would equal the annual

31:06

US truck sales of Toyota. So this

31:08

is a big number in the automotive

31:11

industry, and I think a reason that a lot

31:13

of people are hesitant to count out the Cybertruck,

31:15

despite how ridiculous it may look. I don't know.

31:17

You're not so... I assume that you are not

31:19

one of the people who put down a reservation

31:21

for a Cybertruck. I feel like we need to

31:23

have a moment where you just sort of explain

31:25

to me what the Cybertruck is. Can

31:28

you give me some specs on this thing, some pricing information?

31:30

Because I... I don't know if you

31:32

know this about coming back. I've never bought a truck.

31:34

So I don't really even know. I

31:37

don't even have a frame of reference for understanding. What

31:39

I've heard, though, is that it's actually very expensive. So

31:41

it is available in three different

31:44

models. There is a sort of

31:46

low-end rear-wheel drive model that starts

31:48

at $61,000 in the basic configuration.

31:52

There's an all-wheel drive model that starts at $80,000, and then

31:54

you can get the sort of

31:57

top-of-the-line model, which is being called the

31:59

Cybertruck. Beast which has three motors

32:02

and starts at around $100,000. Now

32:04

see, Google should have named

32:06

DeepMind Ultra Cyber Beast. That

32:08

would have been a good name. Yeah, that's true. Yeah.

32:12

So they did start delivering Cybertrucks to

32:14

initial customers last week, and

32:16

they did a big sort of demo reveal.

32:18

They showed some crash testing. They showed a

32:21

video, as you said, of people shooting bullets

32:23

at the doors of the Cybertruck. It appears

32:25

to be bulletproof. And they showed

32:27

how it compares to a bunch of other trucks in a

32:29

pull test where you basically attach

32:31

a very heavy sled to the

32:34

back of a truck and you try to pull it as far as you

32:36

can. And in this test,

32:38

at least the version that Tesla showed

32:40

off, the Cybertruck beat all of the

32:42

leading pickup trucks, including an F-350. So

32:45

it appears to be a truck with a

32:47

lot of towing capacity, and it's bulletproof if

32:50

you do need to survive a shootout. I

32:52

mean, to me, here's the question, Kevin. If

32:54

this truck was produced by anyone other than Elon

32:57

Musk and Tesla, would we be giving it the

32:59

time of day? No,

33:01

I don't think so. Well, so here, let me say a few

33:04

things about this. So one is, I think

33:07

it looks cool, and I'm sorry about that. And I don't have

33:09

any justification on a moral

33:11

or ethical level for thinking that it

33:13

looks cool. I

33:15

know that you are a

33:18

sort of ... Yeah, it's

33:20

fine to just say that you're having a midlife crisis

33:22

and so you're starting to think that the Cybertruck looks

33:24

cool. That's fine. You can admit that. Well,

33:27

here's what I'll say about it. It

33:29

is different, right? And I think, wow,

33:32

I've never seen someone lower the bar so much during

33:34

a conversation. No, but you know what I mean? You

33:36

just go out on the road and you look at

33:39

all these cars and every car now is a compact

33:41

SUV. Every car looks exactly the

33:43

same to me. It's like, oh, you have a RAV4.

33:46

Cool. But this is a

33:48

car, you would not mistake it for any

33:50

other car. It is a car that would

33:52

not survive the design process at basically any

33:54

of the big car companies. It is only

33:56

something that a truly demented individual such as

33:58

Elon Musk could ... make and put

34:00

into production. And, you know, I like an

34:03

opinionated car design. Yeah. Sue me. No, that's

34:05

fine. I think when like the sort of

34:07

the many years from now when the final

34:09

biography of Elon Musk is written, like cyber

34:12

truck will be a chapter about like a

34:14

sign that we were approaching the end game,

34:16

you know, of like, here is somebody who

34:18

is losing his touch. Yeah, it is clearly

34:21

not something that was designed by committee. So I

34:23

think the question that a lot of people are

34:26

asking about the cyber truck is like, who is

34:28

the market for this, right? Is it pickup

34:30

truck owners who are looking to maybe

34:32

get something electric or upgrade to a

34:34

slightly nicer pickup truck? Is it Elon

34:36

Musk fans who are just going to

34:39

buy whatever the latest Tesla is? Is

34:41

it wealthy tech people who want to,

34:43

you know, own something that looks like

34:45

it drove out of Blade Runner? Like,

34:48

who do you think the target market

34:50

for this is? I would say fugitives.

34:53

I would say carjackers. What

34:55

do you think? People who

34:57

subscribe to X premium, I

34:59

would say are the target audience for this. But

35:02

no, I think there will be a lot

35:04

of people who are interested in this. I

35:06

also am very curious about whether this will

35:08

become sort of a signaling vehicle that will

35:10

say something about you. You know, I can

35:12

not like this is not a neutral car.

35:14

This is not a car that you're supposed

35:16

to see and forget about. You're supposed to

35:18

like ponder it totally. And I'm sure we

35:20

will start seeing these very soon on the

35:22

roads of San Francisco. Although we did try

35:24

to find one this week and we cannot.

35:27

We very much wanted to record this episode

35:29

inside a Cybertruck, but we couldn't find one.

35:31

Yeah, apparently it does have very good noise installation

35:33

inside the cab of a Cybertruck. So maybe next

35:35

year we'll record the podcast from there. Better than

35:37

the inside of an airport? You

35:40

know, maybe. Let's like me to get accosted

35:42

by flight attendants. So

35:45

Casey, we also can't really talk about the

35:48

Cybertruck without talking about Elon Musk and the

35:50

kind of insane couple of weeks

35:52

that he's been having. So last week of

35:54

course he appeared on stage at the deal

35:57

book conference in New York and gave this

35:59

totally unhinged interview. interview to my colleague

36:01

Andrew Ross Sorkin in which he

36:03

told advertisers who are staying away

36:05

from X to quote, go fuck

36:07

themselves. And also said a

36:09

number of inflammatory things about his critics and

36:11

his state of mind. And it was just

36:13

sort of like a glimpse into

36:15

his mind and I would say it was not altogether

36:17

reassuring. It was not, you know, I

36:19

of course enjoyed this, I would say very much

36:22

because I think there is still a contingent of

36:24

folks who want to believe that the Elon Musk

36:26

of 2023 is the Elon

36:29

Musk of 2013 and that, you

36:31

know, he said a couple of kooky things here and

36:33

there, but at his core, he's a billionaire

36:35

genius, Tony Stark, savior of humanity.

36:38

And over and over again, he keeps showing up

36:41

in public to be like, no, I'm actually this

36:43

guy. And we

36:45

got another one of those moments and another group of

36:47

people woke up and they're like, oh, wow, okay, I

36:49

guess he is just really going to be like this

36:51

now forever. Yeah. Yeah. I mean,

36:53

I do think that there is some

36:56

angst among the Tesla owners. I

36:58

know most of whom do

37:00

not support Elon Musk's politics

37:02

or his views on content

37:04

moderation. I've heard from a

37:07

number of people over the past few months

37:09

in my life who say some version of,

37:11

you know, I want to get a Tesla

37:14

for reasons X, Y, or Z. You know,

37:16

they have the most chargers, they have the

37:18

best technology. I really like how it looks.

37:20

It's green and I care about the environment

37:22

and it's the one that sort of fits

37:24

my family's needs. But I don't want to

37:26

give Elon Musk my business. I don't want

37:29

to be driving around in something that makes it

37:31

look like I support him. So do you think that's

37:33

actually going to be a meaningful barrier? Do you think

37:35

there are people who will stay away from the Cybertruck,

37:38

even if it is objectively like a

37:40

good truck just because they hate Elon

37:42

Musk? You know, it's it is hard

37:44

to say because as best as I

37:47

can tell, Tesla has not really suffered

37:49

very much yet because of all of

37:51

Elon's antics. Not only has it not

37:53

suffered, but it is by some accounts

37:55

the best selling car in the world.

37:58

Yeah. And certainly the best selling. electric

38:00

car in the world. Sure, at the same time, I

38:02

just hear anecdotally from folks all the time now

38:05

that they would never buy a Tesla. There's

38:08

actually a great profile in The Times

38:10

this week of Michael Stipe, the great

38:12

singer from REM, and there's an anecdote

38:14

in the story about how a tree

38:16

falls on his Tesla, and he's so

38:18

excited because he didn't want to drive

38:20

an Elon Musk car anymore, and now

38:23

he finally had an excuse. So, look,

38:25

is it possible that this is just

38:27

some very thin layer of coastal elites

38:29

who are turning up their nose at Tesla

38:31

while the rest of America and much of the

38:33

world continues to love to drive them? Possible,

38:36

but the thing that I always just keep in

38:38

the back of my mind is there are a

38:40

lot more electric car companies now than they used

38:42

to be. The state emission standards are going to

38:44

require all new vehicles to be electric, not too

38:46

far into the future, and that's just going to

38:49

create a lot of opportunity for folks who want

38:51

to drive an electric car, who don't have to

38:53

put up with the politics or the perception issues

38:55

that might come from driving a Tesla. So Tesla's

38:57

having its moment in the sun now, and maybe

38:59

the Cybertruck will extend their lead into the future,

39:01

or maybe a few years from now we look back

39:04

and we think, oh yeah, that's when the wheels started

39:06

to come off the wagon. Yeah, or the truck, as

39:08

it were. I did

39:10

see one estimate that Tesla is losing

39:12

tens of thousands of dollars every time

39:14

they sell a Cybertruck because they are

39:16

essentially hand-building these now. They have not

39:18

made it into mass production, and obviously

39:20

it takes some time to ramp up

39:23

production in the numbers that they needed to be. So

39:26

if you are an early Cybertruck buyer, you may

39:28

actually be costing Elon Musk money, so that may be

39:30

one reason to get one. This is the first thing you've said that

39:32

makes you want to buy a Cybertruck. Can

39:36

I ask a question? If this were

39:38

made by some other company, if this

39:40

were made by Ford or GM or

39:42

Chrysler, would you buy one? Would you

39:44

be interested? No, like, I

39:47

don't have a car. I got access

39:49

to Waymo this week, and to me this

39:51

is what is exciting, is like not owning

39:53

a car, is being able to just get

39:55

from point A to point B and not

39:58

worry about the various costs of ownership. any

40:00

of this. So, you know, when I think about

40:03

what I want in this world, it's more public

40:05

transit, it's more walking, it's more biking, and I'll

40:07

say it, it is more autonomous vehicles to get

40:09

me from point A to point B on those

40:11

sort of short trips where transit doesn't make sense.

40:14

So, no, there's nothing about this car

40:16

that makes me want to buy it. But I'm guessing

40:18

that for you the answer is yes. Well, let

40:20

me just stipulate that I am not in the

40:22

market for a very expensive pickup truck. There is

40:25

no version of my life in which I need

40:27

something like that. But I would say like similar

40:29

to the Rivian when I do

40:31

see them driving around on the streets of

40:33

my hometown, I will like turn my head

40:36

and kind of admire them. I do think

40:38

the Cybertruck looks kind of cool. I

40:40

hope that it's sort of a spur to the

40:43

rest of the industry to kind of, I don't

40:45

know, like indulge their worst idea. Yes,

40:48

sketch something on a napkin that looks insane and

40:50

then go make it. It's actually how we came

40:53

up with a lot of this podcast. Yes, true.

40:55

We also shop full of that to make sure it

40:58

was bulletproof. And the hardcore podcast, it turns out, is

41:00

bulletproof, maybe. When we

41:03

go back, what else happened in

41:05

AI this week? An

41:15

SUV this practical shouldn't be this

41:17

enjoyable to drive. One

41:20

of this size shouldn't be able to handle

41:22

so much. One so

41:24

functional shouldn't be this sharp. A

41:27

model this innovative shouldn't also

41:29

be this intuitive. And

41:31

yet the fully electric Audi Q4 e-tron

41:33

is just that. Because how

41:35

you get there matters every single day.

41:38

The fully electric Audi Q4 e-tron.

41:41

Audi, progress you can feel. Learn

41:44

more at audiusa.com/electric.

41:47

All right,

41:49

Casey, there's a lot of stuff happening in AI this

41:51

week that we haven't talked about yet. Really, Kevin? Name

41:53

one thing. Well, we have

41:55

a lot to get through. All right. Which

41:57

is why we are doing... This

42:00

week in AI, play the theme song. This

42:03

week in

42:05

AI. So

42:08

our first story in AI this

42:10

week is about wine fraud. This

42:13

was an article in the New York

42:15

Times by Virginia Hughes titled, Bordeaux wine

42:18

snobs have a point, according to this

42:20

computer model. It's an article

42:22

about a group of scientists who have been

42:24

trying to use AI to understand what the

42:27

wine industry calls terroir. Are you

42:29

familiar with terroir? The people who are really

42:31

into this are known as terroirists, I believe.

42:36

Yes, so this is the word that is

42:38

used in the wine industry to describe the

42:40

specific soil and microclimate that wine grapes are

42:42

grown in. And if you go up to

42:44

Napa and you do wine tastings, they will

42:46

often tell you about, oh, our

42:48

soil is more minerally, and that's why our

42:51

wine tastes better, and things like that. And

42:53

I never knew whether that was real. And

42:55

as it turns out, this is something that

42:57

researchers have also been wondering. So

43:00

these researchers trained an algorithm to

43:02

look for common patterns in the

43:04

chemical fingerprints of different wines. They

43:07

were apparently shocked by the results.

43:09

The model grouped the wines into

43:11

distinct clusters that matched with their

43:13

geographical locations in the Bordeaux region.

43:15

So these researchers, they effectively showed

43:18

that terroir is real. One of

43:20

the scientists said, quote, I have scientific

43:22

evidence that it makes sense to charge

43:24

people money for this because they are

43:26

producing something unique. Wow. Well, this has

43:28

some interesting implications for, if you buy

43:30

some really, really expensive wine, but you

43:32

worry that you've gotten a forgery or

43:34

a fraud, I guess there would maybe

43:36

now be some means by which you

43:38

could test it. Or in the far

43:40

future, you could synthesize wine with maybe

43:43

a higher degree of accuracy because we'll

43:45

be able to sort of catalog these

43:47

chemical footprints. Yeah, so apparently in

43:49

expensive wine collections, fraud is

43:52

fairly common. Producers

43:54

have been adjusting their bottles and

43:56

labels and corks to make these

43:58

wines harder to counterfeit. But this still

44:00

happens and with AI apparently

44:02

this will get much harder because you

44:05

can just have the AI say that's

44:07

not really You know mall back from

44:09

this region. It's actually just like crappy

44:12

supermarket wine from California. Oh, man Well,

44:14

this is just great news for

44:16

wine stumps everywhere. Yes, we celebrate it They've been

44:18

waiting for a break and now they have one

44:20

what else happened this week Kevin Okay, so this

44:23

one is actually something that you wrote about a

44:26

problem with Amazon's Q AI

44:29

Model so Q is a

44:31

chatbot that was released by Amazon

44:33

last week and it's aimed

44:36

at kind of enterprise customers So

44:38

Casey what happened with Q? Yeah, so

44:40

I reported this with my colleague Zoe

44:42

Schoefer at platformer last week Amazon announced

44:44

Q Which is its AI chatbot aimed

44:46

at enterprise customers You can sort of

44:48

think of it as a business version

44:51

of chat GBT And the basic idea

44:53

is that you can use it to

44:55

answer questions about AWS where

44:57

you may be running your applications you can

44:59

edit your source code It will cite sources

45:01

for you and Amazon had made a pretty big

45:03

deal of saying that it built Q to be

45:06

More secure and private and suitable for

45:08

enterprise use than a chat GBT Right

45:10

that this was sort of its big

45:12

marketing pitch around Q was like these

45:14

other chatbots. They make stuff up They

45:16

might be training on your data. You

45:19

can't trust them go with ours instead

45:21

It's much safer for business customers That's

45:23

right And so then of course we

45:25

start hearing about what's happening in the

45:27

Amazon slack where some employees are saying

45:29

this thing is Hallucinating very

45:31

badly. Oh, no, it is leaking

45:33

confidential information And there

45:35

are some things happening that one

45:37

employee wrote quote. I've seen apparent

45:39

Q hallucinations I'd expect to potentially

45:42

induce cardiac incidents in legal You

45:46

know Let's stipulate this stuff is

45:48

very early It's just sort of only barely being

45:50

introduced to a handful of clients The reason that

45:52

Amazon is gonna move slowly with something like this

45:54

is for this exact reason and in fact when

45:57

we asked Amazon What it made of all this

45:59

it basically said, you're just watching the normal

46:01

beta testing process play out. At

46:03

the same time, this is embarrassing. And

46:06

if they could have avoided this moment,

46:08

I think they would have. And I

46:10

think it just underscores how wild it

46:12

is that businesses are starting to use

46:14

this technology at all, given that it

46:16

is so unpredictable and that it could

46:18

cause these cardiac incidents for lawyers at

46:20

these companies. I

46:23

understand why businesses are eager to get

46:25

this stuff to their customers and their

46:27

employees. It is potentially a huge

46:29

time saver for a lot of tasks.

46:32

But there's still so many questions

46:34

and eccentricities around the products themselves.

46:36

They do behave in all these

46:38

strange and unpredictable ways. So I

46:40

think we can expect that the

46:42

lawyers, the compliance departments, and the

46:45

IT departments, any companies that are

46:47

implementing this stuff are going to have a busy 2024. Here's

46:49

my bull case for it, though, which is like, you

46:52

know, if you've worked at any company and you've tried

46:54

to use the enterprise software that they have, like, it's

46:56

usually pretty bad. It barely works. You can barely figure

46:58

it out. It probably gave you the wrong answer

47:00

about something without even being A.I. So

47:02

I think we all assume that these

47:04

technologies will need to hit 100 percent

47:07

reliability before anyone will buy them. In

47:09

practice, I think companies will settle for

47:11

a lot less. Right. They

47:13

don't have to be perfect. They just have to

47:15

be better than your existing crappy enterprise software. A

47:17

low bar indeed. All right. That is Amazon. And

47:20

it's cute, which, by the way, while we're talking

47:22

about bad names for AI models, I literally I was talking

47:24

with an Amazon executive last week and I said, you got to

47:26

rename this thing we can't be naming things after the letter Q

47:28

in the year 2023. We

47:30

will reclaim that letter eventually, but we need to give

47:32

it a couple of years. Yeah,

47:35

the Q in non parallel is too easy. All

47:38

right. This next story was about one of

47:40

my favorite subjects when it comes to A.I.,

47:42

which is jailbreaks and hacks that allow you

47:44

to get around some of the restrictions on

47:47

these models. This one actually came from a

47:49

paper published by researchers at DeepMind, who I

47:51

guess were sort of testing chat

47:53

GPT, their competitor, and found that

47:55

if they asked chat GPT 3.5

47:58

turbo, which is one of company

50:00

you want to patch this stuff as quickly as possible

50:02

and it sounds like that's what open AI has done

50:04

here. Alright great well hopefully we never hear about anything

50:07

like this ever again. Okay can

50:09

we talk about Mountain Dew? Let's talk about

50:11

Mountain Dew. This next one is admittedly a

50:13

little bit of a stunt but I thought

50:15

it was a funny one so I want

50:18

to cover it on the show. Mountain Dew

50:20

this week has been doing something they call

50:22

the Mountain Dew raid in which for a

50:25

few days they had an AI crawl livestreams

50:27

on Twitch to determine whether

50:29

the Twitch streamers had a Mountain Dew

50:31

product or logo visible in their livestream.

50:33

Now Kevin for maybe our international listeners

50:35

or folks who are unfamiliar with Mountain

50:38

Dew how would you describe that beverage?

50:40

Mountain Dew is a military grade stimulant

50:42

that is offered to consumers in American

50:45

gas stations to help them get through long drives

50:48

without falling asleep. Yeah that's right if you've never

50:50

tasted Mountain Dew and are curious just go lick

50:52

a battery. I

50:55

was at a truck stop recently on a road

50:57

trip and do you know how many flavors of

50:59

Mountain Dew there are today? How many are

51:01

there? I would say easily a dozen flavors

51:03

of Mountain Dew. That's innovation it's progress that's

51:05

what this company that's what this country does.

51:08

I said this company and

51:10

that's an interesting slip because

51:12

sometimes I do feel like this world is getting too corporate

51:14

Kevin but look at the end of the day this country

51:16

makes every flavor of Mountain Dew that you can imagine and

51:18

many that you couldn't. Yeah so fridges full

51:20

of Mountain Dew at the retailers of America and

51:22

this isn't an AI that just feels like

51:25

it's a dispatch from a

51:27

dystopian future. Now I think this was sort

51:29

of a marketing stunt I don't think this

51:31

was like a big part of their product

51:34

strategy but with this raid

51:36

AI basically if it if it

51:38

analyzed your Twitch stream and saw

51:40

a Mountain Dew product in it you

51:43

could then be featured on the Mountain

51:45

Dew Twitch channel and also receive a

51:47

one-on-one coaching session with a professional live

51:49

streamer. So

51:52

this document that Mountain Dew released as

51:54

like an FAQ... Their Mountain Doc? It's

52:00

the FA do So

52:06

this is the Mountain Dew I'm reading from

52:08

the Mountain Dew raid Q&A

52:10

it says Mountain Dew

52:12

raid is a first of its kind

52:15

AI capability that rewards streamers for doing

52:17

what they love Drinking Mountain Dew on

52:19

stream and unleashes a combination of rewards

52:21

aimed at building and amplifying each participating

52:23

streamers audience So it

52:25

basically goes out crawls twitch looking for

52:28

streamers who have Mountain Dew products and

52:30

logos on their stream once it identifies the

52:32

presence of Mountain Dew this document says Selected

52:35

streamers will get a chat asking to

52:38

opt in to join the raid Once

52:40

you accept the raid AI will keep monitoring

52:43

your stream for the presence of Mountain Dew

52:45

if you remove your Mountain Dew You'll

52:47

be prompted to bring it back on camera.

52:49

If you don't you'll be removed from our

52:52

participating streamers This

52:55

is like truly the most dystopian use of

52:57

AI that I have heard about like I

52:59

know there are more serious, you know Harms

53:02

that can result from AI but this actually

53:04

does feel like a chapter from a dystopian

53:06

novel like bring your Mountain Dew back On

53:09

camera or you will lose access to your

53:11

entire livelihood render to the Mountain Dew panopticon

53:15

It reminds me of you remember that

53:17

like patent that went viral a few

53:19

years ago where Sony had invented some

53:21

new technology That basically would

53:23

allow them to listen to you in

53:25

your living room Like if your TV

53:28

was playing an ad for McDonald's and

53:30

you wanted it to stop you could

53:32

just sort of yell out McDonald's We

53:38

must prevent that world from coming into existence at

53:40

all costs. Yeah, it reminds me of a few

53:43

years ago We did this um a Demo

53:46

my colleagues and I at the Times were pitched

53:48

on an Angry Birds scooter They told you about this.

53:50

Oh, I think you have This

53:55

was a like this was during the big scooter

53:57

craze of like the the 2018-2019 period.

54:01

And the company that

54:03

makes Angry Birds did a promotional stunt

54:05

where they outfitted one of these electric

54:07

scooters with a microphone. And in order

54:09

to make the scooter go, you had

54:12

to scream into the microphone as loud

54:14

as possible. And the louder you yelled,

54:16

the faster the scooter would go. And

54:18

so I am a sucker for

54:20

a stupid stunt. And so I had them ship two

54:22

of these to us and we drag raced them on

54:25

the Embarcadero in San Francisco, just

54:27

screaming as loud as we could

54:30

into the microphones of our Angry Birds scooters to make them

54:32

go fast. And the nice thing about San Francisco is so

54:34

many other people were screaming. Nobody even paid you any attention.

54:36

It was only the fourth weirdest thing happening on

54:39

the Embarcadero that day. And and

54:41

it was a lot of fun. So

54:43

I support stupid stunts like that. I

54:45

support the Mountain Dew AI. Casey,

54:47

what did you think when you

54:50

saw this Mountain Dew? Well, you

54:52

know, there is something that feels

54:54

like weird futurey about AIs just

54:56

scanning all live media to identify

54:59

products and incentivize and reward

55:01

people for for featuring their

55:03

products. At the same time,

55:05

we're already living in a world where

55:07

on social media, some platforms will automatically

55:09

identify products and will then tag them.

55:11

And then maybe if somebody buys that

55:13

product based on you posting it, you'll

55:15

get a little bit of a kickback.

55:17

So this is just

55:20

kind of the near term future of

55:22

social media is that it is already

55:24

a shopping mall and we are just

55:26

making that shopping mall increasingly sophisticated. If

55:28

you see literally anything on your screen, these

55:30

companies want you to be able to just mash it

55:33

with your paw and have it sent to you. So

55:36

this was the latest instance of

55:38

that. But I imagine we'll see more.

55:40

Totally. And it just strikes me as

55:43

sort of an example of how unpredictable

55:45

the effects of this kind of foundational

55:47

AI technology are. Like when they were

55:49

creating image recognition algorithms a decade ago

55:52

in like the bowels of the Google

55:54

DeepMind research department, like they were probably

55:56

thinking, oh, this will be useful for

55:58

radiologists. for identifying

56:01

pathologies on a scan or

56:03

maybe solving some climate

56:06

problem. And instead, this technology, when it makes its

56:08

way into the world, is in the form of

56:10

the Mountain Dew AI bot that just scours Twitch

56:12

live streams to be able to sell more Mountain

56:14

Dew. You know, I think there actually could be

56:16

a good medical use for this. Did you hear

56:18

this? There was another tragic story this week. A

56:20

second person died after drinking a Panera charged lemonade.

56:22

No! Did you read this? Yeah, so that happened

56:24

again. So I think we should build an AI

56:26

that scans for Panera charged lemonades on these Twitch

56:28

streams, and if it sees one, calls an ambulance.

56:31

Ha ha ha. Ha ha ha. Ha

56:33

ha ha. ♪ Between AI and

56:35

you. ♪ With

56:43

all the chatter about AI, it's hard

56:45

to tell fact from buzzwordy fiction.

56:48

In reality, tangible impact starts when

56:50

AI is used responsibly. Alongside R3

56:53

and the Hope for Justice project,

56:55

Intel is powering AI to be

56:57

used in the fight against human

57:00

trafficking. By aggregating global data, matching

57:02

perpetrators, victims, and addresses, AI links

57:04

cases for investigators, all while reinforcing

57:07

confidentiality. It starts with Intel. Learn

57:10

more at intel.com/stories.

57:16

Before we go, a huge thank you

57:18

to all the listeners who sent in

57:21

hard questions for us. As a reminder,

57:23

hard questions is our advice segment where

57:25

we offer you help with ethical or

57:27

moral dilemmas about technology. We

57:30

still are looking for more of those, so please,

57:32

if you have them, send them to us in

57:34

a voice memo at hardfork at nytimes.com, and

57:37

we'll pick some to play on an upcoming

57:39

episode. And to be clear, Kevin, in addition

57:41

to sort of ethical quandaries, we also want

57:43

the drama. We want something that is like

57:45

happening in your life. Is there a fight

57:47

in your life that people are having over

57:49

technology in some way? Please tell us what

57:51

it is, and we'll see if we can

57:53

help. Yeah, and these don't need to be

57:55

high-minded scenarios about AI wreaking havoc on your

57:57

professional life. It could just be something. See

58:00

from your personal life hot gossip. Yeah,

58:03

spill the tea hard fork at NY times calm Hard

58:06

fork is produced by Rachel Cohen and Davis

58:08

land were edited by Jen point This

58:11

episode was fact-checked by Caitlin love Today's

58:14

show was engineered by Chris wood original

58:17

music by Marian Lozano Sophia

58:19

Landman and Dan Powell our

58:22

audience editor is Nelga Logley video

58:24

production by Ryan Manning and Dylan Bergeson

58:27

Special thanks to Paula Schumann we winged ham

58:30

Kaitlyn Presti and Jeffrey Miranda You

58:32

can email us hard for at NY time Did

58:35

your favorite flavor of Bounty? You

58:58

You You

59:23

Time for a quick break to talk about McDonald's

59:25

know how we make our sausage McMuffin with egg

59:27

It starts with a fresh cracked egg cooked to

59:30

perfection Then we add

59:32

a savory grilled sausage patty American cheese

59:34

and a freshly toasted English muffin Know

59:36

what makes our sausage McMuffin with egg

59:38

even better when you add a caramel

59:40

mocha or refreshing caramel mocha iced coffee

59:43

to it So make your morning better

59:45

by starting with breakfast at McDonald's at

59:47

participating McDonald's