Latest / Chaos Agents / Agentic Romance Scams, Open Source Safety, and Anything Bad on the Internet - With Juliet Shen

Transcript

Agentic Romance Scams, Open Source Safety, and Anything Bad on the Internet - With Juliet Shen · Chaos Agents

0:00
One
0:00
of
0:00
my
0:00
first
0:00
trust
0:00
and
0:00
safety
0:01
jobs
0:01
was
0:01
at
0:01
Tinder,
0:02
and
0:02
I
0:02
was
0:02
looking
0:02
into
0:02
romance
0:03
scams
0:03
all
0:04
the
0:04
time.
0:05
And
0:05
these
0:06
attackers
0:06
are
0:06
patient.
0:07
They're
0:07
willing
0:08
to
0:08
go
0:08
a
0:08
long
0:08
time,
0:09
and
0:09
in
0:09
the
0:10
past,
0:10
that
0:10
required
0:10
a
0:11
human
0:11
to
0:11
do
0:11
it.
0:12
I
0:12
think
0:12
when
0:12
we're
0:12
thinking
0:13
about
0:13
the
0:13
ways
0:13
that
0:14
agentic
0:14
AI
0:15
can
0:15
introduce
0:15
is,
0:15
like,
0:15
now
0:16
you
0:16
can
0:16
just
0:16
have
0:16
an
0:16
agent
0:17
doing
0:17
that,
0:18
and
0:18
it
0:18
can
0:18
do
0:18
that
0:19
with
0:20
50
0:20
or,
0:20
like,
0:20
5,000
0:21
conversations
0:22
at
0:22
the
0:22
same
0:22
time.
0:26
Hello,
0:28
we're
0:28
back.
0:28
Chaos
0:29
Agents,
0:29
the
0:29
podcast
0:30
where
0:30
we
0:30
learn
0:30
about
0:31
AI
0:31
so
0:31
you
0:31
don't.
0:31
You
0:31
don't
0:32
have
0:32
to.
0:33
Becca,
0:33
how
0:33
are
0:33
you
0:33
doing?
0:34
I
0:35
am
0:36
doing
0:38
great.
0:38
And
0:39
medium.
0:40
Medium.
0:40
Just
0:40
because
0:41
the
0:41
overall
0:42
state
0:42
of
0:43
affairs
0:43
in
0:44
the
0:44
world
0:44
and
0:45
America
0:46
is.
0:46
Yeah,
0:48
a
0:48
lot.
0:49
But
0:50
in
0:50
general,
0:50
I'm
0:51
doing
0:51
great.
0:51
How
0:51
are
0:52
you?
0:52
Yeah,
0:53
same
0:53
learning.
0:54
Things
0:55
are
0:55
moving.
0:55
It
0:56
feels
0:56
good
0:56
to
0:56
do
0:56
stuff
0:57
at
0:57
the
0:57
local
0:57
level.
0:57
So
0:58
that's
0:58
what
0:58
I've
0:58
been
0:58
trying
0:59
to
0:59
focus
0:59
on.
0:59
Girl,
1:00
what
1:00
have
1:00
you
1:00
been
1:00
doing
1:01
at
1:01
the
1:01
local
1:01
level?
1:03
So
1:03
I've
1:03
been
1:03
getting
1:04
involved.
1:04
There
1:05
is
1:05
St.
1:06
Luke's
1:06
it's
1:06
the
1:07
only
1:07
church
1:07
in
1:07
New
1:08
York
1:08
City
1:08
that
1:09
did,
1:10
in
1:10
the
1:10
80s,
1:11
did
1:11
services
1:11
for
1:12
folks
1:13
that
1:13
were
1:13
impacted
1:14
by
1:14
HIV.
1:15
They're
1:15
really
1:15
awesome.
1:16
There's
1:16
a
1:16
off
1:17
Broadway
1:17
show
1:17
that
1:17
was
1:17
about
1:18
them
1:18
called
1:19
Saturday
1:19
Church,
1:20
and
1:20
they've
1:21
been
1:21
really
1:21
great.
1:22
And
1:22
so
1:22
I've
1:22
been
1:22
trying
1:22
to
1:23
put
1:23
my
1:23
energy
1:23
into
1:23
that.
1:24
They
1:24
do
1:24
a
1:24
lot
1:24
of,
1:24
like,
1:24
outreach.
1:25
That's
1:26
awesome.
1:26
I
1:26
love
1:27
that.
1:27
Yeah.
1:27
All
1:27
right.
1:28
St.
1:28
Luke's
1:28
they're
1:29
the
1:29
best.
1:29
Shout
1:29
out
1:29
to
1:29
St.
1:30
Luke's
1:30
in
1:30
the
1:30
West
1:30
Village.
1:31
Shout
1:31
out
1:31
to
1:31
St
1:31
Luke's
1:32
oh,
1:32
that
1:32
reminds
1:33
me
1:33
of
1:34
a
1:34
shout
1:34
out.
1:35
Yeah.
1:35
Have
1:36
a
1:36
shout
1:36
out.
1:36
We
1:36
have
1:37
a
1:37
shout
1:37
out.
1:38
We
1:38
want
1:38
to
1:39
give
1:39
a
1:39
shout
1:39
out
1:39
to
1:40
Mario
1:41
Savvy,
1:43
who?
1:45
We
1:45
love
1:45
you,
1:45
Mario.
1:46
Mario
1:47
wrote
1:48
the
1:48
kindest
1:50
shout
1:50
out
1:50
about
1:51
the
1:51
Chaos
1:52
Agents
1:52
podcast,
1:53
and
1:53
Mario's
1:54
not
1:54
even
1:54
related
1:55
to
1:55
either
1:55
of
1:56
us.
1:56
I
1:57
know.
1:57
We're
1:57
both
1:58
like,
1:58
is
1:58
Mario
1:59
your
1:59
best
1:59
friend?
1:59
Like
1:59
what?
2:00
Yeah.
2:00
Is
2:00
your
2:00
cousin?
2:01
You
2:02
paid
2:02
this
2:02
man.
2:04
It
2:04
was
2:04
so
2:04
nice.
2:04
It
2:05
was
2:05
so
2:05
nice.
2:05
When
2:05
you
2:05
put
2:05
something
2:06
in
2:06
the
2:06
Internet,
2:06
like,
2:07
in
2:07
the
2:07
world,
2:07
and
2:08
then
2:08
somewhat
2:08
people
2:09
receive
2:09
it
2:09
well,
2:09
it's.
2:09
It's
2:10
the
2:10
best
2:10
feeling.
2:11
Mario,
2:12
if
2:12
you're
2:12
listening,
2:12
we
2:13
really
2:13
appreciate
2:13
you
2:14
and
2:14
thank
2:14
you
2:14
so
2:15
much
2:15
for
2:15
listening.
2:15
You
2:16
made
2:16
our
2:16
day.
2:17
Okay,
2:17
Sarah,
2:17
what
2:18
are
2:18
you
2:18
learning
2:18
about?
2:18
So
2:19
I
2:19
learned
2:19
about
2:19
two
2:20
things
2:20
this
2:20
week,
2:20
and
2:20
I
2:21
want
2:21
you
2:21
to
2:21
choose
2:21
your
2:21
own
2:22
adventure
2:22
on
2:23
which
2:23
you
2:24
think
2:24
is
2:24
more
2:24
interesting.
2:26
Okay.
2:26
One
2:27
is
2:27
more,
2:27
like,
2:28
businessy,
2:29
and
2:29
one
2:30
is
2:31
more
2:32
techie.
2:33
Actually,
2:34
I
2:34
just
2:34
chose
2:34
for
2:34
you.
2:34
I'm
2:35
gonna
2:35
go
2:35
with
2:35
the
2:35
tech
2:35
one
2:36
because
2:36
here's
2:36
why.
2:37
You
2:38
are
2:38
an
2:38
engineer.
2:39
I
2:39
am
2:39
not.
2:40
I'm
2:40
trying
2:41
to
2:42
learn
2:42
more
2:42
and
2:42
I
2:43
feel
2:43
like
2:43
I
2:43
just
2:43
learned
2:43
something.
2:44
It's
2:44
like
2:44
sort
2:44
of
2:45
somewhat
2:45
of
2:45
a
2:45
core
2:45
concept
2:46
that
2:46
I
2:46
want
2:47
to
2:47
like,
2:47
tell
2:48
you
2:48
what
2:48
I
2:48
learned
2:49
and
2:49
you
2:49
as
2:50
my
2:50
teacher,
2:51
tell
2:51
me
2:51
if
2:52
I
2:52
got
2:52
it
2:52
perfect.
2:53
Okay?
2:54
Okay,
2:54
Professor
2:55
Chips,
2:55
the
2:56
headline
2:56
of
2:57
what
2:57
I
2:57
learned
2:57
this
2:57
week
2:58
is
2:59
that
2:59
Replit
3:00
is
3:00
a
3:01
really
3:01
good
3:01
web
3:02
server,
3:02
but
3:02
it's
3:02
a
3:03
bad
3:03
browser
3:04
host.
3:05
And
3:06
here's
3:07
how
3:07
I
3:08
learned
3:08
that
3:08
and
3:08
learned
3:09
what
3:09
that
3:09
even
3:09
means.
3:10
So
3:11
in
3:11
my
3:11
company,
3:11
Slack,
3:12
I
3:13
created
3:13
a
3:13
private
3:13
Slack
3:14
channel
3:14
that
3:14
only
3:15
I
3:16
am
3:16
a
3:17
member
3:17
of
3:17
called
3:18
Messages
3:18
Go.
3:19
You're
3:19
like,
3:19
hey,
3:19
that
3:19
guy
3:19
has
3:20
her
3:20
name.
3:20
Yeah,
3:21
it's
3:21
literally
3:21
just
3:21
me
3:22
posting
3:22
stuff
3:22
to
3:22
me
3:23
and
3:23
it's
3:23
called
3:24
Becca
3:24
Feed.
3:25
And
3:25
it's
3:26
learnings,
3:27
thoughts,
3:27
whatever.
3:28
Throughout
3:28
the
3:29
day,
3:29
I
3:29
just
3:30
kind
3:30
of
3:30
drop
3:30
in
3:30
there
3:31
and
3:31
I
3:32
built
3:32
a
3:32
Replit
3:33
web
3:33
app
3:33
that
3:34
basically
3:34
takes
3:36
those
3:36
Slack
3:36
posts
3:37
and
3:37
like
3:38
renders
3:38
them
3:39
in
3:39
a
3:39
really
3:39
nice,
3:40
like,
3:40
public
3:41
facing
3:41
ui.
3:42
It's
3:42
almost
3:42
kind
3:43
of
3:43
like
3:43
my
3:43
own,
3:43
like,
3:44
little
3:44
Twitter
3:44
feed
3:45
because
3:45
I
3:45
wanted
3:45
to
3:45
be
3:45
able
3:45
to
3:46
like
3:46
share
3:46
that
3:46
with
3:47
people.
3:47
And
3:47
I
3:48
hate
3:49
X
3:50
and
3:51
don't
3:51
want
3:51
to
3:52
go
3:52
on
3:52
it.
3:53
And
3:53
so
3:54
I
3:54
was
3:55
like,
3:55
I'll
3:55
just
3:55
build
3:55
my
3:55
own.
3:56
But
3:56
then
3:57
I
3:57
was
3:57
like,
3:58
fuck.
3:58
Like,
3:59
I
3:59
work
3:59
in
3:59
vc.
4:00
I
4:00
feel
4:01
the
4:01
need
4:01
to
4:02
be
4:02
active
4:02
on
4:03
X.
4:03
You
4:04
can
4:04
hear
4:04
the
4:04
lack
4:04
of
4:04
enthusiasm
4:05
on
4:05
my,
4:05
in
4:06
my
4:06
voice.
4:07
So
4:07
I
4:07
was
4:07
like,
4:07
how
4:08
can
4:08
I
4:08
figure
4:08
out
4:09
an
4:09
automation
4:10
where
4:10
like,
4:10
I
4:10
can
4:10
just
4:10
keep
4:10
doing
4:11
this
4:11
but
4:11
like,
4:11
I
4:11
never
4:11
have
4:11
to
4:12
touch
4:12
X?
4:13
So
4:13
I
4:13
built
4:14
an
4:14
automation
4:14
where
4:15
when
4:15
one
4:16
of
4:16
my
4:16
posts
4:16
in
4:16
the
4:16
feedback
4:17
is
4:18
tweet,
4:18
character,
4:19
length,
4:20
it
4:20
automatically
4:21
cues
4:21
it
4:22
up
4:22
and
4:22
post
4:22
it
4:22
on
4:22
my
4:23
Twitter
4:23
as
4:23
if
4:24
I
4:24
had
4:24
posted
4:24
it.
4:25
Cool.
4:25
But
4:25
for
4:26
the
4:26
stuff
4:26
that's
4:26
too
4:26
long,
4:27
I
4:28
was
4:28
like,
4:28
what
4:29
should
4:29
I
4:29
do
4:29
with
4:29
that?
4:29
And
4:30
I
4:30
came
4:30
up
4:30
with
4:30
a
4:30
couple
4:30
ideas.
4:30
I
4:31
was
4:31
like,
4:31
maybe
4:31
I
4:31
should
4:31
summarize
4:32
it.
4:32
And
4:32
then
4:32
like,
4:32
that
4:32
didn't
4:33
work
4:33
super
4:33
well.
4:33
And
4:34
then
4:34
I
4:34
was
4:34
like,
4:34
you
4:34
know
4:34
what
4:34
would
4:35
be
4:35
a
4:35
cool
4:36
experiment?
4:37
What
4:37
If
4:37
I
4:37
could
4:38
use
4:38
Playwright
4:39
and
4:39
Chromium
4:40
to
4:41
screenshot
4:42
the
4:42
post
4:43
and
4:44
tweet
4:44
it
4:44
as
4:45
an
4:45
image
4:46
automatically.
4:46
Screenshot,
4:47
like
4:47
in
4:48
your
4:48
fancy
4:48
ui?
4:49
Correct.
4:50
So
4:51
in
4:51
the
4:51
ui,
4:52
like
4:52
screenshot
4:53
what
4:53
the
4:54
message
4:55
looks
4:55
like
4:56
and
4:56
like,
4:57
exactly
4:57
if
4:58
it's
4:58
longer.
4:58
Yeah,
4:59
okay,
4:59
cool.
5:00
Smart.
5:00
Okay,
5:01
thanks.
5:01
I
5:02
sort
5:02
of
5:02
brainstormed
5:03
that
5:03
with
5:03
Gemini.
5:04
Gemini.
5:04
And
5:05
I
5:05
came
5:05
up
5:05
with,
5:05
okay,
5:06
so
5:06
with
5:07
Gemini,
5:08
sort
5:08
of
5:09
went
5:09
about
5:09
figuring
5:09
out
5:10
how
5:10
to
5:10
implement
5:10
this
5:11
and
5:11
then
5:12
got
5:12
to
5:12
work
5:12
with
5:12
my
5:12
Replit
5:13
agent
5:13
in
5:14
like
5:14
implementing
5:15
that
5:15
framework.
5:16
And
5:16
I
5:16
got
5:17
it
5:17
sort
5:17
of
5:17
working
5:18
in
5:18
dev.
5:19
But
5:19
then
5:19
when
5:19
I
5:20
pushed
5:20
my
5:21
code
5:21
to
5:21
production,
5:22
it
5:22
wasn't
5:23
working
5:24
in
5:24
production,
5:25
and
5:25
I
5:25
couldn't
5:25
figure
5:26
out
5:26
why.
5:26
I
5:26
kept
5:26
going
5:27
back
5:27
and
5:27
forth
5:27
with
5:28
the
5:28
replit
5:28
agent.
5:28
I
5:29
was
5:29
like,
5:29
this
5:29
isn't
5:29
working,
5:29
this
5:30
isn't
5:30
working.
5:30
And
5:30
it
5:31
would
5:31
try
5:31
something
5:31
and
5:31
nothing
5:32
would
5:32
get
5:33
fixed.
5:33
And
5:34
then
5:34
I
5:34
figured
5:35
out
5:35
why.
5:35
And
5:36
this
5:36
is
5:36
the
5:36
thing
5:36
I
5:37
learned
5:37
that
5:37
I
5:37
want
5:38
Professor
5:38
Chips
5:39
to
5:39
evaluate
5:40
me
5:40
on.
5:40
Okay,
5:41
great.
5:41
So
5:41
what
5:41
I
5:41
learned
5:42
is
5:42
I
5:43
was
5:43
running
5:43
into
5:44
that
5:44
Playwright's
5:45
browser
5:46
installation
5:46
was
5:47
failing
5:47
during
5:48
the
5:48
build
5:48
process
5:49
of
5:49
going
5:49
to
5:49
production
5:50
because
5:50
it
5:51
was
5:51
trying
5:51
to
5:51
like,
5:52
basically
5:53
install
5:54
playwright
5:55
libraries
5:55
at
5:55
like
5:55
the
5:56
root
5:56
level.
5:57
And
5:57
replit's
5:58
build
5:58
environment
5:59
does
6:00
not
6:00
allow
6:00
that.
6:00
The
6:01
replit
6:01
build
6:01
environment
6:02
does
6:02
not
6:02
allow
6:03
installs.
6:03
Replit's
6:05
production
6:06
code
6:06
is
6:06
like
6:07
a
6:07
very
6:07
locked
6:08
down
6:08
container.
6:09
It's
6:09
like
6:09
a
6:09
sandbox.
6:10
And
6:10
that.
6:10
Yeah,
6:10
you
6:11
can't
6:11
do
6:11
like
6:11
a
6:11
bunch
6:11
of
6:12
stuff.
6:12
Okay,
6:12
so
6:13
exactly
6:13
what
6:13
you
6:13
just
6:14
said,
6:14
I
6:14
was
6:14
like,
6:15
yeah,
6:15
I've
6:15
heard
6:16
these
6:16
terms
6:17
over
6:17
the
6:17
course
6:17
of
6:17
my
6:18
time
6:18
in
6:18
tech,
6:18
sandbox,
6:19
container,
6:20
blah,
6:20
blah,
6:20
blah.
6:21
And
6:21
like,
6:21
I
6:22
loosely
6:23
understood
6:23
what
6:23
it
6:24
meant,
6:24
but
6:24
like,
6:24
not
6:24
really.
6:25
And
6:25
so
6:26
this
6:26
is
6:26
what
6:26
I
6:27
finally,
6:27
like,
6:27
learned.
6:28
So
6:29
here's
6:29
my
6:30
attempt
6:30
at
6:30
explaining
6:31
what
6:31
a
6:31
container
6:32
is.
6:32
A
6:32
container
6:33
is
6:33
kind
6:34
of
6:34
like
6:35
a
6:35
hotel
6:36
room
6:36
for
6:36
your
6:37
code.
6:37
When
6:38
you
6:38
stay
6:38
in
6:38
a
6:38
hotel,
6:38
you
6:39
have
6:39
everything
6:39
you
6:39
need.
6:40
You've
6:40
got
6:40
bed,
6:41
bathroom,
6:41
electricity,
6:43
but
6:43
you
6:43
can't
6:44
knock
6:44
down
6:45
walls
6:46
and
6:47
rewire
6:48
the
6:48
electricity
6:49
or
6:49
change
6:50
the
6:50
plumbing
6:50
in
6:51
your
6:51
hotel
6:51
room.
6:52
And
6:52
essentially
6:53
a
6:53
container
6:53
works
6:54
the
6:54
same
6:54
way.
6:54
It's
6:54
like,
6:55
here's
6:55
the
6:55
box.
6:56
Stay
6:56
within
6:57
these
6:57
bounds.
6:58
We've
6:58
already
6:58
got
6:58
everything
6:58
you
6:59
need.
6:59
But
6:59
like,
6:59
don't
7:00
try
7:00
and
7:00
like,
7:00
rebuild
7:01
anything
7:01
here.
7:02
Yeah,
7:02
and
7:02
that's
7:03
absolutely
7:03
correct,
7:04
I'm
7:04
guessing
7:05
right
7:05
now.
7:05
But
7:05
I
7:06
think
7:06
the
7:06
difference
7:06
between
7:06
a
7:07
sandbox
7:07
and
7:07
a
7:07
container,
7:08
because
7:09
I
7:09
think
7:09
the
7:10
word
7:10
container
7:11
specifically
7:12
applies
7:14
to
7:15
things
7:15
like
7:16
kubernetes
7:17
where
7:17
it's
7:18
like
7:18
data.
7:19
It's
7:19
like
7:19
a
7:19
data
7:20
container.
7:20
You
7:20
know
7:20
what,
7:21
I'm
7:21
gonna
7:21
actually.
7:22
Yeah,
7:23
yeah,
7:23
look
7:23
this
7:23
up
7:24
because,
7:24
you
7:24
know,
7:24
like,
7:25
docker.
7:26
Isn't
7:26
a
7:26
docker
7:27
a
7:27
container?
7:28
Yes.
7:29
Okay.
7:30
And
7:30
is
7:30
that
7:31
different
7:31
than
7:32
a
7:32
data
7:33
container?
7:33
No,
7:34
that's
7:34
exactly
7:34
what
7:35
that
7:35
data
7:35
container
7:35
is.
7:36
So
7:36
the
7:36
container
7:37
is
7:37
exactly
7:37
what
7:38
you're
7:38
talking
7:38
about.
7:39
Like,
7:39
it
7:39
shares
7:40
the
7:40
host
7:40
OS
7:41
kernel,
7:41
but
7:42
you
7:42
don't
7:42
have
7:43
access
7:43
to
7:44
the
7:44
actual
7:44
operating
7:45
system
7:45
or
7:46
like
7:46
labeled
7:47
mechanism
7:47
on
7:48
your
7:48
operating
7:48
system,
7:49
where
7:49
sandbox
7:50
is
7:50
more
7:50
like
7:51
not
7:51
based
7:52
on
7:52
a
7:52
specific
7:53
technology.
7:54
Okay.
7:55
And
7:55
my
7:56
question
7:56
for
7:57
myself
7:57
was,
7:57
like,
7:57
why
7:58
do
7:58
containers
7:59
exist?
8:00
And
8:01
from
8:01
my
8:01
understanding,
8:02
it's
8:02
because,
8:03
like,
8:03
before
8:04
containers,
8:05
developers
8:05
used
8:05
to
8:05
be
8:06
like,
8:06
I
8:06
built
8:07
this
8:07
thing,
8:07
it
8:07
works
8:07
on
8:07
my
8:07
machine,
8:08
and
8:08
then
8:09
I
8:09
hand
8:10
it
8:10
off
8:10
to
8:10
you
8:11
and
8:11
it
8:11
crashes
8:12
because
8:13
you
8:13
have
8:13
a
8:13
different
8:14
version
8:14
of
8:14
Linux
8:15
or
8:15
we're
8:15
missing
8:16
a
8:16
specific
8:16
file.
8:17
And
8:17
so
8:17
what
8:18
a
8:18
container
8:18
does
8:18
is,
8:18
it
8:19
like,
8:19
packages
8:19
your
8:20
code
8:20
together,
8:20
the
8:20
specific
8:21
files
8:21
and
8:21
things
8:21
it
8:22
needs
8:22
to
8:22
run
8:22
so
8:22
that
8:23
it's,
8:23
like,
8:23
portable
8:24
and
8:24
isolated
8:25
and
8:25
you,
8:25
like,
8:26
can
8:27
distribute
8:27
it
8:28
or
8:28
share
8:28
it.
8:29
Is
8:29
that
8:29
essentially
8:30
correct?
8:31
So
8:31
if
8:31
you
8:32
take
8:32
a
8:32
step
8:32
back,
8:33
I
8:33
think
8:33
one
8:34
of
8:34
the
8:34
big
8:34
things
8:34
about
8:34
containers
8:35
is
8:35
we
8:35
all
8:35
used
8:36
to
8:36
have
8:36
to
8:36
have
8:36
our
8:36
own
8:37
full
8:37
servers.
8:38
Like,
8:38
in
8:38
order
8:39
to
8:39
get
8:39
something
8:39
on
8:39
the
8:40
Internet,
8:40
you
8:40
had
8:40
to
8:41
have
8:41
your
8:41
own
8:41
server
8:42
to
8:42
host
8:42
it.
8:43
You
8:43
had
8:43
to
8:43
own
8:44
a
8:44
server
8:44
where,
8:44
like,
8:45
I
8:45
would
8:45
take
8:45
Becca's
8:46
website
8:46
and
8:46
host
8:46
it
8:47
on
8:47
my
8:47
server.
8:47
I
8:47
would.
8:48
But
8:48
then
8:48
what
8:48
containers
8:49
did
8:49
too
8:49
is
8:49
allow
8:50
us
8:50
to
8:50
use
8:50
the
8:51
same,
8:51
same
8:51
server
8:52
like
8:52
you
8:53
and
8:53
I.
8:53
But
8:53
I
8:54
don't
8:54
see
8:54
your
8:54
stuff
8:55
and
8:55
you
8:56
don't
8:56
see
8:56
my
8:56
stuff.
8:57
And
8:57
so
8:58
neither
8:58
of
8:58
us
8:58
need
8:58
to
8:59
actually
8:59
own
9:00
the
9:00
server.
9:00
So
9:01
a
9:01
container
9:01
allows,
9:02
like,
9:02
exactly
9:02
what
9:03
you're
9:03
talking
9:03
about.
9:03
Like,
9:04
if
9:04
I
9:04
install
9:05
XYZ
9:05
in
9:06
my
9:06
container
9:07
and
9:07
tell
9:07
Sarah
9:07
about
9:08
it,
9:08
then
9:08
she
9:09
can
9:09
install
9:09
XYZ
9:10
and
9:11
then
9:11
have
9:11
the
9:11
same
9:11
environment
9:12
that
9:12
I
9:12
do.
9:13
But
9:13
also
9:14
it
9:14
allows,
9:14
like,
9:15
us
9:15
to
9:16
use
9:16
the
9:16
same
9:16
environment
9:17
and
9:17
me
9:17
not
9:18
see
9:18
your
9:18
things
9:19
and,
9:19
and
9:20
have
9:20
that
9:20
security
9:20
boundary.
9:21
Oh,
9:21
that
9:22
makes
9:22
a
9:22
ton
9:22
of
9:22
sense.
9:23
Okay,
9:23
yeah,
9:24
interesting.
9:25
So,
9:25
like,
9:26
then
9:26
my
9:26
next
9:27
question,
9:28
which
9:28
was,
9:29
okay,
9:29
hang
9:30
on.
9:30
I
9:30
install
9:31
libraries
9:32
in
9:32
replit
9:33
projects
9:33
all
9:34
the
9:34
time.
9:34
How
9:35
do
9:35
I
9:35
know
9:36
what
9:36
libraries
9:37
I
9:37
can
9:37
install
9:38
and
9:38
what
9:38
is
9:38
like,
9:38
at
9:39
the
9:39
root
9:39
level?
9:40
And
9:41
what
9:41
I
9:42
think
9:42
I
9:42
understand
9:43
is,
9:43
like,
9:44
there's
9:44
like
9:44
different
9:45
types
9:45
of,
9:46
let's
9:46
say,
9:46
libraries.
9:47
There's
9:47
stuff
9:48
that
9:48
lives
9:48
in,
9:48
like,
9:49
your
9:49
backpack.
9:50
Like
9:50
it's
9:50
a
9:50
file.
9:51
So
9:51
like
9:52
Express
9:53
or
9:53
Pandas,
9:54
which
9:55
are
9:55
just
9:55
like
9:55
files
9:56
that
9:56
are
9:56
written
9:56
in
9:56
like,
9:56
JavaScript
9:57
or
9:57
Python,
9:57
and
9:58
you
9:58
can
9:58
just
9:58
add
9:58
them
9:58
as
9:59
files
9:59
to
9:59
your
9:59
project
9:59
directory
10:01
versus,
10:02
like,
10:02
libraries
10:03
that
10:03
are
10:03
like,
10:04
at
10:04
the
10:04
operating
10:05
system
10:05
level,
10:06
where
10:07
it's
10:07
like,
10:07
infrastructure
10:09
like
10:09
Chromium
10:09
is
10:10
not
10:10
just
10:10
a
10:10
library.
10:11
It's
10:11
like
10:11
a
10:11
really
10:12
big,
10:12
complex
10:12
piece
10:12
of
10:13
software
10:13
I
10:13
now
10:14
know
10:14
that
10:14
needs
10:15
to
10:15
talk
10:15
directly
10:15
to
10:16
the
10:16
computers,
10:16
like
10:17
video
10:17
drivers,
10:18
font
10:18
engines,
10:19
security
10:20
layers,
10:21
their
10:21
system
10:22
dependencies.
10:23
And
10:23
like,
10:24
that
10:24
lives
10:25
like,
10:26
in
10:26
the
10:26
building,
10:27
not
10:27
just
10:27
like
10:28
in
10:28
your
10:28
folder.
10:29
That's
10:29
a
10:29
really
10:30
good
10:30
point.
10:30
Yeah,
10:30
like,
10:31
it's
10:31
the
10:31
difference
10:31
between
10:32
like
10:32
a
10:33
package
10:33
like
10:33
Pandas
10:34
or
10:34
something
10:34
like
10:34
that
10:35
is
10:35
what
10:35
you're
10:35
saying.
10:36
Just
10:36
JavaScript.
10:36
So
10:36
that
10:37
could
10:37
be
10:37
just
10:37
like,
10:38
I'm
10:38
bringing
10:38
my
10:38
own
10:39
pillow
10:39
to
10:39
this
10:39
hotel
10:40
room,
10:40
but
10:40
something
10:41
like
10:41
binary
10:42
and
10:42
so
10:42
like
10:42
a
10:43
binary
10:43
gets
10:43
installed
10:44
on
10:44
the
10:45
computer
10:45
and
10:45
it
10:45
has
10:46
access
10:46
to
10:46
file
10:47
system.
10:47
The
10:47
reason
10:48
why
10:48
it
10:48
won't
10:48
let
10:48
you
10:49
install
10:49
Chromium
10:50
is
10:50
not
10:50
because
10:50
of
10:50
Chromium,
10:51
but
10:51
it's
10:51
so
10:52
you
10:52
can't
10:52
install
10:53
like
10:53
beck
10:54
as
10:54
bad
10:54
software,
10:55
you
10:55
know,
10:56
like
10:56
that's
10:56
going
10:56
to
10:57
wipe
10:57
action.
10:58
Yeah,
10:59
exactly,
10:59
exactly.
11:00
So
11:00
that
11:00
in
11:00
that
11:00
analogy,
11:01
it's
11:01
kind
11:01
of
11:01
like
11:02
I
11:02
can't
11:02
install
11:03
my
11:03
new
11:03
TV
11:03
in
11:04
the
11:04
wall
11:04
in
11:05
my
11:05
hotel
11:05
room.
11:06
Mm,
11:06
okay,
11:07
interesting.
11:08
Thank
11:08
you
11:08
for
11:08
helping
11:09
me
11:09
understand
11:09
that
11:10
better.
11:10
Yeah,
11:10
yeah,
11:11
it's
11:11
really,
11:11
it's,
11:12
it's.
11:12
And
11:12
also
11:13
hot
11:13
tip
11:13
from
11:14
me,
11:14
the.
11:15
Well,
11:15
no,
11:15
this
11:16
won't
11:16
affect
11:16
you.
11:16
Actually
11:17
one
11:17
thing
11:18
I've
11:18
run
11:18
into
11:19
give.
11:19
Us
11:19
the
11:19
hot
11:19
tip.
11:20
The
11:20
hot
11:20
tip
11:20
is
11:21
often
11:21
if
11:22
it's
11:22
since
11:22
it's
11:22
your
11:23
website,
11:23
this
11:23
won't
11:23
happen.
11:24
But
11:24
often
11:24
when
11:25
you're
11:25
trying
11:25
to
11:25
use
11:26
do
11:26
screenshots
11:27
on
11:28
a
11:28
platform
11:28
like
11:28
Replay
11:29
or
11:29
Lovable
11:30
or
11:30
like
11:30
anything
11:31
that
11:31
allows
11:31
you
11:31
to
11:31
like
11:32
host
11:32
code
11:32
that
11:32
you
11:33
run
11:33
if
11:33
it's
11:34
on
11:34
aws.
11:35
Most
11:35
large
11:36
websites
11:37
look
11:37
for
11:38
like
11:38
browsers
11:39
on
11:39
AWS
11:40
and
11:40
block
11:41
them
11:41
in
11:41
some
11:42
way
11:42
because
11:42
they
11:43
know
11:43
it's
11:43
probably
11:44
like
11:44
screen
11:45
scraping
11:45
technology,
11:46
it's
11:47
probably
11:47
a
11:47
bot.
11:48
So
11:48
if,
11:49
if
11:49
you
11:49
ever
11:49
need
11:50
to
11:50
do
11:50
screenshots,
11:51
you
11:51
either
11:51
need
11:52
to
11:52
do
11:52
it
11:53
on
11:53
your
11:53
local
11:53
machine
11:54
or
11:54
you
11:54
need
11:55
to
11:55
use
11:55
there's
11:56
like
11:56
rotating
11:56
IP
11:56
addresses
11:57
that
11:57
you
11:57
can
11:57
buy.
11:58
Okay,
11:58
so
11:59
then
11:59
actually
12:00
on
12:00
this
12:00
note,
12:01
like
12:01
knowing
12:02
what
12:02
my
12:02
project
12:03
is
12:03
and
12:03
like
12:04
sort
12:04
of
12:04
what
12:04
I
12:04
want
12:04
to
12:05
do
12:05
here,
12:06
like
12:07
high
12:07
level
12:07
at
12:07
least,
12:07
and
12:08
we're
12:08
just
12:08
brainstorming
12:08
here,
12:09
like,
12:09
how
12:09
would
12:09
you
12:09
go
12:10
about
12:10
building
12:10
this
12:11
or
12:11
suggesting
12:11
I
12:12
build
12:12
this?
12:12
Yeah,
12:12
actually
12:13
this
12:13
is
12:13
a
12:13
good
12:13
one.
12:14
I
12:14
would
12:14
install
12:14
the
12:15
program
12:15
on
12:15
your
12:15
local
12:16
machine
12:16
because
12:17
you
12:17
can
12:17
have
12:17
Chromium,
12:18
like
12:19
you
12:19
can
12:19
install
12:20
Chromium
12:20
on
12:20
your
12:21
computer
12:21
and
12:22
there's
12:22
something
12:22
called
12:22
a
12:22
headless
12:23
browser.
12:23
And
12:24
what
12:24
a
12:24
headless
12:24
browser
12:25
does
12:25
is
12:25
like
12:25
it
12:25
won't
12:26
pull
12:26
up
12:26
a
12:26
browser.
12:27
You
12:27
won't
12:27
see
12:27
see
12:27
it.
12:28
So
12:28
it'll
12:28
just
12:28
be
12:28
running
12:29
in
12:29
the
12:29
background.
12:29
And
12:30
that's
12:30
not
12:30
gonna
12:30
cost
12:31
you
12:31
like
12:31
a
12:31
lot
12:31
of
12:31
bandwidth
12:32
to
12:32
just
12:32
like,
12:33
how
12:33
are
12:33
you
12:33
doing
12:34
it?
12:34
Are
12:34
you
12:34
doing
12:34
it
12:35
like
12:35
ever
12:35
so
12:35
often
12:36
you
12:36
check
12:36
your
12:36
site
12:36
to
12:37
see
12:37
if
12:37
there's
12:37
a
12:37
new
12:37
one.
12:37
Yeah.
12:38
So
12:38
basically
12:39
here's
12:39
how
12:40
it
12:40
currently
12:40
works.
12:41
You
12:41
know,
12:41
I'll
12:42
post
12:42
max
12:43
three
12:43
or
12:43
four
12:43
times
12:44
a
12:44
day.
12:44
It's
12:44
not
12:44
like
12:45
super,
12:45
super
12:45
frequent.
12:46
And
12:46
each
12:46
time
12:47
I
12:47
post,
12:47
a
12:48
web
12:48
hook
12:48
is
12:48
sent
12:49
to
12:49
a
12:49
zap
12:50
I
12:50
have
12:50
set
12:51
up
12:51
in
12:51
Zapier,
12:52
you
12:52
know,
12:52
has
12:52
a
12:52
payload
12:53
with
12:53
like
12:53
the
12:54
text
12:54
of
12:54
the
12:55
post
12:56
and
12:56
then
12:56
it
12:56
will
12:56
count
12:57
the
12:57
number
12:57
of
12:57
characters.
12:58
And
12:58
if
12:58
it's
12:58
like
12:58
under
12:59
50
12:59
characters,
13:00
it'll
13:00
say,
13:01
okay,
13:01
great,
13:02
just
13:02
make
13:03
this
13:03
into
13:03
a
13:03
tweet,
13:03
add
13:04
it
13:04
to
13:04
my
13:04
buffer
13:05
Twitter
13:05
queue,
13:06
and
13:06
it
13:06
will
13:06
tweet
13:07
it.
13:07
If
13:07
it's
13:08
over
13:08
250
13:08
characters,
13:09
then
13:09
it'll
13:09
route
13:09
it
13:09
through
13:10
this,
13:10
like,
13:10
other
13:10
path
13:11
that.
13:12
What
13:12
I
13:12
initially
13:12
was
13:12
trying
13:13
to
13:13
do
13:13
was
13:13
have
13:14
code
13:15
in
13:15
my
13:15
replit
13:16
web
13:16
app
13:17
that
13:17
for
13:17
each
13:17
post
13:18
to
13:18
create
13:19
a
13:19
screenshot
13:20
and
13:21
host
13:21
that
13:22
image
13:22
in
13:23
my
13:23
production
13:23
database
13:24
of
13:24
my
13:25
web
13:25
app.
13:26
So
13:26
every
13:26
post
13:27
has
13:28
a
13:28
image
13:29
associated
13:29
with
13:30
it
13:30
of
13:30
the
13:31
post,
13:31
so
13:32
that
13:32
when
13:32
that
13:32
payload
13:33
gets
13:33
sent
13:33
to
13:33
Zapier
13:34
and
13:34
it
13:35
gets
13:35
routed
13:35
down
13:35
this
13:35
other
13:36
path,
13:36
it
13:36
like,
13:36
has
13:37
the
13:37
image
13:37
already.
13:38
If
13:38
it's
13:38
in
13:38
that
13:39
category,
13:39
I
13:40
would
13:40
approach.
13:40
This
13:41
actually
13:41
is
13:42
when
13:43
I
13:43
post,
13:44
I
13:44
would
13:45
ping
13:45
my
13:45
local
13:45
machine,
13:46
listen
13:46
to
13:46
that
13:46
Slack
13:47
channel
13:47
to
13:48
see
13:48
if,
13:48
like,
13:48
something
13:49
comes
13:49
up,
13:49
and
13:49
then
13:50
run
13:50
that
13:51
process
13:51
for
13:52
getting
13:52
a
13:52
screenshot
13:53
after,
13:53
like,
13:53
if
13:54
it
13:54
sees
13:54
an
13:54
update
13:54
with
13:55
your
13:55
Slack
13:55
channel.
13:56
But
13:56
that
13:57
means
13:57
your
13:57
computer's
13:57
always
13:58
on.
13:58
There
13:59
has
13:59
to
13:59
be
13:59
a
13:59
service
13:59
somewhere
14:00
that
14:00
takes
14:00
HTML
14:01
CSS
14:01
and
14:02
creates
14:02
an
14:02
image
14:02
for
14:02
you.
14:03
Okay.
14:03
That
14:04
is
14:04
potential.
14:05
There
14:05
also,
14:05
I
14:06
think,
14:06
is
14:07
like
14:07
something
14:08
else
14:09
that
14:10
Gemini
14:11
told
14:11
me
14:11
about.
14:12
There's
14:12
something
14:12
called
14:13
screenshot1,
14:13
which
14:14
is
14:14
like
14:14
a
14:14
separate
14:15
service
14:15
that's
14:15
like
14:16
an
14:16
API
14:17
that.
14:17
It's
14:17
called
14:18
the
14:18
Screenshot
14:18
API
14:19
for
14:19
Developers,
14:19
and
14:20
it
14:20
renders
14:20
screenshots
14:21
in
14:21
one
14:21
simple
14:22
API
14:22
call
14:22
instead
14:22
of
14:23
managing
14:23
browser
14:23
clusters.
14:24
Oh,
14:24
great.
14:25
It'll
14:25
do
14:25
it
14:25
for
14:26
you.
14:26
That's
14:26
what
14:27
it
14:27
says,
14:27
based
14:28
on
14:28
me
14:28
reading
14:29
their
14:29
website
14:29
for
14:29
about
14:30
three
14:30
seconds.
14:30
But
14:31
to
14:31
your
14:31
point,
14:32
then
14:32
maybe
14:32
there's
14:33
other
14:33
added
14:33
complexities
14:34
around,
14:34
like
14:35
the
14:35
AWS
14:35
issue
14:36
you
14:36
were
14:36
talking
14:36
about
14:37
or,
14:37
like,
14:37
I
14:37
don't
14:37
know.
14:37
Well,
14:38
I'll
14:38
have
14:38
to
14:38
explore
14:39
that.
14:39
Yeah,
14:39
Yeah.
14:40
I
14:40
think
14:40
that
14:40
shouldn't
14:40
be
14:40
the
14:40
case
14:41
with
14:41
your
14:41
website.
14:41
So
14:41
that
14:42
actually
14:42
might
14:42
be
14:42
an
14:42
easier
14:42
thing
14:43
so
14:43
you
14:43
don't
14:43
have
14:43
to
14:43
run
14:44
the
14:44
whole,
14:44
like,
14:44
browser
14:45
process
14:45
yourself.
14:46
Okay,
14:46
I'm
14:46
going
14:46
to
14:47
play
14:47
around
14:47
with
14:47
that
14:48
and
14:49
I'll
14:49
report
14:49
back
14:50
next.
14:50
Great.
14:51
Amazing
14:51
episode.
14:53
Thank
14:53
you
14:53
for
14:53
the
14:53
brainstorming.
14:54
Yeah,
14:55
awesome.
14:55
Today
14:55
we're
14:56
talking
14:56
about
14:56
trust
14:57
and
14:57
safety
14:58
with
14:58
a
14:58
former
14:59
coworker
14:59
of
14:59
yours.
15:00
When
15:00
did
15:00
you
15:00
guys
15:00
work
15:01
together?
15:01
Yeah,
15:02
so
15:02
my
15:03
job
15:03
in
15:03
product
15:04
before
15:04
I
15:04
started
15:04
working
15:04
at
15:05
VC
15:05
was
15:05
at
15:06
Grindr,
15:07
which
15:07
I
15:08
was
15:08
so
15:08
fortunate
15:08
to
15:09
work
15:09
with
15:09
an
15:09
amazing
15:10
group
15:10
of
15:11
people,
15:11
all
15:11
of
15:11
whom
15:12
were
15:12
so
15:12
good
15:13
at
15:13
their
15:13
jobs.
15:13
And
15:14
one
15:14
of
15:15
the
15:15
product
15:16
directors
15:17
that
15:17
I
15:18
learned
15:19
the
15:19
most
15:19
from
15:20
was
15:20
Juliet
15:20
Shen,
15:21
who
15:21
was
15:22
sort
15:22
of
15:22
head
15:22
of
15:22
Trust
15:22
and
15:22
Safety
15:23
at
15:23
Grindr,
15:23
has
15:24
a
15:24
ton
15:24
of
15:24
experience,
15:25
is
15:25
now
15:25
building
15:26
open
15:26
Source,
15:27
trust
15:27
and
15:27
safety,
15:28
AI
15:28
stuff.
15:28
So
15:29
that's
15:29
amazing.
15:30
Buy
15:30
any
15:30
trust
15:31
and
15:31
safety
15:31
person,
15:31
you
15:31
know
15:32
coffee,
15:32
because
15:32
they're
15:32
keeping
15:32
you
15:33
safe
15:33
on
15:33
the
15:33
Internet.
15:37
All
15:37
right,
15:38
everyone,
15:39
so
15:39
excited
15:40
for
15:41
our
15:41
guest
15:42
on
15:43
the
15:43
POD
15:43
today
15:44
to
15:44
give
15:44
you
15:44
a
15:44
little
15:45
bit
15:45
of
15:45
a
15:45
TLDR
15:46
on
15:46
her
15:47
amazing
15:47
background.
15:48
She
15:48
essentially
15:49
built
15:49
out
15:50
trust
15:50
and
15:50
safety
15:51
platform
15:51
at
15:51
Snap.
15:53
Ever
15:53
heard
15:53
of
15:53
it?
15:54
And
15:54
then
15:55
also
15:55
at
15:55
Grindr,
15:55
which
15:56
is
15:56
where
15:56
we
15:56
met,
15:56
and
15:56
I
15:56
had
15:57
the
15:57
pleasure
15:57
of
15:57
working
15:58
with
15:58
her
15:58
and
15:58
learning
15:58
from
15:58
her.
15:59
And
15:59
then
15:59
after
16:00
that,
16:00
she
16:00
went
16:00
on
16:01
to
16:01
tackle
16:02
spam
16:02
and
16:02
abuse
16:03
at
16:03
Google,
16:04
did
16:04
some
16:05
research
16:05
on
16:05
trust
16:05
and
16:06
safety
16:06
at
16:07
Columbia,
16:08
and
16:08
now
16:09
she
16:09
is
16:10
in
16:10
her
16:11
entrepreneur
16:12
era,
16:12
building
16:13
open
16:13
source
16:14
software
16:15
for
16:15
trust
16:15
and
16:15
safety
16:16
as
16:16
a
16:16
co
16:16
founder
16:17
and
16:17
head
16:17
of
16:17
product
16:17
at
16:18
Roost,
16:18
which
16:18
is
16:18
a
16:19
nonprofit.
16:19
So
16:20
welcome,
16:21
Juliette,
16:21
to
16:21
Chaos
16:22
Agents.
16:22
Thank
16:23
you
16:23
both
16:23
so
16:23
much
16:23
for
16:24
having
16:24
me.
16:24
I'm
16:24
really
16:24
excited
16:25
to
16:25
be
16:25
here.
16:25
Thinking
16:26
we
16:26
could
16:26
start
16:27
pretty
16:28
foundational
16:28
here.
16:29
And
16:30
can
16:30
you
16:30
explain
16:31
what
16:32
is
16:32
trust
16:33
and
16:33
safety?
16:34
Yeah,
16:34
for
16:34
sure.
16:35
And
16:35
this
16:35
is
16:35
something
16:35
that
16:36
comes
16:36
up
16:36
actually,
16:36
often
16:37
just
16:37
because
16:38
I
16:38
think
16:38
that
16:39
it's
16:39
in
16:39
the
16:39
world
16:40
of
16:40
tech,
16:40
it's
16:40
like
16:40
a
16:40
relatively
16:41
new
16:41
term.
16:42
And
16:42
by
16:42
new,
16:42
I
16:42
mean
16:42
like
16:43
10
16:43
to
16:43
20
16:43
years
16:44
old,
16:44
but
16:44
it
16:45
hasn't
16:45
really
16:45
picked
16:46
up
16:46
steam
16:46
outside
16:47
of
16:47
the
16:47
actual
16:47
industry.
16:48
So
16:49
when
16:49
I'm
16:49
at
16:49
parties,
16:50
I
16:50
like
16:50
to
16:50
explain
16:50
trust
16:51
and
16:51
safety
16:51
as
16:51
like,
16:52
anything
16:52
bad
16:52
that
16:53
happens
16:53
on
16:53
the
16:53
Internet,
16:54
that's
16:55
trust
16:55
and
16:55
safety.
16:56
But
16:56
I
16:56
think
16:57
when
16:57
we're
16:57
actually
16:57
thinking
16:58
about
16:58
the
16:58
definition,
16:59
it's
16:59
a
16:59
pretty
16:59
broad
16:59
term
17:00
referring
17:00
to
17:00
both
17:01
kind
17:01
of
17:01
the
17:01
teams
17:02
and
17:02
functions
17:03
at
17:03
different
17:03
companies
17:03
and
17:04
platforms,
17:05
as
17:05
well
17:05
as
17:06
the
17:06
technology
17:06
that's
17:07
used
17:07
to
17:07
kind
17:07
of
17:07
protect
17:08
users
17:09
from
17:09
each
17:09
other
17:09
online
17:10
and
17:10
from
17:11
automated
17:12
stuff
17:12
online.
17:13
So,
17:13
yeah,
17:13
anything
17:14
bad
17:14
that
17:14
happens
17:14
on
17:14
the
17:14
Internet,
17:15
but
17:15
the
17:15
specific
17:16
term
17:16
for
17:17
everyone
17:17
who's
17:17
trying
17:17
to
17:17
protect
17:18
people
17:18
from
17:18
it.
17:19
And
17:19
can
17:20
you
17:20
give
17:21
a
17:21
couple
17:21
examples
17:22
of
17:22
what
17:23
that
17:23
looks
17:23
like
17:23
in
17:23
practice,
17:24
especially
17:24
from
17:24
a
17:25
product
17:25
perspective?
17:26
Maybe
17:26
based
17:26
on
17:27
some
17:27
of
17:27
your
17:27
experience
17:27
from
17:28
Snap
17:28
or
17:29
Grindr?
17:29
Yeah,
17:30
definitely.
17:30
So
17:31
trust
17:31
and
17:32
safety
17:32
was
17:32
first
17:32
named
17:33
as
17:33
a
17:33
term
17:33
at
17:34
ebay,
17:34
looking
17:35
at
17:35
marketplace
17:36
integrity.
17:36
So,
17:36
like,
17:37
fake
17:37
reviews,
17:37
people
17:38
selling
17:38
fake
17:38
products,
17:39
and
17:39
so
17:39
that
17:39
was
17:40
innately
17:40
an
17:41
impact
17:41
on
17:41
the
17:42
product.
17:42
How
17:42
can
17:43
they
17:43
kind
17:43
of,
17:43
like,
17:43
protect
17:44
their
17:44
marketplace
17:44
by
17:45
adding
17:45
in
17:45
different
17:45
guardrails,
17:46
different
17:46
features
17:47
that
17:47
make
17:47
it
17:47
harder
17:48
for
17:49
scammers
17:49
or
17:49
bad
17:50
actors
17:50
that
17:50
kind
17:50
of
17:50
exploit
17:51
that
17:51
platform?
17:52
I
17:52
think
17:52
in
17:52
today,
17:53
you
17:53
see
17:53
this
17:53
kind
17:54
of
17:54
appear
17:54
in
17:55
a
17:55
lot
17:55
of
17:55
consumer
17:55
technology.
17:56
So
17:57
there's
17:57
kind
17:57
of
17:57
the
17:58
more
17:58
preventative
17:58
safety
17:59
that
17:59
you
17:59
can
18:00
kind
18:00
of
18:00
weave
18:00
into
18:00
product.
18:01
And
18:01
I
18:02
feel
18:02
strongly
18:02
that
18:03
you're
18:03
a
18:03
really
18:03
good
18:04
product
18:04
manager
18:05
if
18:05
you're
18:05
also
18:05
including
18:06
that
18:06
as
18:06
part
18:06
of
18:06
your
18:07
product
18:07
design.
18:08
You're
18:08
thinking
18:08
through
18:08
all
18:09
the
18:09
worst
18:09
case
18:09
scenarios
18:10
of
18:10
like,
18:10
maybe
18:10
you
18:11
built
18:11
this
18:11
really
18:11
fun
18:11
feature
18:12
and
18:12
you
18:13
never
18:13
intended
18:13
for
18:13
it
18:13
to
18:14
have
18:14
this
18:14
completely
18:15
weird
18:15
edge
18:15
case
18:16
that
18:16
allows
18:16
people
18:16
to
18:17
harass
18:17
each
18:17
other.
18:18
But
18:18
you
18:18
should
18:18
be
18:18
thinking
18:19
about
18:19
that
18:19
as
18:19
you're
18:20
designing
18:20
and
18:20
thinking,
18:20
thinking
18:20
through
18:21
the
18:21
user
18:21
journeys.
18:22
So
18:22
that's
18:23
actually
18:23
a
18:23
concept
18:23
called
18:23
safety
18:24
by
18:24
design.
18:24
It's
18:25
like,
18:25
how
18:25
do
18:25
you
18:25
actually
18:25
bake
18:26
in
18:26
the
18:26
best
18:26
practices
18:27
of
18:27
keeping
18:28
people
18:28
safe
18:28
directly
18:29
into
18:29
that
18:30
product
18:30
experience?
18:31
And
18:31
it
18:31
could
18:31
be
18:31
something
18:32
like
18:32
maybe
18:33
adding
18:33
a
18:33
pause
18:34
to
18:34
when
18:35
you're
18:35
sending
18:35
messages
18:36
on
18:36
a
18:36
dating
18:36
app,
18:36
adding
18:37
a
18:37
little
18:37
intervention.
18:37
It's
18:38
like,
18:38
hey,
18:38
are
18:38
you
18:39
sure
18:39
you
18:39
want
18:39
to
18:39
say
18:39
that?
18:40
It
18:41
could
18:41
be
18:41
things
18:41
like
18:42
it
18:42
prioritizes
18:43
content
18:44
or
18:44
video
18:44
that's
18:45
shot
18:45
directly
18:45
from
18:46
the
18:46
camera
18:46
versus
18:47
uploaded
18:47
from
18:47
a
18:47
camera
18:48
poll.
18:48
Those
18:48
are
18:48
all
18:48
kind
18:49
of
18:49
like
18:49
small
18:49
product
18:50
decisions
18:50
that
18:51
have
18:51
a
18:51
huge
18:51
impact
18:52
on
18:52
the
18:53
level
18:53
of
18:53
risk
18:53
or
18:53
danger
18:54
that
18:54
a
18:54
user
18:54
can
18:55
have
18:55
online.
18:56
And
18:56
then
18:57
there's
18:57
kind
18:57
of
18:57
like
18:57
the
18:57
other
18:57
side
18:58
of
18:58
trust
18:58
and
18:58
safety,
18:58
which
18:59
is
18:59
bad
18:59
things
19:00
will
19:00
always
19:00
happen,
19:01
especially
19:01
if
19:01
there's
19:02
any
19:02
platform
19:02
that
19:03
allows
19:03
for
19:03
user
19:03
generated
19:04
content.
19:05
So
19:05
anyone
19:05
posting
19:06
a
19:06
message,
19:06
anyone
19:07
uploading
19:07
a
19:07
photo,
19:08
bad
19:08
stuff
19:08
is
19:09
going
19:09
to
19:09
happen
19:09
because
19:09
people
19:10
are
19:10
weird
19:10
and
19:10
people
19:10
like
19:11
being
19:11
weird
19:11
on
19:11
the
19:11
Internet.
19:12
So
19:13
I
19:13
think
19:13
on
19:13
the
19:13
product
19:13
side
19:14
there
19:14
you
19:14
can
19:14
think
19:14
through
19:15
kind
19:15
of
19:15
like,
19:15
what
19:16
can
19:16
users
19:16
do
19:17
if
19:17
they
19:17
encounter
19:18
something
19:18
that
19:18
makes
19:19
them
19:19
uncomfortable,
19:20
that
19:20
makes
19:20
them
19:20
feel
19:21
like
19:21
they
19:21
are
19:21
in
19:21
danger,
19:22
or
19:22
they
19:22
see
19:22
something
19:23
that
19:24
they
19:24
really
19:24
did
19:24
not
19:24
want
19:25
to
19:25
see?
19:26
Are
19:26
there
19:26
pathways
19:27
that
19:27
are
19:27
intuitive
19:28
for
19:28
them
19:28
to
19:28
kind
19:29
of
19:29
let
19:29
the
19:29
company
19:30
know,
19:30
like,
19:30
hey,
19:31
this
19:31
thing
19:31
happened
19:32
to
19:32
me,
19:32
I
19:32
need
19:32
to
19:33
tell
19:33
you
19:33
about
19:33
it
19:33
so
19:33
that
19:34
you
19:34
can
19:34
actually
19:34
investigate
19:35
and
19:35
do
19:35
something
19:36
about
19:36
it.
19:36
What
19:37
I
19:37
also
19:37
like
19:37
is
19:38
can
19:38
you
19:38
build
19:38
in
19:38
controls
19:39
that
19:40
give
19:41
the
19:41
user
19:41
a
19:41
lot
19:41
more
19:42
autonomy
19:42
and
19:42
empower
19:43
them
19:43
to
19:43
make
19:43
decisions
19:44
on
19:44
their
19:44
own?
19:45
Grindr
19:45
did
19:45
this
19:46
really
19:46
well,
19:46
where
19:46
they
19:47
have
19:47
these
19:47
features
19:47
where
19:47
you
19:48
can
19:48
mute
19:48
someone
19:48
that
19:48
you're
19:49
having
19:49
conversation
19:49
with,
19:50
you
19:50
can
19:50
ban
19:51
them
19:51
yourself,
19:52
you
19:52
can
19:52
kind
19:52
of
19:52
control
19:53
your
19:53
own
19:53
experience
19:54
in
19:54
addition
19:54
to
19:55
flagging
19:55
it
19:55
to
19:55
the
19:56
company
19:56
to
19:56
take
19:57
a
19:57
closer
19:57
look.
19:57
And
19:58
then
19:58
there's
19:58
like
19:58
all
19:58
the
19:59
sausage
19:59
making
19:59
of
20:00
trust
20:00
and
20:00
safety
20:00
work,
20:01
which
20:01
is
20:01
like,
20:01
on
20:02
the
20:02
company
20:02
side,
20:03
what
20:03
are
20:03
the
20:03
types
20:04
of
20:04
technologies
20:04
and
20:04
tools
20:05
that
20:05
the
20:06
internal
20:06
teams
20:07
have
20:07
to
20:07
look
20:07
into
20:08
abuse,
20:09
Whether
20:10
one
20:10
of
20:10
the
20:10
users
20:11
flagged
20:11
it
20:11
for
20:11
the
20:12
company,
20:12
or
20:13
if
20:13
they
20:13
are
20:13
using
20:13
some
20:14
kind
20:14
of
20:14
classifier
20:15
or
20:15
model
20:16
to
20:16
try
20:16
to
20:16
proactively
20:17
find
20:17
something.
20:18
What
20:18
does
20:18
that
20:19
machine
20:19
look
20:19
like?
20:20
That's
20:20
really
20:20
neat.
20:20
I
20:21
feel
20:21
like
20:21
I
20:21
was
20:21
Talking
20:22
about
20:22
this
20:22
the
20:22
other
20:22
day,
20:22
and
20:23
I'm
20:23
sure
20:23
you've
20:23
seen
20:24
this,
20:24
I
20:24
feel
20:24
like
20:25
anytime
20:25
you
20:25
build
20:25
something,
20:26
even
20:26
when
20:26
you're
20:27
trying
20:27
to
20:27
help
20:27
people,
20:27
like
20:27
think
20:28
about
20:28
Roblox
20:28
like
20:29
the
20:29
most
20:29
like
20:29
innocent
20:30
thing,
20:30
kids
20:31
are
20:31
just
20:31
going,
20:31
having
20:32
the
20:32
best
20:32
time.
20:33
And
20:33
I
20:33
think
20:34
people
20:34
build
20:34
things
20:35
and
20:35
they
20:35
get
20:36
them
20:36
out
20:36
there
20:36
and
20:37
then
20:37
all
20:37
of
20:37
a
20:37
sudden
20:37
they're
20:38
like,
20:38
oh,
20:38
I
20:38
never
20:38
intended
20:39
for
20:40
this
20:40
crazy
20:41
use
20:41
case
20:41
that
20:41
like
20:42
so
20:42
weird
20:43
that
20:43
someone
20:43
figured
20:44
out
20:44
how
20:44
to
20:44
be
20:44
abusive
20:45
on
20:45
this
20:45
platform
20:46
or
20:46
how
20:46
to
20:46
take
20:46
advantage
20:46
of
20:47
someone
20:47
on
20:47
this
20:47
platform.
20:48
Because
20:48
of
20:48
your
20:48
background
20:49
in
20:49
trust
20:49
and
20:49
safety,
20:50
do
20:50
you
20:50
feel
20:50
like
20:50
you
20:50
think
20:50
about
20:51
building
20:51
differently?
20:52
Of
20:52
like,
20:52
do
20:53
you
20:53
just
20:53
assume,
20:53
okay,
20:54
if
20:54
this
20:54
is
20:55
successful,
20:55
someone's
20:56
going
20:56
to
20:56
use
20:56
it?
20:56
Weird.
20:57
How
20:57
can
20:57
I
20:57
make
20:58
sure
20:58
that
20:58
doesn't
20:58
happen?
20:58
Or
20:58
how
20:59
can
20:59
I
20:59
make
20:59
sure
20:59
they're
20:59
protected?
21:00
Yeah,
21:00
it's
21:00
definitely
21:01
something
21:01
I
21:01
think
21:01
about
21:01
when
21:02
I'm
21:02
building
21:02
products
21:03
and
21:03
also
21:03
when
21:03
I'm
21:03
talking
21:04
to
21:04
other
21:04
PMs.
21:04
You
21:05
know,
21:05
I
21:05
think
21:05
especially
21:05
if
21:05
you're
21:06
part
21:06
of
21:06
a
21:06
larger
21:06
company,
21:07
you're
21:07
interfacing
21:08
with
21:08
other
21:08
PMs
21:09
who
21:09
are
21:09
building
21:09
stuff
21:09
that
21:10
are
21:10
totally
21:10
in
21:10
a
21:11
different
21:11
scope
21:11
or
21:11
different
21:12
domain
21:12
or
21:12
different
21:12
surface,
21:13
right.
21:14
And
21:14
they
21:14
might
21:14
not
21:15
know
21:15
the
21:15
worst
21:16
case
21:16
scenario
21:16
of
21:16
like,
21:16
oh,
21:17
like,
21:17
you
21:17
know,
21:17
this
21:17
new
21:18
feature
21:18
is
21:18
coming
21:18
out.
21:19
I
21:19
don't
21:19
know
21:19
how
21:19
to
21:19
protect
21:20
it
21:20
because
21:20
I
21:20
don't
21:21
have
21:21
this
21:21
paranoid
21:21
brain
21:22
that's
21:22
constantly
21:23
thinking
21:23
about
21:23
the
21:23
ways
21:23
that
21:24
people
21:24
can
21:24
misuse
21:25
things.
21:25
I
21:25
think
21:26
that
21:26
it's
21:26
important
21:27
to
21:27
not
21:27
over
21:28
index
21:28
on
21:28
those
21:28
things.
21:29
And
21:29
that's
21:29
a
21:29
really
21:30
fast
21:30
way
21:30
of
21:31
kind
21:31
of
21:31
like
21:31
preventing
21:31
innovation
21:32
and
21:32
scale
21:33
of
21:33
the
21:33
product.
21:33
But
21:34
you
21:34
have
21:34
to
21:34
consider
21:34
and
21:35
factor
21:35
that
21:35
in.
21:36
It's
21:36
an
21:36
edge
21:36
case.
21:37
But
21:37
there
21:37
needs
21:37
to
21:37
be
21:38
a
21:38
solution
21:38
for
21:38
that
21:38
edge
21:39
case.
21:39
And
21:40
I
21:40
think
21:40
especially
21:40
from
21:41
my
21:41
time
21:41
doing
21:41
a
21:41
lot
21:41
of
21:42
this
21:42
trust
21:42
and
21:42
safety
21:42
work
21:43
over
21:43
the
21:43
years,
21:43
there's
21:44
this
21:44
instinct
21:44
and
21:45
there's
21:45
these
21:45
kind
21:45
of
21:46
developed
21:46
standards
21:46
of
21:46
like,
21:47
what
21:47
are
21:47
the
21:47
harms
21:47
that
21:47
are
21:48
absolutely
21:48
unacceptable
21:49
that
21:49
we
21:49
absolutely
21:50
need
21:50
to
21:50
reduce
21:51
the
21:51
risk
21:51
of?
21:52
And
21:52
what
21:52
are
21:52
ones
21:52
where
21:52
it's
21:53
like,
21:53
okay,
21:53
we
21:53
have
21:53
a
21:53
way
21:53
to
21:54
handle
21:54
that.
21:54
We
21:54
have
21:55
a
21:55
way
21:55
to
21:55
either
21:55
direct
21:55
that
21:56
to
21:56
a
21:56
team
21:57
or
21:57
direct
21:57
that
21:58
to
21:58
like
21:58
a
21:58
safety
21:59
model.
21:59
But
21:59
if
21:59
that
22:00
happens,
22:00
we
22:00
have
22:00
a
22:01
plan.
22:01
The
22:01
worst
22:02
thing
22:02
is
22:02
if
22:02
you
22:02
build
22:02
a
22:03
product
22:03
and
22:03
you
22:03
don't
22:03
have
22:03
a
22:04
plan
22:04
for
22:04
those
22:04
things.
22:05
There's
22:05
something
22:05
that's
22:06
been
22:06
top
22:06
of
22:06
mind
22:06
for
22:06
me
22:06
actually
22:07
recently
22:07
because
22:08
of
22:08
a
22:08
movie
22:08
I
22:08
watched
22:09
a
22:09
couple
22:09
of
22:09
nights
22:09
ago.
22:10
I'm
22:10
curious
22:10
if
22:10
you
22:10
guys
22:10
have
22:11
seen
22:11
this.
22:11
It's
22:12
with
22:12
Lili
22:13
Reinhart.
22:14
Do
22:14
you
22:14
know
22:14
What
22:14
I'm
22:14
talking
22:15
about.
22:15
Yes.
22:16
Yes,
22:16
I
22:16
do.
22:16
I
22:16
think
22:17
it's
22:17
called,
22:17
like,
22:17
American
22:18
Sweatshop
22:18
or
22:19
something.
22:19
American
22:19
Sweatshop,
22:20
Yes.
22:20
Thank
22:20
you.
22:20
Oh,
22:21
what
22:21
is
22:21
it?
22:21
So
22:21
I
22:21
haven't
22:22
seen
22:22
this.
22:22
Okay,
22:23
so
22:23
basically
22:24
she
22:25
plays
22:25
a
22:26
woman
22:26
who
22:26
works
22:27
in
22:27
a
22:28
content
22:29
moderating,
22:30
like,
22:30
center,
22:31
where,
22:32
you
22:32
know,
22:32
the
22:33
folks
22:33
who
22:33
are
22:34
like,
22:34
the
22:34
third
22:34
parties,
22:35
where
22:35
a
22:35
platform
22:36
like
22:36
Facebook
22:36
or
22:36
YouTube
22:37
or
22:37
whatever
22:37
is
22:37
like,
22:37
hey,
22:38
anything
22:38
that
22:38
gets
22:39
flagged
22:39
as
22:39
potentially,
22:40
like,
22:40
against
22:40
our
22:41
guidelines,
22:42
a
22:42
human
22:42
needs
22:43
to
22:43
watch
22:43
and
22:43
then
22:44
determine
22:44
if
22:44
it's,
22:45
like,
22:45
good
22:45
or
22:45
bad.
22:46
And
22:46
I'm
22:46
sure
22:47
you
22:47
guys
22:47
have
22:47
seen
22:48
things
22:48
about
22:49
the
22:49
impact,
22:50
psychologically
22:52
and
22:52
mental
22:52
health
22:53
wise,
22:53
that
22:53
that
22:54
has
22:54
on
22:55
folks
22:55
who
22:55
really
22:55
work
22:56
in
22:56
places
22:56
like
22:56
that.
22:57
My
22:57
thought
22:57
when
22:58
I
22:58
was
22:58
watching
22:58
that
22:58
was,
22:59
wow,
23:00
like,
23:00
this
23:00
is
23:00
potentially
23:01
a
23:01
great
23:02
area
23:03
to
23:03
apply
23:03
AI
23:04
where
23:04
you
23:05
can
23:05
have
23:05
vision
23:06
models
23:07
instead
23:07
of
23:07
human
23:08
beings
23:09
essentially
23:09
watching
23:10
these
23:10
videos,
23:10
determining
23:11
what's
23:11
going
23:11
on
23:12
and
23:12
coming
23:12
up
23:12
with
23:12
the
23:13
answer.
23:13
And
23:13
thus
23:13
you're
23:14
not
23:14
exposing
23:16
actual
23:16
people's
23:17
brains
23:17
to,
23:17
like,
23:18
trauma
23:18
and
23:19
the
23:19
impact
23:20
of
23:20
this.
23:20
I'm
23:21
just,
23:21
like,
23:21
curious
23:21
in
23:22
general
23:22
about
23:22
your
23:23
take
23:23
on
23:23
that
23:23
and
23:24
what
23:24
you're
23:25
seeing
23:25
in
23:25
the
23:25
space.
23:26
That
23:26
movie
23:27
definitely
23:27
made
23:27
the
23:27
rounds
23:28
in
23:28
all
23:28
the
23:28
trust
23:29
and
23:29
safety
23:29
kind
23:29
of
23:29
like
23:30
group
23:30
chats
23:30
and
23:30
slack
23:31
communities
23:31
I'm
23:31
in.
23:32
I
23:32
think
23:32
that
23:32
there
23:33
have
23:33
been,
23:33
like,
23:33
two
23:33
pieces
23:34
of,
23:34
like,
23:34
pop
23:35
culture
23:35
media
23:35
that
23:36
talk
23:36
about
23:36
the
23:36
content
23:37
moderator
23:37
experience.
23:38
There's
23:38
American
23:39
Sweatshop,
23:39
and
23:40
then
23:40
there's
23:40
this
23:40
play
23:41
that
23:41
was
23:41
on
23:41
Broadway
23:41
called
23:42
Job.
23:42
And
23:43
so,
23:43
like,
23:43
small
23:43
spoiler
23:44
warning
23:44
for
23:44
anyone
23:44
who's
23:44
listening
23:45
or
23:45
watching.
23:45
But,
23:45
like,
23:46
there's
23:46
this
23:46
emerging
23:47
trope
23:47
that
23:47
content
23:48
moderators
23:48
all
23:49
develop
23:49
psychosis
23:50
and
23:50
want
23:51
to
23:51
murder
23:51
someone.
23:51
And
23:52
I
23:52
don't
23:52
love
23:52
that.
23:55
Fair
23:56
way
23:56
to
23:56
call
23:56
that
23:57
out.
23:58
Not
23:58
a
23:58
huge
23:59
fan
23:59
of
23:59
that.
24:00
You
24:00
know,
24:00
could
24:00
we
24:00
pull
24:01
off
24:01
the
24:01
greatest
24:01
heist
24:01
in
24:02
history?
24:02
Probably.
24:03
But,
24:03
you
24:03
know,
24:04
the
24:04
murder
24:04
part,
24:04
maybe
24:05
not.
24:05
But,
24:05
you
24:06
know,
24:06
jokes
24:06
aside,
24:07
I
24:07
think
24:07
that
24:08
that
24:08
content,
24:08
like,
24:09
and
24:09
what
24:09
you
24:09
said,
24:10
Becca,
24:10
like,
24:10
it
24:10
pulls
24:11
at
24:11
a
24:11
few
24:11
different
24:11
points.
24:12
One
24:12
is
24:13
doing
24:13
that
24:13
work
24:13
is
24:14
deeply
24:14
traumatizing.
24:15
I
24:15
was
24:15
looking
24:15
around
24:16
to
24:16
see
24:16
if
24:16
I
24:16
have
24:16
this
24:16
book
24:17
to
24:17
show
24:17
you,
24:17
but
24:17
there's
24:17
this
24:18
book
24:18
by
24:18
this
24:19
researcher
24:20
named
24:20
Dr.
24:20
Sarah
24:21
T.
24:21
Roberts.
24:21
It's
24:22
called
24:22
behind
24:22
the
24:22
Screen,
24:23
and
24:23
it's
24:23
really
24:23
focused
24:24
entirely
24:24
on
24:25
those
24:25
kind
24:26
of
24:26
content
24:26
moderation
24:26
centers
24:27
and
24:27
really
24:27
diving
24:28
into,
24:28
like,
24:28
what
24:28
is
24:28
the
24:28
human
24:29
cost
24:29
of
24:29
doing
24:29
this
24:30
type
24:30
of
24:30
commercial
24:31
content
24:31
moderation
24:32
at
24:32
scale?
24:33
And
24:33
at
24:33
the
24:34
same
24:34
time,
24:34
there
24:35
are
24:35
AI
24:35
models
24:35
that
24:36
can
24:36
do
24:37
those
24:37
things
24:37
of,
24:37
like,
24:38
finding
24:38
blood
24:39
in
24:39
the
24:39
content.
24:40
You
24:40
know,
24:40
like,
24:40
that
24:40
is
24:40
something
24:41
that
24:41
computer
24:41
vision
24:42
is
24:42
pretty
24:42
good
24:42
at.
24:43
You
24:43
know,
24:43
whether
24:43
it
24:43
can
24:44
separate
24:44
between
24:44
like,
24:45
here's
24:45
a
24:45
plastic
24:45
surgery
24:46
video
24:46
or
24:46
a
24:47
pimple
24:47
popping
24:47
video
24:47
that
24:48
people
24:48
want
24:48
to
24:48
watch
24:49
versus
24:49
here's
24:49
a
24:49
gory
24:50
video
24:50
that
24:50
people
24:51
don't
24:51
want
24:51
to
24:51
watch.
24:51
Right.
24:52
That's
24:52
a
24:52
little
24:52
bit
24:52
harder.
24:53
But
24:53
I
24:53
think
24:53
with
24:53
this
24:54
new
24:54
generation
24:54
of
24:55
AI
24:55
models,
24:55
that
24:55
context
24:56
can
24:56
be
24:56
added.
24:57
I
24:57
think
24:57
a
24:57
lot
24:58
of
24:58
us
24:58
who
24:58
have
24:58
been
24:59
in
24:59
the
24:59
trust
24:59
and
24:59
safety
25:00
space
25:00
for
25:00
a
25:00
long
25:00
time
25:01
really
25:01
recognize
25:02
the
25:02
toll
25:03
that
25:03
it's
25:03
taken.
25:03
You
25:03
know,
25:04
I
25:04
myself
25:04
have
25:04
struggled
25:05
with
25:05
mental
25:05
health
25:05
over
25:06
the
25:06
years.
25:06
You
25:06
know,
25:06
I
25:07
often
25:07
joke
25:07
is,
25:08
hey,
25:08
maybe
25:09
we
25:09
don't
25:09
need
25:09
new
25:09
members
25:10
of
25:10
the
25:10
complex
25:11
PTSD
25:11
club.
25:12
Maybe
25:12
we
25:12
can
25:12
close
25:13
that
25:13
membership
25:13
door
25:14
in
25:14
the
25:14
context
25:14
of
25:15
trust
25:15
and
25:15
safety
25:16
and
25:16
deploy
25:17
AI
25:17
in
25:17
the
25:17
places
25:17
where
25:18
it
25:18
actually
25:18
makes
25:18
sense.
25:19
Because
25:19
I
25:19
think
25:20
there's
25:20
a
25:20
delicate
25:20
line
25:20
here,
25:21
right?
25:21
I
25:21
think
25:21
that
25:21
there's
25:22
a
25:22
lot
25:22
of
25:22
people
25:22
who
25:23
think
25:23
that
25:23
AI
25:24
will
25:24
replace
25:25
these
25:25
jobs
25:25
completely.
25:26
And
25:26
I
25:26
don't
25:27
think
25:27
that's
25:27
true
25:27
from
25:27
a
25:28
technical
25:28
standpoint.
25:28
I
25:28
don't
25:29
think
25:29
these
25:29
models
25:29
can
25:29
actually
25:30
take
25:30
in
25:30
the
25:30
complex
25:31
nuance
25:31
of
25:32
some
25:32
of
25:32
these
25:32
situations.
25:33
And
25:33
then
25:33
two,
25:33
I
25:33
do
25:33
think
25:34
there's
25:34
still
25:34
so
25:35
much
25:35
where
25:36
humans
25:36
need
25:36
to
25:37
kind
25:37
of
25:37
like
25:37
apply
25:37
the
25:37
human
25:38
judgment
25:38
and
25:38
apply
25:39
their
25:39
own
25:39
lived
25:39
experience.
25:40
But
25:40
for
25:41
things
25:41
like,
25:41
you
25:42
know,
25:42
CSAM
25:42
that.
25:43
Can
25:43
you
25:43
just
25:44
define
25:44
that
25:44
acronym
25:45
real
25:45
quick?
25:45
CSAM
25:46
stands
25:46
for
25:46
child
25:47
sexual
25:47
abuse
25:47
material.
25:48
I
25:48
think
25:49
in
25:49
years
25:49
past
25:49
people
25:50
called
25:50
it
25:50
child
25:50
pornography,
25:51
but
25:51
we've
25:52
been
25:52
trying
25:52
to
25:52
move
25:52
away
25:52
from
25:53
that
25:53
term
25:53
because
25:53
pornography
25:54
really
25:54
should
25:54
be
25:55
referring
25:55
to
25:55
consenting
25:56
adults.
25:57
And
25:57
children
25:57
cannot
25:58
consent.
25:58
So
25:58
CSAM
25:59
is
26:00
one
26:00
of
26:00
the
26:00
biggest
26:01
things
26:01
that
26:01
trust
26:02
and
26:02
safety
26:02
teams
26:03
have
26:03
to
26:03
deal
26:03
with
26:03
is
26:04
like
26:04
wherever
26:04
someone
26:05
could
26:05
upload
26:05
media,
26:06
there
26:06
is
26:06
most
26:06
likely
26:07
going
26:07
to
26:07
vcsam.
26:08
So
26:08
there
26:08
are
26:09
established
26:09
ways
26:10
to
26:10
kind
26:10
of
26:10
like
26:10
detect
26:11
it
26:11
and
26:12
not
26:12
expose
26:13
it
26:13
to
26:13
either
26:14
the
26:14
moderator
26:14
to
26:14
more
26:14
users.
26:15
And
26:16
the
26:16
reason
26:16
for
26:16
that
26:17
is
26:17
every
26:17
time
26:17
someone
26:18
sees
26:18
that,
26:18
that
26:19
survivor
26:19
who's
26:20
in
26:20
the
26:20
media
26:20
gets
26:20
revictimized
26:21
because
26:21
someone
26:22
is
26:22
looking
26:22
at
26:22
that
26:22
again.
26:23
So
26:23
there
26:24
are
26:24
some
26:24
things
26:24
where
26:24
it
26:25
does
26:25
make
26:25
sense
26:25
to
26:25
try
26:26
to
26:26
use
26:26
AI
26:26
on.
26:27
I'll
26:27
come
26:27
back
26:27
to
26:28
CSAM
26:28
because
26:28
that's
26:28
kind
26:29
of
26:29
a
26:29
sensitive
26:29
one
26:30
around
26:30
AI
26:30
usage
26:31
and
26:31
that.
26:31
But
26:31
things
26:32
like
26:32
Gore,
26:32
things
26:32
like
26:32
beheading,
26:33
people
26:33
don't
26:34
need
26:34
to
26:34
keep
26:34
seeing
26:35
that.
26:35
And
26:35
I
26:35
do
26:35
think
26:36
that's
26:36
a
26:36
huge
26:37
area
26:37
where
26:37
AI
26:37
can
26:38
help.
26:38
I
26:38
imagine
26:39
there's
26:39
a
26:39
lot
26:39
of
26:39
ethics
26:39
and
26:40
AI
26:40
conversations
26:41
happening
26:41
in
26:42
the
26:42
trust
26:42
and
26:42
safety
26:43
community.
26:43
Like
26:44
right
26:44
now
26:44
we're
26:45
in
26:45
a
26:45
world
26:45
where
26:45
all
26:45
the
26:46
AIs
26:46
that
26:46
we
26:46
generally
26:47
interact
26:47
with
26:47
are
26:48
very
26:48
friendly
26:48
and
26:48
warm
26:49
and
26:49
are
26:49
just
26:50
here
26:50
to
26:50
Support
26:50
us.
26:51
And
26:51
there
26:51
already
26:51
is
26:52
nefarious
26:52
AI,
26:53
like
26:53
when
26:54
it
26:54
comes
26:54
to
26:54
people
26:54
that
26:55
are
26:55
big
26:55
butchering
26:56
scams
26:56
and
26:56
things
26:56
like
26:57
that.
26:57
Is
26:57
that
26:57
conversation
26:58
happening
26:58
right
26:59
now
26:59
in
26:59
the
26:59
trust
26:59
and
26:59
safety
27:00
community?
27:00
How
27:01
are
27:01
people
27:01
thinking
27:01
about
27:01
that
27:02
and
27:02
how
27:02
to
27:02
approach
27:02
it?
27:03
A
27:03
theme
27:03
that
27:04
I've
27:04
heard
27:04
come
27:04
up
27:05
is
27:05
about
27:05
how
27:05
kind
27:06
of
27:06
the
27:06
harms
27:06
that
27:07
took
27:07
place
27:07
on
27:07
the
27:07
Internet
27:08
in
27:08
like
27:08
you
27:09
know,
27:09
early
27:09
mid-2000,
27:10
the
27:10
social
27:11
media
27:11
era,
27:12
they're
27:12
still
27:12
around,
27:13
but
27:13
they're
27:13
scaled
27:14
at
27:14
a
27:15
huge,
27:15
huge
27:16
rate
27:16
now
27:16
because
27:16
of
27:17
AI.
27:17
You
27:17
know,
27:17
a
27:17
lot
27:18
of
27:18
the
27:18
telltale
27:18
signals,
27:19
for
27:19
example,
27:20
on
27:20
dating
27:20
apps,
27:20
like
27:21
it
27:21
used
27:21
to
27:21
be
27:21
pretty
27:22
easy
27:22
to
27:22
spot
27:22
a
27:22
scammer
27:23
because
27:23
of
27:23
the
27:23
way
27:23
that
27:24
they
27:24
message
27:24
people,
27:25
because
27:25
you
27:25
know,
27:25
they'll
27:26
start
27:26
a
27:26
message
27:26
with
27:26
like,
27:26
hello
27:27
dear.
27:27
But
27:28
with
27:28
AI,
27:28
it's
27:28
like,
27:29
you
27:29
know,
27:29
they,
27:29
that
27:29
becomes
27:30
a
27:30
lot
27:30
more
27:30
like
27:30
natural
27:31
language.
27:32
It's
27:32
harder
27:32
to
27:32
spot
27:33
the
27:33
scammers
27:33
because
27:34
they're
27:34
all
27:34
being
27:35
run
27:35
through
27:35
AI
27:36
now.
27:36
And
27:36
of
27:37
course
27:37
scaling
27:37
it,
27:37
it's
27:38
a
27:38
lot
27:38
easier
27:38
to
27:38
create
27:38
botnets,
27:39
to
27:39
just
27:39
automate
27:40
anything
27:40
bad
27:41
you
27:41
want
27:41
to
27:41
do.
27:41
And
27:41
at
27:42
the
27:42
same
27:42
time
27:42
we
27:43
not
27:43
only
27:43
have
27:43
to
27:43
deal
27:44
with
27:44
these
27:44
larger
27:45
tidal
27:45
waves
27:45
of
27:45
the
27:46
old
27:46
harms,
27:46
but
27:46
there
27:47
are
27:47
brand
27:47
new
27:47
harms
27:48
that
27:48
AI
27:49
does
27:49
introduce.
27:50
I
27:50
think
27:51
we
27:51
have
27:51
yet
27:51
to
27:51
really
27:52
understand
27:52
the
27:53
impact
27:53
of
27:53
agentic
27:54
abuse
27:54
and
27:54
agentic
27:55
risk.
27:55
And
27:56
I
27:56
think
27:56
that
27:57
when
27:57
we
27:57
talk
27:57
about
27:57
open
27:58
source
27:58
AI,
27:59
there's
27:59
some
27:59
really
27:59
powerful
28:00
benefits
28:00
there.
28:00
And
28:01
of
28:01
course
28:01
there
28:01
are
28:01
some
28:01
trade
28:02
offs
28:02
on
28:02
the
28:03
benefit
28:04
side.
28:04
There's
28:05
so
28:05
much
28:05
more
28:05
transparency
28:06
to
28:06
understand
28:07
what
28:07
training
28:07
data
28:08
is
28:08
in
28:08
these
28:08
models,
28:09
how
28:09
are
28:09
they
28:09
developed,
28:10
who
28:10
developed
28:11
it,
28:11
what
28:11
was
28:11
that
28:12
methodology
28:12
and
28:13
how
28:13
did
28:13
they
28:13
really
28:13
perform.
28:14
I
28:14
think
28:14
that
28:15
is
28:15
inherently
28:16
safer
28:16
than
28:17
black
28:17
box
28:17
algorithms
28:18
or
28:18
models
28:18
where
28:19
you
28:19
don't
28:19
really
28:19
know
28:19
what's
28:19
happening
28:20
under
28:20
the
28:20
hood.
28:20
And
28:21
of
28:21
course,
28:21
by
28:22
making
28:22
it
28:22
democratic,
28:23
by
28:23
making
28:23
it
28:23
pluralist,
28:24
by
28:24
making
28:24
it
28:25
easier
28:25
to
28:25
access
28:26
this
28:26
type
28:26
of
28:26
cutting
28:27
edge
28:27
technology,
28:28
with
28:28
scale
28:29
comes
28:29
abuse.
28:29
There
28:30
will
28:30
be
28:30
people,
28:30
people
28:31
who
28:31
want
28:31
to
28:31
exploit
28:32
open
28:32
source
28:32
models.
28:32
But
28:33
I
28:33
do
28:33
think
28:33
that
28:34
if
28:34
we
28:34
push
28:35
towards
28:35
a
28:35
more
28:35
open
28:36
society
28:36
and
28:36
develop
28:37
AI
28:37
in
28:37
the
28:37
open,
28:38
there's
28:38
more
28:39
opportunities
28:39
for
28:40
collective
28:40
vigilance
28:41
to
28:41
understand
28:42
what
28:42
are
28:42
the
28:42
vulnerabilities
28:43
of
28:43
these
28:43
different
28:44
models
28:44
and
28:45
to
28:45
understand
28:45
how
28:46
they
28:46
could
28:46
be
28:47
potentially
28:47
abused.
28:48
I
28:49
want
28:49
to
28:49
definitely
28:50
get
28:50
into
28:50
like
28:51
on
28:51
the
28:52
point
28:52
of
28:52
open
28:52
source
28:53
models
28:53
and
28:54
trust
28:54
and
28:54
safety
28:54
and
28:54
like
28:55
what
28:55
you're
28:55
working
28:55
on
28:55
now.
28:55
But
28:55
you
28:56
just
28:56
said
28:56
something
28:56
that
28:57
really
28:57
piqued
28:57
my
28:57
curiosity
28:58
that
28:58
I
28:58
wanted
28:59
to
28:59
ask
28:59
you
28:59
about,
28:59
which
28:59
is
28:59
you
29:00
said
29:00
that
29:00
AI
29:01
models
29:01
introduce
29:02
new
29:03
types
29:04
of
29:04
harm
29:05
or
29:05
like
29:05
new
29:06
ways
29:06
of
29:07
facilitating
29:08
Abuse
29:08
online.
29:08
I'm
29:09
curious
29:09
if
29:09
you
29:09
could
29:10
like
29:10
tell
29:10
us
29:11
what
29:11
that
29:11
looks
29:12
like
29:12
and
29:12
like
29:12
both
29:13
from
29:13
the
29:13
perspective
29:13
of
29:14
like
29:14
you
29:14
as
29:14
a
29:14
technology
29:15
builder,
29:15
but
29:15
also
29:16
for
29:16
all
29:16
of
29:16
us
29:17
as
29:17
like
29:18
just
29:18
consumers
29:19
of
29:19
media
29:19
and
29:20
like
29:20
users.
29:20
Like,
29:21
what
29:21
kind
29:21
of
29:21
stuff
29:22
should
29:23
people
29:23
just
29:23
be
29:24
aware
29:24
is
29:25
like
29:25
potentially
29:26
happening
29:26
on
29:27
platforms
29:27
now?
29:28
I
29:28
think
29:28
that's
29:29
a
29:29
lot
29:29
of
29:29
the
29:29
things
29:29
where
29:30
in
29:30
prior
29:30
trust
29:31
and
29:31
safety
29:31
experience
29:32
we
29:32
kind
29:33
of
29:33
like
29:33
thought
29:33
about
29:33
things
29:33
in
29:34
like
29:34
broad
29:34
binaries.
29:35
One
29:35
is
29:35
like,
29:36
okay,
29:36
this
29:36
is
29:36
automated
29:36
abuse,
29:37
this
29:37
is
29:37
spam.
29:38
These
29:38
are
29:38
scams,
29:39
they're
29:39
bots,
29:39
they're
29:40
romance
29:40
scammers.
29:40
And
29:41
then
29:41
there's
29:41
kind
29:41
of
29:41
like
29:41
the
29:42
longer
29:42
running,
29:42
like
29:43
user
29:43
to
29:43
user
29:44
harms.
29:44
And
29:45
I
29:45
said
29:45
romance
29:45
scams
29:46
for
29:46
bots,
29:46
but
29:46
it
29:46
actually
29:46
belongs
29:47
more
29:47
in
29:47
that
29:47
kind
29:47
of
29:47
human
29:48
to
29:48
human
29:48
harm.
29:49
Like
29:49
these
29:49
are
29:49
going
29:49
to
29:49
be
29:49
longer
29:50
term
29:50
interactions
29:51
where
29:52
typically.
29:52
Did
29:52
you
29:53
say
29:53
romance
29:53
scams?
29:54
Romance
29:55
scams,
29:55
yes.
29:55
Okay,
29:57
so
29:57
this
29:57
is
29:57
like
29:57
when
29:58
someone's
29:58
like,
29:58
hey,
29:59
like,
29:59
like
30:00
trying
30:00
to
30:00
like
30:00
cozy
30:01
up
30:01
someone
30:01
and
30:01
then
30:01
being
30:01
like,
30:02
I
30:02
really
30:02
need
30:02
like
30:02
$5,000
30:04
for
30:05
my
30:06
whatever.
30:07
And
30:07
they're
30:07
like,
30:07
oh
30:08
sure
30:08
honey,
30:08
yeah.
30:09
Yeah,
30:09
yeah,
30:09
yeah,
30:09
yeah,
30:10
it's
30:10
pig
30:10
butchering.
30:11
But
30:11
it's
30:11
been
30:11
around
30:11
for
30:12
ages.
30:12
You
30:12
know,
30:12
like,
30:13
one
30:13
of
30:13
my
30:13
first
30:13
trust
30:13
and
30:14
safety
30:14
jobs
30:14
was
30:14
at
30:14
Tinder
30:15
and
30:15
I
30:15
was
30:15
looking
30:15
into
30:16
romance
30:16
scams
30:16
all
30:17
the
30:17
time.
30:18
And
30:18
these
30:19
attackers
30:19
are
30:20
patient,
30:20
they're
30:20
willing
30:21
to
30:21
go
30:21
a
30:21
long
30:22
time.
30:22
And
30:22
in
30:23
the
30:23
past
30:23
that
30:23
required
30:24
a
30:24
human
30:24
to
30:24
do
30:24
it.
30:25
I
30:25
think
30:25
when
30:25
we're
30:25
thinking
30:26
about
30:26
the
30:26
ways
30:27
that
30:27
agenda
30:27
AI
30:28
can
30:28
introduce,
30:28
like
30:29
now
30:29
you
30:29
can
30:29
just
30:29
have
30:29
an
30:29
agent
30:30
doing
30:30
that,
30:31
and
30:31
it
30:31
can
30:31
do
30:32
that
30:32
with
30:33
50
30:33
or
30:33
like
30:34
5,000
30:34
conversations
30:35
at
30:35
the
30:35
same
30:35
time,
30:36
you
30:36
know,
30:36
so
30:37
the
30:37
scale
30:37
of
30:37
the
30:38
kind
30:38
of
30:38
like
30:38
previously,
30:39
like
30:39
complex,
30:40
hard
30:40
to
30:40
find,
30:40
human
30:41
to
30:41
human
30:41
harm
30:42
is
30:42
now
30:42
scaled
30:43
rapidly
30:43
through
30:43
AI.
30:44
That's
30:45
wild.
30:46
One
30:46
thing
30:46
we've
30:46
heard
30:47
in
30:47
the
30:47
news
30:47
a
30:47
lot
30:47
recently
30:48
is
30:48
consent.
30:49
You
30:49
know,
30:50
like
30:50
consent
30:50
in
30:51
AI,
30:51
consent
30:52
with
30:52
imagery.
30:53
We've
30:53
been
30:53
seeing
30:54
a
30:54
lot
30:54
of
30:54
people
30:54
talking
30:55
about
30:55
like
30:55
the
30:56
ability
30:56
for
30:57
AI
30:58
to
30:58
create
30:59
images
30:59
of
30:59
people
31:00
that
31:00
they
31:00
didn't
31:00
consent
31:01
to
31:01
that
31:01
might
31:02
be
31:02
exploitative.
31:03
Is
31:04
that
31:04
kind
31:04
of
31:04
what
31:04
you're
31:05
thinking
31:05
of,
31:05
like
31:05
being
31:05
able
31:06
to
31:06
fight
31:06
at
31:07
that
31:07
scale
31:07
or,
31:08
and
31:08
I
31:08
don't
31:08
even
31:09
know
31:09
fight
31:09
is
31:09
the
31:09
right
31:10
term,
31:10
maybe
31:10
identify.
31:11
But
31:11
what
31:11
are
31:12
you
31:12
thinking
31:12
about
31:12
when
31:12
you're
31:12
approaching
31:13
what
31:13
you're
31:13
building
31:13
with
31:13
AI?
31:14
So
31:14
I'm
31:15
really
31:15
glad
31:15
you
31:15
asked
31:16
because
31:16
you're
31:17
right.
31:17
You
31:17
know,
31:17
I
31:17
think
31:17
that
31:18
the
31:18
bad
31:18
actors
31:19
and
31:19
abusers
31:20
and
31:20
like
31:20
these
31:20
networks
31:21
of
31:21
people
31:21
who
31:22
are
31:22
making
31:23
a
31:23
living
31:23
or
31:24
just
31:24
being
31:24
entertained
31:25
from
31:25
causing
31:25
chaos
31:26
and
31:26
Harm,
31:26
they
31:27
talk
31:27
to
31:27
each
31:27
other,
31:27
they
31:28
share
31:28
practices.
31:29
If
31:29
there's
31:29
a
31:29
new
31:29
kind
31:30
of
31:30
agentic
31:31
AI,
31:31
if
31:31
there's
31:31
a
31:31
new
31:31
open
31:32
source
31:32
model
31:32
that
31:32
is
31:33
really
31:33
powerful
31:33
tool
31:34
for
31:34
them
31:34
to
31:34
cause
31:35
harm,
31:35
they
31:35
distribute
31:36
it
31:36
among
31:36
their
31:36
networks.
31:37
We
31:37
don't
31:38
really
31:38
do
31:38
that
31:38
on
31:38
the
31:39
platform,
31:39
like
31:39
the
31:39
defender
31:40
protector
31:40
side,
31:41
because
31:42
company
31:42
privacy,
31:43
everyone's
31:43
got
31:43
their
31:44
own
31:44
proprietary
31:44
technology.
31:45
So
31:45
a
31:45
lot
31:46
of
31:46
what
31:46
we're
31:46
trying
31:46
to
31:46
do
31:46
at
31:46
Roost
31:47
is
31:47
how
31:48
can
31:48
we
31:48
take
31:48
those
31:49
core
31:49
building
31:50
blocks
31:50
of
31:50
safety
31:50
technology
31:51
and
31:51
put
31:51
them
31:52
in
31:52
the
31:52
public
31:52
domain?
31:53
How
31:53
can
31:53
we
31:53
create
31:53
a
31:53
public
31:54
commons
31:55
and
31:55
create
31:55
really
31:55
robust,
31:56
high
31:57
quality,
31:57
production
31:58
grade
31:58
safety
31:59
technology
31:59
that
31:59
can
32:00
kind
32:00
of
32:00
be
32:00
like
32:01
collaboratively
32:02
worked
32:02
on?
32:02
Because
32:03
I
32:03
think
32:03
that
32:03
the
32:04
more
32:04
eyes
32:04
on
32:04
something,
32:05
the
32:05
more
32:05
secure
32:06
it
32:06
becomes,
32:07
more
32:07
people
32:07
can
32:07
notice
32:08
vulnerabilities
32:08
or
32:09
risks.
32:09
And
32:10
these
32:10
are
32:10
the
32:11
primitives,
32:11
right?
32:12
Everyone
32:12
can
32:12
still
32:13
build
32:13
their
32:13
own
32:13
complex
32:14
internal
32:15
systems,
32:15
but
32:16
the
32:16
core
32:16
functions
32:17
don't
32:17
need
32:17
to
32:17
be
32:18
built
32:18
and
32:18
rebuilt
32:19
over
32:19
and
32:19
over
32:19
again.
32:19
Actually,
32:20
at
32:20
Roost,
32:20
we're
32:21
working
32:21
on
32:21
open.
32:22
We
32:22
publish
32:22
our
32:22
project
32:23
roadmap
32:23
on
32:24
GitHub
32:24
for
32:24
public
32:24
feedback
32:25
and
32:25
it
32:25
actually
32:26
states
32:26
that
32:27
right
32:27
now
32:28
we're
32:28
trying
32:28
to
32:28
get
32:28
those
32:29
MVP
32:29
projects
32:30
out.
32:30
Things
32:31
like
32:31
a
32:31
rules
32:31
engine,
32:32
an
32:33
event
32:33
streaming
32:33
investigation
32:34
tool,
32:35
a
32:35
scalable
32:36
review
32:36
console,
32:37
so
32:37
you
32:37
can
32:37
actually
32:37
take
32:37
a
32:38
look
32:38
at
32:38
what's
32:38
going
32:38
on.
32:38
Hash
32:39
matching.
32:40
But
32:40
then
32:40
later
32:40
on,
32:41
our
32:41
roadmap
32:41
going
32:41
this
32:42
year
32:42
and
32:42
next
32:42
year,
32:43
we
32:43
do
32:43
want
32:43
to
32:44
introduce
32:44
those
32:45
kind
32:45
of
32:45
AI
32:46
components
32:46
to
32:47
our
32:47
tools
32:47
because
32:48
we
32:48
think
32:48
that
32:48
equipping
32:49
the
32:49
teams
32:50
that
32:50
are
32:50
fighting
32:51
the
32:51
abuse
32:51
with
32:52
the
32:52
same
32:52
level
32:52
of
32:53
technology
32:53
and
32:54
functionality
32:54
is
32:55
really
32:55
important
32:55
for
32:56
us
32:56
to
32:56
stay
32:56
ahead
32:56
or
32:57
at
32:57
least
32:57
keep
32:57
up
32:58
with
32:58
what
32:58
the
32:58
bad
32:58
actors
32:59
are
32:59
doing.
32:59
So
33:00
does
33:00
this
33:00
mean
33:00
that
33:01
essentially
33:01
your
33:01
primary
33:03
customers
33:03
or
33:04
users
33:04
are
33:05
trust
33:06
and
33:06
safety
33:06
teams
33:07
at
33:07
technology
33:08
companies
33:08
who
33:09
are
33:09
building,
33:09
or
33:10
not
33:10
even
33:10
trust
33:10
and
33:10
safety
33:11
teams
33:11
necessarily,
33:11
but
33:12
just
33:12
any
33:12
product
33:12
team
33:13
who's
33:13
building
33:14
something
33:14
that
33:14
wants
33:14
to
33:15
make
33:15
sure
33:15
that
33:15
they
33:16
have
33:17
the
33:17
right
33:17
protections
33:18
in
33:18
place,
33:18
especially
33:19
if
33:19
they're
33:20
building
33:20
something
33:20
that's
33:20
like
33:20
an
33:21
AI
33:21
native
33:21
product,
33:22
they
33:22
go
33:23
to
33:23
Roost
33:23
and
33:24
you
33:24
guys
33:24
essentially
33:25
are
33:25
arming
33:25
them
33:25
with
33:25
like
33:26
the
33:26
toolkit
33:27
they
33:27
need.
33:28
Yeah,
33:28
pretty
33:28
much.
33:29
You
33:29
know,
33:29
I
33:29
think
33:29
trust
33:30
and
33:30
safety
33:30
teams
33:31
are
33:31
of
33:31
course
33:31
like
33:31
the
33:31
people
33:32
who
33:32
are
33:32
going
33:32
to
33:32
be
33:32
most
33:32
familiar
33:33
with
33:33
our
33:33
tools.
33:34
But
33:34
what
33:34
we've
33:35
really
33:35
found
33:35
is
33:35
that
33:35
like
33:36
startup
33:36
founders,
33:36
developers,
33:37
hobby
33:38
developers,
33:38
people
33:38
who
33:39
want
33:39
to
33:39
create
33:39
the
33:39
next
33:40
unicorn
33:41
AI
33:41
consumer
33:42
app,
33:42
those
33:42
are
33:42
the
33:42
ones
33:43
who
33:43
are
33:43
kind
33:43
of
33:43
like
33:43
really
33:44
curious
33:44
and
33:44
interested
33:44
in
33:45
Roost
33:45
because
33:46
maybe
33:46
they're
33:46
a
33:46
small
33:47
team,
33:47
maybe
33:47
they
33:47
don't
33:48
have
33:48
someone
33:48
who
33:48
have
33:48
safety
33:49
experience,
33:49
maybe
33:50
they
33:50
don't
33:50
have
33:50
A
33:50
dedicated
33:51
trust
33:51
and
33:51
safety
33:51
person.
33:52
But
33:52
they
33:52
do
33:52
want
33:52
to
33:53
make
33:53
sure
33:53
their
33:53
users
33:54
have
33:54
a
33:54
good
33:54
user
33:54
experience.
33:55
Or,
33:56
you
33:56
know,
33:56
like,
33:56
maybe
33:56
they're
33:57
interested
33:57
in
33:57
complying
33:58
with
33:58
like,
33:59
legislation
33:59
that
34:00
deals
34:00
with
34:00
online
34:00
safety
34:01
and
34:01
content
34:01
moderation.
34:02
So
34:03
we
34:03
want
34:03
to
34:03
equip
34:03
them
34:04
with
34:04
those
34:04
building
34:04
blocks.
34:05
And
34:05
honestly,
34:06
some
34:06
of
34:06
the
34:06
smaller
34:07
and
34:07
medium
34:08
companies
34:08
as
34:08
well,
34:09
where
34:09
they
34:09
may
34:09
have
34:10
established
34:10
teams,
34:10
but
34:11
maybe
34:11
they
34:11
only
34:11
have
34:11
like
34:12
one
34:12
person
34:12
on
34:12
that
34:12
team.
34:13
They
34:13
also
34:13
need
34:14
those
34:14
powerful
34:15
tools.
34:15
We
34:16
really
34:16
want
34:16
to
34:16
take
34:17
kind
34:17
of
34:17
like
34:17
the
34:18
same
34:18
technology
34:18
that's
34:19
been
34:19
scaled
34:20
and
34:20
built
34:20
at
34:20
the
34:20
big
34:21
companies
34:21
and
34:22
make
34:22
it
34:22
equally
34:23
accessible
34:23
and
34:23
free
34:24
to
34:24
use
34:24
for
34:25
smaller
34:25
developers.
34:26
Yeah,
34:27
that
34:27
makes
34:27
sense.
34:27
I
34:27
think
34:28
one
34:28
of
34:28
the
34:28
concepts
34:29
I
34:29
learned
34:29
from
34:30
a
34:30
trust
34:30
and
34:30
safety
34:31
expert,
34:31
something
34:31
that
34:31
really
34:32
carried
34:32
me
34:32
and
34:33
I
34:33
think
34:33
about
34:33
often.
34:34
Cesar.
34:34
I
34:34
used
34:34
to
34:35
work
34:35
with
34:35
him
34:35
at
34:35
Stack
34:36
Overflow
34:36
and
34:37
he
34:37
talked
34:37
to
34:37
me
34:37
a
34:37
lot
34:38
about
34:38
impact
34:39
versus
34:39
intent
34:40
and
34:40
how,
34:40
you
34:40
know,
34:41
like,
34:41
for
34:41
example,
34:42
you
34:42
know,
34:42
if
34:42
you're
34:42
walking
34:43
down
34:43
the
34:43
street
34:43
and
34:44
you
34:44
trip
34:44
and
34:44
spill
34:45
your
34:45
hot
34:45
coffee
34:45
on
34:45
someone,
34:46
you
34:46
burn
34:46
them.
34:47
And
34:47
you're
34:48
not
34:48
someone
34:48
that
34:48
wants
34:48
to
34:49
burn
34:49
people
34:49
with
34:49
coffee.
34:50
You're
34:50
just
34:50
like
34:50
walking
34:51
down
34:51
the
34:51
street
34:51
having
34:51
a
34:51
nice
34:52
day.
34:52
However,
34:52
your
34:53
impact
34:53
is
34:53
now
34:53
this
34:54
person
34:54
has
34:54
been
34:54
burned
34:55
with
34:55
coffee
34:55
and
34:56
what
34:56
responsibility
34:57
do
34:57
you
34:57
have?
34:58
Of
34:58
course
34:58
you're
34:58
going
34:58
to
34:58
like,
34:58
help
34:59
them
34:59
clean
34:59
up
35:00
and
35:00
like
35:00
do
35:00
what
35:00
you
35:00
can.
35:01
And
35:01
when
35:02
it
35:02
comes
35:02
to
35:02
building
35:02
things,
35:03
it's
35:03
very
35:03
similar.
35:04
Right.
35:04
Like
35:05
we
35:05
talked
35:05
about
35:05
that
35:05
a
35:05
little
35:05
bit
35:05
earlier
35:06
of
35:06
like,
35:06
I
35:07
just
35:07
want
35:07
to
35:07
build
35:07
a
35:07
fun
35:08
game
35:08
for
35:08
people
35:08
or
35:08
I
35:09
just
35:09
want
35:09
people
35:09
to
35:09
be
35:10
friends
35:10
online
35:10
or
35:10
I
35:11
want
35:11
them
35:11
to
35:11
collaborate
35:11
in
35:12
this
35:12
way.
35:12
But
35:13
then
35:13
that
35:13
impact
35:13
can
35:14
be
35:14
like
35:14
somewhat
35:15
a
35:15
bad
35:15
actor
35:16
uses
35:16
it
35:16
as
35:16
a
35:16
way
35:17
to
35:17
take
35:17
advantage
35:17
of
35:17
people.
35:18
And
35:18
then
35:19
what
35:19
responsibility
35:19
do
35:20
the
35:20
founders
35:20
have
35:21
there?
35:21
And
35:21
open
35:22
source
35:22
tools,
35:22
like
35:22
what
35:22
you're
35:23
talking
35:23
about
35:23
is,
35:23
I
35:23
imagine
35:24
so
35:24
beneficial
35:24
to
35:25
people
35:25
who
35:25
are
35:25
thinking
35:26
about
35:26
their
35:26
impact
35:27
early.
35:27
Of
35:27
we
35:28
want
35:28
to
35:28
make
35:28
sure
35:29
that
35:29
we're
35:30
prepared
35:31
should
35:31
our
35:31
tool
35:31
be
35:31
used
35:32
for
35:32
something
35:32
nefarious.
35:33
That's
35:33
great,
35:34
thanks.
35:34
Yeah,
35:34
and
35:35
also
35:35
I
35:35
think
35:35
open
35:35
source
35:36
makes
35:36
it
35:36
customizable.
35:37
Every
35:37
platform
35:38
is
35:38
going
35:38
to
35:38
have
35:38
their
35:38
own
35:39
norms.
35:39
Becca,
35:40
you'll
35:40
appreciate
35:40
this,
35:41
but
35:41
I
35:41
always
35:41
use
35:41
Grindr
35:42
as
35:42
an
35:42
example
35:42
where
35:43
like
35:43
when
35:44
we
35:44
were
35:44
developing
35:44
like
35:45
nude
35:45
photo
35:46
detection
35:46
or
35:46
like
35:47
dick
35:47
pic
35:47
detection,
35:48
you
35:48
know,
35:48
we
35:48
want
35:48
to
35:48
make
35:49
sure
35:49
if
35:49
they
35:49
want
35:49
them,
35:50
they
35:50
can
35:50
get
35:50
them.
35:50
If
35:50
they
35:51
don't
35:51
want
35:51
them,
35:51
they
35:51
don't
35:51
get
35:51
them.
35:52
Everyone
35:52
should
35:52
be
35:53
able
35:53
to
35:53
get
35:53
all
35:54
the
35:54
dick
35:54
pics
35:54
they
35:55
want
35:55
consensually.
35:58
Yeah,
35:58
that's
35:58
great.
35:59
Yes.
36:00
And
36:00
so
36:00
like
36:00
Grindr
36:01
concepts
36:02
and
36:02
the
36:02
way
36:02
we
36:02
thought
36:02
about
36:03
nudity
36:03
was
36:03
just
36:04
different
36:04
from
36:04
a
36:05
lot
36:05
of
36:05
other
36:05
companies.
36:07
For
36:07
the
36:07
photos
36:08
that
36:08
people
36:08
upload
36:09
to
36:09
their
36:09
profile
36:09
on
36:09
Grindr,
36:10
their
36:11
butts
36:11
were
36:11
okay.
36:12
We
36:12
were
36:12
okay
36:12
with
36:13
rear
36:13
nudity.
36:14
But
36:14
what
36:14
we
36:14
found
36:15
is
36:15
that
36:15
a
36:15
lot
36:15
of
36:15
the
36:16
third
36:16
party
36:16
classifiers
36:17
that
36:17
we
36:17
wanted
36:17
to
36:17
build,
36:18
they
36:18
see
36:18
a
36:19
but
36:19
and
36:19
they're
36:19
like,
36:19
that's
36:19
nudity.
36:20
And
36:20
we're
36:20
like,
36:20
no,
36:21
no,
36:21
no,
36:21
but
36:21
it's
36:21
okay.
36:22
We
36:22
like
36:22
the
36:22
buts.
36:23
And
36:24
we
36:24
weren't
36:24
able
36:24
to
36:25
customize
36:25
that
36:26
for
36:26
the
36:26
Grindr
36:26
use
36:27
case
36:27
and
36:27
for
36:27
our
36:28
community,
36:29
who
36:29
said
36:29
that's
36:29
okay.
36:30
With
36:31
open
36:31
source
36:31
technology,
36:31
whether
36:32
it's
36:32
a
36:32
model
36:32
or
36:32
a
36:32
classifier,
36:33
you
36:33
can
36:33
tune
36:34
or
36:34
more
36:34
like
36:34
infrastructure
36:35
based
36:35
tools,
36:36
like
36:36
what
36:36
roost
36:36
is
36:36
doing.
36:37
That
36:38
level
36:38
of
36:38
customization,
36:39
I
36:39
think
36:39
is
36:39
what's
36:40
going
36:40
to
36:40
be
36:41
really,
36:41
really
36:41
helpful
36:42
for
36:42
founders
36:43
and
36:43
developers
36:44
to
36:44
make
36:45
sure
36:45
that
36:45
their
36:46
tools
36:46
meet
36:46
their
36:47
community
36:47
and
36:47
their
36:47
user
36:48
base's
36:48
needs.
36:49
Great
36:49
example.
36:49
Love
36:50
the
36:50
butts
36:50
example.
36:51
Yeah,
36:51
I'm
36:52
actually
36:52
very
36:52
curious.
36:52
In
36:53
that
36:53
case,
36:53
you
36:54
can
36:54
define
36:54
nudity
36:55
and
36:55
to
36:55
your
36:56
point
36:56
of
36:57
the
36:57
pictures,
36:57
like
36:57
you
36:57
should
36:58
have
36:58
all
36:58
the
36:58
pictures
36:58
that
36:58
you
36:59
consent
36:59
to.
37:00
What
37:00
does
37:00
it
37:00
look
37:00
like
37:01
to
37:01
build
37:01
in
37:01
safeties
37:02
for
37:02
making
37:03
sure
37:03
that
37:03
people
37:04
are
37:04
like,
37:05
this
37:05
is
37:05
this
37:05
content
37:06
that
37:06
you
37:06
want?
37:07
Or
37:07
this
37:07
is
37:08
this
37:08
content
37:08
you
37:08
very
37:09
much
37:09
don't
37:09
want?
37:10
In
37:10
that
37:10
case.
37:11
Yeah.
37:12
So
37:12
I'll
37:12
use
37:12
another
37:13
example.
37:13
So
37:14
using
37:14
Blue
37:14
sky
37:15
as
37:15
an
37:15
example,
37:15
there's
37:16
a
37:16
lot
37:16
of
37:16
people
37:16
on
37:16
Blue
37:17
sky
37:17
who
37:17
build
37:17
these
37:18
labelers
37:18
and
37:19
there's
37:19
labelers
37:19
for
37:19
like
37:20
NSFW
37:20
content.
37:21
Maybe
37:21
you
37:21
want
37:22
to
37:22
see
37:22
a
37:22
certain
37:22
type
37:22
of
37:22
NSFW
37:23
content
37:24
and
37:24
you
37:24
don't
37:24
want
37:24
to
37:24
see
37:25
like
37:25
another
37:25
type.
37:25
Maybe
37:26
you
37:26
love
37:26
furry
37:26
NSFW
37:27
content,
37:27
you
37:28
don't
37:28
want
37:28
to
37:28
see
37:28
real
37:28
people.
37:29
And
37:29
so
37:29
it's
37:30
all
37:30
about
37:30
opting
37:30
into
37:31
those
37:31
labels.
37:31
And
37:32
so
37:32
I
37:32
think
37:32
the
37:33
underlying
37:33
technology
37:33
is
37:34
most
37:34
likely
37:34
the
37:34
same.
37:35
It's
37:35
doing
37:35
some
37:35
kind
37:35
of
37:36
classification
37:36
to
37:36
say,
37:37
is
37:37
this
37:37
NSFW
37:38
content?
37:39
Cool.
37:39
What
37:39
kind
37:39
is
37:40
it?
37:40
And
37:40
those
37:41
classifiers
37:41
you
37:41
can
37:41
tune
37:42
on
37:42
yourself,
37:42
you
37:43
can
37:43
look
37:43
at
37:43
different
37:43
signals
37:44
to
37:44
try
37:44
to
37:44
figure
37:44
out
37:44
is
37:45
it
37:45
2D,
37:45
is
37:45
it
37:45
3D?
37:46
Whatever.
37:47
And
37:47
this
37:47
is
37:47
me
37:48
naively
37:48
thinking
37:48
about
37:49
how
37:49
labels
37:49
and
37:49
technology
37:49
work.
37:50
But
37:50
there
37:50
should
37:51
be
37:51
a
37:51
simple
37:51
flag
37:52
of
37:52
saying
37:53
accepts
37:53
NSFW
37:54
photos
37:54
or
37:55
opted
37:55
into
37:55
this
37:55
labeler
37:56
to
37:56
say,
37:56
I'm
37:57
okay
37:57
with
37:57
seeing
37:57
this.
37:58
And
37:58
it
37:58
can
37:58
be
37:58
the
37:58
same
37:59
with
37:59
Grindr
37:59
or
37:59
any
38:00
platform
38:00
that
38:00
has
38:01
just
38:01
like
38:01
that
38:01
flag
38:01
set.
38:02
Like,
38:02
I'm
38:02
okay
38:02
with
38:03
this
38:03
or
38:03
I'm
38:03
not
38:03
okay
38:04
with
38:04
this.
38:04
Yeah,
38:05
Reddit
38:05
does
38:05
this
38:05
really
38:06
well.
38:06
Yeah,
38:06
yeah,
38:07
yeah.
38:08
It's
38:08
like
38:08
kind
38:08
of
38:09
like
38:09
what
38:09
you
38:09
were
38:09
saying
38:09
about
38:10
like
38:10
the
38:10
balance
38:11
between
38:11
the
38:12
platform
38:12
or
38:13
product
38:13
handling
38:14
trust
38:14
and
38:14
safety
38:15
for
38:15
the
38:15
user
38:15
and
38:16
giving
38:16
the
38:16
user
38:16
autonomy
38:17
to
38:18
participate
38:19
in
38:20
expressing
38:20
what
38:20
they
38:21
do
38:21
and
38:21
don't
38:21
want.
38:22
I
38:22
was
38:22
wondering
38:22
if
38:22
we
38:22
could
38:23
talk
38:23
a
38:23
little
38:23
bit
38:23
about
38:24
like
38:24
some
38:24
of
38:24
the
38:25
specific
38:27
tools
38:27
that
38:28
you
38:28
guys
38:28
are
38:28
building
38:29
at
38:29
Roost.
38:30
And
38:30
like,
38:31
I
38:31
think
38:31
you're
38:31
a
38:31
great
38:31
person
38:32
to
38:32
talk
38:32
about
38:32
this
38:33
because
38:33
you
38:34
straddled
38:34
this
38:34
line
38:34
between
38:35
technical
38:36
and
38:37
product.
38:37
But
38:38
like,
38:38
so
38:38
you
38:38
don't
38:38
have
38:38
to
38:38
go
38:39
super
38:39
deep
38:39
into
38:39
the
38:39
weeds.
38:40
But
38:40
like
38:40
what
38:40
are
38:41
some
38:41
of
38:41
the
38:41
things
38:42
in
38:42
the
38:42
Roost
38:43
toolbox?
38:44
Yeah,
38:45
so
38:45
we
38:45
have
38:45
two
38:46
flagship
38:47
software
38:47
products
38:48
that
38:48
we
38:49
are
38:49
just
38:49
launched.
38:50
So
38:50
we
38:50
actually
38:50
just
38:51
publicly
38:51
announced
38:51
the
38:52
V
38:52
one
38:52
of
38:53
our
38:53
first
38:53
project
38:54
yesterday.
38:55
And
38:55
this
38:55
is
38:55
called
38:55
Osprey.
38:56
You'll
38:57
notice
38:57
that
38:57
I
38:57
have
38:57
named
38:58
all
38:58
of
38:58
our
38:58
products
38:58
after
38:59
birds
38:59
or
38:59
chicken
39:00
related
39:00
words
39:00
because
39:01
my
39:01
humor
39:01
is
39:02
that
39:02
of
39:02
like
39:02
a
39:02
65
39:03
year
39:03
old
39:03
man.
39:05
So
39:06
Osprey
39:06
is
39:06
a
39:07
rules
39:07
engine
39:07
and
39:08
it's
39:08
like
39:08
an
39:09
investigation
39:09
tool.
39:10
And
39:10
so
39:10
it
39:11
comes
39:11
with
39:11
this
39:11
interface
39:12
where
39:12
like
39:13
if
39:13
you
39:13
have
39:13
someone
39:14
who
39:14
wants
39:14
to
39:14
kind
39:14
of
39:14
like
39:15
look
39:15
into
39:15
a
39:15
signal,
39:16
let's
39:16
say,
39:16
you
39:17
know,
39:17
this
39:17
URL
39:18
is
39:18
going
39:18
to
39:18
a
39:19
really
39:19
gross
39:19
video
39:20
that
39:20
you
39:20
don't
39:21
want
39:21
on
39:21
your
39:21
platform,
39:21
but
39:22
you
39:22
don't
39:22
know
39:22
how
39:22
many
39:23
times
39:23
it's
39:23
been
39:23
posted.
39:24
You
39:24
don't
39:24
know
39:24
who's
39:24
sharing
39:25
it.
39:25
You
39:25
don't
39:25
know
39:25
like
39:25
the
39:26
scope
39:26
or
39:26
variable
39:27
reality
39:27
of
39:27
this
39:27
thing
39:27
on
39:27
your
39:27
platform.
39:28
And
39:28
you
39:28
don't
39:29
know
39:29
if
39:29
it's
39:29
just
39:29
like
39:29
people
39:30
sharing
39:30
it
39:30
or
39:30
if
39:31
there's
39:31
like
39:31
coordinated
39:32
networks
39:32
of
39:32
bots
39:33
that
39:33
are
39:33
distributing
39:34
this
39:34
video.
39:34
You
39:35
could
39:35
use
39:35
Osprey
39:36
to
39:37
query
39:37
the
39:37
data.
39:37
And
39:38
so
39:38
whatever
39:38
data
39:38
you
39:39
put
39:39
into
39:39
Osprey,
39:39
it's
39:40
all
39:40
self
39:40
hosted
39:40
on
39:40
your
39:40
own
39:40
infrastructure.
39:41
So
39:41
all
39:42
the
39:42
stuff
39:42
just
39:42
stays
39:42
within
39:43
your
39:43
own
39:43
system.
39:43
It's
39:44
another
39:44
beauty
39:44
of
39:44
open
39:45
source.
39:46
You
39:46
could
39:46
look
39:46
through
39:47
whatever
39:47
information
39:47
you
39:48
think
39:48
is
39:48
valuable
39:49
to
39:49
investigate
39:50
and
39:50
find
39:50
those
39:51
networks.
39:51
You
39:52
can
39:52
find
39:52
those
39:52
new
39:52
patterns
39:53
of
39:53
abuse,
39:53
new
39:54
signals,
39:54
and
39:54
then
39:54
you
39:54
can
39:55
create
39:55
rules
39:55
to
39:56
then
39:56
automatically
39:57
label
39:57
if
39:57
you
39:57
just
39:57
want
39:58
to
39:58
keep
39:58
an
39:58
eye
39:58
on
39:58
it.
39:58
Or
39:59
you
39:59
can
39:59
attach
39:59
those
39:59
labels
40:00
to
40:00
an
40:00
action
40:01
and
40:01
then
40:01
take
40:01
automated
40:02
action.
40:03
So
40:03
you're
40:03
fighting
40:03
automated
40:04
abuse
40:04
with
40:04
automated
40:05
tools.
40:05
But
40:06
that
40:06
open
40:06
ended
40:07
investigation
40:07
piece
40:08
is
40:08
super,
40:08
super
40:08
important.
40:09
A
40:09
lot
40:10
of
40:10
times
40:10
on
40:10
platforms
40:11
if
40:11
there's
40:12
like
40:12
a
40:12
current
40:12
event,
40:13
something
40:13
really
40:14
bad
40:14
happens
40:15
and
40:15
you
40:15
want
40:15
to
40:15
know,
40:15
is
40:16
my
40:16
platform
40:16
affected?
40:17
Is
40:17
this
40:17
person
40:17
on
40:18
my
40:18
platform?
40:18
What
40:19
are
40:19
they
40:19
doing
40:19
on
40:19
my
40:20
platform?
40:20
They
40:21
need
40:21
a
40:21
tool
40:21
like
40:21
Osprey
40:22
to
40:22
go
40:22
figure
40:23
that
40:23
out.
40:23
And
40:24
you
40:24
can
40:24
tell
40:24
based
40:25
on
40:26
the
40:26
data
40:27
like
40:27
whether
40:27
something
40:28
is
40:28
being
40:28
spread
40:29
by
40:29
a
40:29
network
40:29
of
40:30
bots
40:30
versus
40:31
people.
40:31
Depending
40:32
on
40:32
the
40:32
sophistication
40:33
of
40:33
the
40:33
bots.
40:33
But
40:33
yeah,
40:34
a
40:34
lot
40:34
of
40:34
times
40:34
it's
40:35
kind
40:35
of
40:35
like,
40:35
oh,
40:36
you
40:36
know,
40:36
like,
40:36
is
40:36
there
40:36
a
40:37
pattern
40:37
to
40:37
all
40:37
the
40:37
display
40:38
names
40:38
that
40:39
are
40:39
of
40:39
the
40:39
accounts
40:40
posting
40:40
this
40:40
thing?
40:41
How
40:42
frequently
40:42
is
40:42
this
40:42
thing
40:43
being
40:43
posted?
40:43
If
40:44
one
40:44
account
40:44
is
40:44
posting
40:45
the
40:45
same
40:45
video,
40:45
like
40:46
every
40:46
5
40:47
milliseconds,
40:48
that's
40:48
probably
40:48
not
40:48
a
40:48
human.
40:50
Yep.
40:51
So
40:51
we
40:52
have
40:52
Osprey
40:52
and
40:53
then
40:53
there
40:53
was
40:53
another
40:54
product.
40:54
What
40:55
part
40:55
are
40:55
we
40:55
talking
40:55
about?
40:56
It's
40:57
confusing.
40:57
I
40:57
kind
40:58
of
40:58
regret
40:58
it
40:58
because,
40:58
like,
40:59
it's
40:59
Coop.
40:59
So
41:00
it's
41:00
about
41:00
chickens,
41:01
but
41:01
it's
41:01
not
41:01
a
41:01
type
41:02
of
41:02
chicken.
41:03
We
41:03
call
41:03
it
41:03
Coop
41:04
because
41:04
a
41:04
lot
41:04
of
41:04
things
41:04
can
41:05
go
41:05
into
41:05
the
41:05
coop.
41:05
You
41:06
know,
41:06
you
41:06
might
41:06
have
41:06
different
41:06
types
41:07
of
41:07
chickens
41:07
and
41:07
you
41:07
want
41:07
them
41:08
all
41:08
live
41:08
in
41:08
a
41:08
coop.
41:09
But
41:09
Coop
41:09
is
41:09
our
41:10
flexible
41:10
review
41:10
tool.
41:11
And
41:11
this
41:11
is
41:11
actually
41:12
a
41:12
really
41:12
interesting
41:12
kind
41:13
of
41:13
origin
41:13
story
41:13
for
41:14
both
41:14
projects
41:14
too.
41:15
Osprey
41:15
was
41:15
originally
41:16
developed
41:16
by
41:16
Discord
41:17
and
41:17
that's
41:18
what
41:18
they
41:18
used
41:18
in
41:18
production
41:19
to
41:19
fight
41:20
spam
41:20
and
41:20
abuse
41:21
and
41:21
do
41:21
these
41:21
investigations
41:22
at
41:22
Discord.
41:23
They
41:23
donated
41:24
that
41:24
source
41:24
code
41:24
to
41:25
Roost
41:25
and
41:25
then
41:25
open
41:26
sourced
41:26
it
41:26
with
41:26
us.
41:27
So
41:27
we
41:27
work
41:27
directly
41:28
with
41:28
our
41:28
engineering
41:28
team
41:29
to
41:29
generalize
41:30
all
41:30
of
41:31
it.
41:31
So
41:31
it's
41:32
being
41:32
used
41:32
in
41:32
production
41:33
at
41:33
Bluesky.
41:34
Discord
41:34
has
41:34
pulled
41:35
the
41:35
open
41:35
source
41:35
version
41:36
of
41:36
Osprey
41:37
back
41:37
internally.
41:37
So
41:38
they're
41:38
using
41:38
it
41:38
in
41:38
production
41:39
and
41:39
then
41:39
Matrix
41:40
is
41:40
also
41:41
going
41:41
to
41:41
start
41:41
deploying
41:42
it
41:42
this
41:42
week.
41:43
So
41:43
that's
41:43
one
41:43
origin
41:43
story
41:44
for
41:44
Coop.
41:45
We
41:45
as
41:45
Roost
41:45
acquired
41:46
the
41:46
IP
41:47
of
41:47
a
41:47
trust
41:48
and
41:48
safety
41:48
startup
41:49
that
41:49
was
41:50
no
41:50
longer
41:50
operating.
41:51
And
41:52
so
41:52
we
41:53
took
41:53
the
41:53
IP
41:54
and
41:54
we've
41:54
been
41:54
spending
41:55
the
41:55
past
41:55
six
41:55
to
41:56
seven
41:56
months
41:56
refactoring
41:57
all
41:57
of
41:57
it
41:58
to
41:58
turn
41:58
it
41:58
from
41:59
like
41:59
a
41:59
SaaS
41:59
startup
42:00
into
42:00
an
42:00
open
42:01
source
42:01
project.
42:02
So
42:02
that's
42:03
something
42:03
we're
42:03
really
42:03
excited
42:04
to
42:04
release.
42:04
The
42:04
startup's
42:05
name
42:05
is
42:05
Cove,
42:06
so
42:06
if
42:06
you
42:06
want
42:06
to
42:06
learn
42:07
more,
42:07
it's
42:07
getcove.com
42:08
but
42:08
Coop
42:09
is
42:09
like,
42:09
like
42:09
it's
42:10
a
42:10
classic
42:10
review
42:10
tool.
42:11
It
42:11
could
42:11
be
42:11
used
42:11
for
42:12
user
42:12
reports
42:12
if
42:13
you
42:13
have
42:13
like
42:14
a
42:14
classifier
42:14
running,
42:15
anything
42:16
you
42:16
want
42:16
to
42:16
proactively
42:16
detect
42:17
or
42:18
like
42:18
just
42:18
like
42:18
review
42:19
at
42:19
scale,
42:20
you
42:20
can
42:20
do
42:20
it
42:20
in
42:20
Coop
42:21
and
42:21
you
42:21
can
42:21
customize
42:22
whatever
42:22
actions
42:23
you
42:23
take.
42:23
And
42:23
then
42:24
more
42:24
importantly,
42:25
and
42:25
I
42:25
said
42:25
earlier
42:26
I
42:26
would
42:26
get
42:26
back
42:26
to
42:26
kind
42:26
of
42:27
like
42:27
CSAM
42:27
AI
42:28
tools.
42:29
Koop
42:29
is
42:30
going
42:30
to
42:30
be
42:30
the
42:30
first
42:31
kind
42:31
of
42:31
like
42:32
child
42:32
safety
42:33
end
42:33
to
42:33
end
42:34
system.
42:34
We
42:34
are
42:34
building
42:35
in
42:35
the
42:35
ability
42:36
to
42:36
do
42:36
hash
42:37
matching
42:38
for
42:38
known
42:38
csam.
42:39
So
42:39
this
42:39
is
42:39
stuff
42:39
that
42:40
everyone
42:40
already
42:40
knows
42:41
is
42:41
csam,
42:41
we
42:42
don't
42:42
need
42:42
to
42:42
look
42:42
at
42:42
it.
42:42
We
42:42
just
42:42
look
42:43
at
42:43
that
42:43
digital
42:43
fingerprint.
42:44
You
42:44
send
42:44
those
42:44
hashes
42:45
for
42:45
review
42:46
in
42:46
coop
42:47
and
42:47
then
42:47
we've
42:47
actually
42:47
integrated
42:48
with
42:48
NCMEC's
42:49
reporting
42:49
API.
42:50
So
42:50
that
42:50
way
42:51
the
42:51
reports
42:51
go
42:51
to
42:52
the
42:52
right
42:52
place.
42:52
And
42:53
NCMEC
42:53
is
42:53
the
42:53
national
42:54
center
42:54
for
42:54
Missing
42:55
and
42:55
Exploited
42:55
Children
42:56
that
42:56
handles
42:56
all
42:57
of
42:57
these
42:57
types
42:57
of
42:57
cases.
42:58
That's
42:59
really
42:59
effective.
43:00
I
43:00
imagine
43:00
there's
43:01
a
43:01
manual
43:01
process
43:02
and
43:02
automating
43:03
that
43:03
process
43:04
I'm
43:04
sure
43:04
will
43:05
hopefully
43:05
help
43:06
a
43:06
lot
43:06
of
43:06
people.
43:07
That's
43:07
really
43:07
neat.
43:08
Yeah.
43:08
And
43:08
there's
43:09
some
43:09
cases
43:09
that
43:10
can't
43:10
be
43:10
automated
43:11
because
43:11
you
43:12
do
43:12
need
43:12
that
43:12
manual
43:12
review
43:13
to
43:13
make
43:13
sure
43:13
that
43:14
it
43:14
is
43:14
what
43:14
you're
43:14
looking
43:15
for.
43:15
But
43:15
detection
43:16
piece,
43:16
doing
43:17
that
43:17
through
43:18
automated
43:19
hash
43:19
matching,
43:20
that's
43:20
something
43:20
that
43:20
can
43:21
definitely
43:21
make
43:22
a
43:22
lot
43:22
of
43:22
workflows
43:23
a
43:23
lot
43:23
faster
43:23
than
43:24
just
43:25
waiting
43:25
for
43:25
someone
43:25
to
43:26
report
43:26
it
43:26
to
43:26
you.
43:27
This
43:27
has
43:27
been
43:28
such
43:28
a
43:28
fascinating
43:29
conversation.
43:29
Thank
43:30
you
43:30
so
43:30
much
43:31
for
43:32
joining
43:33
us.
43:33
Is
43:33
there
43:34
anything
43:35
that,
43:36
whether
43:36
it's
43:36
like
43:37
resources,
43:38
content,
43:39
whatever
43:40
that
43:40
you're
43:40
like,
43:41
if
43:41
you
43:41
work
43:41
in
43:41
product,
43:42
if
43:42
you
43:42
work
43:42
in
43:43
tech,
43:43
like,
43:43
you
43:43
need
43:44
to
43:44
be
43:44
educated
43:46
on
43:46
this
43:46
and
43:46
like
43:46
you
43:47
need
43:47
to
43:47
read
43:47
this
43:48
book
43:48
or
43:48
you
43:48
need
43:48
to
43:48
watch
43:49
this
43:50
documentary
43:50
or
43:51
whatever,
43:51
like
43:52
what's
43:52
your
43:52
stack
43:53
that
43:53
you're
43:53
like,
43:53
everyone
43:54
should
43:55
look
43:55
at
43:55
this.
43:56
Oh,
43:56
okay,
43:57
I've
43:57
got
43:57
a
43:57
few
43:58
for
43:58
you.
43:58
So
43:59
one
43:59
is,
43:59
is
43:59
the
44:00
Trust
44:00
and
44:00
Safety
44:01
Professional
44:01
association.
44:02
They're
44:03
just
44:03
tspa.org
44:05
they
44:05
have
44:05
kind
44:05
of
44:06
like
44:06
aggregated
44:07
all
44:07
the
44:07
expertise
44:08
and
44:08
knowledge
44:08
from
44:09
everyone
44:09
in
44:10
Trust
44:10
and
44:10
Safety
44:11
and
44:11
created
44:11
this
44:12
really
44:12
amazing
44:13
curriculum
44:13
and
44:14
library.
44:14
So
44:15
if
44:15
you
44:15
want
44:15
to
44:15
learn
44:15
about
44:16
like
44:16
the
44:16
risks
44:17
of
44:17
different
44:17
types
44:17
of
44:17
products
44:18
and
44:18
features
44:18
and
44:19
how
44:19
to
44:19
like
44:20
defend
44:20
against
44:20
them
44:20
and
44:21
you
44:21
know,
44:21
adjust
44:22
the
44:22
design
44:22
to
44:23
reduce
44:23
that
44:23
risk.
44:23
Amazing
44:24
resource.
44:25
If
44:25
you
44:25
just
44:25
want
44:25
to
44:25
learn
44:25
more
44:25
about
44:26
trust
44:26
and
44:26
safety
44:26
in
44:26
general
44:27
and
44:27
you
44:28
know,
44:28
what
44:28
are
44:28
the
44:28
different
44:29
types
44:29
of
44:29
technologies,
44:30
what
44:30
are
44:30
different
44:30
teams?
44:30
What
44:31
do
44:31
they
44:31
do
44:31
do?
44:32
Super,
44:32
super
44:32
helpful.
44:33
So
44:33
tspa.org
44:34
is
44:34
like
44:35
always
44:35
my
44:35
number
44:35
one.
44:35
Go
44:36
to
44:36
the
44:36
second
44:37
one.
44:37
For
44:37
product
44:38
builders
44:38
and
44:38
designers
44:39
and
44:39
founders
44:39
especially,
44:40
I
44:40
like
44:41
the
44:41
Pro
44:41
Social
44:41
Design
44:42
network.
44:43
And
44:43
so
44:43
this
44:43
is
44:44
a
44:44
really
44:44
cool
44:45
library
44:45
of
44:46
essentially
44:46
like
44:46
best
44:47
practices
44:48
for
44:48
healthy
44:49
product
44:49
design.
44:49
How
44:50
do
44:50
you
44:50
tweak
44:50
that
44:50
product
44:50
design
44:51
of
44:51
like,
44:51
whether
44:51
it's
44:52
a
44:52
chat
44:52
platform,
44:53
live
44:53
streaming,
44:54
posting
44:55
media,
44:55
like,
44:55
what
44:56
are
44:56
the
44:56
things
44:56
you
44:56
can
44:57
incorporate
44:57
into
44:57
the
44:58
design
44:59
that
44:59
can
44:59
encourage
45:00
healthier
45:00
pro
45:01
social
45:01
behavior
45:02
among
45:02
people
45:02
and
45:03
reduce
45:03
harms
45:04
and
45:04
risks
45:04
too?
45:05
On
45:05
the
45:05
builder
45:06
side,
45:06
I
45:06
will
45:07
kind
45:07
of
45:07
like
45:07
selfishly
45:08
Recommend
45:08
the
45:08
Roost
45:09
GitHub.
45:10
In
45:10
addition
45:10
to
45:10
our
45:10
software
45:11
projects,
45:11
we
45:11
also
45:12
have
45:12
been
45:12
putting
45:12
together
45:13
this
45:14
directory
45:14
or
45:15
list
45:15
of
45:16
all
45:16
the
45:16
free
45:16
open
45:17
source
45:18
technologies
45:18
that
45:19
currently
45:19
exist
45:20
for
45:20
online
45:20
safety.
45:21
Most
45:21
of
45:21
them
45:22
are
45:22
not
45:22
built
45:22
by
45:22
us,
45:22
but
45:23
I
45:23
couldn't
45:23
find
45:24
a
45:24
mapping
45:24
of
45:25
what
45:25
is
45:25
out
45:25
there.
45:26
If
45:26
I'm
45:26
a
45:26
small
45:27
team,
45:27
if
45:27
I
45:27
want
45:27
to
45:28
use
45:28
free
45:28
open
45:28
source
45:29
stuff
45:29
stuff,
45:29
what
45:30
can
45:30
I
45:30
use?
45:30
And
45:31
so
45:31
that
45:31
includes
45:31
like
45:32
AI
45:32
guardrails,
45:33
it
45:34
includes
45:34
like
45:34
data
45:35
sets,
45:35
it
45:35
includes
45:36
review
45:36
tools,
45:37
hash
45:37
matching
45:38
algorithms,
45:39
different
45:39
types
45:39
of
45:39
classifiers,
45:40
PII
45:41
models.
45:42
Like
45:42
all
45:42
of
45:42
that
45:42
stuff
45:43
is
45:43
in
45:43
the
45:43
awesome
45:44
safety
45:44
tools
45:45
Repo
45:45
in
45:45
the
45:45
roost
45:46
GitHub.
45:46
It
45:47
just
45:47
sounds
45:47
like
45:47
if
45:47
you're
45:48
building
45:48
an
45:48
AI
45:48
tool
45:48
now,
45:49
you
45:49
should
45:49
be
45:49
checking
45:49
out
45:50
Roost
45:50
as
45:50
like
45:50
one
45:51
of
45:51
the
45:51
first
45:51
things
45:51
before
45:52
you
45:52
roll
45:53
it
45:53
out
45:53
there.
45:53
Because
45:53
I
45:53
think
45:54
people
45:54
really
45:54
don't
45:54
know
45:55
the
45:55
impact
45:55
that
45:55
their
45:56
work
45:56
can
45:56
have.
45:56
And
45:57
if
45:57
you're
45:57
thinking
45:57
about
45:58
these
45:58
things
45:58
early,
45:58
it's
45:59
so
45:59
much
45:59
easier
45:59
than
46:00
all
46:00
of
46:00
a
46:00
sudden
46:00
ending
46:01
up
46:01
with
46:01
a
46:01
mess.
46:02
Oh,
46:02
well
46:03
said
46:03
everyone.
46:04
Get
46:04
on
46:05
over
46:05
to
46:05
Roost.
46:05
Okay,
46:06
thank
46:06
you
46:06
both
46:06
so
46:07
much
46:07
for
46:07
having
46:07
me.
46:08
This
46:08
has
46:08
been
46:08
super
46:09
fun.
46:09
Thanks
46:09
for
46:09
letting
46:09
me
46:10
yap
46:10
our
46:10
pleas.