0:00

Whether you're starting or scaling your

0:02

company's security program, demonstrating top-notch security

0:04

practices and establishing trust is more

0:07

important than ever. Vanta automates compliance

0:09

for SOC 2, ISO 27001, and

0:11

more. With Vanta, you can streamline

0:14

security reviews by automating questionnaires and

0:16

demonstrating your security posture with a

0:18

customer-facing trust center. Over 7,000 global

0:21

companies use Vanta to manage risk

0:23

and prove security in real-time. Get

0:25

a thousand dollars off Vanta when

0:28

you go to vanta.com slash daily.

0:30

vanta.com/hard fork for a thousand dollars

0:32

off. I just got my weekly,

0:34

you know, I set up ChatGPT to

0:36

email me a weekly affirmation before we

0:39

start taping because you can do that

0:41

now with the tasks feature. Yeah, people

0:43

say this is the most expensive way

0:46

to email yourself for a reminder. So

0:48

what sort of affirmation did we get?

0:50

Today it said, you are an incredible

0:52

podcast host, sharp, engaging, and completely in

0:55

command of the mic. You're taping today

0:57

is going to be phenomenal, and you're

0:59

going to absolutely kill it. Wow, and

1:02

that's why it's so important that ChatGPT

1:04

can't actually listen to podcasts, because I

1:06

don't think it would say that if

1:08

it actually ever hurt us. It would

1:11

say, just get this over with it.

1:13

Get on with it! I'm Kevin

1:15

Roos, a tech columnist at the New

1:17

York Times. I'm Casey Noon from Platformer.

1:19

And this is Hard Fork. This

1:21

week, we go deeper on Deep Seek.

1:23

China Talks, Jordan Schneider joins us to

1:26

break down the race to build powerful

1:28

AI. Then, hello, operator. Kevin and I

1:30

put open AI's new agent software to

1:32

the test. And finally, the train is

1:34

coming back to the station for a

1:36

round of hot mess express. Well

1:46

Casey it is rare that we spend two

1:48

consecutive episodes of this show talking

1:50

about the same company, but I

1:52

think it is fair to say

1:54

that what is happening with Deep

1:56

Seek has only gotten more interesting

1:58

and more confusing. Yeah, that's right.

2:00

It's hard to remember a story

2:02

in recent months, Kevin, that has

2:05

generated quite as much interest as

2:07

what is going on with Deep

2:09

Seek. Now, Deep Seek for anyone

2:11

catching up is this relatively new

2:13

Chinese AI startup that released some

2:15

very impressive and cheap AI models

2:17

this month that lots of Americans

2:19

have started downloading and using. Yeah,

2:21

so some people are calling this

2:23

a Sputnik moment for the AI

2:25

industry when kind of every nation

2:27

perks up and starts, you know,

2:29

paying attention at the same time to

2:31

the AI arms race. Some people are saying

2:34

this is the biggest thing to happen in

2:36

AI since the release of China. GPT, but

2:38

Casey, why don't you just catch us up

2:40

on what has been happening since

2:43

we recorded our emergency podcast episode

2:45

just two days ago? Well, I

2:47

would say that there have probably

2:49

been three stories, Kevin, that I

2:51

would share to give you a

2:53

quick flavor of what's been going

2:55

on. One, a market research firm

2:57

says Deep Seek was downloaded 1.9

3:00

million times on iOS in recent

3:02

days, and about 1.2 million times

3:04

on the Google Play store. The

3:06

second thing I would point out

3:08

is that Deep Seek has been

3:10

banned by the US Navy over

3:12

security concerns, which I think is

3:14

unfortunate, because what is a submarine

3:16

doing, if not Deep Seeking? It

3:18

was also banned in Italy, by

3:20

the way, after the data protection

3:22

regulator made an inquiry. And finally,

3:24

Kevin, Open AI says that there

3:26

is evidence that Deep Seek distilled

3:29

its models. Distillation is kind of

3:31

the AI lingo or euphemism for

3:33

they used our API to try

3:35

to unravel everything we were doing

3:37

and use our data in ways

3:39

that we don't approve of. And

3:41

now Microsoft and Open AI are

3:43

now jointly investigating whether Deep Seek abused

3:45

their API. And of course we can

3:47

only imagine. how open AI is feeling

3:50

about the fact that their data might

3:52

have been used without payment or consent.

3:54

Oh yeah, must be really hard to

3:57

think that someone might be out there

3:59

trading AI. on your data without permission.

4:01

And I want to acknowledge that literally every

4:03

single user, a blues guy, already made this

4:05

joke, but they were all funny, and I'm

4:07

so happy to repeat it here on Hard

4:10

Fork this week. Now Kevin, as always, when

4:12

we talk about AI, we have certain disclosures

4:14

to make. New York Times Company is currently

4:16

suing Open AI and Microsoft over copyright violations

4:18

alleged related to the use of their copyrighted

4:21

data to train AI models. I think that

4:23

was good. It was very good. And I'm

4:25

in love with a man who works in

4:27

anthropic. But that said Kevin, we have even

4:29

further we want to go into the

4:31

Deep Seek story and we want to

4:33

do it with the help of Jordan

4:35

Schneider. Yes, we are bringing in the

4:37

big guns today because we wanted to

4:39

have a more focused discussion about Deep

4:41

Seek that is not about, you know,

4:43

the stock market or how the American

4:45

AI companies are reacting to this, but

4:48

is about one of the biggest sets

4:50

of questions that all of this raises,

4:52

which is what is China up to

4:54

with deep seek and AI more broadly?

4:56

What are the geopolitical implications of

4:58

the fact that Americans are now

5:00

obsessing over this Chinese-made AI app?

5:02

What does it mean for deep

5:04

seeks prospects in America? What does

5:07

it mean for their prospects in

5:09

China? And how do it's all

5:11

this fit together from the Chinese

5:13

perspective? So... Jordan Schneider is our

5:16

guest today. He's the founder and

5:18

editor-in-chief of China Talk, which is

5:20

a very good newsletter and podcast

5:22

about US-China tech policy. He's been

5:24

following the Chinese AI ecosystem for

5:27

years. And unlike a lot of American

5:29

commentators and analysts who were sort of

5:31

surprised by Deep Seek and what they

5:33

managed to pull off over the last

5:35

couple weeks, I'll say it. I was

5:38

surprised. Yeah, me too. But Jordan has

5:40

been following this company for a long

5:42

time and a big... focus of China

5:44

Talk, his newsletter and podcast, has been

5:46

translating literally what is going on in

5:49

China into English, making sense of it

5:51

for a Western audience, and keeping tabs

5:53

on all the developments there. So perfect

5:55

guest for this week's episode, and

5:57

I'm very excited for this conversation.

6:00

Yes, I have learned a lot

6:02

from China Talk in recent days

6:04

as I've been boning up on

6:06

Deep Seek, so we're excited

6:08

to have Jordan here,

6:10

and let's bring him in.

6:13

Joiner Schneider, welcome to Hard

6:15

Fork! Oh my God, such a huge

6:17

fan. This is such an honor. We're so

6:19

excited. I have learned truly so much from

6:21

you this week. And so when we were

6:24

talking about what to do this week, we

6:26

just looked at each other and said, we

6:28

have got a see of Jordan can come

6:30

on this podcast. Yeah. So this has been

6:32

a big week for Chinese tech policy, maybe

6:35

the biggest week for Chinese tech policy, at

6:37

least that I can remember. I realized that

6:39

something important was happening last weekend when I

6:41

started getting texts from like all of my...

6:44

non-tech friends being like, what is going on

6:46

with Deep Seek? And I imagine you

6:48

had a similar reaction because you

6:50

are a person who does constantly

6:52

pay attention to Chinese tech policy.

6:54

So I've been running China Talk for

6:56

eight years and I can get my

6:58

family members to maybe read like one

7:01

or two editions a year and the

7:03

same exact thing happened with me Kevin

7:05

where all of a sudden I got

7:07

oh my god Deep Seek like it's

7:10

on the cover of the New York

7:12

Post Jordan you're so clairvoyance like maybe

7:14

I should read you more. I'm like

7:16

okay thanks mom appreciate that. Yeah, so

7:18

I want to talk about Deep Seek

7:20

and what they have actually done here,

7:23

but I'm hoping first that you can

7:25

kind of give us the basic lay

7:27

of the land of the sort of

7:29

Chinese AI ecosystem, because that's not an

7:31

area where Casey or I have spent

7:34

a lot of time looking, but tell

7:36

us about Deep Seek and sort of

7:38

where it sits in the overall Chinese

7:40

industry. So Deep Seek is a really

7:42

odd... It was born out of

7:45

this very successful quant hedge fund.

7:47

The CEO of which basically after

7:49

ChatGPT was released was like, okay,

7:52

this is really cool. I want

7:54

to spend some money and some

7:57

time and some compute and hire

7:59

some. fresh young graduates to see if

8:01

we can give it a shot to

8:03

make our own language models. And so

8:06

a lot of companies are out there

8:08

building their own large language models. What

8:10

was the first thing that happened that

8:12

made you think, oh, this one, this

8:14

company is actually making some interesting

8:16

ones. Sure. So there are lots

8:18

and lots of very money to

8:20

Chinese companies that have been trying

8:22

to follow a similar path after

8:24

chat. giant players like Ali Baba,

8:27

Tencent, Bight Dance, Huawei even,

8:29

trying to, you know, create

8:31

their own open AI, basically.

8:33

And what is remarkable is

8:35

the big organizations can't quite

8:37

get their head around creating

8:39

the right organizational institutional structure

8:41

to incentivize this type of

8:43

collaboration and research that leads

8:45

to real breakthroughs. So, you

8:47

know, Chinese firms have been

8:49

releasing models for years now,

8:51

but deep seek seek because.

8:53

because of the way that

8:55

it structured itself and the

8:57

freedom they had not necessarily being

8:59

under a direct profit motive, they

9:01

were able to put out some

9:03

really remarkable innovations that caught the

9:05

world's attention, you know, starting maybe

9:08

late December, and then, you know,

9:10

really blew everyone's mind with the

9:12

release of the R1 chatbot. Yeah,

9:14

so let's talk about R1 in

9:16

just a second, but one more

9:18

question for you, Jordan, about Deep

9:20

Seek. What? do we know about

9:22

their motivation here? Because so much

9:25

of what has been puzzling American

9:27

tech industry watchers over the last

9:29

week is that this is not

9:31

a company that has sort of

9:34

an obvious business model connected to

9:36

its AI research. We know why

9:38

Google is developing AI because it

9:41

thinks it's going to make the

9:43

company Google much more profitable. We

9:45

know why Open AI is developing

9:47

advanced AI models. It does not

9:50

seem obvious to me, and I

9:52

have not read anything from people

9:54

involved in Deep Seek, about why

9:56

they are actually doing this and

9:58

what their ultimate... goal is. So

10:01

can you help us understand

10:03

that? We don't have a lot

10:05

of data, but my base case,

10:07

which is based on two extended

10:09

interviews that the Deep C CEO

10:11

released, which we've translated on China

10:14

Talk, as well as just like

10:16

what Deep C employees have been

10:18

tweeting about in the West, and

10:20

then domestically, is that their dreamers.

10:22

I think the right mental model

10:24

is open AI, you know, 2017

10:26

to 2022. Like, I'm sure you

10:28

could ask the same thing, like,

10:30

what the hell are they doing?

10:33

literally said, I have no idea

10:35

how we're ever going to make

10:37

money, right? And here we are

10:39

in this grand new paradigm. So

10:41

I really think that they do

10:43

have this like vision of AGI

10:45

and like, look, we'll build it

10:47

and we'll make it cheaper for

10:49

everyone, you know, we'll make it

10:52

cheaper for everyone, and like, we'll

10:54

figure it for everyone, we'll figure

10:56

it out, you know, we'll make

10:58

it cheaper for everyone, you know,

11:00

we'll figure it in deep. but

11:02

bite dance or Ali or Tencent or

11:04

Hualway and the government's going to start

11:06

to pay attention in a way which

11:09

it really hasn't over the past few

11:11

years. Right and I want to I want to

11:13

drill down a little bit there because

11:15

I think one thing that most listeners

11:17

in the West do know about Chinese

11:19

tech companies is that many of them

11:21

are sort of inextricably linked

11:24

to the Chinese government that the

11:26

Chinese government has. access to user

11:28

data under Chinese law, that these

11:30

companies have to follow the Chinese

11:32

censorship guidelines. And so as soon

11:34

as Deep Sikhs started to really

11:36

pop in America over the last

11:39

week, people started typing in things

11:41

to Deep Sikhs model, like tell

11:43

me about what happened at Tiananmen

11:45

Square or tell me about Xi

11:47

Jinping or tell me about the

11:49

great leap forward. And it just

11:52

sort of wouldn't do it at all.

11:54

And so people I think saw that and

11:56

said, oh, this is... This is like every

11:58

other Chinese company that has this. sort of

12:00

hand-in-glove relationship with the Chinese ruling

12:03

party, but it sounds from what

12:05

you're saying like Deep Seek has

12:07

a little bit more complicated a

12:10

relationship to the Chinese government

12:12

than maybe some other better-known

12:15

Chinese tech companies. So explain

12:17

that. Yeah, I mean I think

12:19

it's it's it's important like the

12:21

mental model you should have for

12:23

these CEOs are not like people

12:25

who are dreaming to spread Xi

12:27

Jinping thought. Like what they want

12:29

to do is compete. with Mark

12:31

Zuckerberg and Sam Altman and show

12:33

that they're like really awesome and

12:35

great technologists. But the tragedy is,

12:37

is let's take bite dance for

12:39

example, you can look at Jianguming,

12:41

their CEOs, Weibo posts from 2012,

12:43

2013, 2014, which are super liberal

12:45

in a Chinese context, saying like, you

12:47

know, we should have freedom of expression,

12:49

like we should be able to do

12:52

whatever we want. And the early years

12:54

of bite dance, there was a lot

12:56

of relatively more subversive content on the

12:58

platform. you sort of saw like real

13:00

poverty in China, you saw off-color jokes,

13:02

and then all of a sudden in

13:04

2018, he posts a letter saying, I

13:07

am really sorry, like, I need to

13:09

be part of this sort of like

13:11

Chinese national project and like better adhere

13:13

to, you know, modern Chinese socialist values

13:15

and I'm really sorry and it won't

13:17

ever happen again. You know, the same thing

13:19

happened with DD, right? Like they don't really

13:22

want to have... to do anything with politics

13:24

and then they get on someone's side

13:26

and all of a sudden they get

13:28

zapped. DD is of course the big

13:30

Chinese ride share company. Correct. Yeah. What

13:32

did DD do? So they listed on the

13:34

Western Stock Exchange after the Chinese government

13:36

told them not to and then they

13:38

got taken off app stores and it

13:41

was a whole giant. nightmare like they

13:43

had to sort of go through their

13:45

rectification process. So point being with Deep

13:47

Seek right is like now they are

13:49

whether they like it or not going

13:52

to be held up as a national

13:54

champion and that comes with a lot

13:56

of headaches and responsibilities from you know

13:58

potentially giving the Chinese more access, you

14:00

know, having to fulfill government contracts,

14:02

which like honestly are probably really

14:05

annoying for them, to do in

14:07

sort of distracting from the broader

14:09

mission they have of developing and

14:11

deploying this technology in the widest

14:13

range possible. But like Deep Seek

14:15

thus far has flown under the

14:17

radar, but that is no longer

14:19

the case and things are about

14:21

to change for them. Right. And

14:23

I think that was one of

14:25

the surprising things about Deep Seek. for

14:27

the people I know, including you, who

14:29

follow Chinese tech policy, is, you know,

14:31

I think people were surprised by the

14:33

sophistication of their models, and we talked

14:36

about that on the emergency pod that

14:38

we did earlier this week, and how

14:40

cheaply they were trained. But I think

14:42

the other surprise is that they were

14:44

released as open-source software, because, you know,

14:46

one thing that you can do with

14:48

open-source software is download it, hosted in

14:51

another country, removed some of the guardrails

14:53

and the censorship filters that might have

14:55

been part of the original model, But

14:57

by the way, it turned out there weren't even

14:59

really guardrails on the on the V3

15:01

model, right? That it had not been

15:03

trained to avoid questions about Tiananmen Square

15:05

or anything. So that was another really

15:07

unusual thing about this. Right. And one

15:09

thing that we know about Chinese technology

15:12

products is that they don't tend to

15:14

be released that way. They tend to

15:16

be hosted in China and overseen by

15:18

Chinese teams who can make sure that

15:20

they're not out there talking about Tiananmen

15:22

Square. Is the open source nature of what

15:24

Deep Seek has done here part of

15:26

the reason that you think there might

15:28

be conflict looming between them and the

15:30

Chinese government? You know, honestly, I think

15:32

this whole askate about Tiananmen stuff is

15:35

a bit of a red herring on

15:37

a few dimensions. So first, one of

15:39

these like arguments that there's a little

15:41

sort of confusing to me is like

15:43

folks used to say, oh, like the

15:46

Chinese models are going to be lobotomized

15:48

and like they will never be as

15:50

smart as the Western ones because like

15:52

they have to be politically correct. I

15:54

mean, look, if you ask Claude to say

15:57

racist things, it won't. And Claude's still

15:59

pretty smart. a bit of a red

16:01

herring when talking about sort of long-term

16:03

competitiveness of Chinese and Western models. Now,

16:05

you asked me like, oh, so they

16:07

released this model globally and it's open

16:09

source, maybe someone in the Chinese government

16:11

would be uncomfortable with the fact that

16:14

people can get a Chinese model to

16:16

say things that would get you thrown

16:18

in jail if you posted them online

16:20

in China. It's going to be a

16:22

really interesting calculus for the Chinese government

16:24

to make, because on the one hand,

16:26

this is the most positive shine that

16:28

Chinese AI has got globally in the

16:31

history of Chinese AI. So they're

16:33

gonna have to navigate this and

16:35

it might prompt some uncomfortable conversations

16:37

and bring regulators to a place

16:39

they wouldn't have otherwise landed.

16:41

Now, Jordan, I want to ask you

16:43

about something that people have been talking

16:46

about and speculating about in relationship to

16:48

the Deep Seek news for the last

16:50

week or so, which is about chip

16:53

controls. So we've talked a little bit

16:55

on the show earlier this week about

16:57

how Deep Seek managed to put together

17:00

these models. using some of these kind

17:02

of second-rate chips from invidia that are

17:04

allowed to be exported to China. We've

17:06

also talked about the fact that you

17:09

cannot get the most powerful chips legally

17:11

if you are a Chinese tech

17:13

company. So there have been some

17:15

people, including Elon Musk and other

17:18

American tech luminaries, who have said,

17:20

oh, well, Deep Seek has this

17:22

sort of secret stash of these

17:24

banned chips that they have smuggled

17:26

into the country. and that actually

17:28

they are not making due with

17:31

kind of the Kirkland signature chips

17:33

that they say they are. What

17:35

do we know about how true

17:37

that is? So, did Deep Seek

17:39

have band ships? It's kind of

17:41

impossible to know. This is a question

17:43

more for the US intelligence community than

17:45

like Jordan Schneider on Twitter. But I

17:48

do think that it is important to

17:50

understand that the delta between what you

17:52

can get in the West and what

17:54

you can get in China is actually

17:56

not that big. And, you know, we're

17:58

talking about training a lot, but on

18:00

the infrared side, trying to can still

18:02

buy this H20 chip from a video, which

18:04

is basically world-class at like deploying

18:06

the AI and letting everyone use it.

18:08

So does this mean that we should

18:11

just give up? I don't think so.

18:13

Compute is going to be in a

18:15

core input, regardless of how much model

18:17

distillation you're going to have in the

18:19

future. There have been a lot of

18:22

quotes even from the deep seek founder

18:24

basically saying like the one thing that's

18:26

holding us back are these export controls.

18:29

Right. Okay, I want to ask a big

18:31

picture question. Sure. I think

18:33

that a reason that people have

18:35

been so fascinated by this Deep

18:37

Seek story is that at least

18:40

for some folks, it seems to

18:42

change our understanding of where China

18:44

is in relation to the United

18:47

States when it comes to developing

18:49

very powerful AI. Jordan, what is

18:51

your assessment of what? the V3

18:54

and R1 models mean? And to

18:56

what extent do you think the

18:58

game has actually changed here? I'm

19:00

not really sure the game has

19:03

changed so much. Like Chinese engineers

19:05

are really good. I think it

19:07

is a reasonable base case that

19:09

Chinese firms will be able to

19:11

develop comparable or fast follow on the

19:13

model side. But the real sort of

19:16

long-term competition is not just going to

19:18

be on developing the models, but deploying

19:20

them and deploying them at scale. And

19:22

that's really where compute comes in, and

19:25

that's why expert controls are going to

19:27

continue to be a really important piece

19:29

of America's strategic arsenal when it comes

19:31

to making sure that the 21st century

19:34

is defined by the US and our

19:36

friends as opposed to China and theirs.

19:38

Right. So it's one thing to have

19:40

a model that is about as capable

19:42

as the models that we have here

19:44

in the United States. It's another thing

19:46

to have the energy to actually let

19:49

everyone use them as much as they

19:51

want to use them. What you're saying

19:53

is no matter what Deep Seek may

19:55

have invented here, that fundamental dynamic has

19:57

not changed. China simply does not have

19:59

nearly the... amount of compute that the United

20:01

States has. As long as we don't

20:03

screw up export controls. So I

20:05

think the sort of base case

20:08

for me is that if the

20:10

US stays serious about holding a

20:12

line on semiconductor manufacturing equipment and

20:14

export of AI chips, then it

20:17

will be incredibly difficult for the

20:19

Chinese sort of broader semiconductor and

20:21

AI ecosystem to leap ahead much

20:24

less kind of like fast follow

20:26

beyond being able to develop comparable

20:28

models. I'm feeling good as long

20:30

as you know, Trump doesn't make

20:33

some like crazy trade for, you

20:35

know, soybeans in exchange for ASML

20:37

EU machines. That would really break

20:39

my heart. I want to inject

20:41

kind of a note of skepticism

20:44

here because I buy everything you're

20:46

saying about how Deep Seek's progress

20:48

has been sort of bottlenecked by

20:51

the fact that it can't get

20:53

these very powerful American AI chips

20:56

from companies like invidia. But

20:58

I also am hearing... people who

21:00

I trust say things that make

21:02

me think that actually the bottleneck

21:05

may not be the availability of

21:07

chips that maybe with some of

21:09

these algorithmic efficiency breakthroughs that Deep

21:12

Seek and others have been making,

21:14

it might be possible to run

21:16

a very very powerful AI model

21:19

on a conventional piece of hardware

21:21

on a Mac book even. And

21:23

I wonder about How much of

21:26

this is just like AI companies

21:28

in the West trying to cope,

21:30

trying to make themselves feel better,

21:32

trying to reassure the market that

21:35

they are still going to make

21:37

money by investing billions and billions

21:39

of dollars into building powerful AI

21:41

systems? If these models do

21:43

just become sort of lightweight commodities

21:46

that you can run on a

21:48

much less powerful cluster of computers,

21:50

or maybe on one computer, doesn't

21:53

that just mean we

21:56

can't? control the proliferation

21:58

of them at like this is

22:00

one potential future and maybe that potential

22:03

future like went up 10 percentage

22:05

points of likelihood of like you

22:07

being able to fit the biggest

22:09

badest smartest most fast efficient AI

22:11

model on something that you that

22:13

can sit in your home but

22:15

I think there are lots of

22:17

other futures in which sort of

22:19

the world doesn't necessarily play out

22:21

that way and look in video

22:23

went down 15% it didn't go

22:26

it didn't go down 95% like

22:28

I think if we're really in

22:30

that world where chips don't matter

22:32

because everything can be shrunk down

22:34

to kind of consumer grade hardware

22:36

then the sort of reaction that

22:38

I think you would have seen

22:40

in the stock market would have

22:42

been even more dramatic than the

22:44

kind of freak out we saw

22:46

over this week so we'll see

22:48

I mean it would be a

22:50

really remarkable kind of democratizing thing

22:52

if that was the future we

22:54

ended up living in, but it

22:56

still seems pretty unlikely to my

22:58

history major brain here. I would

23:00

also just point out, Kevin, that

23:02

when you look at what Deep

23:04

Seek has done, they have created

23:07

a really efficient version of a

23:09

model that American companies themselves had

23:11

trained like nine to 12 months

23:13

ago. So they sort of caught

23:15

up very quickly. And there are

23:17

fascinating technological innovations in what they

23:19

did. But in my mind, these

23:21

are still primarily optimization. Like for

23:23

me, what would tip me over

23:25

into like, oh my gosh, America

23:27

is losing this race is China

23:29

is the first one out of

23:31

the gate with a virtual co-worker, right?

23:34

Or like a truly phenomenal agent. Some

23:36

sort of leap forward in the technology

23:38

as opposed to we've caught up really

23:41

quickly and we've figured out something more

23:43

efficiently. Are you saying it differently than

23:45

that? I mean, I guess I just

23:48

don't know what like a six-month lag

23:50

would buy us if it does take

23:52

six months for the Chinese AI companies

23:55

like Deep Seek to sort of catch

23:57

up to the state of the art.

23:59

I was struck by Adari

24:01

Amade, who's the CEO of

24:04

Anthropic, wrote an essay just

24:06

today about Deep Seek and

24:08

Export Controls. And in it,

24:10

he makes this point about

24:12

the sort of difference between

24:15

living in what he called

24:17

a unipolar world, where one

24:19

country or one block of

24:21

countries has access to something

24:23

like an AGI or an

24:25

ASI, and the rest of

24:28

the world doesn't. versus the

24:30

situation where China gets there roughly

24:32

around the same time that we

24:34

do. And so we have this

24:36

bipolar world where two blocks of

24:38

countries, the East and the West,

24:40

basically have access to this equivalent

24:42

technology. And so- And of course

24:44

in a bipolar world, sometimes we're

24:47

very happy and sometimes we're very

24:49

sad. Exactly. So I just think

24:51

like, whether we get there six months

24:53

ahead of them or not. I just

24:55

feel like there isn't that much of

24:57

a material difference. But Jordan, maybe I'm

24:59

wrong. Can you make the other side

25:01

of that? That it really does matter. I'm

25:03

kind of there. I, you know, I'll

25:06

take a little bit of issue with

25:08

what Darrio says. And I think, you

25:10

know, what one of the lessons that

25:12

Deep Sea shows is we should expect

25:14

a base case of Chinese model makers

25:16

being able to fast follow the innovations,

25:18

which, by the way, Casey. actually do

25:20

take those giant data centers to run

25:22

all the experiments in order to find

25:24

out, you know, what is this sort

25:26

of future direction you want to take

25:28

your model? And what sort of AI

25:30

is going to come down to is

25:32

not just creating the model, not just

25:34

sort of like Dario envisioning the future

25:36

and then all of a sudden like

25:39

things happen. Like there's gonna be a

25:41

lot of messiness in the implementation and

25:43

there are gonna be sort of like

25:45

teachers unions who are upset that AI

25:47

comes in the classroom and there are

25:49

gonna be like all these regulatory pushbacks

25:51

and a lot of societal reorganization which

25:53

is gonna need to happen just like

25:55

it did during the industrial revolution. So

25:57

look model making is a frontier of

25:59

competition. Compute access is a frontier of

26:01

competition, but there's also this broader like

26:04

how will a society kind of adopt

26:06

and cope with all of this new

26:08

future that's going to be thrown in

26:10

our faces over the coming years. And

26:12

I really think it's that just as

26:15

much as the model development and the

26:17

compute, which is going to determine which

26:19

countries are going to gain the most

26:21

from what AI is going to offer

26:23

us. Yeah. Well, Jordan, thank you

26:26

so much for joining and

26:28

explaining all of this to

26:30

us. I feel more enlightened.

26:32

Me too. Oh, my pleasure.

26:34

My chain of thought has

26:36

just gotten a lot longer.

26:39

That's an AI joke. Let

26:41

me come back. Kevin, there's

26:43

an agent at our door.

26:45

Is it Jerry McGuire? No,

26:47

it's an AI one. Oh,

26:49

okay. Whether you're starting or

26:54

scaling

26:56

your

26:59

company's

27:01

security

27:06

program,

27:10

Demonstrating top-notch security practices and establishing

27:12

trust is more important than ever.

27:14

Vanta automates compliance for SOC 2,

27:17

ISO 27001, and more. With Vanta,

27:19

you can streamline security reviews by

27:21

automating questionnaires and demonstrating your security

27:24

posture with a customer-facing trust center.

27:26

Over 7,000 global companies use Vanta

27:29

to manage risk and prove security

27:31

in real-time. Get a thousand dollars

27:33

off Vanta when you go to

27:36

vanta.com/hard fork. That's vanta.com/hard fork for

27:38

a thousand dollars off. The New

27:40

York Times app has all this stuff that you

27:42

may not have seen. The way the tabs are

27:45

at the top with all of the different

27:47

sections. I can immediately navigate to something

27:49

that matches what I'm feeling. Play

27:51

wordle or connections and then swipe

27:53

over to read today's headlines. There's

27:55

an article next to a recipe

27:57

next to a recipe, next to

27:59

games. It's just easy to

28:01

get everything in one place.

28:04

This app is essential. The

28:06

New York Times app, all

28:08

of the times, all in

28:10

one place. Download it now

28:12

at nytimes.com slash app.

28:15

operator information. Give me Jesus

28:17

on the line. Do you

28:19

know that one? No. Do

28:21

you know operator by Jim

28:23

Croci? No. operator. Oh, won't

28:26

you help me post this

28:28

call? Well Casey call

28:30

your agent because today we're talking

28:32

about AI agents Why do I need to

28:35

call my agent? I don't know I

28:37

just sounded good. Okay, well I appreciate

28:39

the effort, but yes Kevin because For

28:41

months now, the big AI labs have

28:43

been telling us that they are going

28:46

to release agents. This year, agents of

28:48

course, being software that can essentially use

28:50

your computer on your behalf or use

28:53

a computer on your behalf. And the

28:55

dream is that you have sort of

28:57

a perfect virtual assistant or co-worker. You

28:59

name it. If they are somebody who

29:02

might work with you at your job,

29:04

the AI labs are saying, we are

29:06

building that for you. Yeah, so last

29:09

year toward the end of the year

29:11

we started to see kind of these

29:13

demos, these these previews that companies like

29:15

Anthropic and Google were working on. Anthropic

29:18

released something called computer use, which was

29:20

an AI agent, a sort of very

29:22

early preview of that. And then Google

29:24

had something called Project Mariner that I

29:27

got a demo of, I believe in

29:29

December, that was basically the same thing,

29:31

but their version of it. And then

29:33

just last week, Open AI announced that

29:36

it was launching. which is its first

29:38

version of an AI agent, and unlike

29:40

Anthropic and Google's, which you either had

29:42

to be a developer or part of

29:44

some early testing program to access, you

29:46

and I could try it for ourselves

29:48

by just upgrading to the $200 a

29:51

month pro subscription of ChatGPT. Yeah, and

29:53

I will say that as somebody who's

29:55

willing to spend money on software all

29:57

the time, I thought, am I really

29:59

about... to spend $200 to do this,

30:01

but in the name of science, Kevin,

30:03

I had to. At this point, I

30:06

am spending more on AI subscription products

30:08

than on my mortgage. I'm pretty sure

30:10

that's correct. But it's worth it. We

30:12

do it for journalism. We

30:14

do. So we both spent a couple

30:16

of days putting operator through its paces,

30:18

and today we want to talk a

30:21

little bit about what we found. Yeah,

30:23

so would you just explain like what?

30:25

operator is and how it works.

30:27

Yeah, sure. So operator is a separate

30:29

sub domain of chat GPT. You know,

30:31

sometimes the chat GPT will just let

30:34

you pick a new model from a

30:36

drop-down menu. But for operator, you got

30:38

to go to a dedicated site. Once

30:40

you do, you'll see a very familiar

30:42

chatbot interface, but you'll see different kinds

30:45

of suggestions that reflect some of the

30:47

partnerships that Open AI has struck up.

30:49

So for example, they have partnerships with

30:51

open table and stub hub hub and

30:54

all recipes. meant to give you an

30:56

idea of what operator can do. And

30:58

frankly Kevin, not a lot of

31:00

this sounds that interesting, right? Like

31:02

the suggestions are on the the

31:04

order of suggest a 30-minute meal

31:06

with chicken or reserve a table

31:08

for eight or find the most

31:10

affordable passes to the Miami Grand

31:13

Prix. Again, so far, kind of

31:15

so boring. What is... different about

31:17

operator though is that when you

31:19

say okay find the most affordable

31:21

passes to the Miami Grand Prix

31:23

when you hit the enter button

31:25

it is going to Open up

31:27

its own web browser and it's

31:29

going to use this new model that

31:32

they have developed to try to actually

31:34

go and get those passes for you.

31:36

Yeah, so this is an important thing

31:38

because I think, you know, when people

31:40

first heard about this, they thought, okay,

31:43

this is an AI that kind of

31:45

takes over your computer, takes over your

31:47

web browser, that is not what operator

31:49

does. Instead, it opens a new browser

31:51

inside your browser and that browser is

31:53

hosted on open AI servers. It doesn't

31:56

have your bookmarks and stuff like

31:58

that saved, but you can take

32:00

it over from the autonomous AI agent if

32:02

you need to click around or do something

32:04

on it. But it basically exists. It's like

32:06

a it's a browser within a browser. Yeah.

32:08

So the one of the ideas on operator

32:11

is that you should be able to leave

32:13

it on supervised and just kind of go

32:15

do your work while it works. But of

32:17

course it is very fun initially at least

32:19

to watch the computer try to use itself.

32:21

And so I sat there in front of

32:23

this browser within a browser within a browser

32:25

and I watched. type the, you know, URL,

32:27

navigate to a website, and, you know, in

32:30

the example I just gave, actually

32:32

search for passes to the Miami

32:34

Grand Prix. Yeah, and it's interesting

32:36

on a slightly more technical level,

32:38

because until now, if an AI...

32:41

system like a chat GPT wanted

32:43

to interact with some other website,

32:45

it had to do so through

32:47

an API, right? APIs, application program

32:49

interfaces are sort of the way

32:52

that computers talk to each other,

32:54

but what operator does is essentially

32:56

eliminate the need for APIs because

32:58

it can. just click around on

33:00

a normal website that is designed

33:03

for humans and behave like a

33:05

human and you don't need a

33:07

special interface to do that. Yeah,

33:09

and now some people might hear

33:11

that, Kevin, and start screaming because

33:13

what they will say is APIs

33:15

are so much more efficient than

33:17

their operator is doing here. APIs

33:19

are doing here. APIs are very

33:21

structured. They're very fast. They let

33:23

computers talk to each other without

33:25

having to, for example, open up

33:27

a browser. have to be built.

33:29

There is a finite number of

33:32

them. The reason that Open AI

33:34

is going through this exercise is

33:36

because they want a true general-purpose

33:38

agent that can do anything for

33:40

you, whether there is an API

33:42

for it or not. And maybe

33:44

we should just pause for a

33:46

minute there and zoom out a

33:49

little bit to say, why are

33:51

they building? That's like, what is

33:53

the long-term vision here? Sure. So

33:55

the vision is to create virtual

33:57

co-workers, Kevin. create some kind of

33:59

digital. entity that you can just hire

34:01

as a co-worker. The first ones they'll

34:04

probably be engineers because these systems are

34:06

already so good at writing code, but

34:08

eventually they want to create virtual consultants,

34:10

virtual lawyers, virtual doctors, you name it.

34:12

Virtual podcast hosts? Let's hope they don't

34:14

go that far. But everything else is

34:16

on the table. And if they can

34:19

get there, presumably there are going to

34:21

be huge profits in it for them.

34:23

They're going to potentially be huge productivity

34:25

gains for companies. And then there is,

34:27

of course, the question of, well, what

34:29

does this mean for human beings? And

34:31

I think that's somewhat workier. Right. And

34:34

I think there's also, it also helps

34:36

to justify the cost of running these

34:38

things because $200 a month is a

34:40

lot to pay for a remote worker.

34:42

And if you could, say, use the

34:44

next version of operator, or maybe two

34:47

or three versions from now, to say,

34:49

replace a customer service agent or someone

34:51

in your billing department, that actually starts

34:53

to look like a very good deal.

34:55

Absolutely, or even if I could bring

34:57

it into the realm of journalism, Kevin,

34:59

if I had a virtual research assistant

35:02

and I said, hey, I'm going to

35:04

write about this today, go pull all

35:06

of the most relevant information about this

35:08

from the past couple of years and

35:10

maybe organize it in such a column

35:12

based. off of it, like yeah, that's

35:14

absolutely worth $200 a month to me.

35:17

Okay, so Casey, walk me through something

35:19

that you actually asked operator to do

35:21

for you and what it did autonomously

35:23

on its own. Sure. I'll maybe give

35:25

like two examples, like a pretty good

35:27

one and maybe a not so good

35:30

one. Pretty good one was, and this

35:32

was it actually suggested by operator. I

35:34

used trip advisor to look up walking

35:36

tours in London that I might want

35:38

to do. I'm not actually going to

35:40

London. Oh, so you lied to the

35:42

AI? And not for the first time.

35:45

But here's what I'll say. If anybody

35:47

wants to break heaven in London, I'll

35:49

get in touch. We love the city.

35:51

Yep. So I said, OK, operator, sure,

35:53

let's do it. Let's find. me some

35:55

walking tours. I clicked that it opened

35:58

a browser. It went to TripAdvisor, it

36:00

searched for Luden Walking Tours, it read

36:02

the information on the website, and then

36:04

it presented it to me, did that

36:06

within a couple of minutes. Now, on

36:08

one hand, could I have done that

36:10

just as easily by Google? Could I

36:13

probably have done it even faster if

36:15

I'd done it myself? Sure. But if

36:17

you're just sort of interested in the

36:19

technical feat that is getting one of

36:21

these models to open a browser to

36:23

navigate to a website, navigate to a

36:25

website, read it and share information, computer

36:28

using itself and you know going around

36:30

like typing things and selecting things from

36:32

drop-down menus yeah it's sort of like

36:34

you know if you think it is

36:36

cool to be in a self-driving car

36:38

like this is that but for your

36:41

web a self-driving browser it is a

36:43

self-driving browser so that's the good example

36:45

yes what was another example so another

36:47

example and this was something else that

36:49

open AI suggested that we try was

36:51

to try to use operator to buy

36:53

groceries and they have a partnership with

36:56

instakart And so I thought, okay, they're

36:58

gonna have like sort of dialed this

37:00

in so that there's a pretty good

37:02

experience. And so I said, okay, let's

37:04

go ahead and buy groceries and I

37:06

went to operator and I said something

37:09

like, hey, can you help me buy

37:11

groceries on Instagram? And it said, sure.

37:13

And here's what it did. It opened

37:15

up, Instagram, in a browser, so far,

37:17

so good. And then it started searching

37:19

for milk in stores located in Des

37:21

Moines, Iowa. Now, you do not live

37:24

in Des Moines, Iowa, so why did

37:26

it think that you did? As best

37:28

as I can tell, the reason it

37:30

did this is that Instacart defaults to

37:32

searching for grocery stores in the local

37:34

area and the server that this instance

37:36

of operator was running on was in

37:39

Iowa. Now, if you were designing a

37:41

grocery product like Instacart, and Instacart does

37:43

this, when you first sign on and

37:45

say you're looking for groceries, it will

37:47

say, quite sensibly, where are you. Instagram

37:49

might also offer suggestions for things that

37:52

you might want to buy. It does

37:54

not just assume that you want milk.

37:56

Wow, I'm just picturing like a house

37:58

in Des Moines Iowa where there's just

38:00

like a palette of milk being delivered

38:02

every day from all these poor operator

38:04

users. Yes. So I thought, okay, whatever,

38:07

you know, this thing makes mistakes. Let's

38:09

hope that it gets on the right

38:11

track here. And so I tried to

38:13

pick the grocery store that I wanted

38:15

it to shop at, which is, you

38:17

know, in San Francisco where I live,

38:20

and it entered that grocery store's address

38:22

as the delivery address. So like it

38:24

would try to deliver groceries presumably from

38:26

Des Moines Iowa to my grocery store,

38:28

which is not what I wanted. And

38:30

it actually could not. solve this problem

38:32

without my help. I had to take

38:35

over the browser, log into my Instacart

38:37

account, and tell it which grocery store

38:39

that I wanted to shop it. So

38:41

already, all of this has taken at

38:43

least 10 times as long as it

38:45

would have taken me to do this

38:47

myself. Yeah, so I had some similar

38:50

experiences. The first thing that I had

38:52

operator tried to do for me was

38:54

to buy a domain name and set

38:56

up a web server for a project

38:58

that you and I are working on

39:00

that we can't really talk about yet.

39:03

Secret project. Secret project. And so I

39:05

said to operator, I said, go research

39:07

available domain names related to this project,

39:09

buy the one that costs less than

39:11

$50. And then by hosting it. and

39:13

set it up and configure all the

39:15

DNS settings and stuff like that. Okay,

39:18

so that's like a true multi-step project

39:20

and something that would have been legitimately

39:22

very annoying to do yourself. Yes, you

39:24

know, that would have taken me, I

39:26

don't know, half an hour to do

39:28

on my own, and it did take

39:30

operator some time. Like I had to

39:33

kind of like set it and forget

39:35

it, and like I, you know, got

39:37

myself a snack and a cup of

39:39

coffee, and then when I came back,

39:41

it had done most of these tasks.

39:43

the browser and enter my credit card

39:46

number I had to give it some

39:48

details about like my address for the

39:50

sort of registration for the domain name

39:52

I had to pick between the various

39:54

hosting plans that were available on this

39:56

website but It did 90% of the

39:58

work for me. And I just had

40:01

to sort of take over and do

40:03

the last mile. And this is really

40:05

interesting because what I would assume was

40:07

it would get like, I don't know,

40:09

5% of the way and it would

40:11

hit some hicup and it just wouldn't

40:14

be able to figure something out until

40:16

you came back and saved it. But

40:18

it sounds like from what you're saying

40:20

was, it was somehow able to work

40:22

around whatever unanswered questions there were and

40:24

still get a lot done while you

40:26

weren't paying attention. It felt a little

40:29

bit like training like a very new

40:31

very insecure intern because like it at

40:33

first it would keep prompt me be

40:35

like well do you want a.com or

40:37

a dot net? And eventually you just

40:39

have to prompt it and say, like,

40:41

make whatever decisions you want. Like, wait,

40:44

you said that to it. Yes, I

40:46

said, like, only ask for my intervention

40:48

if you can't progress any farther, otherwise

40:50

just make the most reasonable decision. You

40:52

said, I don't care how many people

40:54

you have to kill. Just get me

40:57

this domain. And it said, understood, sir.

40:59

Yeah, and I'm now wants it in

41:01

42 states. Anyway, that was one thing

41:03

that operator did for me that was

41:05

pretty impressive. That feels like a grand

41:07

success compared to what I got operator

41:09

to do. Yeah, it was pretty impressive.

41:12

I also had to send lunch to

41:14

one of my coworkers, Mike Isaac, who

41:16

was hungry, because he was on deadline,

41:18

and I said go to DoorDash and

41:20

get Mike some lunch. It did initially

41:22

mess up that. process because it decided

41:25

to send him tacos from a taco

41:27

place which you know is great and

41:29

it's a taco place I know it's

41:31

very good but I said order enough

41:33

for two people and sort of ordered

41:35

two tacos and this is one of

41:37

those places where the tacos are quite

41:40

small operator said get your portion size

41:42

under control America yeah so then I

41:44

had to go in and say does

41:46

that sound like enough food operator and

41:48

it said actually now that you mentioned

41:50

I should probably order more wait no

41:52

so here's a question so in these

41:55

cases is the first step that you

41:57

log into your account because it doesn't

41:59

have any of your payment details or

42:01

anything so at what point are you

42:03

actually sort of teaching at that it

42:05

depends on the website so sometimes you

42:08

can just say up front like here

42:10

is my email address or here is

42:12

my login information and it will sort

42:14

of you know log you in and

42:16

do all that. Sometimes you take over

42:18

the browser. There are some privacy features

42:20

that are probably important to people where

42:23

it says Open AI says that it

42:25

does not take screenshots of the browser

42:27

while you are in control of it

42:29

because you might not want your credit

42:31

card information getting sent to open AI

42:33

servers or anything like that. So sometimes

42:36

it happens at the beginning of the

42:38

process, sometimes it happens like when you're

42:40

checking out at the end. And so

42:42

were you taking it over to log

42:44

in or were you saying, I don't

42:46

care, and you just like we're giving

42:48

operator your door dash password and plain

42:51

text? I was taking it over. Okay,

42:53

smart, smart, smart. So. Those were the

42:55

good things I also this was a

42:57

fun one. I I wanted to see

42:59

if operator could make me some money

43:01

So I said go take a bunch

43:03

of online surveys because you know there

43:06

are all these websites where you can

43:08

like get a couple cents for like

43:10

filling out an online survey Something that

43:12

most people don't know about Kevin is

43:14

he devotes 10% of his brain at

43:16

any given time to thinking about schemes

43:19

to generate and it's one of my

43:21

favorite Aspects of your personality that I

43:23

feel like doesn't get exposed very much,

43:25

but this is truly the most rusian

43:27

approach to using operator I can imagine

43:29

So I can't wait to find out

43:31

how this went. Well the most rusian

43:34

approach may might have been what I

43:36

tried just before this which was to

43:38

have it go play online poker for

43:40

me But it did not do it.

43:42

It said I can't help with gambling

43:44

or lottery related activities. Okay woke AI

43:46

Does the Trump administration know about this?

43:49

But it was able to actually fill

43:51

out some online surveys for me and

43:53

it earned a dollar and twenty cents.

43:55

Is that right? Yeah, in about 45

43:57

minutes. So if you had it going

43:59

all month, presumably you could maybe eke

44:02

out the $200 to cover the cost

44:04

of operator, pro? Yes, and I'm sure

44:06

I spent hundreds of dollars worth of

44:08

GPU computing power just to be able

44:10

to make that dollar and 20 cents.

44:12

But hey, it worked. But hey, it

44:14

worked. So those were some of the

44:17

things that I tried. There were some

44:19

other things that it just would not

44:21

do for me, no matter how hard

44:23

I tried. One of them. So one

44:25

of them was to. I was trying to

44:27

update my website and put some

44:30

links to articles that I'd written

44:32

on my website. And what I

44:34

found after trying to do this

44:36

was that there are just websites

44:38

where operator is not allowed to

44:40

go. And so when I said

44:42

to operator, go pull down these

44:44

New York Times articles that I

44:47

wrote and put them onto my

44:49

website, it said I can't get

44:51

to the New York Times website. I'm

44:53

going to guess you expected that to

44:55

happen. Well, I thought maybe it has

44:58

some clever work around, and maybe I

45:00

should alert the lawyers at the New

45:02

York Times, if that's the case. But

45:04

no, I assumed that if any website

45:07

were to be blocking the open

45:09

AI web crawlers, it would be

45:11

the New York Times. Yeah. There

45:13

are other websites that have also

45:15

put up similar blockades to prevent

45:17

operator from crawling them, read it,

45:19

you cannot go on to with

45:21

operator, YouTube, you cannot go on

45:24

to with operator, various other websites,

45:26

go daddy for some reason, did

45:28

not allow me to use operator

45:30

to buy a domain name there,

45:32

so I had to use another

45:34

domain name site to do that.

45:36

So right now there are some

45:38

pretty janky... parts of operator. I would

45:40

not say that most people would get

45:42

a lot of value from using it.

45:45

But what do you think? Well, I

45:47

do think that there is something just

45:49

undeniably cool about watching a computer

45:51

use itself. Of course, it can

45:54

also be quite unsettling. A computer

45:56

that can use itself can cause

45:58

a lot of harm. think that

46:00

it can do a lot of

46:02

good and so it was fun

46:04

to try to explore what some

46:06

of those things could be. And

46:08

to the extent that operator is

46:11

pretty bad at a lot of

46:13

tasks today, I would point out

46:15

that it showed pretty impressive gains

46:17

on some benchmarks. So there is

46:19

one benchmark, for example, that Anthropic

46:21

used when they unveiled computer use

46:23

last year, and they scored 14.9%

46:25

on something called OS World, which

46:27

is an evaluation for testing agent,

46:29

so not great. Just three months

46:31

later, Open AI said that its

46:33

Kua model scored 38.1% on the

46:35

same evaluation. And of course, we

46:37

see this all the time in

46:39

AI where there's just this very

46:42

rapid progress on these benchmarks. And

46:44

so on one hand, 38.1% is

46:46

a failing grade on basically any

46:48

test. On the other hand, if

46:50

it improves at the same rate

46:52

over the next three to six

46:54

months, you're gonna have a computer

46:56

that is very good at using

46:58

itself, right? So that I just

47:00

think is worth noting. Yes, I

47:02

think that's plausible. We've obviously seen

47:04

a lot of different AI products

47:06

over the last couple of years

47:08

start out being pretty mediocre and

47:10

get pretty good within a matter

47:12

of months. But I would give

47:15

one cautionary note here, and this

47:17

is actually the reason that I'm

47:19

not particularly bullish about these kind

47:21

of browser using AI agents. I

47:23

don't think the internet is going

47:25

to sit still and allow this

47:27

to happen. The internet is built

47:29

for humans to use, right? It

47:31

is every news publisher that shows

47:33

ads on their website, for example,

47:35

prices those ads based on the

47:37

expectation that humans are actually looking

47:39

at them. If browser agents start

47:41

to become more popular and all

47:43

of a sudden 10 or 20

47:46

or 30 percent of the visitors

47:48

to your website are not actually

47:50

humans but are instead operator or

47:52

some similar system, I think that

47:54

starts to break the... assumptions that

47:56

power the economic model of a

47:58

lot of the internet. Now is

48:00

that still true if we find

48:02

that the agents actually get persuaded

48:04

by the ads and that if

48:06

you send operator to buy door

48:08

dash and it sees an ad

48:10

for McDonald's it's like you know

48:12

what that's a great idea I'm

48:14

gonna ask Kevin if he actually

48:16

wants some of that. Totally Totally,

48:19

I actually think you're joking, but

48:21

I actually think that is a

48:23

serious possibility here is that people

48:25

who, you know, build e-commerce sites,

48:27

Amazon, etc. start to put in

48:29

basically signals and messages for browser

48:31

agents to look at on their

48:33

website to try to influence what

48:35

it ends up buying. And I

48:37

think you may start to see

48:39

restaurants popping up in certain cities

48:41

with names like operator, pick me,

48:43

or order from this one, Mr.

48:45

That's maybe a little extreme, but

48:47

I do think that there's going

48:49

to be a backlash among websites,

48:52

publishers, e-commerce, vendors, as these agents

48:54

start to take off. I think

48:56

that that is reasonable. I'll tell

48:58

you what I've been thinking about

49:00

is how do we turn this

49:02

tech demo into a real product?

49:04

And the main thing that I

49:06

noticed when I was testing operator

49:08

was there is a difference between

49:10

an agent that is using a

49:12

browser and an agent that is

49:14

using your browser. When an agent

49:16

is able to use your browser,

49:18

which it can't right now, it's

49:20

already logged into everything. It already

49:23

has your payment details. It can

49:25

do everything so much faster and

49:27

more seamlessly and without as much

49:29

hand-holding. Of course, there are also

49:31

so many more privacy and security

49:33

risks that would come from entrusting

49:35

an agent with that kind of

49:37

information. So there is some sort

49:39

of chasm there that needs to

49:41

be closed and I'm not quite

49:43

sure how anyone does it. but

49:45

I will tell you I do

49:47

not think the future is opening

49:49

up these virtual browsers and me

49:51

having to enter all of my

49:53

login and payment details every single

49:56

time I want to do anything

49:58

on the internet because truly I

50:00

would rather just do it myself.

50:02

Right. I also think there's just

50:04

a lot more potential for harm

50:06

here. A lot of AI safety

50:08

experts I've talked to are very

50:10

worried about this because what you're

50:12

essentially doing is letting the AI

50:14

models make their own decisions and

50:16

actually carry out tasks. And so

50:18

you can imagine a world where

50:20

an AI agent that's very powerful,

50:22

a couple versions from now, decides

50:24

to start. doing cyber attacks because

50:27

maybe some malevolent user has told

50:29

it to make money and it

50:31

decides that the best way to

50:33

do that is by hacking into

50:35

people's crypto wallets and stealing their

50:37

crypto. Yeah, so those are the

50:39

kinds of reasons that I am

50:41

a little more skeptical that this

50:43

represents a big breakthrough, but I

50:45

think it's really interesting and it

50:47

did give me that feeling of

50:49

like, wow, this could get really

50:51

good, really fast, and if it

50:53

does, the world will look very

50:55

different. When we come back, Kevin,

50:57

back that caboose up. It's time

51:00

for the Hot Mess Express. You

51:02

know, Roose Caboose was my nickname

51:04

in middle school. Kevin Caboose. Choo-choo.

51:19

Whether you're starting or scaling your

51:21

company's security program, demonstrating top-notch security

51:23

practices and establishing trust is more

51:25

important than ever. Vanta automates compliance

51:28

for SOC 2, ISO 27001, and

51:30

more. With Vanta, you can streamline

51:32

security reviews by automating questionnaires and

51:34

demonstrating your security posture with a

51:36

customer-facing trust center. Over 7,000 global

51:39

companies use Vanta to manage risk

51:41

and prove security in real-time. Get

51:43

a thousand dollars off Vanta when

51:45

you go to vanta.com/hard fork. That's

51:47

vanta.com/hard fork for a thousand dollars

51:50

off. I'm Julie Turkwits. I'm a

51:52

reporter at the New York Times

51:54

to understand changes in migration. traveled

51:56

to the Darien Gap. Thousands have

51:58

been risking their lives to pass

52:01

through the border of Colombia and

52:03

Panama in the hopes of making

52:05

it to the United States. We

52:07

interviewed hundreds of people to try

52:09

and grasp what's making them go

52:12

to these lengths. New York Times

52:14

journalists spend time in these places

52:16

to help you understand what's really

52:18

happening there. You can support this

52:20

kind of journalism by subscribing to

52:23

the New York Times. Well

52:26

Casey we're here wearing our trained

52:29

conductor hats and my child's train

52:31

set is on the table in

52:33

front of us Which can only

52:36

mean one thing we're going to

52:38

train a large language model. Nope.

52:40

That's not what that means. It

52:43

means it's time to play a

52:45

game of the hot mess express

52:48

Paws for theme song Hot

52:53

Mess Express, Kevin is our segment where

52:55

we run through some of the messiest

52:57

recent tech stories and deploy our official

53:00

hot mess thermometer to tell you just

53:02

how messy we think things have gotten.

53:04

And Kevin, you better sit down for

53:07

this one. This is about a messy

53:09

week. Sure has. So why don't we

53:11

go ahead? Fire up the Hot Mess

53:14

Express and see what is the first

53:16

story coming down the charts. I hear

53:18

a faint chugga, chugga in my headphones.

53:21

Oh, it's pulling into the station. Casey,

53:23

what's the first cargo that our hot

53:25

mess express is carrying? All right, Kevin,

53:27

this first story comes to us from

53:30

the New York Times, and it says

53:32

that Fable, a book app, has made

53:34

changes after some offensive AI messages. Now

53:37

Casey, have you ever heard of Fable,

53:39

the book app? Well, not until this

53:41

story, Kevin, but I am told that

53:44

it is an app for sort of

53:46

keeping track of what you're reading, not

53:48

unlike a good reads, but also for

53:51

discussing what you're reading, and apparently this

53:53

app also offers some AI chat. Yeah,

53:55

you can have AI sort of summarize

53:57

the things that you're reading in a

54:00

personalized way, and this story said that

54:02

an... addition to spitting out bigoted and

54:04

racist language, the AI Inside Fable's book

54:07

app had told one reader who had

54:09

just finished three books by black authors,

54:11

quote, your journey dives deep into the

54:14

heart of black narratives and transformative tales,

54:16

leaving mainstream stories gasping for air. Don't

54:18

forget to surface for the occasional white

54:21

author, okay? And another personalized AI summary

54:23

that Fable Produce told another reader that

54:25

their book choices were, quote, making me

54:27

wonder if you're ever in the mood

54:30

for a straight cis white man's perspective.

54:32

And if you are interested in a

54:34

straight cis white man's perspective, follow Kevin

54:37

Roos on x.com. Now, Kevin, why do

54:39

we think this happened? I don't know,

54:41

Casey. This is a headscratcher for me.

54:44

I mean, we know that these apps

54:46

can can spit out. things that is

54:48

just sort of like part of how

54:51

they are trained and part of what

54:53

we know about them. I don't know

54:55

what model Fable was using under the

54:57

hood here, but yeah, this seems not

55:00

great. Well, it seems like we've learned

55:02

a lesson that we've learned more than

55:04

once before, which is that large language

55:07

models are trained on the internet, which

55:09

contains near infinite racism, and for that

55:11

reason, it will actually produce racism when

55:14

you ask it questions. In this case,

55:16

they were not successful. Fable's head of

55:18

community, Kim Marsh Alley, has said that

55:21

all features using AI are being removed

55:23

from the app and a new app

55:25

version is being submitted. to the app

55:27

store. So you always hate it when

55:30

the first time you hear about an

55:32

app is that they added AI and

55:34

it made it super racist and they

55:37

have to redo the app. Now Casey,

55:39

one more question before we move on.

55:41

Do you think this poses any sort

55:44

of competitive threat to Grock, which until

55:46

this story was the leading racist AI

55:48

app on the market? I do think

55:51

so. And I have to admit that

55:53

all the folks over at Grock are

55:55

breathing a sigh of relief now that

55:57

they have once again claimed the mantle.

56:00

All right. Casey, how hot is this

56:02

mess? Well Kevin, in my opinion, if

56:04

your AI is so bad that you

56:07

have to remove it from the app

56:09

completely, that's a hot mess. Yeah, I

56:11

rate this one a hot mess as

56:14

well. All right, next stop. Amazon pauses

56:16

drone deliveries after aircraft crashed in rain.

56:18

Casey, this story comes to us from

56:21

Bloomberg, which had a different line of

56:23

reporting than we did just a few

56:25

weeks ago on the show about Amazon's

56:27

drone program, Prime Air. Casey, what happened

56:30

to Amazon Prime Air? Well... If you

56:32

heard the episode of Hard Fork where

56:34

we talked about it, Amazon Prime Air

56:37

delivered us some Brazilian bumbum cream and

56:39

it did so without incident. However, Bloomberg

56:41

reports that Amazon has had to now

56:44

pause all of their commercial drone deliveries

56:46

after two of its latest models crashed

56:48

in rainy weather at a testing facility.

56:51

And so the company says it is

56:53

immediately suspending drone deliveries in Texas and

56:55

Arizona and will now fix the aircraft's

56:57

software. Kevin, how did you react to

57:00

this? deliveries before they fixed the software

57:02

because these things are quite heavy, Casey.

57:04

I would not want one of them

57:07

to fall in my head. I wouldn't

57:09

either. And I have to tell you,

57:11

this story gave me the worst kind

57:14

of flashbacks because in 2016, I wrote

57:16

about Facebook's drone, Echila, and its first,

57:18

what the company told me, had been

57:21

its first successful test flight in its

57:23

mission to deliver internet around the world

57:25

via drone. What the company did not

57:27

tell me when I was interviewing its

57:30

executives, including Mark Zuckerberg. the plane had

57:32

crashed after that first flight. And so

57:34

I was small detail. I'm sure it

57:37

was an innocent omission from their briefing.

57:39

Yes, I'm sure. Well, it was Bloomberg

57:41

again who reported, you know, a couple

57:44

months after I wrote this story that

57:46

the Facebook drone had crashed. I was

57:48

of course, hugely embarrassed and, you know,

57:51

wrote a bunch of stories about this.

57:53

But anyways, it really should have occurred

57:55

to me when we were out there

57:57

watching the Amazon drone that this thing

58:00

was also probable. secretly crashing and we

58:02

just hadn't found out about it yet

58:04

and indeed we now learn it is.

58:07

So here's my vow to you Kevin

58:09

as my friend and my co-host. If

58:11

we ever see a company fly anything

58:14

again we have to ask them. Did

58:16

this thing actually crash? Yeah. I'm tired

58:18

of being burned. Now Casey, we should

58:21

say, according to Bloomberg, these drones reportedly

58:23

crash in December. We visited Arizona to

58:25

see them in very early December. So

58:27

most likely, you know, this all happened

58:30

after we saw them. But I think

58:32

it's a good idea to keep in

58:34

mind that as we're talking about these

58:37

new and experimental technologies. that many of

58:39

them are still having the kinks worked

58:41

out. All right, Kevin, so let's get

58:44

out the thermometer. How hot of a

58:46

mess is this? I would say this

58:48

is a moderate mess. Look, these are

58:51

still testing programs. No one was hurt

58:53

during these tests. I am glad that

58:55

Bloomberg reported on this. I'm glad that

58:57

they've suspended the deliveries. These things could

59:00

be quite dangerous flying through the air.

59:02

I do think it's one of a

59:04

string of reported. incidents with these drones.

59:07

So I think they've got some quality

59:09

control work ahead of them and I

59:11

hope they do well on it because

59:14

I want these things to exist in

59:16

the world and be safe for people

59:18

around them. All right. I will agree

59:21

with you and say that this is

59:23

a warm mess and hopefully you can

59:25

get straightened out over there. Let's see

59:27

what else is coming down the tracks.

59:30

Fitbit has agreed to pay $12 million

59:32

for not quickly reporting burn risk with

59:34

watches. Kevin, do you hear about this?

59:37

I did. This was the fitbit. Devices

59:39

were like literally burning people. Yes, from

59:41

2018 to March of 2022, Fitbit received

59:44

at least a hundred and seventy four

59:46

reports globally of the lithium ion battery

59:48

in the Fitbit Ionic watch overheating, leading

59:51

to a hundred and eighteen reported injuries,

59:53

including two cases of third degree burns

59:55

and four of second degree burns. That

59:57

comes from the New York Times. Deal

1:00:00

Hassan, Kevin, I thought these things were

1:00:02

just supposed to burn calories. Well, it's

1:00:04

like I always say, exercising is very

1:00:07

dangerous and you should never do it.

1:00:09

And this justifies my decision not to

1:00:11

wear a fit bit. To me, the

1:00:14

biggest surprise of this story was that

1:00:16

people were wearing fit bits from March

1:00:18

2018 to 2022. I thought every fitbit

1:00:21

had been purchased by like 2011 and

1:00:23

then put in a drawer never to

1:00:25

be heard again. So what is going

1:00:27

on with these sort of late stage

1:00:30

fitbit buyers? I'd love to find out.

1:00:32

But of course. we feel terrible for

1:00:34

everyone who was burned by a fit

1:00:37

bit and it's not gonna be the

1:00:39

last time technology burns you. I mean

1:00:41

realistically. That's true. You know? It's true.

1:00:44

Now what kind of mess is this?

1:00:46

I would say this is a hot

1:00:48

mess. This is an officially hot, literally

1:00:51

hot, they're hot. Here's my sort of

1:00:53

rubric. If technology physically burns you, it

1:00:55

is a hot mess. If you have

1:00:57

physical burns on your body, what other

1:01:00

kind of mess could it be? Okay,

1:01:02

next stop on the Hot Mess Express.

1:01:04

Google says it will change Gulf of

1:01:07

Mexico to Gulf of America in Maps

1:01:09

app after government updates. Casey, have you

1:01:11

been following this story? I have, Kevin,

1:01:14

every morning when I wake up I

1:01:16

scan America's maps and I say, what

1:01:18

has been changed? And if so, has

1:01:21

it been changed for political reasons? And

1:01:23

this was probably one of the biggest

1:01:25

examples of that we've seen. Yeah, so

1:01:27

this was an interesting story that came

1:01:30

out. in the past couple of days.

1:01:32

Basically, after Donald Trump came out during

1:01:34

his first days in office and said

1:01:37

that he was changing the name of

1:01:39

the Gulf of Mexico to the Gulf

1:01:41

of America and the name of Denali,

1:01:44

the Mountain in Alaska, to Mount McKinley,

1:01:46

Google had to decide, well, when you

1:01:48

go on Google Maps and look for

1:01:51

those places, what should it call them?

1:01:53

It seems to be saying that it

1:01:55

is going to take inspiration from the

1:01:57

Trump administration and update the names of

1:02:00

these places in the maps app. Yeah,

1:02:02

and look, I don't think Google really

1:02:04

had a choice here. that the company

1:02:07

has been on Donald Trump's bad side

1:02:09

for a while, and if it had

1:02:11

simply refused to make these changes, it

1:02:14

would have sort of caused a whole

1:02:16

new controversy for them. And it is

1:02:18

true that the company changes place names

1:02:21

when governments change place names, right? Like

1:02:23

Google Maps existed when Mount McKinley was

1:02:25

called Mount McKinley, and President Obama changed

1:02:27

it to Janali, and Google updated the

1:02:30

map. Now it's changed back. They're doing

1:02:32

the same thing. Kevin, I think there's

1:02:34

room for Donald Trump to have a

1:02:37

lot of fun with the company. Yeah,

1:02:39

what can you do? Well, you could

1:02:41

call it the Gulf of Gemini isn't

1:02:44

very good and just see what would

1:02:46

happen. Because they would kind of have

1:02:48

to just change it. Can you imagine

1:02:51

every time you opened up Google Maps

1:02:53

and you looked at the Gulf of

1:02:55

Mexico slash America and just said the

1:02:57

Gulf of Gemini is not very good?

1:03:00

You know, I hate to give Donald

1:03:02

Trump any ideas, but I don't know.

1:03:04

I think this is a mild mess.

1:03:07

I think this is a tempest in

1:03:09

a teapot. I think that this is

1:03:11

the kind of update that companies make

1:03:14

all the time. Because places change names

1:03:16

all the time. Let's just say it.

1:03:18

Well, Kevin, I guess I would say

1:03:21

that one is a hot mess. Because

1:03:23

if we're just going to start renaming

1:03:25

everything on the map, that's just going

1:03:27

to get extremely confusing for me to

1:03:30

follow. I got places to go. You

1:03:32

go to like three places. Yeah, and

1:03:34

I use Google Maps to get there.

1:03:37

And I need them to be named

1:03:39

the same thing that they were yesterday.

1:03:41

I don't think they're gonna change the

1:03:44

name of Barry's boot camp. All right,

1:03:46

final stop on the Hot Mess Express.

1:03:48

Casey, bring us home. All right. Kevin,

1:03:51

and this is some sad news. Another

1:03:53

Waymo was vandalized. This is from one-time

1:03:55

hard-for-guess Andrew J. Hawkins at The Virgin.

1:03:57

He reports that this Waymo was vandalized

1:04:00

during an illegal street takeover near the

1:04:02

Beverly Center in LA. Video from Fox

1:04:04

11 shows a crowd of people basically

1:04:07

dismantling the driverless. piece by piece and

1:04:09

then using the broken pieces to smash

1:04:11

the windows. Kevin, what did you make

1:04:14

of this? Well Casey, as you recall,

1:04:16

you predicted that in 2025, Waymo would

1:04:18

go mainstream and I think there is

1:04:21

no better proof that that is true

1:04:23

than that people are turning on the

1:04:25

Waymo's and starting to beat them up.

1:04:27

Yeah, I, you know, look, I don't...

1:04:30

know that we have heard in the

1:04:32

interviews from why these people were doing

1:04:34

this. I don't know if we should

1:04:37

see this as like a reaction against

1:04:39

AI in general or of Waymo's specifically.

1:04:41

But I always find it like weird

1:04:44

and sad when people attack Waymos because

1:04:46

they truly are safer cars than every

1:04:48

other car. Well, not if you're going

1:04:51

to be riding in them and people

1:04:53

just going to start like beating the

1:04:55

car, then they're not safer. No, but

1:04:57

you know, that's only happened a couple

1:05:00

times that we're aware of. Right. Yeah.

1:05:02

So yeah, this story is sad to

1:05:04

me. Obviously people are reacting to Waymos.

1:05:07

Maybe they have sort of fears about

1:05:09

this technology or think it's going to

1:05:11

take jobs. or maybe they're just pissed

1:05:14

off and they wanna break something. But

1:05:16

don't hurt the way most people, in

1:05:18

part, because they will remember. They will

1:05:21

remember. They will remember, and they will

1:05:23

come for you. I'm not sure that

1:05:25

that's true, but I think we should

1:05:27

also note that Waymo only became officially

1:05:30

available in LA in November of last

1:05:32

year. And so part of this just

1:05:34

might be a reaction to the newness

1:05:37

of it all and people getting a

1:05:39

little carried away, just sort of curious,

1:05:41

what will happen if we try to,

1:05:44

you know, destroy this thing? Will it

1:05:46

deploy defensive measures and so on? So

1:05:48

they're going to have to put flame

1:05:51

throwers on them. I'm just calling it

1:05:53

right now. one was. I think this

1:05:55

one is a is a lukewarm mess

1:05:57

that has the potential to escalate. I

1:06:00

don't want this to happen. I sincerely

1:06:02

hope this does not happen, but I

1:06:04

can see as Waymo start. being rolled

1:06:07

out across the country, that some people

1:06:09

are just going to lose their minds.

1:06:11

Some people are going to see this

1:06:14

as the physical embodiment of technology invading

1:06:16

every corner of our lives, and they

1:06:18

are just going to react in strong

1:06:21

and occasionally destructive ways. I'm sure that

1:06:23

Waymo has gamed this all out. I'm

1:06:25

sure that this does not surprise them.

1:06:27

I know that they have been asked

1:06:30

about what happens if Waymo's start getting

1:06:32

vandalized and they presumably have plans to

1:06:34

deal with that, including prosecuting the people

1:06:37

who are doing this. But yeah, I

1:06:39

always go out of my way to

1:06:41

try to be nice to Waymo's. And

1:06:44

in fact. Some other Waymo news this

1:06:46

week, Jane Manchin Wong, the security researcher,

1:06:48

reported on X recently that Waymo is

1:06:51

introducing or at least testing a tipping

1:06:53

feature. And so I'm gonna start tipping

1:06:55

my Waymo just to make up for

1:06:57

all the jerks in LA who are

1:07:00

vandalizing them. It looks like the tipping

1:07:02

feature, by the way, will to be

1:07:04

to tip a charity and that Waymo

1:07:07

will not keep that money. At least

1:07:09

that's what's been reported. No, I think

1:07:11

it's going to the flame thorough. for

1:07:14

taking this journey with me. Whether

1:07:32

you're starting or scaling your company's security

1:07:34

program, demonstrating top-notch security practices and establishing

1:07:36

trust is more important than ever. Vanta

1:07:39

automates compliance for SOC 2, ISO 27001,

1:07:41

and more. With Vanta, you can streamline

1:07:43

security reviews by automating questionnaires and demonstrating

1:07:45

your security posture with a customer-facing trust

1:07:48

center. Over 7,000 global companies use Vanta

1:07:50

to manage risk and prove security in

1:07:52

real-time. Get a thousand dollars off Vanta

1:07:54

when you go to... of.com/Hard Fork.

1:07:57

That's Vanta.com slash Hard

1:07:59

Fork for ,000 dollars off. Hard

1:08:01

Fork is produced by Rachel Fork

1:08:04

is produced by We're

1:08:06

and Whitney Jones. by

1:08:08

edited this week

1:08:10

by Rachel Dry by Ena

1:08:13

by Today's show Today's

1:08:15

show was engineered by

1:08:17

Dan Powell. Original

1:08:19

music by Wong and Dan and

1:08:22

Dan Powell. Our executive

1:08:24

producer is Jen Our Our

1:08:26

audience editor is Gololioli. Video production

1:08:28

Ryan Manning and and Chris Shot.

1:08:30

You can watch this

1:08:32

whole episode on on YouTube.com slash

1:08:34

slash Hard Special thanks to to

1:08:36

Paul Shuman, Puing Tam, Dahlia Hidad, and Jeffrey Miranda.

1:08:39

You You can email us

1:08:41

at Hard Fork at y.com with what

1:08:43

you're calling the Gulf

1:08:45

of Mexico. of Mexico.