0:01

You know, it's like every man of a certain age, all

0:03

he wants to do is just sit on the couch and

0:05

watch the History Channel and read about how the Americans fought

0:08

World War II. And like watching my own dad do this

0:10

growing up, I thought, oh, that's like, that's so nice, you

0:12

know, how nice that that story ended

0:14

so positively. And like he fast forward to 2024, it's

0:16

like, is World War II over? No, we're still fighting

0:18

it out of the comments section. Yep.

0:26

I'm Kevin Arus, a tech columnist at The

0:28

New York Times. I'm Casey Newton from Platformer.

0:30

And this is Hard Fork. This week, how

0:32

Substax Nazi problem left me with a hard

0:35

question of my own. Then, The Wall Street

0:37

Journal's Kirsten Grind joins us to talk about

0:39

her reporting on Elon Musk's drug

0:41

use. And finally, how

0:43

AI helped researchers discover a

0:45

new class of antibiotics. It's

0:48

a drug show. So, Casey, on

0:50

this show, we sometimes answer

0:54

hard questions from our

0:56

listeners about ethical dilemmas

1:08

that they are dealing with in their

1:10

lives. And this week,

1:12

you have actually been dealing with your own

1:15

very hard question. It's about

1:17

Substax and Nazis and the

1:19

future of Platformer, your newsletter

1:21

business. And so I want

1:23

to talk about that this week. And I want

1:25

to just make clear that we're not just talking

1:27

about this, because this is a thing that has

1:30

happened to you. But I think it really is

1:32

sort of a microcosm for some of these larger

1:34

debates that we cover on this show about free

1:36

speech and content moderation and the role of internet

1:39

platforms in policing the public square. Well, I want

1:41

to talk about Elon Musk's drug use, Kevin, but

1:43

I'm open to any questions you have about me.

1:47

OK. And I

1:49

think we should just also acknowledge that

1:51

this is going to feel a little

1:53

weird, because even though we are both

1:55

journalists who have covered content moderation by

1:57

big tech platforms and weighed in on.

2:00

on various controversies involving people like

2:02

Alex Jones or whatever, this

2:04

is an instance in which you are actually

2:07

directly involved in the controversy, not

2:09

only because you are covering it and have become

2:11

sort of part of the news story, but because

2:13

you run a business on Substack and are sort

2:15

of directly financially involved in this story. Yes, that's

2:17

very true. And listeners should just sort of like

2:19

keep that in mind as we are talking about

2:22

this. It's like, yeah, this is a weird case

2:24

where we're actually talking about my business. I think

2:26

if you sort of bracket out all of the

2:28

business implications for me, though, there is still

2:30

a really important story to be told about

2:33

how the modern internet should work and what

2:35

people should be allowed to say there. Yeah,

2:37

and it goes to one of the questions

2:39

that we have asked on this show before,

2:41

which is like, when do you know that

2:43

it is time to leave a platform? And

2:46

how do you draw your own

2:48

personal line for what is and

2:50

isn't acceptable on the internet? Absolutely.

2:53

So let's just talk about the

2:56

nuts and bolts of what has happened. Over

2:58

the last few weeks, Substack has

3:00

been fielding lots of criticism about

3:02

its content moderation policies and

3:05

specifically how it treats

3:07

pro-Nazi content. And we'll

3:09

get into what we mean by pro-Nazi content

3:11

in a minute, but this is sort of

3:14

something that has flared up for them in

3:16

the past and that flare up again recently

3:18

and came to a head just the other

3:20

day when Substack announced that it would take

3:22

down some newsletters that promoted Nazi ideas and

3:25

ideology, but wouldn't make changes

3:27

to its broader content moderation policy,

3:29

which it has described as decentralized

3:32

and hands off. And

3:34

Casey, you as a Substack partner, I guess,

3:36

you publish on Substack and have, since you

3:38

started your newsletter, you have ended up in

3:40

the middle of this story. So I think

3:42

we should start with just like, what is

3:45

your news? What do you have to announce

3:47

today? So I've

3:49

decided this week that Platformer is going

3:51

to move off of Substack. So by

3:53

next week, we will have a new

3:56

website and we'll no longer be part

3:58

of that network. Yeah. Can

4:00

you just tell the story of

4:02

your relationship with Substack, maybe starting

4:04

from when it started, when

4:07

you started Platformer? Yeah. So, you know, Substack

4:09

has been around since 2017. And

4:12

it was actually around the time that Substack

4:15

started that I started to write another newsletter

4:17

for The Verge on another platform because Substack

4:19

didn't exist yet. But in 2020, I left

4:21

to start Platformer, my own email newsletter. And

4:24

Substack was the best tool to do that at the

4:26

time. They made it very, very simple to do so.

4:29

It was very fast. And so since

4:31

October 2020, I've been there. If

4:33

you're not familiar with Substack, the basic idea is that

4:35

while anyone can set up a free newsletter and send

4:38

it out to as many people as you can get

4:40

to subscribe to you, if you want to build a

4:42

business, you just connect your Substack to a Stripe account,

4:44

and then you can sell subscriptions. So, you know, in

4:46

my case, people pay 10 bucks a month or 100

4:48

bucks a year. And in exchange for that, you

4:51

get three newsletters every week. So that's

4:53

how the business works. And

4:55

for some number of people, it

4:58

has been amazing. It's built these

5:00

incredible businesses. And I

5:02

think beyond that, Substack has also created

5:04

a really large cultural footprint, right? It's

5:07

not just journalists like me who are on there.

5:09

There are a lot of artists, you know, like

5:11

the novelist George Saunders is on there. You know,

5:14

some of my favorite cooking writers are on the

5:16

platform. Some of my favorite musicians, like Patti Smith,

5:18

are on the platform. So at

5:20

a time when the media industry is contracting and

5:23

it often feels really scary and bad, Substack

5:25

has been this real bright spot where if you

5:27

go there, chances are you'll find something there that

5:29

is like really cool, that's really well suited to

5:31

your interest. Yeah, I'd say it's been one of

5:33

the biggest changes in the media ecosystem over the

5:36

past few years. It's just this

5:38

sort of transition where a lot of journalists

5:40

and writers and creators of all sorts have

5:42

decided to sort of hang out their own

5:44

shingle, set up a Substack and start charging

5:46

people directly for it rather than sort of

5:49

joining some larger media company. Yeah, that's right.

5:51

And my understanding is that Substack takes like

5:53

a 10% cut of everything

5:55

that you and other Substack creators who

5:57

have paid newsletters charge customers on the

5:59

platform. the platform. That's right. So

6:01

I think it's fair to say that

6:03

in addition to musicians and

6:06

cooking writers and journalists, Substack has

6:08

also become home to sort of

6:11

this like alternative media network, these

6:13

people who are sort of dissatisfied

6:15

or disgruntled with sort of mainstream

6:18

media, people like Glenn Greenwald and

6:20

Matt Taibbi and Barry Weiss were

6:23

sort of like these dissenters

6:25

from sort of media orthodoxy and sort

6:27

of more right wing folks have set

6:30

up on Substack and for them it

6:32

sort of seemed like this was the way

6:34

to sort of avoid censorship, right? If you

6:37

were on Substack rather than working at a

6:39

big media institution, no one could tell you

6:41

what to publish and not publish. And for

6:43

them that was part of the appeal. Yeah.

6:45

And this was something that the founders of

6:47

Substack really touted when people would ask them

6:49

about it. You know, they very much leaned

6:52

into the idea that there was too much

6:54

orthodoxy in the mainstream media and that Substack

6:56

would be a place where people could come

6:58

and say just about anything and that Substack

7:00

was always going to take a really laissez-faire

7:02

approach to moderating that stuff. Yeah. And it

7:05

always seemed like that was

7:07

a premonition to me. Like when Substack's

7:09

executives in the early days started coming

7:11

out and saying like, we're not going

7:13

to ban basically anything because my understanding

7:16

is like they have a content moderation

7:18

policy, but it's very, as you said,

7:20

laissez-faire. It's very permissive when it comes

7:22

to what they will and won't host

7:25

on their platform. And for

7:27

me, that was always, it reminded me of

7:29

the so-called Nazi bar problem, which Mike Masnick,

7:31

the tech blogger has written a lot about.

7:33

And it's basically this sort of perennial thorny

7:36

issue that online platforms face

7:38

when they're tasked with dealing

7:40

with Nazis or people

7:43

with other hateful speech. And

7:45

the Nazi bar problem is sort of this

7:48

maybe apocryphal story about like

7:50

a bar owner who, you know, sees a

7:52

guy come in wearing like Nazi regalia and

7:54

just says like, no, you got to get

7:56

out. You're out. No questions

7:58

asked. And someone else. at the bar is

8:00

like, why'd you do that? The guy was just trying to

8:02

have a drink. Why would you kick him out? And he

8:04

basically explains, look, it starts with one

8:07

Nazi. And then that Nazi gets

8:09

allowed to have a drink at the bar and

8:11

then he brings his friends back. And

8:14

pretty soon you're a Nazi bar and they

8:16

are so entrenched and established that it becomes

8:18

very hard to get rid of them. And

8:21

the point is that you shouldn't be allowed to run

8:23

like this is America, you can have a Nazi bar.

8:25

It's just, you shouldn't be confused

8:27

about what you are if you're letting in

8:30

Nazis. Yeah. And you know, I appreciate that

8:32

analogy. I think it has its limits in

8:34

this case, because on the internet, there just

8:36

are Nazis in most places that will show

8:38

up on any platform. And I think just

8:41

because there are three or four of them

8:43

doesn't mean that you're running a Nazi bar,

8:45

it just means that you have a place

8:47

on the internet, you know, at the same

8:50

time, this did eventually snowball for reasons I'm

8:52

sure we'll get into. Yeah, so let's talk

8:54

about that. So when did the permissive content

8:56

moderation policies of sub stack become an issue

8:58

for you? Well, so in November, a journalist

9:01

named Jonathan M. Katz, who was also my

9:03

college classmate, hi, Jonathan, he wrote a story

9:05

for the Atlantic saying sub stack has a

9:07

Nazi problem. And he went through, he said

9:09

he'd identify 16 cases where

9:11

he felt like there were Nazis on

9:13

the platform and suggested sub stack ought to

9:15

do something about it. Sub stack was pretty

9:18

quiet during that period. But then a

9:20

group of 247 sub stack writers sends this

9:22

open letter asking sub stack, are you planning

9:24

to do anything about this? Does this content

9:27

violate your policies? And then

9:29

on December 21, Hamish McKenzie is

9:31

one of the co founders of sub

9:33

stack wrote a blog post in which

9:35

he said that at sub stack, they

9:37

don't like Nazis. But given that they

9:40

exist, sub stack did not believe that

9:42

censorship was the best approach. And it

9:44

did not believe that demonetizing them, you

9:46

know, preventing them from selling subscriptions was

9:48

the best approach. And so sub stack

9:50

said, if you know, anyone could be

9:53

found to be directly inciting violence, they

9:55

would be removed, but nothing else would

9:57

be removed. I read that

9:59

as a statement essentially declaring that

10:01

Nazis were welcome to sell subscriptions on Substack,

10:03

and that's when I thought, okay, I actually

10:05

have a problem now. Now, I want to

10:08

talk more about your decision, but first, can

10:10

we clarify what we mean by Nazi content?

10:12

Yeah. Because that word, I think, gets thrown

10:14

around a lot and has been

10:16

used to mean lots of different things. So

10:19

what was the content that Jonathan M.

10:21

Katz identified that you took issue with?

10:23

Well, so this was the exact question

10:25

that I had, because the Atlantic article

10:27

doesn't actually link to any of the

10:29

Nazi blogs for good reasons. You don't

10:31

want to give undue amplification to extremist

10:33

material. But I thought, well, if I'm

10:35

going to have to make a decision about my business,

10:37

then the first thing I need to do as a

10:39

reporter is to examine the problem myself, right? So

10:42

I reached out to a number of journalists and

10:44

researchers and asked them to share with me what

10:46

they viewed was the worst of the worst

10:48

content on Substack, and I wound up with

10:50

about 40 different publications that had been flagged

10:52

to me. And together with my colleagues at

10:55

Platformer, Zoe Schiffer and Lindsay Chu, we spent

10:57

a few days just going through those. And

11:00

what I was looking for when I was

11:02

planning to try to flag some things for

11:04

Substack were just literal 1930s Nazis. I

11:07

was looking for people who were

11:09

praising Hitler, who were using Nazi

11:11

iconography like the swastika, who were

11:13

talking about the virtues of German

11:15

national socialism. And I

11:17

decided to myself that if I find any of

11:19

that to send a subject, that's all I'm going

11:21

to send them, because they made it very clear

11:23

that they're not going to do anything about right

11:25

wing extremism generally. But I do want to know

11:27

what they have to say about the literal Nazis.

11:29

Right. The sort of clear cut,

11:32

like you are wearing a swastika or you are

11:34

declaring your affinity for Adolf

11:36

Hitler. Like that's the that's sort of the bar

11:38

that you were looking for. Yeah, that's right. And

11:40

what did you find? Yeah. So

11:43

of all of those, we found six

11:45

things that had not yet been removed by the time

11:48

that we submitted them that we thought to sort of

11:50

clearly met the definition of pro Nazi content. We

11:53

submitted this to Substack and we're just waiting.

11:55

And at this point, there was a little

11:57

bit of drama, Kevin, which is I

11:59

had never intended platform

14:00

in the United States that says we are a welcome

14:02

place for Nazis to monetize. Like this was an extremely

14:04

unusual position for anyone to take. So I was just

14:06

hoping that Substack would come along and say what all

14:08

the other platforms say. And they

14:10

kind of didn't. They said that essentially,

14:13

thank you for raising these

14:15

things. We removed five of

14:17

the six. And if

14:19

other people flag similar content to us,

14:21

we will review it on a case

14:23

by case basis. That's what they said,

14:26

which was a very long way

14:28

of saying we're not changing our policies. And

14:30

I think can be fairly interpreted to

14:33

say there will be some content

14:36

that is widely viewed as praising Nazis

14:38

that Substack for whatever reason is not

14:40

going to remove. They did not offer

14:42

me any kind of additional clarity on

14:44

that point. And so that was kind

14:46

of a disappointing thing. And so I write this news

14:48

story that sort of like they've taken down some material.

14:51

I didn't want to get into like exactly how many

14:53

blogs. This sort of came back

14:55

to bite me, because eventually the whole thing

14:57

did get out. And so now

14:59

there are sort of like two sides who are

15:01

mad at me, you have this sort of, like

15:04

the free Spreech Brigade, that's like, you're

15:06

making them out of a molehill. Oh, you found

15:09

five Nazi blogs. What's the big deal, right? And

15:11

then on the sort of more liberal side that

15:13

wants to see stronger action, they say wait, like

15:15

after this whole thing, Substacks only taking down five

15:17

Nazi blogs, they didn't even take down the sixth

15:19

Nazi blog, they're not even saying they're going to

15:21

change their policy. And so

15:23

the situation just became like even

15:26

more polarized. And again, like, this

15:29

whole thing for me did not start as like,

15:31

I'm going to have the last word about the

15:33

quality of content on Substack, it started in the

15:35

spirit of inquiry of like, well, if I

15:37

find some Nazi blogs, will this company take

15:39

them down? Because to me, that's the first

15:41

step to decide whether to do anything else.

15:43

So I just want to say that because

15:45

it's really unfortunate to me that there's been

15:47

so much focus on the specific number of

15:49

Nazi blogs I reported, when again, we found

15:51

dozens of blogs with some like really disturbing

15:53

material that it's now clear will just be

15:55

up there forever. Yeah, I was I was

15:57

looking through some of these Well,

18:01

you know, if it's if they're using like

18:03

open source software and they find some sort

18:05

of web provider that will accept them I

18:08

think the answer is like basically yes, right if we're

18:10

gonna have an internet that is open to all then

18:12

yes Nazis should be able to send out an email

18:14

newsletter. Okay, that's like that's just something that's gonna happen

18:18

I think where it starts to get more

18:20

complicated is when you

18:22

have built recommendation algorithms

18:24

and other tools that surface this

18:26

content to other people who were not

18:28

looking for it And help these

18:31

folks build audiences and it might be helpful

18:33

to talk about how sub stack has evolved

18:35

over the past couple of years Right because

18:38

when it started it was just kind

18:40

of dumb email infrastructure You sign up

18:42

for your account you start sending out

18:44

your emails and you're not the emails

18:47

Never come anywhere near what I'm doing at platformer,

18:49

right? It's your own thing You're just using their

18:51

infrastructure at that point I probably don't make a

18:53

fuss about it because again, this is just kind

18:55

of the cost of doing business on the internet

18:58

But then sub stacks start to do a few things like

19:00

they'll start sending you a personalized digest based on

19:02

the stuff that you're reading That says you might

19:04

want to read these other publications, right? They

19:07

build this social network called notes where anyone

19:09

could publish anything to it looks a lot

19:11

like Twitter And so if you're a Nazi

19:13

and you want to get some attention You can just start putting

19:15

stuff right in that feed and if I don't

19:17

block it I might see it and so

19:19

now there's a chance that my post and

19:21

platformer are showing right up next to Nazi

19:23

post Well, that doesn't feel good. But also

19:25

because these things are getting all this algorithmic

19:27

amplification It means that their audiences can grow

19:29

It means they can make a

19:31

lot more money than they might otherwise

19:33

and all of a sudden the platform

19:35

is in position of being a sort

19:38

of unwitting assistant Fundraiser and growth hacker

19:40

to people who I believe are very

19:42

dangerous So to me that

19:44

was the kind of threshold that this cross

19:46

where I thought I have to have an

19:48

opinion about this, right? So this is a

19:50

more nuanced argument than I think some people

19:52

like to portray it as which is like

19:54

There's this free speech brigade that wants like,

19:56

you know, every social platform and and website

19:58

to be open to any kind of speech,

20:01

no matter how offensive or potentially

20:03

harmful. And then there are these

20:05

kind of internet hall monitors who

20:07

walk around websites like flagging stuff

20:10

that they find objectionable and saying,

20:12

like, you have to take this

20:14

down. Like, what you're saying is

20:16

there actually are some layers of

20:18

the internet that maybe shouldn't be

20:20

censored, but that when you start

20:22

moving into more recommending content, using

20:24

algorithms to filter and rank content

20:26

for people, showing content to people

20:28

who maybe didn't go looking for

20:30

it, that's where you start to take on

20:32

more responsibility for moderating what's on your service.

20:34

Is that what I'm hearing you say? That's

20:37

right, Kevin. And this is a story we

20:39

know so well. What was your last podcast

20:41

about rabbit hole? It's about this exact same

20:44

phenomenon on YouTube, right? It's not about like

20:46

Nazis and direct monetization, but it is about

20:48

people who are discovering stuff via an algorithm

20:50

on YouTube that is potentially drawing them into

20:53

an ideology that could radicalize them and lead

20:55

to some kind of harm. Right? And when

20:57

I think about the past decade on the

20:59

internet, I think about some of

21:01

the harmful characters who have appeared,

21:03

folks like Alex Jones, folks like

21:06

the QAnon movement. When these

21:08

started, these were just individual posts

21:10

on web pages, posts here on

21:12

a social network, but they were

21:14

able to harness the power of those

21:16

recommendation algorithms to grow large audiences.

21:18

And in the case of Alex

21:20

Jones, really enact real harm against real

21:23

people in ways that platforms just

21:25

kind of took too long to

21:27

catch up to. So as

21:29

I sat with this problem, I just

21:31

kept thinking, I know what

21:33

happens next. I know what happens when

21:36

you build this plumbing into

21:38

your little infrastructure company, you

21:40

help things grow that might

21:42

not otherwise grow. And

21:45

if you're someone like me and you don't want

21:47

that to happen, you know, my only real alternative

21:49

is to just sit back and wait for it

21:51

to happen and then say, well, now that it's

21:53

happened, I can leave, you know, and that doesn't

21:55

feel like a very satisfying solution. Yeah, I mean,

21:57

it's been very bizarre to watch this because as

21:59

you eliminating

24:00

that kind of thinking was

24:02

essentially like their number one priority

24:05

was just like making sure that

24:07

didn't happen. So they really want

24:09

to be a home for the

24:11

absolute maximum amount of speech. And

24:13

that's how it was communicated to me. Now, you know,

24:15

my criticism of that would be, while

24:17

I do believe that the founders are sincere,

24:19

it is also true that that is the

24:21

absolute cheapest way to run a platform. When

24:24

you say almost anything goes, that means you

24:26

don't have to hire content moderators. It means

24:28

you don't have to hire policy people. It

24:30

doesn't mean you have to do these tedious

24:32

ongoing reviews of what is on the platform

24:34

and whether your policies should change. Companies like

24:37

Meta spend conservatively hundreds

24:39

of millions of dollars on this stuff, right?

24:41

And Substack is, you know, relatively small. And

24:43

I can imagine why it wouldn't want to

24:45

do that. But at the end of the

24:47

day, I will give the founders credit for

24:49

the fact that when we had these discussions,

24:51

they wanted to talk about them on the

24:53

principles. Yeah. And did

24:55

Substack actually give you a comment

24:58

about their decision to take

25:00

down these five Nazi blogs?

25:02

Yes. So I

25:05

will read most of the statement that Substack

25:07

sent to me. They said, if and when

25:09

we become aware of other content that violates

25:11

our guidelines, we will take appropriate action. Relatedly,

25:13

we've heard your feedback about Substack's content moderation

25:16

approach, and we understand your concerns and those

25:18

of some other writers on the platform. We

25:21

sincerely regret how this controversy has affected writers

25:23

on Substack. We appreciate the input from everyone.

25:25

Writers are the back one of Substack. And

25:27

we take this feedback very seriously. We're actively

25:29

working on more reporting tools that can be

25:31

used to flag content that potentially violates our

25:33

guidelines. And we will continue working on tools

25:35

for user moderation so Substack users can set

25:37

and refine the terms of their own experience

25:39

on the platform. Okay. So

25:41

that's Substack's position. And we've now heard

25:43

your position. I want to just raise

25:45

a few potential objections to this. If

25:47

people are thinking, well, Casey seems to

25:49

be doing this rashly or blowing things

25:51

out of proportion. I want

25:54

to just give you the chance to respond. So

25:56

I'll put on my free speech warrior hat and

25:58

just run through some of So, Objection 1

26:01

would be the sort of sub-stack

26:03

argument, right? That censoring content on

26:05

the internet, it doesn't make extremists

26:07

go away. We know this because

26:09

platforms have been, you know, demonetizing

26:11

and deplatforming extremist content for years

26:13

now, and these people have not

26:16

gone away. In fact, some

26:18

people would argue that this kind of censorship

26:20

actually makes extremist views worse.

26:22

It gives them a cause

26:24

that they can claim they are being

26:26

martyred for and sort of rallies support

26:28

that way, the way that people like Alex

26:30

Jones have for years. So what's your response to that? Sure.

26:33

So that may be true, but getting rid

26:35

of extremism was not the job that I

26:37

gave sub-stack. I'm not asking sub-stack, make racism

26:39

or extremism go away. I'm asking them for

26:41

a place where I can run my business

26:44

and not have my posts appear next to

26:46

Nazis, right? Because that's not good for

26:48

my business. You know, dozens of people have

26:50

canceled their paid subscriptions to platformer, and many of

26:52

them said, hey, I really like what you do,

26:54

but I can't justify giving you money when I

26:57

know it's going to build Nazi monetization infrastructure. So

26:59

it is absolutely the case that there is

27:01

a demand side for extremism, and that has

27:03

to be solved at a society level, but

27:05

platforms can also do their part to not

27:07

help those movements grow and make money. Right.

27:10

I wonder what you would say to this other objection that

27:12

sub-stacks have raised. They pointed out

27:14

that the number of subscribers that

27:16

the sub-stacks or pro-Nazi sub-stacks had

27:19

was tiny. It was like, you

27:21

know, these were not big, thriving

27:23

publications. These were publications with a

27:25

handful of subscribers. None of

27:27

them had made any money. And so basically,

27:30

this is a tempest in a teapot. Yeah.

27:32

And here is a case where I think everyone is just

27:34

going to have to make up for themselves how big they

27:37

think the problem has to be before they take action. I

27:39

think some people will decide it's going to need to be

27:41

a lot worse than this before it rises to the

27:43

level of my attention. For me, I

27:45

just thought I have seen this movie before. I

27:47

know where this is going. And

27:49

if this is eventually going to lead

27:52

me to have to leave the platform, I would rather just

27:54

do it now and move all my life. Got it. Then

27:56

there's the objection that I actually raised with you when you

27:58

told me last week that you were considering moving

28:00

off sub-stack, I think my response was,

28:03

well, aren't there Nazis on every platform?

28:05

Like you're on Instagram, you're on threads,

28:07

you're on Facebook, you used to be

28:09

on X, you use YouTube. All these

28:12

platforms, if you looked hard enough, would

28:14

have some number of Nazis on

28:16

them using them to spread their message and

28:18

potentially even to make money. So

28:21

basically, there are no pure platforms.

28:24

Absolutely true. And I think what I would

28:26

say is that the platforms that I'm on

28:29

and I'm spending time on, while it's true

28:31

that there are bad things on there, for

28:33

the most part, these platforms at least have

28:35

policies against it. When stuff is flagged to

28:37

them, they do remove it. And

28:40

they don't have to be dragged kicking and

28:42

screaming into doing that. They don't ask their

28:44

own user base to be volunteer content moderators

28:46

for them all the time. So

28:48

to me, that was kind of the bare minimum that

28:50

I was looking for is like, well, is there at

28:52

least an affirmative policy that Nazis are banned here? And

28:54

then maybe we can figure something out. I

28:56

would also say that like, it's different when you're

28:59

running a business on the platform, right? Because

29:02

I am not just having to act on my

29:04

own principles here. I have employees who have opinions

29:06

and we were sort of aligned on this, we

29:08

had to talk it through together. And

29:10

I have customers who are very

29:12

principled. And I should say, because I write a

29:15

lot about content moderation, a lot of my customers

29:17

work in trust and safety and content moderation. They

29:19

have heard the arguments that sub-stack is making before,

29:21

like potentially at their own platforms when their own

29:23

platforms were younger and more naive. And this

29:25

stuff just doesn't fly with them. Okay? They

29:28

just do not accept the arguments that are

29:30

being given to them. So I am in

29:32

the unusual position of having a very savvy

29:34

audience that is very sensitive to this subject.

29:36

And that just made me have to take

29:39

it more seriously than like, can I have

29:41

an Instagram account? I

29:43

want to bring up this last objection

29:45

that I've heard, which was made by

29:47

among other people, Ben Thompson, who writes

29:49

the Stratecory newsletter. It seemed

29:51

like he basically had two qualms with what

29:53

you've done here. One of them was sort

29:55

of this, you know, slippery slope argument about

29:58

once you start arguing for that platform. should take

30:00

down Nazi content, if they do that,

30:03

then you sort of start asking them

30:05

to take down other sorts of objectionable

30:07

content, maybe stuff that's like opposed to

30:09

vaccines or questioning the origins of the

30:11

coronavirus pandemic, things that are

30:14

just sort of controversial and not literal

30:16

Nazis and that there's sort of a

30:18

slippery slope effect there. But he also

30:20

raised the objection that you were essentially

30:23

aiming your guns at the wrong

30:25

part of the stack, as it

30:27

were. And in particular,

30:29

he singled out this line in a story that

30:31

you wrote where you said that you were going to

30:33

be contacting Stripe about this substat

30:36

content moderation issue. Now Stripe is

30:38

a payments processing company. And so

30:40

when people sign up to subscribe

30:42

to a substat publication, Stripe is

30:44

the company that actually takes their

30:46

credit card information and charges that

30:48

credit card. And they also have

30:50

content moderation guidelines for the types

30:52

of payments that they will process.

30:55

And so basically, Ben Thompson said

30:57

by going to Stripe, you were

30:59

essentially escalating this beyond a

31:01

level of reasonable disagreement. Yeah, I think it

31:03

is a fair criticism. I do think that

31:06

if sub sec had said, like, yes, we

31:08

are going to affirmatively say that Nazis are

31:10

banned on this platform, and we will proactively

31:12

remove them, then absolutely, the next week, there

31:14

would be calls to do sort of like

31:16

the next level up. Fortunately, for me, it

31:18

never got that far, because they never made

31:21

the affirmative argument that Nazis were banned. And

31:23

I could just sort of, you know, walk

31:25

away and not not have to wonder about

31:27

that anymore. The thing about the slippery slope

31:29

argument, Kevin, is that it presupposes that

31:31

if we just drew one hard line,

31:33

we could stop talking about the boundaries

31:35

of speech forever. That's not how society

31:37

works. We are constantly renegotiating the boundaries

31:40

for speech of social norms of mores.

31:42

These change all the time. That's what

31:44

society is. It is an ongoing conversation

31:46

about how to be. So the idea

31:48

that you could just write one rule

31:50

and keep it forever is a libertarian

31:52

fantasy. Now, on the stripe side of

31:54

it all, I will admit that was

31:56

me being a little edgy. But

32:00

like, here's the thing. I

32:02

also approached Stripe in a spirit of journalistic

32:04

inquiry. And the inquiry was this. Stripe has

32:06

a policy that says that you're not allowed

32:08

to use their services to like fundraise for

32:11

violent causes. Nazism was one of the

32:13

most famous violent causes of all time. And

32:15

so I thought it was worth sending them an

32:17

email to say, hey, one

32:19

of your customers is saying that Nazis are

32:21

free to set up shop and monetize here.

32:24

Is that consistent with your policies? I sent

32:26

that email. I did not get a response.

32:28

So if Stripe had said, yes, that's fine,

32:30

I would not have led a parade down Main Street like

32:32

calling for the end of Stripe. But I did think it

32:34

was worth sending an email just to ask if it was

32:36

true. Yeah. And I imagine that some

32:38

people will hear about your decision to leave Substack and

32:40

say, well, what more does Casey

32:42

want? Like they took down the Nazi blogs that

32:44

he flagged to them. What's the

32:46

problem here? Sure. So what I was looking

32:49

for was a couple of things. You know, one was

32:51

just to say, like, you know, Nazis

32:53

are not allowed. We will proactively

32:55

monitor for this content. Here's how

32:57

we're going to define what we

32:59

view as Nazis. Right. And

33:01

then I also wanted them to look at that in

33:03

that recommendations infrastructure, because again, that's really the difference here.

33:05

You know, we will be on a new platform next

33:07

week at platformer and there will probably

33:10

be Nazis who are using that infrastructure to send

33:12

emails. The difference is going to be it is

33:14

not attached to a social network that was built

33:16

by our provider. Right.

33:18

There will not be these digest

33:21

emails like recommending the Nazi blogs

33:23

along with mine. Right. So

33:25

if something like once they get their hands around that, they

33:27

would need to come in and they would say that certain

33:30

publications are eligible for promotion and

33:32

recommendation and others are not. YouTube

33:35

did this. Meta has done this. Again, a lot of

33:37

this is just very standard stuff that happens at every

33:39

other platform that I write about. It is Substack that

33:41

is the outlet here. So that's what I wanted to

33:43

see. And this became very clear to me over the

33:45

past couple of weeks that nothing like that is coming.

33:48

Yeah. Casey, I want to say

33:50

something sincere to you, which

33:52

I know is terrifying. I'm

33:55

really proud of you for

33:57

making this stand. agree

34:00

about the finer points of online content moderation

34:02

and what shouldn't, shouldn't be allowed. But

34:05

at the end of the day, this is a judgment call

34:07

that you made. And it

34:09

just comes down to who do you want to do

34:11

business with? What kinds of businesses

34:13

do you want to enrich with

34:15

your labor? And

34:17

given that we're in this sort

34:20

of age of rising anti-Semitism and

34:23

increased polarization of all kinds, like, do you

34:25

really want to be giving

34:27

10% of your revenue to a company that

34:31

will not say, like, we don't want to

34:33

do business with literal actual 1930s Nazis? And

34:39

I just, I think about decisions a

34:41

lot through the framework of like, when

34:43

I'm old, and I'm explaining to my

34:45

grandkids decisions that I made earlier in

34:47

my life, like, will I be proud

34:49

of having made the decision that I

34:51

made? Or will I be ashamed of

34:53

the decision that I made and wish

34:55

that I could redo it? And

34:58

like, whatever happens with platformer on your

35:00

new provider, I just think that this

35:02

is going to be a decision that you feel good

35:04

about. And so I'm proud to

35:06

be your friend and your co-host. And I think we

35:08

should also just declare once and for all that

35:11

the Hard Fork podcast is anti-Nazi. This is

35:13

a Nazi free zone. And there is really

35:15

no wiggle room here. Okay. Stop

35:18

sending us your pictures, Nazis, because you're not coming here.

35:21

Woo! When

35:25

I come back, Elon Musk is officially

35:27

on drugs. We'll talk to the reporter

35:29

who nailed that story down right

35:32

after the break. Hey,

35:54

it's Anna Martin from the New York Times, and

35:56

I'm here to tell you about something for New

35:58

York Times news subscribers. And honestly, if you're

36:01

a podcast fan, you're going to want this. It's

36:03

an app called New York

36:05

Times Audio where you can get the

36:08

latest dispatch. It's 10 a.m. in Kiev.

36:10

It's a really loud night. Perfect your

36:12

technique. A splash of soy sauce and then

36:14

a lot of red pepper flakes. Or contemplate

36:16

the future. A computer program is passing the

36:19

bar exam. And we are over here pretending

36:21

not to be amazed by that. It has

36:23

exclusive shows from The New York Times.

36:25

It's the headlines storytelling from serial productions

36:28

in this American life. Fact to a

36:30

fiasco involving a village marauding Visigoth and

36:32

some oil sports from the athletic and

36:34

those big moments, he puts the team

36:37

on his back. And narrated articles from

36:39

the times and beyond. In recent years,

36:41

the unexpected sounds of ice have periodically

36:43

gone viral. New York Times Audio. Download

36:46

it now at NY times.com

36:49

audio. Casey,

36:55

let's talk about drugs. Let's talk about drugs.

36:58

Now, we have gotten at least

37:00

one note from a listener who says, you

37:02

guys seem to talk about drugs a lot

37:04

on the show. You make a lot of

37:06

jokes about magic mushrooms in particular. What's going

37:08

on with that? I mean, I I can

37:11

only say that it is relevant to understanding

37:13

Silicon Valley. This will be my answer. It's

37:15

true. We are tech reporters. We cover an

37:17

industry and a culture out here in San

37:19

Francisco, in the Bay Area. And, you know,

37:22

drugs are part of that culture. So we

37:24

are going to talk about drugs in this segment.

37:26

If you are a parent who doesn't want your kid

37:28

to listen to something like that, you know,

37:30

this one might not be for you. Yeah. So

37:34

this week, the Wall Street Journal reported

37:36

that Elon Musk, the world's wealthiest person,

37:38

has, quote, used LSD,

37:41

cocaine, ecstasy and psychedelic mushrooms

37:43

often at private parties around

37:45

the world where attendees sign non-disclosure

37:47

agreements or give up their

37:49

phones to enter. This was

37:51

a very juicy story that ran

37:53

the other day. And

37:56

it included some actual on the record

37:58

details about. Elon Musk's

38:00

drug use over the years, the journal reported

38:02

that in 2018, Elon Musk took multiple

38:06

tabs of acid at a party he

38:08

hosted in Los Angeles. The

38:11

story also reported that Elon Musk had

38:13

taken magic mushrooms at an event in

38:15

Mexico and took ketamine

38:17

recreationally in 2021 with his

38:19

brother in Miami at

38:21

a house party. And

38:24

that, you know, this has not gone

38:26

down well with members of his orbit

38:28

who are increasingly worried about his erratic

38:30

and potentially drug-fueled behavior. That's right. And what

38:33

I would say is what made this story

38:35

interesting to us, Kevin, is not that Elon

38:37

Musk has been spotted doing drugs a handful

38:40

of times over the years. It is rather

38:42

that people close to him seem both

38:44

very concerned about his behavior in some cases,

38:46

and were willing to talk about it on

38:48

the record with the Wall Street Journal, which

38:51

I just think suggests that this has

38:53

become a serious issue and given Musk's power

38:55

and influence in the world is really worth

38:57

trying to understand. Yeah. So when

38:59

we talk about drugs in this conversation, we

39:02

are not really talking about all the

39:04

classes of illegal drugs. We're talking specifically

39:06

about the ones that are

39:08

popular with people in the

39:10

tech industry, things like psychedelics,

39:12

things like MDMA, things like

39:14

ketamine. These are

39:16

the drugs that Elon Musk has

39:18

been spotted using, according to the

39:21

journal, and that I would say

39:23

are some of the most popular drugs out here

39:25

in the Bay Area among people who work in

39:27

the tech industry. Yeah, I think that's right. So

39:29

today to talk about Elon Musk's drug use and

39:31

this kind of wider phenomenon of drug use

39:33

in Silicon Valley, we've invited

39:35

Kirsten Grind. Kirsten is an

39:37

enterprise reporter at the Wall Street Journal. She

39:39

reports on tech companies and their executives, and

39:42

she was one of the people, along with

39:44

Emily Glazer, her colleague, who broke this story

39:46

about Elon Musk's drug use. She's

39:49

written a lot about this topic over the years,

39:51

and I'm really excited to talk to her about

39:53

her reporting. I answer you. Kirsten

40:03

Grind, welcome to Hard Fork. Thanks so much

40:05

for having me. Hey, Kirsten. So

40:07

you've been reporting on drug

40:10

use in Silicon Valley for a

40:13

while now, and I want to ask you about

40:15

some of the stories you've reported on. But

40:18

I want to zoom in on this

40:20

recent story about Elon Musk and his

40:22

drug use, which I would say

40:25

has been something that a lot of reporters have been

40:27

kind of gossiping about, you know,

40:29

off the record at the bar for many

40:31

months now. Yes. I want

40:33

to say, Kirsten, I was so jealous of this

40:35

story because I have heard so many whispers about

40:37

this stuff. And I have tried to get some

40:40

of this stuff on the record and absolutely failed.

40:42

So when I saw your story come out,

40:44

it was appointment television for me. I

40:46

dropped everything and like inhaled it all

40:48

in one go. Oh, thank you, guys.

40:50

That's so kind. So very

40:52

impressive story. What was the start

40:54

of this? When did you get interested in this

40:56

particular story? Yeah. So at the

40:58

journal, I'm an enterprise reporter. So what

41:00

that basically means for nonjournalists is

41:03

I jump around from topic to

41:05

topic. And so as you pointed

41:07

out, for some reason, the last

41:09

few years, I've kind of been

41:11

on this whole billionaire tech drug

41:13

scene beat, I guess,

41:15

which is really funny because I'm

41:17

the most boring person ever, actually.

41:20

But OK, so now I'm an expert in like

41:22

ketamine and cocaine and all of this. And

41:26

we had done some stories on Elon. We

41:29

reported last year on his Texas plans and

41:31

some of this other stuff. But

41:34

we had been hearing same as

41:36

you guys about this for a

41:38

long time. And so

41:40

we just wanted to definitively know, is

41:42

this man using drugs? I

41:46

think it's a very important question for someone

41:49

of his stature and his

41:51

power and his companies

41:54

with billions of dollars in government contracts.

41:56

So we just kind of went down

41:58

that road. We'll get into

42:00

all the implications of it, but maybe just to start

42:02

with, can you give us an overview of Elon Musk's

42:04

drug use? So he

42:07

is a pretty heavy, I

42:09

would say recreational drug user.

42:13

We have him on, you know, have

42:15

tried a bunch of different drugs. Ketamine,

42:17

we had reported last year. That's sort

42:19

of the popular, by the way, as

42:21

you guys probably know, Silicon Valley drug

42:23

of the moment. Cocaine,

42:26

ecstasy, LSD. And

42:30

a lot of it, one of the

42:32

reasons why it's been kind of kept

42:34

under wraps in a way is because

42:36

a lot of this is happening at

42:38

these private, high-end parties where you often

42:40

have to sign an NDA. A

42:43

lot of the people I spoke to, the

42:45

parties were in different countries. It's

42:47

not just like going out here

42:49

in the valley. And

42:52

so it's been pretty, you

42:54

know, we have examples going back

42:56

years of this. And

42:58

I would talk about my time at these parties,

43:01

but I did unfortunately sign the NDA. So I

43:03

will have to kind of pass on that. But,

43:05

you know, knowing what I

43:07

know about, you know, how these kinds of

43:09

stories come together and the standards at the

43:11

Wall Street Journal as well as the New

43:13

York Times, like it's hard to get a

43:16

story like this into print because you can't

43:18

just go on one sort of anonymous source

43:20

or two anonymous sources. So you actually have

43:22

talked to multiple people who have firsthand accounts

43:24

of witnessing Elon Musk doing these drugs. Is

43:26

that correct? Oh my gosh. We

43:29

had to have people who have witnessed

43:31

his drug use. And I

43:34

cannot even begin to go into the

43:36

rigor of our process for getting a

43:39

story like this into the newspaper to

43:42

give you an idea. Like it was much easier

43:44

writing my last book. Then to

43:47

publish this one story. Yeah.

43:49

Yeah. I mean, for good

43:52

reason, right? Like we don't just, as

43:54

you guys know, like these newspapers don't

43:56

just willy-nilly publish something like this. I

43:58

mean, I think Elon's. followers would

44:00

like to think that, but no, we

44:02

spend a lot of time making sure

44:05

everything's right, we have the sourcing, we

44:07

have it all lined up for sure.

44:09

Right, and so we actually asked months

44:12

ago, Walter Isaacson, who's Elon Musk's biographer,

44:14

about Elon's alleged drug use, and

44:16

what he responded to us was that, you

44:18

know, he knew that Elon Musk

44:20

had been taking ketamine for

44:23

depression, and ketamine, like some of

44:25

these other drugs, is commonly used

44:27

for sort of mental health treatment,

44:29

and so he seemed to think that this was

44:31

all above board, but what you reported was not

44:34

that, so just walk us through some of the

44:36

specifics around the drug use that you have reported

44:38

on with Elon Musk. So it's a lot of

44:40

partying, right? One thing that's

44:43

interesting with Elon, but with a

44:45

lot of these guys, they're using

44:47

psychedelics at parties, but

44:49

also for quote-unquote treatment, right? But

44:51

they're treating themselves, so that's kind

44:54

of the problem, right? So even

44:56

at parties, I think,

44:58

and I'm not saying this specific to Elon

45:00

necessarily, but I think in their heads they're

45:02

saying, oh, if I take

45:04

mushrooms, that's actually healthier than having like

45:07

five shots, or doing a line

45:09

of cocaine, and so with

45:11

Elon, you

45:14

know, he's used a bunch of different drugs,

45:16

but this ketamine is one that a lot

45:18

of people are using at the moment. And

45:21

I think that, you know, what

45:23

you said really speaks to the

45:25

cultural change around drug use in

45:27

Silicon Valley, and of course there

45:29

has basically always been drugs

45:32

in Silicon Valley. LSD is a huge

45:34

part of the story of Steve Jobs

45:36

and Apple, and yet

45:38

at the same time, you know, like Kevin

45:40

and I are around the same age. We

45:42

grew up in a dare America, you know,

45:45

a dare to raise drugs, just say no,

45:47

sort of all of that, and

45:49

it sort of seemed like the only accepted

45:52

drug to do was like alcohol, right? But

45:54

you fast forward to today, you can order

45:56

ketamine off Instagram in a sort of mental

45:58

health context, you can walk... down Posto

46:00

Street and buy mushrooms from

46:02

a, quote, church. Right? So

46:05

the vibe here is just very different than

46:07

I think. And if you have not spent

46:10

time in San Francisco recently, it

46:12

might shock you just how common some of this

46:14

stuff is. Absolutely. You know, I've obviously spent a

46:16

lot of time thinking about this and it really

46:18

goes back to that, I think,

46:21

whole Silicon Valley mentality where it's sort

46:23

of like, I can disrupt,

46:26

you know, myself. I can take charge of

46:28

my own healthcare. And so I think

46:30

in their heads, they're thinking,

46:32

ketamine can be used legitimately

46:35

for mental health treatments. And

46:37

some of these other drugs can be

46:39

used in a good way too, but

46:41

I'm gonna do it myself. Like, nevermind

46:43

like that doctor that's administering it. Right?

46:46

So that's where they're at. Right. Yeah.

46:49

I'll confess that when I first saw some of

46:51

the headlines that you and other general reporters were

46:53

pointing out about sort of drug use in Silicon

46:55

Valley and about Elon Musk, actually, my

46:57

first thought was sort of like, why do I

47:00

care about this? Like, you know,

47:02

these are adults, they're making decisions

47:04

about their own, you know, substance

47:06

use. Some of these drugs,

47:08

as you mentioned, like do have sort of

47:10

demonstrated effects for mental health and are, you

47:12

know, maybe legalized for use

47:14

in the coming years. And

47:17

we live in Silicon Valley where drugs

47:19

have been around forever. So why is

47:21

this such a problem for Elon Musk

47:23

in particular? A hundred percent. And you

47:25

can imagine we had like many conversations

47:27

about this too, right? The

47:29

reason it's very important for Elon

47:31

in particular isn't just because he's

47:34

the world's richest person or the

47:36

world's most powerful person or because

47:38

he runs Twitter or whatever, X,

47:40

sorry. It's

47:42

because in particular, he's running

47:44

six companies, one of them,

47:47

the publicly traded Tesla, where

47:49

he's supposed to be reporting

47:51

to investors, but especially SpaceX,

47:54

which has billions of dollars

47:56

in government contracts. And those

47:58

government contracts aren't like. Yeah,

48:00

if you do a little cocaine on the

48:02

weekend, it's all good. Those are

48:04

like you cannot do illegal

48:07

substances Ever like

48:09

we're not talking about lines at your

48:11

desk. It's like you cannot go to

48:13

Burning Man and do Ecstasy

48:15

or whatever you're doing there, right?

48:18

They're extremely strict and you

48:20

know as you guys I'm sure well

48:22

No, when all he did was smoke

48:24

a little marijuana five years ago on

48:26

Joe Rogan taxpayers footed

48:28

the bill for a 5-million-dollar

48:30

NASA review of his drug

48:32

use and that was just like I think

48:35

one puff And you're just like how did it how

48:37

did it cost the taxpayers 5 million dollars to just

48:39

watch one episode of the Joe Rogan thing? Yeah,

48:44

so they had to do a

48:46

whole drug review of SpaceX employees

48:49

SpaceX employees were subjected to

48:52

Random drug tests for some period of

48:54

time. There's not a lot We

48:57

know about like what went into

48:59

that review But Elon talked about

49:01

this after and some podcasts and

49:03

about how he had not apparently

49:05

Realized the effect this would have on

49:08

SpaceX. So they had to do this

49:10

whole review and Taxpayers

49:12

basically footed the bill Wow

49:15

Congrats taxpayers So

49:17

I think that's an important point about the difference between

49:19

sort of Elon Musk doing this and any sort of

49:22

other You know private citizen who

49:24

does not have government contracts or a

49:26

security clearance But you

49:28

also reported that his drug use has caused

49:30

concern among the board members of his company.

49:33

So tell us about that That's right. So

49:35

that's the second important point. This is not

49:37

the journal like judging Elon Musk. This

49:39

is us saying Listen, it has

49:41

gotten to the point where even Leaders

49:44

at his two largest companies including

49:46

some directors the directors who aren't

49:48

the ones doing the drugs along

49:51

with him Are also

49:53

concerned about this right? And so

49:55

that's that's really the whole point

49:57

of the story like they've had

49:59

years concern, they don't know how

50:01

to handle it. When they're really concerned, they

50:03

kind of go over to Kimball Musk, his

50:05

brother, and are sort of like, hey,

50:08

like, is he getting enough sleep? You

50:10

know, they don't even say drug use

50:12

because that can end up in board

50:14

meeting minutes, right? I thought this was

50:16

so interesting, the way that even those

50:18

who are placed in positions to have

50:21

some measure of authority to serve as

50:23

a check on him, they are terrified

50:25

of just saying what is plain to

50:27

everyone in his orbit, which is just

50:29

that he is on drugs a lot.

50:31

Absolutely. I mean, not to excuse them,

50:33

but you can see this really

50:35

challenging position they're in because first

50:38

of all, we need to say

50:40

Tesla and SpaceX are doing great.

50:42

Tesla especially performing super well. So

50:44

on the first of all, it's

50:46

like, what, who are we to

50:48

complain about that? Like if even

50:51

I think even Elon himself said something

50:53

like this on Twitter after like, if

50:55

I'm using drugs, like I should keep

50:57

doing it. I'm doing a great job.

50:59

That's, that's exactly the position they're

51:02

in. Yeah. And I think there's

51:04

sort of never been a problem with a drug

51:06

user who's sort of in a good run and

51:08

decides to just do more drugs. That's never ended

51:10

badly for anyone who's ever done drugs. Right.

51:14

I wanted to ask about one director

51:16

in particular who you report stepped down,

51:18

Linda Johnson Rice stepped down from Tesla

51:20

decided not to stand for reelection in

51:23

2019 in part because of the drug

51:25

use. I wonder if you could share

51:27

any more of that story. And also

51:29

I have to say reading that that

51:31

does not seem like somebody who was

51:34

worried that he was doing ketamine every once in a while

51:36

at a party. No, I mean, again, that

51:38

was a, the ketamine issue

51:40

is a lot more recent. I would say

51:42

it's been in recent months that people are

51:44

much more worried about ketamine and that, that

51:46

kind of tracks as well with like the

51:49

ketamine popularity growing generally. We

51:51

should say that ketamine is

51:53

legal. It's legal, but

51:55

it's like a gray area legal. And

51:57

I also want to be clear. that

52:00

most people are doing this through dealers

52:02

or you know randomly through Instagram. Yeah

52:04

an online pill mill type of thing

52:06

Yeah, but back to your question I

52:09

mean there's not a ton more I

52:11

can share about what's in the story

52:13

But I would say for a Tesla

52:15

director to step down before their three-year

52:17

term right two years That's

52:20

that's really saying something. Yeah, and this

52:22

is a woman who's very well respected

52:24

respected right in the industry as being

52:27

on many boards and Corporations,

52:30

etc. Yeah, this doesn't sound like somebody who just heard

52:32

that Elon had done mushrooms a couple times at a

52:34

party and said I'm out Of here. Yeah. Yeah, and

52:36

one of the things that is often said about Elon

52:39

Musk's drug use by people who are sort of gossiping

52:41

about It is that it's it's changing his behavior that

52:43

he part of the reason and the

52:45

explanation for why he's been so erratic in

52:48

the past few years and has made all these

52:50

controversial decisions about X and and just

52:52

sort of the Personality that he's adopted

52:55

that this can also be traced to

52:57

his drug use and I

52:59

wonder what you think of that And

53:01

if there are any specific examples of

53:03

behavior that you've reported that has been

53:05

specifically linked to drug use. So I

53:08

Have a lot of in my head, you

53:10

know and also from just knowing The

53:13

drug use situation now instances where

53:15

I've seen him where I think

53:18

You know, maybe that doesn't matter though

53:21

Like in that in this story one

53:23

point we really try to bring up is

53:25

this exact thing that you mentioned?

53:27

He's acting erratically. He's acting strangely.

53:30

Is that just Elon the genius?

53:32

the the guy who said he

53:34

is autistic or

53:36

is he actually on something and so

53:39

this is one reason we brought up

53:42

this example from 2017 where he's speaking

53:44

at SpaceX and Hilariously

53:46

SpaceX has since released that video

53:49

and I would encourage anyone

53:51

to go look at it because in our

53:53

Reporting the executives were all

53:56

worried after that. He was

53:58

on drugs now We don't know

54:00

if he was, and we say that in the

54:02

story. We do not know, right? But

54:05

they're like, is that drugs or

54:07

is that his erotic behavior? And this

54:09

is something that everyone around him

54:11

has struggled with for years. Yeah, this

54:13

is often a question I ask after

54:15

Casey says something stupid on the podcast.

54:20

Or is it the drugs? What exactly is in this tea? Now

54:24

Elon Musk and his camp have responded to

54:26

this story. His lawyer, Alex Spyro,

54:28

told you that parts of this story were

54:30

false, although he didn't specify what exactly was

54:33

false. He also said that

54:35

Elon Musk is, quote, regularly and randomly

54:37

drug tested at SpaceX and has never

54:39

failed a test. So I'm curious what

54:41

you make of that statement and what you know about

54:44

these drug tests, like what are they testing for? How

54:46

often do they have to take them? And

54:49

if it's true that he's never failed a drug test,

54:51

how do you square that with what's in your story?

54:54

So I would first of all say, as

54:56

you guys probably know as journalists, that's not

54:58

necessarily a denial. That's what we call a

55:00

non-denial denial. Okay. And

55:04

I think Matt Levine even pointed that

55:06

out, like in a hilarious way. But

55:09

a note about these drug tests. I wish

55:11

I could tell you more about them. They

55:13

are apparently extremely secretive. So we do

55:16

not know how often

55:18

he's tested, when even what

55:21

drugs are being

55:23

tested for. Generally, I've

55:25

learned that psychedelics aren't

55:27

usually in a test. I want

55:29

to be clear. I don't know if they're testing

55:31

Elon for psychedelics. That's

55:33

the point. We kind of don't know. Then I

55:36

think Elon came out after and said, I

55:38

was tested for three years. So

55:40

I don't know if that means he's

55:42

not been tested the last couple years,

55:44

three years since the Joe Rogan incident

55:47

in 2018. So there's

55:49

just a lot we don't know about these

55:51

drug tests. So reporters have

55:53

been trying to nail down this story about Elon Musk

55:55

and his drug use for years. You

55:58

were actually able to get people. on the

56:00

record talking about it who have firsthand

56:02

encounters with his drug use. Why

56:05

do you think people are willing to open

56:07

up now? I'm so glad you asked that

56:09

question. It was, I have, through

56:11

this whole thing, often thought about

56:13

people's motivations because a lot of

56:15

the times people talk to reporters

56:17

because they're exposing something bad or,

56:20

you know, they're unhappy with how something's

56:22

going. But in this case, you're asking

56:24

people to describe someone in a lot

56:26

of times, someone who they admirers, drug

56:29

use, and they want to be in

56:31

that crowd that's getting into that NDA

56:33

party and all of this. So I

56:36

would say that, you know, without going

56:38

too much into it, a lot of

56:40

the motivation here, well, some of the

56:42

motivation at least, is from people who

56:44

have concern, right? It's

56:47

not just people who, you

56:49

know, saw him one time at a party.

56:52

I mean, definitely I've talked to some of

56:54

those, but there's also just

56:56

a general concern out there. Not people who are

56:58

necessarily trying to get him in trouble. No, not

57:00

at all. People who are trying to maybe get

57:02

him help. That's right. And not

57:04

even just with Elon, but in this reporting in

57:07

general, I found that people

57:09

who are willing to talk about someone

57:11

else's drug use, especially someone in a

57:13

position of power, are doing it because

57:16

they're worried. Yeah, right. And

57:18

so just to take kind of the devil's

57:20

advocate position here. And

57:22

that the drugs are good? No. But,

57:25

you know, I've heard and I've seen since your

57:27

reporting came out, some people just saying like, well,

57:30

the proof is in the pudding, right? His

57:32

companies are doing great. Like he

57:34

has the best rockets. He has the best

57:36

selling car in the world. His behavior is

57:38

unimpeachable, a model of integrity and kindness. But

57:41

you know what I'm saying? Like, like, if

57:43

if these drugs were really hurting him, wouldn't

57:45

it be showing up in the performance of

57:48

his companies? And if it's

57:50

not showing up in the performance of his companies, why

57:52

is it any of our business what he's doing in his

57:54

free time? Well, let's let's take SpaceX

57:56

out of that out of this for a second, because it's

57:59

just a full violation of his SpaceX

58:02

contract. So let's just maybe look at

58:04

Tesla. I think it's

58:06

a great question because Tesla is performing

58:08

really well, right? And so I think

58:11

for directors or other

58:13

executives to reach that level

58:15

of kind of concern about

58:18

his behavior, that's what to look at

58:20

there. You know, they're not they're not

58:22

bringing it up just because they think

58:24

he's had a bad day or something

58:27

like that. Also, like Tesla is

58:29

in part kind of a meme stock. Like,

58:31

yes, the car company itself is performing well

58:33

in the world. Yes. And part of that

58:35

is just because there's a huge fandom around

58:37

Elon Musk, who thinks he's a cool dude

58:39

and likes to see him do stuff. So

58:41

the fact that Elon Musk is on drugs

58:43

all the time, I could see how that

58:45

would make the stock price of Tesla go

58:47

up because it means that Tesla stockholders are

58:49

gonna say, Cool, bro. Yeah, I also wonder

58:51

what you think of the Matt Levine point

58:53

that he made in his newsletter this week,

58:55

which is that Elon Musk is

58:57

in some ways too big to fail a

58:59

drug test. Yeah, that was a great line.

59:02

But also like, you know, and he basically

59:04

says, Look, if you're NASA, or

59:06

you're, you're in the Defense Department,

59:08

and you find out that Elon Musk has

59:10

done drugs, maybe he did drugs in front

59:12

of you, what are you gonna do? Like,

59:14

are you gonna are you gonna, you know,

59:16

put your payload into orbit with someone else's

59:18

inferior rockets? And I thought that was

59:20

a really interesting point. Like, even if it

59:22

is true that he's doing all these drugs, and they're

59:24

getting, you know, in the way of his performance, and

59:27

directors of his companies are growing concerned about it,

59:29

like, what are we

59:31

supposed to do about it? Well, that

59:34

that is the thing. I mean, SpaceX

59:36

is so intertwined with the US government.

59:38

I mean, they are the space program,

59:40

right? So I mean, I don't have

59:42

any insight, knowledge, inside knowledge, but who

59:45

know, who knows what they're going to

59:47

do, or if they can do anything.

59:50

And even on his boards, like

59:52

as we've reported, they've just kind of tiptoed

59:54

around it. So he could be

59:57

too big to fail a drug test. And we just do this all

59:59

the time, right? I mean, And like this is the

1:00:01

troublesome thing about having somebody who is this rich

1:00:03

and powerful and it seems like there just is

1:00:05

no check on his power. Think about how many

1:00:07

times in the past he has done something, he

1:00:10

has broken some law, he's violated some SEC regulation,

1:00:12

and it just seems like everyone throws up their

1:00:14

hands and say, well, what are you going to

1:00:16

do? Like, we don't have any legal system to

1:00:19

take it. He's a genius. Yeah. Yeah. He's

1:00:21

a genius. And also we have no legal protections that

1:00:24

would actually check him. Yeah. Yeah.

1:00:27

And that's why I came out a little

1:00:29

bit because the use of drugs and particularly

1:00:31

of psychedelics is sort of this hidden force

1:00:33

in Silicon Valley that many

1:00:35

people in positions of authority in

1:00:37

the tech industry specifically are fans

1:00:40

of these drugs for legitimate mental

1:00:42

health issues and productivity, but also

1:00:44

for partying and that there's a

1:00:47

sense in which the

1:00:49

drugs are sort of a hidden mover in the

1:00:51

tech industry today. And I wonder what your thoughts

1:00:53

are on that, having spent so much time reporting

1:00:55

on this. Yeah, I have spent a lot of time

1:00:57

on it. First of all,

1:00:59

I want to say I actually

1:01:02

totally agree with the research behind.

1:01:04

I've interviewed a lot of doctors

1:01:06

and like legitimate medical professionals who

1:01:08

are working to make ketamine, you

1:01:10

know, ecstasy, psilocybin, all of those

1:01:13

legal and helpful for post-traumatic stress

1:01:15

syndrome, depression, all of this. So

1:01:17

that is definitely happening and is legit.

1:01:21

I do think that a lot

1:01:23

more people are using psychedelics,

1:01:25

you know, a lot of

1:01:27

tech executives who we probably know

1:01:29

than we know and that it's

1:01:32

way more common. Just no

1:01:34

one still wants to talk about it because

1:01:36

it's illegal, you know, but a

1:01:38

lot of these people are funding some of

1:01:40

these organizations where they're

1:01:43

trying to push for legality

1:01:45

and research medical cures

1:01:47

in part, I think, because it

1:01:50

could help them if done in the

1:01:52

right way. And right now they're doing

1:01:54

it illegally. I will say after your story

1:01:56

came out, I want to put this to you in the interest of

1:01:58

fairness. A friend of mine who who works in

1:02:00

the tech industry texted me and

1:02:03

said, why is the Wall Street

1:02:05

Journal talking

1:02:07

about this like it's the end of

1:02:09

the world? Why are we getting this

1:02:11

story that's sort of talking about how

1:02:13

illegal all of these drugs are? And

1:02:16

this is just what people do in

1:02:18

society and they're only

1:02:20

making a big deal out of this because it's Elon Musk.

1:02:22

What do you say to that? I have

1:02:24

heard that from about 10,000 of Elon's fans as

1:02:29

well over the last few days. So I've

1:02:31

definitely heard that. I mean, I

1:02:33

just have to keep going back to the

1:02:35

fact that he is pretty

1:02:38

much the most powerful person in

1:02:40

this country and all his businesses

1:02:42

are integrated with our infrastructure. He

1:02:44

has billions of dollars in government

1:02:47

contracts. And again, like

1:02:49

even if he's holding it together now,

1:02:52

I'm not saying anything's gonna happen, but

1:02:54

it's something we need to know about the

1:02:56

health of one

1:02:58

of our most powerful people in this country. And

1:03:00

I would just say as a gossipy person who

1:03:02

loves mess, thank you so much for reporting this

1:03:04

story. And I hope you do so much more.

1:03:06

Don't worry if other people think it's important or

1:03:09

not because I'm living for it, Kirsten. Okay, thank

1:03:11

you, Casey. Well, yesterday I

1:03:13

was accused of eating live babies by

1:03:15

one of Elon's followers. Go on. Oh,

1:03:18

it was, I almost wanna read

1:03:20

you guys this. It was

1:03:22

a new low. It was

1:03:25

like Kirsten Grind eats live

1:03:27

babies. Well, it sounds like that

1:03:29

person might've been doing some recreational drugs before

1:03:32

they took the message. And are

1:03:34

you denying on the record that you

1:03:36

eat live babies? I am denying that

1:03:38

on the record, you guys. Just

1:03:41

have to check in the interest of being cripe-like.

1:03:44

Is that even the record's rate? Definitely. I mean,

1:03:46

as you guys know, like covering

1:03:48

Elon Musk comes with hearing

1:03:51

from his many

1:03:53

thousands of fans. Yeah, yeah.

1:03:55

Million probably. Yeah. Yeah,

1:03:57

well, Kirsten Grind, thank you so much for coming on. Thank

1:04:00

you guys so much for having me. When

1:04:06

we come back, we're going to talk about drugs

1:04:08

again. Surprise! But

1:04:12

this time we're talking about the

1:04:14

other kind of drugs, the prescription

1:04:16

ones that AI is helping researchers

1:04:19

discover to treat serious illnesses. So

1:04:29

Casey, as we were sort of planning out

1:04:32

some of our goals for the podcast

1:04:34

this year, one of the topics that

1:04:36

I really wanted to spend more time

1:04:38

talking about is AI and your

1:04:41

coke is like perched at a

1:04:43

very precarious angle. That's amazing that

1:04:45

that didn't spill. I

1:04:48

know. It was like your coke can was

1:04:50

literally like leaned against your laptop at a

1:04:52

45 degree angle in a

1:04:55

way that suggested that you were trying to

1:04:57

play some kind of Daredevil game whereby it

1:04:59

was going to spill on yourself. That was

1:05:01

like the old story, but Prince like Jesus

1:05:03

was carrying me right then. Like I didn't

1:05:05

know it, but he was carrying me and

1:05:07

that's why it didn't spill. Thank you Jesus.

1:05:09

Okay, so Casey,

1:05:12

one of the stories that I have been

1:05:14

sort of devoting more time to trying to

1:05:16

follow recently is what's happening with

1:05:18

AI in the field of medicine. Yes,

1:05:20

because this is a story that I

1:05:22

think everyone who is optimistic about AI

1:05:24

touts as kind of the highest and

1:05:26

best use of this technology. If you

1:05:29

want AI to go faster, this is

1:05:31

one of the best reasons that you

1:05:33

could want it to go faster is

1:05:35

we could discover more drugs more quickly.

1:05:37

Yeah, so this kind of thing is

1:05:39

what a lot of people in tech

1:05:41

and biotech are very excited about. They

1:05:43

say AI is going to be

1:05:45

radically transformative. It's going to help

1:05:47

us, you know, discover new treatments

1:05:50

for cancer and Alzheimer's

1:05:52

disease and heart disease and all

1:05:54

these deadly and debilitating illnesses

1:05:56

and basically AI is going to

1:05:58

serve turbocharts. this

1:06:00

entire field of medicine. And

1:06:03

so I wanted to start covering this

1:06:05

in more detail in 2024, because

1:06:08

there's just a ton of money

1:06:10

and attention and hype and

1:06:13

real promise in the intersection of AI and

1:06:15

medicine. That's right, Kevin. And not only is

1:06:17

there promise, but we are just now starting

1:06:19

to see the fruits of these labors. And

1:06:21

this has gone beyond the realm of, Oh,

1:06:23

wouldn't it be cool if AI could discover

1:06:25

a drug, we are starting to see the

1:06:27

science that, Oh, my gosh, this stuff actually

1:06:29

works. Yeah, this is something that I really

1:06:31

didn't appreciate until I started looking into this.

1:06:33

There's this big healthcare conference, the JP Morgan

1:06:35

Healthcare Conference, which is a sort of a big

1:06:37

deal in that world is happening

1:06:39

in San Francisco this week. And I've

1:06:42

just been reading some of the stuff

1:06:44

coming out of that conference. And it

1:06:46

is remarkable how much of the discussion

1:06:48

in healthcare and medicine today is about

1:06:51

AI, and particularly this use of AI

1:06:53

to discover new drugs. So

1:06:55

I've just had my kind

1:06:57

of antennas up for interesting

1:06:59

and novel stories related to

1:07:01

AI and drugs of the

1:07:04

medical variety. And one

1:07:06

of these stories popped up last

1:07:08

month, researchers at MIT and Harvard

1:07:10

published a paper in the science

1:07:12

journal Nature. They claim to

1:07:14

have discovered an entire class of drugs

1:07:16

using AI and confirmed that these drugs

1:07:19

were successful at combating a type of

1:07:21

bacteria called MRSA. Yeah. And when I

1:07:23

hear the word MRSA, it's always in

1:07:26

the context of why you never want

1:07:28

to be hospitalized. Because apparently in hospitals,

1:07:30

this is a drug resistant infection that

1:07:32

spreads around there can be very difficult

1:07:35

for our existing medicines to treat. And

1:07:37

so it's the exact sort of thing

1:07:39

that we could use some help from

1:07:41

AI and solve. And as it turns

1:07:43

out, AI is already helping researchers trying

1:07:45

to figure out what kinds of chemicals

1:07:48

could be helpful in combating MRSA. And

1:07:50

this is an area where we already

1:07:53

have some evidence that AI is accelerating

1:07:55

discovery. So to talk about this discovery,

1:07:57

we've invited one of the lead authors

1:07:59

of this. Nature study Felix Wong

1:08:01

to join us. Felix is a

1:08:03

postdoc in the lab of James

1:08:05

J. Collins at MIT where he

1:08:07

worked on this research alongside a

1:08:09

big team of scientists. He's also

1:08:12

the co-founder of a drug discovery

1:08:14

startup called Integrated Biosciences and we're

1:08:16

going to talk to him today

1:08:18

about how AI helps make this

1:08:20

discovery possible. Felix

1:08:29

Wong, welcome to Hard Fork. Thank you

1:08:31

for having me. Hi Felix. So we

1:08:34

are interviewing you today because something very

1:08:36

exciting happened just before the holiday break

1:08:38

which is that a research team that you

1:08:41

are on announced that you had used AI

1:08:43

to discover a new class of antibiotics that

1:08:45

could be effective against MRSA and

1:08:47

I also read in the coverage of

1:08:50

this research that there hasn't really been

1:08:52

a new class of antibiotics discovered in

1:08:54

60 years. So why is

1:08:56

that? Why is it hard to

1:08:58

discover new antibiotics using conventional methods?

1:09:00

Yeah so there is a bit of

1:09:03

hype to that statement. So there have

1:09:05

been new antibiotics as well as a

1:09:07

few new classes of antibiotics discovered in

1:09:09

the past 60 years but certainly not

1:09:12

a lot and in fact most of

1:09:14

the clinically used antibiotics that we use

1:09:16

today were discovered in the 1960s and

1:09:19

when we kind of discovered those antibiotics

1:09:21

just by looking at soil bacteria. Turns

1:09:24

out that the bacteria growing in soil weighs

1:09:26

warfare on each other and you can just

1:09:28

kind of take their weapons and use

1:09:30

them as antibiotics. Once this

1:09:32

pipeline really dried up there's just been a

1:09:34

dearth of new drug candidates coming out again

1:09:36

because we've already exhausted kind of this natural

1:09:38

source of antibiotics. Yeah we were really good

1:09:40

at drugs in the 60s but after that

1:09:42

it really seems like America lost its way.

1:09:44

So help me understand here because I hear

1:09:47

a lot about you know the use of

1:09:49

AI to discover new drugs and

1:09:51

I want to talk about your specific discovery

1:09:53

process but I also just want to like

1:09:55

understand at a very broad level what does

1:09:58

it mean to say that AI can help? help

1:10:00

us discover new drugs. Right, because it's not just

1:10:02

going to chat GPT and saying, hey, got an

1:10:04

idea for a new drug? Right, yeah. Yeah, so

1:10:06

of course one can do that. Go to some

1:10:08

LLM and ask for an idea for a new

1:10:10

drug. The question is, is it accurate? And is

1:10:13

it actually worth following up on whatever the LLM

1:10:15

says? In the case of drug

1:10:17

discovery, things are a bit more niche than

1:10:19

LLMs. So it's not like we're training a

1:10:21

general purpose model in order to just write

1:10:24

us poetry or write us emails or whatever.

1:10:27

It's really about training very specialized models

1:10:29

in order to make very specific predictions

1:10:31

as to whether or not a new

1:10:33

chemical might have antibacterial activity. And so

1:10:36

tell us about the nature of that

1:10:38

predictive step. How is it predicting? Yeah,

1:10:40

so as drug discoverers, what we do

1:10:43

is find needles in large haystacks. And

1:10:45

at least in our work, which is

1:10:47

quite typical of these machine learning drug

1:10:50

discovery approaches, the first step is we

1:10:52

need to get training data. And the

1:10:54

best way to do this is empirically.

1:10:56

So in our case, for instance, we

1:10:59

screened 39,000 compounds. So

1:11:01

one by one in a test tube, we

1:11:04

looked at things including, does the compound

1:11:06

affect MRSA? Does the compound become toxic

1:11:08

to human cells, which you don't want

1:11:10

because in that case, bleach

1:11:12

might also be an effect of antibiotic, right?

1:11:14

You had 39,000 different test tubes, each with

1:11:16

a little thing in it? That's basically correct.

1:11:19

So the only kind of quantification there is

1:11:21

that everything is stored for kind of compactness

1:11:23

in place. You could probably fit it in

1:11:25

just a stack of plates here in the

1:11:27

corner of this room. So

1:11:30

when we do the hard fork novel

1:11:32

pathogen creation process, that will be a

1:11:34

very compact storage facility. We

1:11:36

have a Thanksgiving episode this year when we

1:11:38

create a novel bioweapon. So I

1:11:41

think I can follow the story now here because you

1:11:43

conduct these 39,000 plus tests. And

1:11:46

I'm going to guess that some of these

1:11:48

compounds that you test seem more promising than

1:11:50

others. And so you're able to feed this

1:11:52

into your system. And then it can just

1:11:54

start to make predictions by saying, well, this

1:11:56

was more promising than that one. And so

1:11:58

here are a bunch of compounds. that look

1:12:00

like this one that was more promising. And so

1:12:02

let's look into this a little bit more. That's

1:12:04

true with two caveats. So step two is kind

1:12:06

of the model training. And that's where we dump

1:12:08

in all of the data to kind of these

1:12:10

graph neural networks, which are a type of deep

1:12:13

learning model. So the main

1:12:15

thing about deep learning models and one of the

1:12:17

key innovations of our study is really

1:12:20

that up until now, they've been known as black

1:12:22

box. We don't know how the heck is coming

1:12:24

at its predictions. It also means that if it's

1:12:26

inside of a plane that falls out of a

1:12:28

sky, it will survive. Just

1:12:30

ignore K. OK, I'm sorry. Please, please,

1:12:32

just continue with the science. Go ahead.

1:12:34

The concept is similar in the sense

1:12:37

that we wanted to kind of open

1:12:39

up and make sense of what the

1:12:41

model is doing. We don't necessarily have

1:12:43

to reverse engineer the model. But can

1:12:45

we get to a point where at

1:12:47

least we can be like, ah, this

1:12:49

is what the model is looking for.

1:12:51

Can we identify patterns, say, of chemical

1:12:53

substructures and small molecules? And then can

1:12:55

we use this to guide drug discovery?

1:12:58

So one of the kind of key

1:13:00

things about this approach, we kind

1:13:02

of developed this additional kind

1:13:04

of module, if you will, to the

1:13:06

AI model. And what that

1:13:09

module does is it employs a

1:13:11

type of search called Monte Carlo

1:13:13

tree search. That's a word salad.

1:13:15

But the main idea for that

1:13:17

is that we use the same

1:13:19

algorithm as AlphaGo. AlphaGo, the

1:13:21

deep mind algorithm that was able to

1:13:23

beat the best human go players, go

1:13:25

the board game. Yeah.

1:13:27

What was the moment where you're

1:13:29

fiddling around with your 39,000 plates and you say, wait

1:13:31

a minute, how do they beat that board game again? Yeah.

1:13:35

Exactly. So the moment here for

1:13:38

us was when we applied this

1:13:40

Monte Carlo tree search, this AlphaGo

1:13:42

kind of algorithm, to kind of

1:13:45

identifying new chemical substructures that are

1:13:47

predicted to underlie new classes of

1:13:49

antibiotics. We can now actually confidently

1:13:52

say which parts of a chemical

1:13:54

substructure account for its predicted

1:13:56

antibiotic activity. I see. So after you

1:13:58

get the suggestion for. hey, this is

1:14:00

a promising compound. You have a process that

1:14:03

lets you say, okay, why was this thing

1:14:05

promising? Exactly. And this is quite different from

1:14:07

how we've been using AI in the past,

1:14:09

where AI has really just been, at least

1:14:11

in many drug discovery instances, trained on model,

1:14:14

applied to predict some new stuff, and then

1:14:16

you validate some new stuff, great, call it

1:14:18

a day, go home, or maybe

1:14:20

go to the patent office, whatever it might be. In

1:14:23

this case, because we have this

1:14:25

explainable approach to AI, we can

1:14:27

now identify not just single compounds,

1:14:29

but entire classes of compounds, and that's

1:14:31

what's really salient. So instead of finding a

1:14:34

needle in a haystack, which was the old

1:14:36

approach, you're essentially finding little piles of needles

1:14:38

in the haystack. Yeah, we're finding sewing kits.

1:14:40

Right. But are you

1:14:42

saying that the same technology that helped AlphaGo

1:14:45

discover new moves in a board game just

1:14:47

sort of mapped neatly to discovering new chemical

1:14:49

compounds? That's correct. That's something magical about this,

1:14:51

is that, in a sense, the underlying question

1:14:53

is the same. In the case of AlphaGo,

1:14:56

it was kind of looking at the search

1:14:58

space of all possible moves, and

1:15:00

then predicting or anticipating the opponent's

1:15:02

moves. In our case, for

1:15:04

chemical structures, it was looking at

1:15:06

the combinatorial search space of which subset

1:15:09

of a chemical structure actually accounted

1:15:11

for its predicted activity by the

1:15:14

model. That's crazy. That's wild. Yeah,

1:15:16

I mean, and that's like a big reason that

1:15:18

I think people are so optimistic about AI for

1:15:20

drug discovery, is that it turns

1:15:23

out that some of these other problems that

1:15:25

people have been using AI to address, like

1:15:27

playing a board game, or

1:15:29

predicting the next word in a sentence, turn

1:15:31

out to also be very valuable for

1:15:33

other kinds of basic scientific research. Yeah.

1:15:36

And is that a kind of prediction that a

1:15:38

researcher, like a human, could do, but it would

1:15:40

just take them forever? Or is this just fundamentally

1:15:43

like a new kind of ability? This is fundamentally

1:15:45

different. So what a human might do is, because

1:15:47

we do not have any first principles kind of

1:15:49

approach to understanding what or not this new compound

1:15:51

might work, what a human might do would just

1:15:54

be to brute force screen them and

1:15:56

say, well, maybe I invest a few hundred million

1:15:59

into this project. by all of these

1:16:01

millions of compounds and then just brute force them

1:16:03

all. But the main idea of

1:16:05

kind of this machine learning approach is that

1:16:07

it can enable us to now start to

1:16:09

generalize beyond our training data set and look

1:16:12

for maybe often subtle patterns

1:16:14

in the arrangements of atoms and bonds

1:16:16

in a chemical structure in a way

1:16:18

that humans just can't do. You show

1:16:21

me a lot of pictures of

1:16:23

the chemical structures of beta-lactams and quinolones

1:16:25

and other known antibiotics. I can't really

1:16:27

point you to this new class of

1:16:30

antibiotics that we discovered and described. So as

1:16:32

I mentioned, the main prediction step here in step

1:16:34

three is an absence of single hits now.

1:16:37

It's of entire chemical substructures

1:16:39

that define hundreds, if

1:16:41

not thousands, of different chemical compounds.

1:16:43

Right. You're discovering like a

1:16:45

new class of potential drugs, not just like one or two.

1:16:47

Exactly. Yeah, exactly. So

1:16:49

you get back this list. You

1:16:51

shove all this stuff into this neural network. You get back

1:16:54

this list of a bunch of

1:16:56

compounds that might be helpful against MRSA. I

1:16:59

assume then you have to actually go figure out

1:17:01

whether they actually are helpful against MRSA. Oh,

1:17:03

yeah, exactly. So the first aha moment

1:17:06

for us was to actually get this list

1:17:08

in the first place. We had no guarantees

1:17:10

that anything would actually even give us an

1:17:12

output. So we were quite surprised and elated

1:17:14

really when we actually got something from the

1:17:17

algorithm identifying new structural classes of

1:17:19

putative antibiotics, in this case putative, because

1:17:22

as you mentioned, Kevin, we still have to validate

1:17:24

them. So in the end, what we actually did

1:17:26

was we bought around 280 compounds that had high

1:17:29

predicted antibiotic activity, several

1:17:32

of which were also predicted to underlie a new

1:17:34

class of antibiotics. Right now, is there a company

1:17:36

that'll just make any compound for you and sell

1:17:38

it to you? Yeah, can you just go on

1:17:40

Amazon and buy some compounds? Yeah, in fact, not

1:17:42

Amazon, unfortunately, otherwise, you know. You could

1:17:44

get the free delivery with Prime. Exactly. You could

1:17:47

get free delivery with Prime. You can

1:17:49

do garage experiments as well. But

1:17:51

in our case, there are actually

1:17:53

commercially available compounds from synthesis

1:17:56

suppliers as well as chemical

1:17:58

suppliers. many of which

1:18:00

are well known in the field. Great. So

1:18:02

then you have to test these things. How do you

1:18:04

test these things? Yeah, so as I mentioned, we bought

1:18:07

around 280 compounds that

1:18:09

had high predicted antibiotic activity, low

1:18:11

predicted toxicity to human cells. And

1:18:13

also they were quite structurally distinct

1:18:16

from known antibiotics. And so that's

1:18:19

one of the main takeaways of

1:18:21

our work is that we found two

1:18:23

compounds that share the same predicted sub-structure

1:18:26

that defines the new structural class

1:18:28

of antibiotics. And we

1:18:30

found that these compounds work. But

1:18:33

in the end, one of the main

1:18:35

experiments to do is really, does it work

1:18:37

for treating a mouse model

1:18:39

in vivo? A mouse?

1:18:42

A mouse model. So for

1:18:44

instance, like what we did in our work

1:18:46

was we had two mouse models. One was

1:18:48

where we just scraped off the skin of

1:18:50

mice. And then we infected that skin with

1:18:52

MRSA. And that was a topical model in

1:18:54

which you can just apply a cream on

1:18:57

the wound. The other model was

1:18:59

a systemic model. And this is where things

1:19:01

start to get a bit more interesting because

1:19:03

systemic infections underlie the most deadly bacterial infections,

1:19:06

including those leading to sepsis and

1:19:08

other things. So these two compounds

1:19:11

that you discovered using your neural network, they

1:19:13

actually did cure or treat MRSA

1:19:15

in these mice? That's correct. And that

1:19:17

was our second aha moment. So we

1:19:19

found that administration of one of the

1:19:21

compounds of this structure or class actually

1:19:23

decreased MRSA by over 90% in both

1:19:25

models. Got it. So

1:19:28

the process, I'm just gonna repeat this back one more time, just

1:19:30

to make sure I understand that you acquire the data. Oh, you're

1:19:32

gonna do this at home later? Yeah, I am. Yes,

1:19:35

I need the address of that website that sells

1:19:37

you the novel chemical compounds. So

1:19:40

you get the data, you train the

1:19:42

neural network on that data, you use

1:19:45

this kind of like AlphaGo Monte Carlo

1:19:47

tree search technique to like figure

1:19:50

out what the heck is happening inside

1:19:53

the neural network, why it's giving you back

1:19:55

these predictions. And then you get these suggestions

1:19:57

that says these, these 10

1:19:59

compounds. or these however many compounds might be

1:20:02

effective against MRSA and you go and you

1:20:04

rub some cream onto some mice to see

1:20:06

whether it actually works. Is that more or

1:20:08

less what happens? Yeah. I would

1:20:11

also add that we, in addition to rubbing some

1:20:13

cream on mice, we also inject the mouse with

1:20:15

some compound for the systemic model. So yeah. How

1:20:17

are the mice doing? Well, the

1:20:19

mice unfortunately are currently all dead. We have

1:20:22

to sacrifice all of them in order to

1:20:24

extract the bacteria. Okay. So

1:20:28

not a good day for the mice. But potentially they

1:20:30

are going to be... Their sacrifice was not

1:20:32

in vain. Exactly. Because we are

1:20:34

going to have maybe some drugs that actually do treat

1:20:37

this in humans. And that leads me to my next

1:20:39

question, which is, I've been hearing a lot about AI

1:20:41

drug discovery now for what feels like a couple of

1:20:43

years. I know there are a bunch of companies and

1:20:46

labs out there getting funding to use

1:20:48

AI to discover new drugs for certain

1:20:50

common illnesses. I also know that there

1:20:53

have been some companies that have raised

1:20:55

a bunch of money, used AI to

1:20:57

discover some drugs, and then went through

1:20:59

clinical trials and the drugs didn't

1:21:01

work. Or they didn't work as well

1:21:03

as the AI models predicted that they

1:21:05

would. So is there kind of

1:21:07

a step here that you all are taking?

1:21:10

Like the AI model predicts that these compounds

1:21:12

will work against MRSA, but then when you

1:21:14

go to test it in humans, it actually

1:21:16

doesn't work as well as your model predicted

1:21:19

it would. Is there a danger that there's

1:21:21

sort of some missing middle step there? Yeah,

1:21:23

for sure. And so how I like to

1:21:25

think about this is that AI in general

1:21:28

can help with one of two things. It

1:21:31

can help with discovering new compounds

1:21:33

for basic research and also preclinical

1:21:35

development as we do in our work. And

1:21:38

AI can also inform clinical trials and how do

1:21:40

you administer them. That I'm kind of

1:21:42

less an expert on, so I won't really comment too

1:21:44

much about that. But at least for the

1:21:46

former, using AI to discover

1:21:48

new compounds, basically it kind

1:21:51

of ends at that. We

1:21:53

really use AI as a tool to

1:21:55

discover new compounds that ultimately must be

1:21:57

tested and still rather traditionally.

1:22:00

So as I mentioned, even for antibiotics, we had to run

1:22:02

a battery of traditional

1:22:04

microbiological assays, experiments, to determine

1:22:06

what the mechanism of action

1:22:09

is. We had to... I

1:22:11

mean, AI did not help us with

1:22:13

dissecting the mouse or anything. So all

1:22:15

of that is quite traditional. But for

1:22:18

sure, I think things are still in

1:22:20

early days, as well as AI itself

1:22:22

might best be currently, at least utilized

1:22:24

for searching large surf spaces, as we

1:22:26

kind of mentioned. Right. It sounds like

1:22:28

the main thing that AI brings to

1:22:30

the process of drug discovery is just

1:22:33

being able to kind of shrink the

1:22:35

haystacks, take millions and millions of potential

1:22:37

chemical compounds and give you a list

1:22:39

of the 20 or 30 most

1:22:42

promising ones for treating a given disease.

1:22:44

Exactly. At least personally, that's how

1:22:46

I feel AI has created a

1:22:48

law of value. It's really for

1:22:50

initial stages of drug discovery, where you want

1:22:52

to shrink the haystack in order to make

1:22:55

things a bit more manageable. But once you

1:22:57

find a needle, I mean, there is no

1:23:00

guarantee that that needle is sharp, that you have

1:23:02

a great needle. And so I think

1:23:04

at least today

1:23:06

we still do not have great tools

1:23:08

to inform that process. Actually,

1:23:11

before you got here, Casey did actually

1:23:13

volunteer to be a human guinea pig

1:23:15

for any AI discovered drugs. Yeah, you

1:23:17

bring one of those MRSA syringes with

1:23:19

you. Unfortunately. Maybe we could expedite this

1:23:22

process and potentially sacrifice you

1:23:24

in the name of science. I mean, this

1:23:26

is really interesting, Kevin, because I think it

1:23:28

speaks to a question that we have had

1:23:30

over the past year or so, which is,

1:23:33

what is the ideal relationship between human beings

1:23:35

and artificial intelligence? Right. And what

1:23:37

Felix is describing for us here is

1:23:39

a system where people are able to use

1:23:42

AI to develop greater understanding, essentially working

1:23:45

like not hand in hand. That's like

1:23:47

to anthropomorphic. But they are using this

1:23:49

as a tool to further their own

1:23:51

research. It's not quite a creative tool,

1:23:54

but it is a tool that enables

1:23:56

human beings to be more creative while

1:23:58

deepening their scientific understanding. And this

1:24:00

was a really exciting thing. And to automate a

1:24:03

manual labor process that would take probably

1:24:05

centuries to do by hand, as

1:24:08

I hear you describe it, it's

1:24:10

basically creating a lab with tens

1:24:13

of thousands of scientists worth of labor

1:24:15

that you can use to go through

1:24:17

this huge list of compounds and screen

1:24:19

them all very quickly. Yeah, that's one

1:24:21

way to think about it. And

1:24:24

this idea of scale is quite important because,

1:24:26

at least in our paper, we looked at

1:24:28

12 million compounds in a candidate set. But

1:24:31

in principle, drug-like chemical space, which is

1:24:33

all possible really small

1:24:35

molecule compounds, 10 to the 60, 10

1:24:38

to the 60 compounds, that's basically

1:24:40

infinity for most practical purposes. So

1:24:43

you need a couple of postdocs to get through all that.

1:24:46

I have a last question, which is how close

1:24:48

are we to an AI that could actually

1:24:51

automate the testing part of this? It

1:24:54

seems sort of brutish and antiquated to

1:24:56

have to get a bunch of mice

1:24:58

and inject them with stuff and then

1:25:00

maybe move up to monkeys or some

1:25:02

other animal and then do it in

1:25:04

humans and have this whole long process.

1:25:07

Is there no way that you could

1:25:09

use AI to accurately simulate how

1:25:12

a mouse would react to a given compound?

1:25:14

Or do we still... is this sort of

1:25:16

hand in vivo testing? Do I

1:25:18

use in vivo correctly? That was beautiful.

1:25:20

No, that was great. Wow. I'm

1:25:23

not really a scientist I've ever... I'm so

1:25:25

happy with myself for remembering that fact from

1:25:27

biology class. Or is it

1:25:29

the case... He didn't remember it from... He just said it 10

1:25:31

minutes ago. That's true. That's true.

1:25:33

I'm sorry. So is it actually possible that

1:25:35

we could use AI in that phase of the

1:25:38

testing too? Yeah, that's a great question. So of

1:25:40

course there's a huge AI for science movement of

1:25:42

which this work is part of. I think parts

1:25:45

of science are still way too complex

1:25:47

for us to accurately model. And at

1:25:49

least personally I believe that includes how

1:25:51

do we simulate a whole mouse in

1:25:53

terms of all the organs, physiology, etc.

1:25:55

So I think we are still a

1:25:57

ways off from that. Perhaps

1:26:00

one of the things that we could also

1:26:02

consider is also using AI for robotics. And

1:26:04

so I think that is quite an interesting

1:26:06

field because eventually, if you use AI to

1:26:08

do science, you're going to have to interface

1:26:10

with the physical world. And of

1:26:12

course, that's something that, you know, not a

1:26:15

lot of companies are doing nowadays. So you're

1:26:17

saying it's possible that in a few years

1:26:19

we could have like an army of bacteria

1:26:21

resistant robot mice. It's

1:26:25

possible or I think maybe a

1:26:27

way to look at this

1:26:29

might be, you know, in the short term,

1:26:31

maybe AI could like automate like mouse forms

1:26:33

and like very high throughput experiments with handling

1:26:35

mice, especially if the robotics are correct. But

1:26:37

that would kind of look quite dystopian and

1:26:40

not quite like, you know, AI for science

1:26:42

that we have in mind. Yeah, well, I

1:26:44

just got a new idea for a screenplay.

1:26:46

So the rat in two easy quote we

1:26:48

never knew it on. All right. So

1:26:53

thank you so much for coming out. Part

1:27:28

4 is produced by Rachel Cohn and

1:27:30

Davis Land. We're edited by Jen Poyant.

1:27:32

This episode was fact checked by Mary

1:27:34

Mathis. Today's show was engineered

1:27:36

by Alyssa Moxley. Original

1:27:38

music by Mary Lozano, Diane

1:27:40

Wong, Pat McCusker and Dan

1:27:42

Powell. Our audience editor

1:27:44

is Nell Gologuie. Video production

1:27:46

by Ryan Manning and Dylan Bergison. If

1:27:49

you haven't already followed us on YouTube,

1:27:51

check us out. youtube.com/hard fork. Special

1:27:54

thanks to Paula Schumann, Kewing Tam,

1:27:56

Kate Lapreste and Jeffrey Miranda. As

1:27:59

always, you can... email us at

1:28:01

hardfork at nytimes.com. I

1:28:04

feel like we're wrapping so early. Let's do the show again just for

1:28:06

safety. Oh, let's go

1:28:08

get some sandwiches. I

1:28:12

feel like now that we've done it once, we could like really nail

1:28:14

it on the second go through. Yeah?

1:28:17

Yeah. Okay, all right, let's do it again. Okay.