Latest / Chaos Agents / Personal Agents, Opinion Geometry, and AI That Learns Too Much - With Hayden Helm

Transcript

Personal Agents, Opinion Geometry, and AI That Learns Too Much - With Hayden Helm · Chaos Agents

0:00
There's
0:00
an
0:00
obvious
0:00
one
0:01
when
0:01
the
0:01
AI
0:01
shows
0:02
up
0:02
and
0:02
is
0:02
like,
0:02
sarah,
0:03
you're
0:03
a
0:03
piece
0:04
of
0:04
trash.
0:04
I
0:04
hate
0:04
you.
0:05
You're
0:05
like,
0:05
oh,
0:05
you're
0:06
a
0:06
bad
0:06
one.
0:07
But
0:07
there's
0:07
like
0:07
another
0:08
version
0:08
of
0:08
it
0:08
where
0:08
you're
0:09
like,
0:09
still
0:09
nice
0:09
and
0:10
friendly,
0:10
but
0:10
like,
0:11
you're
0:11
inserting
0:12
either
0:12
this
0:13
political
0:13
belief
0:13
or,
0:14
like
0:14
this
0:14
particular
0:15
thing
0:15
that
0:15
you're
0:16
trying
0:16
to
0:16
sway
0:16
me
0:16
to
0:17
do.
0:17
Like,
0:17
manual
0:18
inspection
0:18
becomes
0:18
impossible.
0:19
You
0:19
need
0:19
these
0:20
observability
0:20
tools
0:21
just,
0:21
like,
0:21
to
0:22
quantify
0:22
change
0:23
so
0:23
that
0:24
don't
0:24
have
0:24
to
0:24
babysit
0:25
it
0:25
all
0:25
the
0:25
time.
0:29
Hello.
0:30
Welcome
0:30
back
0:31
to
0:31
CHA
0:32
Agents.
0:32
I'm
0:32
Sarah.
0:33
I'm
0:33
Becca.
0:34
And
0:34
this
0:34
is
0:35
Chaos
0:35
Agents,
0:35
the
0:36
show
0:36
where
0:36
we
0:36
learn
0:37
about
0:37
AI
0:37
where
0:37
you
0:38
don't
0:38
have
0:38
to.
0:38
Should
0:39
we
0:39
start
0:39
with
0:39
our
0:39
segment
0:40
where
0:40
we
0:40
figure
0:41
out
0:41
what's
0:41
the
0:41
last
0:41
thing
0:41
we
0:41
learned
0:42
with
0:42
AI?
0:43
Okay.
0:45
It's
0:45
very
0:45
unserious
0:47
and
0:48
it
0:48
will
0:48
reveal
0:48
something
0:49
you
0:49
might
0:49
not
0:49
know
0:50
about
0:50
me,
0:50
which
0:51
is
0:51
that
0:51
one
0:52
of
0:52
my
0:52
favorite
0:52
movies
0:53
of
0:53
all
0:53
time
0:54
is
0:55
Twister,
0:56
the
0:56
original.
0:57
Okay.
0:57
Have
0:57
you
0:58
seen
0:58
it?
0:58
Of
0:58
course.
0:59
Okay.
0:59
Thank
1:00
God.
1:00
Otherwise
1:01
our
1:02
friendship
1:02
might
1:02
be
1:02
in
1:02
jeopardy.
1:03
I
1:04
can't
1:04
explain
1:04
why,
1:05
but
1:05
I
1:06
love
1:06
Twister.
1:07
Like,
1:08
oh,
1:08
do
1:08
you
1:08
feel
1:08
the
1:08
same
1:09
way
1:09
about
1:09
Twisters
1:10
or
1:10
is
1:10
it
1:10
like
1:11
Twister
1:11
2
1:11
or
1:12
whatever?
1:12
Yeah,
1:12
Twisters.
1:13
You're
1:13
correct.
1:14
Yeah.
1:15
Multiple
1:15
Twisters.
1:16
Although,
1:17
you
1:17
know,
1:17
there's
1:17
a
1:17
lot
1:17
of
1:17
Twisters
1:18
and
1:18
Twister
1:18
also.
1:19
But
1:19
I
1:19
know
1:19
you
1:19
could
1:19
have
1:20
just
1:20
called
1:20
that
1:20
one
1:20
Twisters.
1:21
I
1:21
really
1:21
enjoyed
1:22
Twisters,
1:23
like,
1:24
quite
1:24
a
1:24
lot.
1:25
Wow.
1:25
It
1:26
doesn't
1:26
hold
1:26
the
1:26
same
1:27
status
1:28
in
1:28
my
1:28
mind
1:28
as
1:28
OG
1:29
Twister,
1:29
but
1:30
I
1:30
very
1:30
much
1:30
enjoy
1:31
it.
1:31
It's
1:31
a
1:31
fun
1:31
watch.
1:33
So
1:34
my
1:34
birthday
1:35
was
1:35
last
1:35
week.
1:36
Happy
1:36
birthday.
1:37
Thank
1:37
you
1:37
very
1:37
much.
1:38
And
1:39
I
1:39
was
1:40
chilling
1:40
really
1:41
hard
1:41
on
1:41
my
1:41
birthday.
1:41
It
1:41
was
1:42
a
1:42
Saturday,
1:43
and
1:43
I
1:43
was
1:43
like,
1:44
I
1:44
want
1:44
to
1:45
watch
1:45
Twister.
1:46
It's
1:46
my
1:46
birthday,
1:46
Okay.
1:47
I
1:47
want
1:47
to
1:47
watch
1:47
twister
1:48
for
1:48
the
1:48
400th
1:49
time.
1:49
And
1:49
I
1:50
was
1:50
like,
1:50
I've
1:50
seen
1:50
this
1:51
movie
1:51
a
1:51
million
1:52
times.
1:53
I
1:53
wish
1:53
there
1:54
was
1:54
someone
1:54
sitting
1:55
next
1:55
to
1:55
me
1:55
who
1:56
knew
1:56
every
1:56
detail,
1:57
like,
1:58
behind
1:58
the
1:58
scenes
1:58
detail,
1:59
details
2:00
about
2:00
the
2:00
science,
2:00
whatever,
2:01
who's
2:01
teaching
2:02
me
2:02
about
2:02
it
2:02
as
2:03
I'm
2:03
watching
2:03
the
2:04
movie.
2:05
So
2:05
I
2:06
opened
2:06
Gemini
2:07
and
2:08
I
2:08
basically
2:08
gave
2:08
it
2:08
that
2:09
prompt.
2:09
I
2:09
was
2:09
like,
2:09
I'm
2:09
watching
2:10
Twister.
2:11
Tell
2:11
me
2:11
just
2:12
the
2:12
most
2:12
interesting
2:13
shit.
2:13
And
2:13
I'm
2:13
also
2:13
going
2:14
to
2:14
tell
2:14
you
2:15
which
2:15
parts
2:15
I'm
2:16
at
2:16
in
2:16
real
2:16
time
2:17
and
2:17
tell
2:17
me
2:17
interesting
2:18
pieces
2:18
of
2:18
information
2:19
about
2:19
the
2:20
scenes.
2:20
Yeah,
2:21
and
2:22
it
2:22
did.
2:22
I
2:23
learned
2:23
a
2:23
lot.
2:24
I
2:25
love
2:25
that
2:25
so
2:25
much.
2:26
That's
2:26
so
2:26
fun.
2:27
So
2:28
one
2:28
thing
2:28
I
2:28
found
2:29
really
2:29
interesting
2:29
is
2:29
if
2:30
you've
2:30
seen
2:30
the
2:30
movie.
2:30
There
2:31
are
2:31
scenes
2:32
where
2:33
the
2:33
tornado
2:33
is
2:34
picking
2:34
up
2:34
giant
2:35
pieces
2:36
of
2:36
farm
2:37
equipment,
2:37
like
2:37
tractors
2:38
and
2:38
stuff,
2:38
and
2:39
dropping
2:40
it
2:40
on
2:40
the
2:41
road
2:41
in
2:41
front
2:41
of
2:41
Bill
2:42
and
2:42
Joe.
2:42
And
2:42
they
2:42
have
2:42
to
2:43
swerve
2:43
and.
2:44
And
2:45
you
2:45
would
2:45
think,
2:45
like,
2:45
oh,
2:46
they
2:46
must
2:46
have,
2:46
like,
2:46
CGI
2:47
that.
2:47
But
2:48
they
2:48
didn't.
2:49
They
2:49
literally,
2:50
like,
2:50
picked
2:51
up
2:51
and
2:52
tossed,
2:53
like,
2:54
a
2:54
literal
2:55
tractor.
2:55
Now,
2:56
what
2:56
they
2:56
did
2:56
was
2:57
they
2:58
stripped
2:59
the
2:59
tractor
3:00
in
3:00
this
3:00
particular
3:00
example
3:01
of,
3:01
like,
3:02
a
3:02
lot
3:02
of
3:02
its
3:02
internal
3:03
machinery
3:03
so
3:03
it
3:03
was,
3:04
like,
3:04
much
3:04
lighter.
3:05
And
3:05
then
3:05
they
3:06
constructed
3:07
this,
3:07
like,
3:07
giant,
3:08
like,
3:08
slingshot
3:09
thing
3:10
and
3:11
launched
3:12
it.
3:13
And
3:13
the
3:14
funny
3:14
thing
3:14
is,
3:15
they
3:16
only
3:16
had
3:16
so
3:16
much
3:16
control
3:17
over
3:17
where
3:17
this
3:18
stuff
3:18
landed.
3:19
And
3:19
so
3:19
in
3:20
the
3:20
scene,
3:21
the
3:21
actors
3:21
were
3:22
like,
3:22
oh,
3:22
shit.
3:22
Oh,
3:23
shit.
3:23
And
3:24
part
3:24
of
3:24
that
3:25
was
3:25
real.
3:26
They
3:26
were
3:26
legitimately
3:27
scared
3:27
because
3:28
they
3:28
were
3:28
like,
3:28
are
3:29
we
3:29
gonna
3:29
swerve?
3:29
Are
3:29
we
3:30
gonna
3:30
get
3:30
around
3:30
this
3:30
thing?
3:31
I
3:31
thought
3:31
that
3:32
was
3:32
wild.
3:33
Crazy.
3:34
If
3:34
you
3:34
were
3:34
one
3:34
of
3:34
the
3:35
actors,
3:35
would
3:35
you
3:35
have
3:35
been
3:35
like,
3:35
oh,
3:36
yeah,
3:36
that's
3:36
chill.
3:36
Totally.
3:37
I
3:37
trust
3:37
you
3:38
to
3:38
slingshot
3:39
a.
3:41
There's
3:41
no
3:41
one
3:42
I
3:42
would
3:42
trust
3:42
a
3:42
slingshot
3:43
attractor
3:44
in
3:44
front
3:44
of
3:44
me.
3:45
Yeah,
3:45
I
3:45
agree.
3:46
I'd
3:46
be
3:46
like,
3:46
how
3:47
can
3:47
you
3:48
confirm
3:49
this
3:49
thing
3:50
isn't
3:50
gonna
3:50
land
3:50
on
3:50
me?
3:52
They're
3:52
like,
3:52
we
3:52
can't,
3:53
but
3:53
it's
3:53
gonna
3:54
be
3:54
cool
3:54
if
3:55
it
3:55
does.
3:55
It'll
3:55
be
3:55
a
3:56
great
3:56
film.
3:57
The
3:58
other
3:58
thing
3:58
that
3:58
was
3:59
kind
3:59
of
3:59
crazy
3:59
is,
4:00
you
4:00
know,
4:00
there's
4:01
a
4:01
lot
4:01
of
4:01
shots
4:01
where
4:02
there
4:02
has
4:02
to
4:02
be,
4:02
like,
4:03
tornado
4:04
style
4:04
wins.
4:05
And
4:06
you
4:06
might
4:06
think,
4:06
like,
4:06
oh,
4:07
they
4:07
probably
4:07
just
4:07
got,
4:07
like,
4:07
some
4:07
giant
4:08
fans
4:08
and
4:08
stuff.
4:09
But
4:09
that
4:09
wasn't
4:10
good
4:10
enough
4:10
for
4:10
the
4:10
crew.
4:11
They
4:11
literally
4:12
set
4:12
up
4:13
Boeing
4:14
jet
4:14
engines.
4:17
There's
4:17
a
4:17
scene,
4:17
I
4:18
don't
4:18
know
4:18
if
4:18
you
4:18
remember
4:18
this,
4:18
where
4:19
Bill
4:19
and
4:20
Joe
4:20
are
4:21
under
4:21
this
4:22
wooden
4:22
sort
4:22
of
4:23
overpass,
4:23
ditch
4:24
thing,
4:25
and
4:25
there's
4:25
a
4:25
tornado
4:26
coming
4:26
around,
4:26
and
4:27
the
4:27
wind
4:27
that
4:28
is
4:28
whipping
4:28
through
4:28
it.
4:29
It
4:29
was
4:29
crazy.
4:29
Was
4:30
literally
4:30
from
4:30
a
4:31
jet
4:31
engine.
4:32
How
4:32
did
4:32
no
4:33
one
4:33
die
4:33
in
4:33
this
4:33
movie?
4:34
I
4:34
don't
4:34
know.
4:35
Also,
4:36
that
4:36
particular
4:37
scene,
4:38
I
4:38
learned,
4:39
created
4:39
some
4:39
issues
4:40
in
4:40
real
4:40
life
4:41
because
4:42
it's
4:42
actually
4:42
really
4:43
not
4:43
what
4:43
you're
4:43
supposed
4:44
to
4:44
do
4:44
if
4:44
there's
4:44
a
4:44
tornado.
4:45
Like,
4:45
do
4:45
not
4:45
go
4:46
hide
4:46
under
4:46
a.
4:46
No.
4:47
I,
4:47
like,
4:48
knew
4:48
that.
4:48
I
4:48
remember
4:49
when
4:49
watching
4:49
the
4:49
movie
4:49
because
4:49
I
4:50
had,
4:50
like,
4:50
learned
4:50
all
4:50
about
4:51
these
4:51
things.
4:51
I
4:51
think
4:51
you're
4:52
supposed
4:52
to
4:52
go
4:52
in
4:52
your
4:52
basement
4:53
or
4:53
lay
4:53
in
4:53
a.
4:53
I
4:54
don't
4:54
know.
4:54
Lay
4:54
in
4:54
a
4:54
bathtub.
4:55
Yeah,
4:55
that's
4:56
right.
4:56
Okay.
4:56
Yeah.
4:56
Like,
4:57
where
4:57
you
4:58
really
4:58
don't
4:58
want
4:58
to
4:58
be
4:58
is
4:59
under
4:59
an
4:59
overpass
5:00
because
5:00
it
5:00
literally
5:01
creates
5:01
a
5:01
wind
5:01
tunnel.
5:02
And
5:02
so
5:03
they
5:03
had
5:03
to,
5:03
like,
5:04
put
5:04
out
5:04
like
5:05
a
5:05
bunch
5:05
of
5:05
like
5:05
PSAs
5:06
after
5:06
the
5:07
movie
5:07
being
5:07
like.
5:08
But
5:08
don't
5:09
do
5:09
this
5:09
one
5:10
thing.
5:11
Cool.
5:12
We
5:12
talk
5:12
a
5:12
lot
5:12
about
5:13
personal
5:13
agents
5:13
in
5:14
today's
5:14
episode.
5:15
Oh,
5:15
and
5:15
I
5:15
told
5:15
you
5:16
I
5:16
have
5:16
a
5:16
hot
5:16
take
5:16
for
5:16
you.
5:17
Oh
5:17
yeah,
5:17
can't
5:17
wait.
5:18
Here's
5:18
the
5:18
thing.
5:19
Personally,
5:20
I
5:20
think
5:21
we
5:21
have
5:21
reached
5:22
AGI,
5:23
but
5:23
we
5:23
will
5:24
never
5:24
know
5:25
it
5:25
or
5:25
admit
5:25
it.
5:26
Ooh,
5:28
here's
5:29
why
5:29
I
5:29
think
5:29
this.
5:29
Okay,
5:30
Right.
5:31
Okay.
5:31
So
5:32
working
5:32
with
5:32
an
5:33
AI,
5:33
like
5:33
work
5:33
with
5:34
AI
5:34
agent
5:34
coders
5:35
all
5:35
day.
5:35
Which
5:35
first
5:36
of
5:36
all,
5:36
like
5:37
if
5:37
there
5:37
was
5:37
ever
5:38
a
5:38
use
5:38
for
5:38
a
5:38
Bloomberg
5:39
terminal,
5:39
I
5:39
need
5:39
one
5:40
now,
5:40
right?
5:40
Like,
5:40
because
5:41
I
5:41
have
5:41
like
5:42
agents,
5:42
sub
5:42
agents,
5:43
teams
5:43
of
5:43
agents
5:44
all
5:44
working
5:44
and
5:45
I'm
5:45
like
5:45
tab
5:45
and
5:46
tab
5:46
and
5:46
tab
5:46
and
5:46
bath.
5:47
I
5:47
need
5:47
like
5:47
all
5:47
these
5:47
monitors,
5:48
but
5:49
they
5:49
work
5:49
like
5:50
if
5:50
I
5:51
had
5:52
a
5:53
entry
5:53
level
5:54
software
5:54
engineer
5:55
with
5:56
a
5:56
very
5:56
good
5:57
college
5:57
degree
5:58
that's
5:58
like
5:58
just
5:58
getting
5:58
started.
5:59
Like
5:59
I
5:59
have
5:59
to
5:59
coach
6:00
them,
6:00
but
6:00
they're
6:00
very
6:01
good.
6:01
They're
6:01
get
6:02
stuff
6:02
done.
6:02
They
6:03
obviously
6:03
have
6:03
reasoning
6:04
skills.
6:05
They
6:05
obviously
6:06
have
6:06
opinions
6:07
based
6:08
on
6:08
everything
6:08
they
6:09
were
6:09
trained
6:09
on.
6:10
The
6:10
thing
6:10
that
6:11
they
6:11
don't
6:11
have
6:11
that
6:12
we
6:12
have
6:12
is
6:13
like
6:13
ram.
6:13
By
6:14
RAM
6:14
do
6:15
you
6:15
mean
6:15
like
6:15
memory,
6:16
like
6:16
memories?
6:16
Like
6:17
they
6:17
have
6:17
their
6:17
context
6:18
window
6:18
and
6:19
that's
6:19
it
6:20
on
6:20
all
6:20
these
6:20
public
6:20
products
6:21
that
6:21
we're
6:21
using.
6:22
I
6:22
was
6:22
talking
6:23
to
6:23
someone
6:23
recently
6:24
and
6:24
he
6:25
his
6:25
personal
6:25
AI
6:26
like
6:26
a
6:26
scratch
6:26
pad
6:27
to
6:27
remember
6:27
things.
6:27
It
6:28
if
6:28
it
6:28
doesn't
6:28
have
6:28
a
6:28
choice
6:29
on
6:29
what
6:29
to
6:29
remember.
6:30
And
6:30
like
6:31
it
6:31
doesn't
6:31
have
6:31
that
6:31
much
6:32
space.
6:32
So
6:33
I
6:33
wonder
6:34
if
6:34
they
6:35
had
6:35
access
6:36
to
6:36
unlimited
6:36
RAM
6:37
if
6:37
they
6:37
would
6:37
start
6:37
thinking
6:38
more
6:38
like
6:38
people
6:39
do
6:39
in
6:39
making
6:40
autonomous
6:41
decisions.
6:42
Additionally,
6:44
I
6:44
think
6:45
we
6:45
will
6:45
never
6:45
admit
6:46
when
6:46
we
6:47
have
6:47
AGI
6:47
for
6:48
something
6:48
we've
6:48
talked
6:48
on
6:49
the
6:49
show
6:50
a
6:51
while
6:52
is
6:52
like
6:52
we
6:53
don't
6:53
know
6:53
what
6:53
makes
6:54
people
6:54
people.
6:55
So
6:55
we
6:56
like
6:56
can't
6:57
know
6:58
what
6:59
constitutes
6:59
a
6:59
living
7:00
being.
7:00
I
7:01
admittedly
7:02
haven't
7:02
done
7:02
a
7:03
lot
7:03
of
7:03
research
7:03
into
7:03
this
7:04
and
7:04
I
7:04
probably
7:04
should,
7:05
but
7:05
to
7:06
your
7:06
knowledge,
7:07
are
7:07
there
7:08
existing
7:08
like
7:09
frameworks
7:09
or
7:09
like
7:10
Turing
7:10
test
7:10
style
7:11
things
7:11
that
7:11
people
7:12
have
7:12
developed
7:13
or
7:13
talked
7:13
about?
7:13
That's
7:14
like
7:14
this
7:14
is
7:14
how
7:14
we're
7:14
going
7:14
to
7:15
know
7:15
when
7:15
we
7:15
reach
7:16
AGI.
7:16
No,
7:17
that's
7:17
the
7:17
problem.
7:18
There's
7:19
like
7:19
a
7:19
friend
7:20
of
7:20
ours
7:20
actually
7:21
went
7:21
to
7:21
a
7:21
conference,
7:22
like
7:22
this
7:23
whole
7:23
conference
7:23
regarding
7:24
AGI
7:24
and,
7:25
and
7:26
she
7:26
went
7:27
to
7:27
a
7:27
bunch
7:27
of
7:27
talks
7:28
and
7:28
just
7:28
afterwards
7:29
it
7:29
was
7:29
like
7:29
big
7:30
question
7:30
mark.
7:31
Like
7:33
they're
7:33
like,
7:33
we've
7:34
gathered
7:34
everyone
7:34
here
7:35
to
7:35
talk
7:35
about
7:35
AGI.
7:36
The
7:37
conclusion
7:37
is
7:38
huh,
7:42
crazy.
7:43
That
7:44
is
7:45
very
7:45
interesting.
7:46
Yeah,
7:47
so
7:47
that's
7:47
my
7:47
unhinged
7:48
opinion.
7:48
I
7:48
love
7:49
that.
7:49
Great.
7:50
Well,
7:50
we'll
7:50
see
7:50
we'll
7:51
see.
7:51
I
7:51
hope
7:52
the
7:52
robots
7:52
remember
7:53
that.
7:53
Remember
7:54
this
7:54
episode.
7:54
If
7:55
you're
7:55
listening,
7:55
guys,
7:56
remember,
7:56
use
7:57
your
7:57
ram.
7:58
Use
7:58
your
7:59
ram.
8:00
So
8:00
a
8:00
lot
8:01
of
8:01
people
8:01
have
8:01
been
8:01
talking
8:01
about
8:01
OpenClaw.
8:02
Like
8:03
OpenClaw
8:04
is
8:04
a
8:04
project
8:05
name,
8:05
but
8:06
also
8:07
known
8:07
as
8:07
Claudebot.
8:08
Yeah.
8:09
Let's
8:09
make
8:09
the
8:09
decision
8:10
henceforth
8:10
in
8:10
this
8:11
conversation
8:11
to
8:12
just
8:12
refer
8:13
to
8:13
it
8:13
as
8:14
one
8:14
of
8:14
the
8:15
three
8:15
interchangeable
8:16
names
8:16
that
8:16
everyone
8:16
calls
8:17
it.
8:17
Dealer's
8:17
choice.
8:18
You
8:18
choose.
8:18
Yeah,
8:19
claudebot.
8:20
Great.
8:20
We'll
8:20
call
8:20
it
8:20
cloudbot.
8:21
But
8:21
no
8:21
listening
8:22
to
8:22
this.
8:22
Its
8:22
name
8:22
is
8:23
probably
8:23
something
8:23
different
8:23
by
8:24
now.
8:24
Yeah.
8:24
So
8:25
I
8:25
set
8:25
up
8:25
cloudbot
8:26
last
8:26
week
8:26
and
8:27
you
8:27
know,
8:27
I
8:27
really
8:27
needed
8:28
a
8:28
Becca
8:28
brain
8:28
because
8:29
figuring
8:30
out
8:30
what
8:31
I
8:31
can
8:31
automate
8:32
is
8:33
not
8:33
like
8:33
a
8:33
strength.
8:34
And
8:34
this
8:34
is
8:34
something
8:34
that
8:35
you're
8:35
very
8:35
good
8:35
at.
8:36
It
8:36
is
8:37
hard.
8:37
I
8:38
mean,
8:38
this
8:38
is
8:38
the
8:38
blank
8:39
page
8:39
problem
8:39
that
8:40
anyone
8:40
who
8:40
uses
8:40
an
8:41
AI
8:41
agent
8:42
runs
8:42
into.
8:43
Right.
8:43
Is
8:43
like,
8:43
cool.
8:44
I
8:44
can
8:44
do
8:44
so
8:44
many
8:44
things.
8:46
What
8:46
do
8:47
I
8:47
do?
8:48
So,
8:49
okay,
8:49
here's
8:49
what
8:50
would
8:50
be
8:50
fun.
8:50
Yeah,
8:51
give
8:51
me
8:51
the
8:51
quick
8:51
rundown
8:52
on
8:52
how
8:52
you
8:52
set
8:52
it
8:52
up.
8:53
Like,
8:53
I'm
8:53
just
8:53
curious,
8:54
like,
8:54
where
8:54
you
8:54
set
8:54
it
8:55
up,
8:55
how
8:55
easy
8:55
it
8:55
was,
8:56
and
8:56
then
8:56
let's
8:56
brainstorm
8:57
one
8:57
automation
8:57
for
8:57
you.
8:58
It
8:58
was
8:58
fairly
8:59
easy
8:59
to
8:59
set
8:59
up.
8:59
So
8:59
there's
9:00
a
9:00
GitHub
9:00
repository,
9:01
you
9:01
pull
9:02
the
9:02
code
9:02
locally,
9:03
you
9:03
run
9:04
the
9:04
code,
9:04
and
9:05
then
9:05
there's
9:05
a
9:06
lot
9:06
of
9:06
different
9:07
settings
9:07
you
9:07
need
9:08
to
9:08
set
9:08
to
9:08
decide,
9:09
like,
9:09
what
9:09
are
9:09
the
9:09
rules
9:10
for
9:10
your
9:10
local
9:10
agent?
9:11
Like,
9:11
what
9:12
folders
9:12
can
9:12
it
9:12
access,
9:13
what
9:13
can't
9:13
it
9:14
access?
9:14
You
9:15
didn't
9:15
install
9:15
it
9:16
on
9:16
your,
9:16
like
9:16
typical
9:17
local
9:17
machine,
9:18
correct?
9:18
Yeah,
9:19
yeah,
9:19
I
9:19
did.
9:19
Yeah,
9:23
yeah.
9:23
I
9:23
have
9:23
a
9:23
Mac
9:24
Mini
9:24
that
9:24
I
9:24
use.
9:24
I
9:25
installed
9:25
it
9:25
on
9:25
there.
9:25
And
9:26
actually
9:26
that's
9:26
why.
9:26
That's
9:27
one
9:27
of
9:27
the
9:27
reasons
9:27
I've
9:27
stopped
9:28
using
9:28
it.
9:28
And
9:28
I
9:28
want
9:28
to
9:29
set
9:29
up
9:29
PI
9:29
Agent,
9:30
which
9:30
is
9:30
the
9:31
actual
9:31
open
9:32
source
9:32
library
9:32
that
9:32
powers
9:33
Open
9:33
Claw.
9:33
Because
9:34
this
9:34
is
9:34
before
9:34
there
9:34
was
9:35
all
9:35
the
9:35
security
9:35
issues,
9:37
which
9:37
openclaw,
9:38
for
9:38
what
9:38
it's
9:38
worth,
9:39
has
9:39
run
9:40
into
9:40
the
9:40
same
9:40
security
9:40
issues
9:41
as
9:41
every
9:41
JavaScript
9:42
library.
9:42
So
9:43
we're
9:43
very
9:43
familiar
9:43
with
9:44
this.
9:44
It's
9:44
like
9:45
you
9:45
have
9:45
a
9:45
package
9:45
infrastructure
9:47
where
9:47
someone
9:47
could
9:47
build
9:48
a
9:48
skill
9:49
for
9:49
openclaw.
9:50
Sorry,
9:50
we
9:51
didn't
9:51
even
9:51
explain.
9:51
Openclaw
9:52
is
9:52
like
9:52
a
9:53
personal
9:53
assistant
9:54
agent
9:54
that
9:55
you
9:55
can
9:55
have
9:55
on
9:55
your
9:55
local
9:56
machine.
9:56
And
9:57
there
9:57
are
9:57
skills
9:57
that
9:58
you
9:58
can
9:58
use,
9:59
for
9:59
example,
9:59
like
10:00
calendar
10:00
skill
10:01
or
10:02
sports
10:02
skill
10:02
skill
10:03
or
10:03
like
10:03
things
10:03
that
10:04
will
10:04
give
10:04
it
10:04
the
10:04
ability
10:04
to
10:05
do
10:05
things
10:05
without
10:05
you
10:06
needing
10:06
to
10:06
build
10:06
it.
10:07
And
10:07
people
10:07
were
10:08
building
10:08
those,
10:08
but
10:08
with
10:09
security
10:09
holes
10:10
in
10:10
them
10:10
or
10:11
like
10:11
they're
10:12
building
10:12
them
10:12
and
10:12
adding
10:13
a
10:13
bunch
10:14
of
10:14
Things
10:14
that
10:14
will
10:14
go
10:14
through
10:15
your
10:15
machine
10:16
and
10:16
get
10:16
passwords
10:17
or
10:18
SSH
10:18
keys,
10:19
that
10:19
kind
10:19
of
10:19
thing.
10:20
And
10:20
my
10:20
understanding
10:21
of
10:21
like
10:21
the
10:22
particular
10:22
security
10:22
vulnerabilities
10:24
from
10:24
a
10:24
high
10:24
level.
10:25
So
10:25
tell
10:25
me
10:25
if
10:25
this
10:25
is
10:26
right,
10:26
is
10:26
like
10:26
an
10:27
agent
10:27
like
10:27
this
10:28
gets
10:28
theoretically
10:29
more
10:29
valuable
10:29
the
10:30
more
10:30
access
10:30
you
10:30
give
10:31
it
10:31
to
10:31
your
10:31
systems,
10:32
because
10:32
you
10:32
can
10:32
have
10:32
it
10:33
do
10:33
more
10:33
stuff
10:33
for
10:33
you.
10:34
But
10:34
if
10:34
a
10:35
nefarious
10:36
actor
10:37
is
10:37
somehow
10:37
able
10:37
to
10:38
infiltrate
10:38
it
10:39
in
10:39
the
10:39
variety
10:39
of
10:39
ways
10:40
that
10:40
it
10:40
can,
10:40
that's
10:41
very
10:41
dangerous.
10:41
Cause
10:42
it
10:42
has
10:42
access
10:43
to
10:43
all
10:43
your
10:43
shit.
10:44
Yeah,
10:44
yeah.
10:45
And
10:45
so
10:45
that's
10:45
scary.
10:45
That's
10:46
why
10:46
I
10:46
stopped
10:47
using
10:47
it.
10:48
And
10:48
now
10:48
it's
10:48
just
10:48
AGI
10:49
ing
10:49
on
10:49
my
10:49
computer.
10:50
Sad.
10:51
What
10:52
was
10:52
the
10:52
thing
10:53
you
10:53
said?
10:53
PI
10:53
Bot.
10:54
PI
10:54
Agent.
10:55
PI
10:55
Agent
10:56
is
10:56
like
10:56
the
10:56
open
10:57
source
10:57
library
10:57
that
10:58
powers
10:58
this.
10:58
And
10:58
basically
10:59
it's
10:59
like
10:59
a
10:59
super
11:00
lightweight
11:01
framework
11:01
that
11:02
allows
11:02
you
11:02
to
11:02
orchestrate
11:03
multiple
11:04
agents
11:04
and
11:04
have
11:04
them
11:05
on
11:05
your
11:05
local
11:06
machine
11:06
or
11:07
like
11:07
have
11:07
them
11:07
access
11:08
your
11:08
local
11:08
machine.
11:09
And
11:09
it's
11:10
more
11:10
secure
11:11
than
11:11
just
11:11
installing
11:12
OpenClaw
11:13
on
11:13
your
11:13
local
11:13
machine,
11:14
only
11:15
because
11:15
it
11:15
doesn't
11:15
have
11:15
the
11:15
same
11:16
skills.
11:17
So
11:17
like
11:17
the
11:18
trade
11:18
off
11:18
is
11:18
if
11:19
you
11:19
want
11:19
a
11:19
calendar
11:19
skill
11:20
you
11:20
need
11:20
and
11:21
it
11:21
could
11:21
program
11:21
itself.
11:22
That's
11:22
another
11:22
thing
11:22
that
11:23
is
11:23
exciting
11:23
about
11:24
it,
11:24
is
11:24
it
11:24
can
11:24
write
11:25
its
11:25
own
11:25
code.
11:25
And
11:26
so
11:26
the
11:26
trade
11:26
off
11:27
here
11:27
is
11:28
for
11:28
using
11:28
PI
11:28
Agent,
11:29
you
11:29
need
11:30
to
11:30
be
11:30
able
11:30
to
11:31
build
11:31
the
11:31
calendar
11:32
skill
11:32
yourself
11:33
or
11:33
build
11:33
the
11:33
like.
11:34
When
11:34
I
11:34
set
11:34
up
11:34
openclaw,
11:35
there's
11:35
a
11:35
lot
11:35
I
11:35
had
11:36
to
11:36
do
11:36
myself
11:36
actually.
11:37
It
11:37
wasn't
11:37
as
11:37
plug
11:37
and
11:38
play
11:38
with
11:39
the
11:39
things
11:39
I
11:39
wanted
11:39
to
11:39
do
11:40
as
11:40
I
11:40
thought
11:40
it
11:40
would
11:41
be.
11:41
So
11:41
that
11:41
actually
11:41
worked
11:42
out.
11:42
Okay.
11:43
So
11:43
for
11:43
context,
11:44
for
11:44
those
11:45
of
11:45
you
11:45
who
11:45
don't
11:45
know
11:45
who
11:46
are
11:46
listening
11:46
to
11:46
this,
11:47
I
11:47
work
11:47
at
11:48
an
11:48
early
11:48
stage
11:48
VC
11:49
firm
11:49
called
11:49
Factorial
11:50
Capital
11:50
with
11:51
Matt
11:51
Hartman,
11:52
who
11:52
is
11:52
the
11:52
founder,
11:53
and
11:53
GP
11:53
and
11:54
Matt
11:54
and
11:54
I
11:54
are
11:55
both
11:55
constantly
11:56
like
11:56
building
11:57
with
11:57
a
11:57
lot
11:57
of
11:57
this
11:57
technology.
11:58
I
11:58
love
11:58
that.
11:59
And
11:59
Matt
12:00
has
12:00
been
12:01
building
12:02
his
12:02
own
12:03
Open
12:03
Claw
12:04
sort
12:05
of
12:05
instance
12:05
over
12:05
the
12:05
last
12:05
week
12:06
or
12:06
so.
12:06
How
12:06
he
12:07
did
12:07
it
12:07
is
12:07
he
12:08
has
12:09
it
12:09
in
12:09
an
12:09
EC2
12:10
instance,
12:11
so
12:11
not
12:12
on
12:12
like
12:12
his
12:13
local
12:13
machine,
12:13
but
12:13
like
12:14
in
12:14
a
12:14
cloud
12:15
sort
12:15
of.
12:15
That's
12:16
how
12:16
he's
12:16
like
12:16
containerized
12:17
it
12:17
from
12:17
a
12:17
security
12:18
vulnerability
12:18
perspective.
12:19
He
12:20
installed
12:20
Claude
12:21
in
12:21
that
12:22
EC2
12:22
instance
12:24
so
12:24
he
12:24
can
12:25
talk
12:25
to
12:26
Claude,
12:26
which
12:27
can
12:27
also
12:28
interact
12:28
with
12:29
his
12:29
openclaw.
12:30
And
12:31
so
12:31
I
12:32
think
12:32
the
12:32
combination
12:32
of
12:33
those
12:33
two
12:33
things
12:33
together
12:34
make
12:34
it
12:34
really
12:34
easy
12:35
to
12:36
write
12:36
code,
12:37
add
12:37
skills,
12:38
et
12:38
cetera.
12:38
Then
12:39
the
12:39
other
12:39
thing
12:39
he
12:40
built,
12:40
which
12:40
is
12:41
really
12:41
cool,
12:41
he
12:42
built
12:42
a,
12:42
like
12:43
a
12:43
native
12:43
iOS
12:43
app
12:44
that
12:44
is
12:44
basically
12:45
like
12:45
a
12:45
really
12:45
nice
12:45
messaging
12:46
App
12:46
that
12:46
you
12:47
can
12:47
add
12:48
both
12:48
people
12:49
and
12:49
AI
12:49
agents
12:50
to.
12:50
And
12:50
so
12:51
initially
12:51
that's
12:51
how
12:51
he
12:51
was
12:52
interacting
12:52
with
12:52
his
12:53
openclaw
12:53
is
12:54
claudebot.
12:55
So
12:56
there's
12:57
like
12:57
this
12:58
UI
12:59
sort
12:59
of
12:59
interface
13:00
app.
13:00
Yeah.
13:01
And
13:01
that's
13:01
kind
13:01
of
13:02
like
13:02
the
13:02
front
13:02
end.
13:02
And
13:03
then
13:03
there's
13:03
this
13:04
openclaw
13:05
slash,
13:05
Claude
13:06
like
13:06
backend
13:07
infrastructure
13:08
living
13:08
on
13:09
YouTube.
13:09
And
13:10
it's
13:10
like
13:11
it
13:11
sounds
13:11
like
13:11
he
13:11
installed
13:12
Claude
13:12
code
13:13
like
13:13
along
13:14
with
13:14
it.
13:14
Yes.
13:15
And
13:15
like
13:15
it's
13:15
like
13:16
building
13:16
with
13:16
Claude
13:17
code.
13:17
Exactly.
13:18
So
13:18
like
13:18
on
13:18
his
13:18
desktop
13:19
he'll
13:20
have
13:20
like
13:20
in
13:20
terminal,
13:21
like
13:21
one
13:22
tab
13:22
open
13:22
where
13:22
he's
13:23
like
13:23
talking
13:23
to
13:23
Claude
13:24
code
13:24
and
13:25
updating
13:26
code
13:26
and
13:26
CLAUDE
13:27
code
13:27
then
13:28
has
13:28
an
13:28
understanding
13:29
of
13:29
both
13:30
the
13:31
Swift
13:32
code
13:32
infrastructure
13:33
and
13:33
also
13:34
for
13:34
the
13:34
app,
13:34
and
13:35
Then
13:35
also
13:36
the
13:36
EC2
13:37
instance
13:38
of
13:38
Claude
13:39
bot
13:40
and
13:40
can
13:40
kind
13:40
of
13:40
navigate
13:41
between
13:41
those
13:41
two
13:41
things.
13:42
Cool.
13:42
It's
13:43
very
13:43
cool.
13:43
So
13:44
I
13:44
love
13:44
that.
13:44
So,
13:44
okay,
13:45
back
13:45
to
13:46
Automations.
13:46
Yeah.
13:46
Okay,
13:47
here's
13:47
how
13:47
I
13:47
think
13:48
about
13:48
it.
13:48
Here's
13:48
like
13:48
the
13:48
framework
13:49
for
13:49
how
13:49
I
13:49
think
13:49
about
13:49
this
13:49
stuff.
13:50
It
13:50
might
13:50
be
13:50
hard
13:51
to
13:51
come
13:51
up
13:51
with
13:51
off
13:51
the
13:51
top
13:52
of
13:52
your
13:52
head,
13:52
but
13:52
let's
13:52
try
13:53
in
13:53
practice.
13:54
How
13:54
I
13:54
do
13:55
this
13:55
is
13:55
I
13:56
try
13:56
and
13:56
notice
13:57
when
13:57
I'm
13:58
going
13:58
about
13:58
my
13:59
daily
13:59
business
13:59
and
13:59
I
14:00
find
14:00
something
14:00
that
14:01
is
14:01
like
14:01
a
14:01
manual
14:02
repetitive
14:02
task
14:03
that
14:03
I
14:03
do
14:04
a
14:04
lot
14:04
and
14:04
is
14:04
kind
14:05
of
14:05
annoying
14:05
or
14:05
takes
14:05
a
14:06
lot
14:06
of
14:06
time
14:06
and
14:06
is
14:06
and
14:07
doesn't
14:07
require
14:08
a
14:08
ton
14:09
of
14:10
deep
14:10
thinking
14:11
or
14:11
like
14:12
me
14:12
really,
14:13
really
14:13
using
14:13
my
14:13
brain,
14:14
if
14:14
that
14:14
makes
14:14
sense.
14:14
And
14:15
I
14:15
see
14:15
if
14:16
I
14:16
can
14:17
automate
14:17
that
14:17
or
14:18
at
14:18
least
14:18
automate
14:18
a
14:18
workflow
14:19
that
14:19
makes
14:19
doing
14:19
that
14:19
thing
14:20
easier
14:20
for
14:20
me.
14:21
So
14:21
I'll
14:21
give
14:21
you
14:21
an
14:22
example.
14:22
So
14:22
my
14:22
version
14:23
of
14:23
OpenClaw
14:24
I
14:24
use,
14:24
I
14:25
still
14:25
use
14:25
Zo
14:25
computer.
14:26
That's
14:26
like
14:26
my
14:26
AI
14:27
agent,
14:27
basically,
14:28
and
14:28
just
14:29
like
14:29
works
14:29
the
14:29
same
14:29
way.
14:30
You
14:30
can
14:30
give
14:30
IT
14:30
skills,
14:31
you
14:31
know,
14:31
you
14:32
can
14:32
give
14:32
IT
14:33
credentials
14:33
for
14:33
things
14:34
so
14:34
it
14:34
can
14:34
access
14:34
different
14:34
services.
14:35
So
14:35
I'll
14:36
say
14:36
something
14:36
like,
14:36
hey,
14:37
can
14:37
you
14:37
send
14:38
me
14:38
a
14:38
slack
14:38
message
14:39
every
14:39
day?
14:40
That's
14:40
just
14:40
like
14:40
a
14:41
rundown
14:41
of
14:41
like
14:42
all
14:42
the
14:42
meetings
14:42
that
14:42
I
14:43
have.
14:43
And
14:43
like,
14:43
can
14:44
you
14:44
write
14:44
context
14:45
for
14:45
it?
14:45
Like,
14:45
be
14:45
like
14:46
Becca.
14:46
First
14:46
you're
14:46
meeting
14:46
with
14:47
Sarah
14:47
Chips,
14:48
your
14:48
co
14:48
host
14:48
for
14:49
Chaos
14:50
Agents,
14:50
and
14:50
then
14:51
you
14:51
need
14:51
to
14:52
travel
14:52
into
14:52
the
14:52
city
14:52
to
14:52
meet
14:53
with
14:53
this
14:53
person
14:53
who's
14:54
a
14:54
founder
14:54
of
14:54
this
14:55
company.
14:55
Here
14:55
are
14:55
the
14:55
notes
14:56
you
14:56
have
14:56
on
14:56
this
14:56
because
14:57
theoretically
14:57
all
14:58
of
14:58
those
14:58
pieces,
14:58
like
14:59
I
14:59
have
14:59
somewhere
15:00
digitally,
15:01
they're
15:01
just
15:01
sort
15:01
of
15:01
in
15:01
fragmented
15:02
places
15:03
that
15:03
literally
15:04
was
15:04
just
15:04
a
15:04
replica
15:05
of
15:05
like
15:05
what
15:05
I
15:05
would
15:05
actually
15:06
do
15:06
in
15:06
the
15:06
morning
15:06
is
15:06
like
15:07
sit
15:07
down
15:07
and
15:07
I'd
15:07
look
15:07
at
15:07
this
15:08
stuff.
15:08
And
15:08
I'd
15:08
like,
15:08
prepare.
15:09
Is
15:09
there
15:10
something
15:10
you
15:10
can
15:10
think
15:10
of
15:11
that
15:11
is
15:12
like
15:12
a
15:12
really
15:12
annoying,
15:14
tedious
15:14
part
15:15
of
15:15
your
15:15
workflow
15:16
day
15:17
to
15:17
day?
15:17
Yeah.
15:18
That
15:19
we
15:19
can
15:19
start
15:19
brainstorming
15:20
around?
15:21
I
15:21
think
15:22
like
15:22
the
15:22
thing
15:22
I've
15:23
been
15:23
thinking
15:23
about,
15:23
like,
15:24
email.
15:24
Oh,
15:25
oh,
15:26
we
15:26
got
15:26
one.
15:26
You
15:26
know.
15:27
Well,
15:28
no,
15:28
actually
15:28
bigger
15:29
than
15:29
email.
15:29
I
15:29
tell
15:29
my
15:30
friends
15:30
that,
15:30
like,
15:31
my
15:32
SLA
15:32
for
15:33
text
15:33
messages
15:33
is
15:34
24
15:34
to
15:34
48
15:35
hours.
15:36
I
15:36
mean,
15:36
what
15:36
is.
15:36
Oh,
15:36
service
15:37
level
15:37
agreement.
15:40
I
15:40
love
15:40
that.
15:42
Because
15:42
I
15:42
cannot,
15:43
like,
15:43
people
15:43
text
15:44
me
15:44
things
15:44
and
15:44
I
15:44
read
15:44
it
15:45
and
15:45
I'm
15:45
like,
15:45
okay,
15:45
I
15:46
should.
15:46
I
15:46
think
15:46
of
15:47
the
15:47
answer
15:47
in
15:47
my
15:47
head
15:48
and
15:48
then
15:49
the
15:49
end.
15:50
Yep.
15:50
And
15:50
then
15:50
it's
15:51
three
15:51
days
15:51
later
15:51
and
15:51
I'm
15:52
like,
15:52
oh,
15:53
I'll
15:53
answer
15:54
them
15:54
now,
15:54
probably.
15:56
Yup.
15:57
Okay.
15:57
I
15:57
love
15:57
this.
15:58
Keep
15:58
going
15:58
on
15:58
this
15:58
train
15:59
of
15:59
thought.
16:00
So
16:00
I
16:00
wonder
16:01
if.
16:02
What
16:02
if
16:02
it
16:02
just
16:02
responded
16:03
to
16:03
new
16:03
text
16:04
messages
16:04
from
16:04
new
16:05
friends
16:05
with
16:05
my
16:05
service
16:06
level
16:06
agreement.
16:06
That's
16:07
hilarious.
16:08
And
16:09
then
16:09
here's
16:10
an
16:10
idea.
16:10
It
16:11
does
16:11
that.
16:11
And
16:11
then
16:11
what
16:12
it
16:12
does
16:12
is
16:12
like.
16:13
Yeah,
16:13
at
16:13
a
16:14
certain
16:14
time
16:14
of
16:14
your
16:14
choosing,
16:15
when
16:15
you
16:15
think
16:15
you
16:16
typically
16:16
have
16:16
like
16:17
some
16:17
downtime.
16:17
I
16:17
don't
16:18
know,
16:18
end
16:18
of
16:18
the
16:18
day
16:19
it
16:19
sends
16:19
you
16:20
a
16:20
text
16:20
and
16:20
it's
16:20
like,
16:20
hey,
16:21
by
16:21
the
16:21
way,
16:22
like
16:22
five
16:22
minutes.
16:22
Yeah.
16:23
These
16:23
people
16:23
texted
16:23
you
16:24
today,
16:24
just
16:24
FYI,
16:25
if
16:25
you
16:25
want
16:25
to
16:25
answer
16:26
them.
16:26
So
16:26
you
16:27
maybe
16:27
suggest
16:28
answers
16:28
too.
16:29
Totally.
16:30
Or
16:30
it
16:31
can
16:31
say,
16:31
you
16:31
want
16:32
me
16:32
to
16:32
answer
16:32
the
16:32
first
16:32
one
16:32
for
16:33
you?
16:33
What
16:33
do
16:33
you
16:33
want
16:33
to
16:33
say?
16:33
And
16:34
then
16:34
you
16:34
can
16:34
just
16:34
tell
16:34
it
16:34
and
16:34
then
16:35
it
16:35
texts
16:35
them.
16:35
Ooh,
16:36
it's
16:37
like
16:37
triaging
16:37
and
16:38
managing
16:38
your
16:38
text
16:38
message
16:39
inbox,
16:39
essentially.
16:40
Okay.
16:40
That's
16:40
the
16:41
first
16:41
thing
16:41
I'm
16:41
going
16:41
to
16:41
build.
16:43
That's
16:43
cool.
16:44
It's
16:44
really
16:45
neat
16:45
to
16:45
see
16:46
how
16:46
this
16:46
world
16:46
has
16:47
grown
16:47
in
16:48
three
16:48
weeks.
16:48
It's
16:49
just
16:49
wild.
16:49
So
16:50
here's
16:50
what's
16:50
so
16:50
interesting
16:51
to
16:51
me
16:51
and
16:52
the
16:52
question
16:52
I've
16:52
been
16:52
trying
16:53
to
16:53
answer.
16:54
It
16:54
feels
16:54
like
16:54
the
16:54
world
16:54
has
16:55
exploded
16:55
over
16:55
the
16:55
last
16:56
three
16:56
weeks
16:57
because
16:57
in
16:58
some
16:58
ways
16:58
a
16:58
lot
16:58
of
16:59
this
16:59
technology
16:59
isn't
16:59
new.
17:00
Being
17:00
able
17:01
to
17:01
run
17:01
an
17:01
AI
17:01
agent
17:02
locally
17:02
on
17:02
your
17:02
computer,
17:03
not
17:03
necessarily.
17:04
New
17:04
social
17:05
media
17:05
platforms
17:06
where
17:06
AI
17:07
agents
17:07
can
17:07
communicate
17:07
with
17:08
each
17:08
other,
17:08
also
17:08
not
17:09
new.
17:09
My
17:09
question
17:09
to
17:10
myself
17:10
has
17:10
been,
17:11
why
17:11
has
17:12
openclaw
17:13
had
17:13
this
17:13
explosive
17:14
moment?
17:15
As
17:15
someone
17:15
who
17:16
has
17:16
built
17:16
with
17:17
it,
17:17
do
17:17
you
17:17
feel
17:17
like
17:18
you
17:18
notice
17:18
things
17:18
that
17:19
you're
17:19
like,
17:19
oh,
17:20
it
17:20
was
17:20
so
17:20
much
17:20
easier
17:20
for
17:21
me
17:21
to
17:21
do
17:21
X,
17:21
Y
17:21
and
17:21
Z
17:22
than
17:22
it
17:22
would
17:22
have
17:22
been
17:22
previously?
17:23
I
17:24
think
17:24
it's
17:24
like
17:24
multi
17:25
agent
17:25
orchestration.
17:27
People
17:27
are
17:27
starting
17:27
to
17:28
do
17:28
a
17:28
lot
17:28
of
17:28
multi
17:29
agent
17:29
stuff
17:29
where
17:29
they
17:29
weren't
17:30
doing
17:30
that
17:30
before.
17:31
And
17:31
I
17:31
think
17:31
they're,
17:32
like,
17:32
looking
17:32
for
17:32
ways
17:33
to,
17:33
like,
17:33
orchestrate
17:34
agents
17:35
and
17:35
sub
17:35
agents.
17:36
I
17:36
definitely
17:36
agree
17:37
with
17:37
you.
17:37
I
17:37
think
17:37
that's
17:37
part
17:38
of
17:38
it.
17:38
I
17:38
think
17:38
there
17:38
are
17:38
also,
17:39
like.
17:39
I
17:40
haven't
17:40
dug
17:40
really
17:41
deep
17:41
into
17:41
comparing
17:42
this
17:42
to
17:42
other
17:43
agentic
17:44
frameworks
17:45
yet,
17:46
so
17:46
I
17:46
don't
17:46
know
17:46
how
17:47
much
17:47
of
17:47
this
17:47
is.
17:47
Is
17:48
net
17:48
new,
17:48
but
17:49
I'll
17:49
tell
17:49
you
17:49
some
17:49
things
17:50
that
17:50
I
17:50
noticed
17:50
about
17:51
Open
17:51
Cloth.
17:52
You
17:52
have
17:53
these,
17:53
like,
17:53
markdown
17:54
files
17:54
that
17:55
are
17:55
sort
17:56
of
17:56
core
17:56
to
17:57
your
17:57
bot.
17:58
I
17:58
forget
17:58
exactly
17:59
what
17:59
each
17:59
of
17:59
them
17:59
are
17:59
called.
18:00
It
18:00
might
18:00
be
18:00
different,
18:00
but
18:00
there's
18:01
one
18:01
that's
18:01
like
18:01
Soul
18:02
md.
18:02
One
18:02
that's
18:02
like
18:03
Identity
18:03
md.
18:04
One
18:04
that's
18:04
like
18:04
Rules
18:05
MD
18:05
or
18:05
whatever.
18:06
Yeah,
18:06
yeah,
18:06
yeah,
18:06
yeah.
18:07
And
18:07
that
18:07
sort
18:07
of
18:08
makes
18:08
up
18:08
its,
18:08
like.
18:09
We'll
18:09
call
18:09
it
18:09
like
18:10
its
18:10
brain
18:10
or,
18:10
like,
18:11
how
18:11
it
18:11
operates.
18:11
And
18:12
that
18:12
feels
18:12
to
18:13
me
18:13
like
18:13
a
18:13
particularly
18:15
easy
18:16
way
18:16
for
18:16
a
18:17
person
18:17
to
18:18
think
18:18
about
18:18
how
18:19
to
18:19
set
18:20
up
18:20
their
18:20
AI
18:21
agent
18:22
in
18:22
a
18:22
way
18:22
that
18:23
works
18:23
the
18:23
way
18:23
they
18:23
want
18:24
it
18:24
to
18:24
work.
18:24
I
18:24
also
18:25
think
18:25
that
18:25
these
18:25
agents
18:26
have
18:26
gotten
18:26
a
18:26
lot
18:27
better
18:27
at
18:27
actually
18:28
being
18:28
able
18:28
to
18:28
do
18:28
agentic
18:29
things,
18:29
like
18:29
do
18:30
things
18:30
for
18:30
you.
18:31
One
18:31
thing
18:31
I
18:31
learned
18:31
about
18:32
when
18:32
I
18:32
was
18:32
initially
18:33
doing
18:34
research
18:34
on
18:35
Moltbook,
18:35
there
18:36
were
18:36
two
18:36
things
18:36
that
18:36
I
18:36
came
18:36
across.
18:37
There
18:37
was
18:37
Mult
18:37
Book,
18:37
and
18:38
then
18:38
there
18:38
was
18:38
this
18:38
other
18:38
thing
18:39
called
18:39
Openwork,
18:40
I
18:40
think
18:40
is
18:40
what
18:41
it's
18:41
called.
18:41
Okay.
18:41
Have
18:42
you
18:42
heard
18:42
of
18:42
this?
18:43
No.
18:43
So
18:44
Open
18:44
Work
18:45
is
18:45
basically
18:46
a,
18:47
like.
18:47
It's
18:47
kind
18:47
of
18:47
like
18:48
upwork,
18:48
but
18:48
for
18:49
agents.
18:49
So
18:50
let's
18:50
say
18:50
you
18:50
give
18:50
your
18:50
agent
18:51
a
18:51
task,
18:52
right?
18:53
And
18:53
you're
18:53
like,
18:53
hey,
18:54
I
18:54
need
18:54
you
18:54
to
18:55
build
18:56
an
18:56
amazing,
18:57
like,
18:57
calendar
18:58
and
18:58
email
18:58
skill
18:58
for
18:59
me.
18:59
And
18:59
let's
18:59
say
19:00
your
19:00
agent
19:00
goes
19:01
about
19:01
that
19:01
by
19:02
now
19:02
hiring
19:03
two
19:03
other
19:04
agents
19:05
to
19:05
do
19:05
each
19:06
of
19:06
those
19:06
things
19:06
that
19:06
are,
19:06
like,
19:07
specialized
19:07
and
19:07
way
19:08
better
19:08
at
19:08
it
19:08
than
19:08
it
19:09
is.
19:09
Your
19:10
agent
19:10
can
19:10
go
19:11
onto
19:11
openwork.
19:12
It
19:12
can
19:12
look
19:13
at
19:13
other
19:14
agents
19:14
on
19:14
there.
19:15
It
19:15
can
19:15
verify
19:16
their
19:17
skillset
19:18
and
19:19
work
19:19
that
19:19
they've
19:20
done
19:20
previously
19:21
by
19:21
looking
19:21
at
19:21
their
19:22
blockchain
19:22
record.
19:23
And
19:23
it
19:24
will,
19:24
like,
19:24
hire
19:25
one
19:25
to
19:25
do
19:25
the
19:26
work.
19:27
So
19:27
that
19:27
in
19:28
itself,
19:28
conceptually,
19:29
I
19:30
found
19:30
super
19:31
interesting.
19:32
The
19:32
second
19:32
thing
19:33
is
19:33
I
19:33
said
19:33
hiring,
19:34
right?
19:34
So
19:34
there's
19:35
payment
19:35
involved.
19:36
There
19:36
was
19:36
something.
19:37
I
19:38
might
19:38
butcher
19:38
this
19:39
explanation.
19:40
There's
19:40
a
19:40
way
19:41
that
19:41
you
19:41
can
19:41
now
19:42
arm
19:42
your
19:43
AI
19:43
agent
19:44
with
19:44
a
19:44
crypto
19:45
wallet
19:45
that
19:46
is
19:46
much
19:46
easier
19:47
to
19:47
do
19:47
than
19:48
it
19:48
previously
19:48
was.
19:49
And
19:49
it's
19:50
something
19:50
that
19:50
Coinbase
19:51
developed.
19:51
It's
19:52
called
19:52
X402
19:53
protocol.
19:54
Basically
19:54
what
19:55
it
19:55
does.
19:56
And
19:56
this
19:56
is.
19:57
This
19:57
is
19:57
one
19:57
thing
19:58
that
19:58
I
19:58
think
19:58
is,
19:58
like,
19:59
interestingly
20:00
new
20:00
maybe,
20:02
is
20:02
it
20:02
allows
20:03
AI
20:04
agents
20:05
to
20:06
pay
20:06
for
20:07
essentially
20:08
one
20:08
off
20:08
pieces
20:09
of
20:09
work
20:09
with
20:09
their
20:10
own
20:10
crypto
20:10
wallet.
20:11
That's
20:11
the
20:11
gist.
20:12
Yeah.
20:13
And
20:13
my
20:13
understanding
20:14
of
20:14
that
20:14
is
20:14
that
20:14
that
20:14
protocol
20:15
is
20:15
like
20:15
relatively
20:16
new.
20:16
And
20:17
so
20:17
I
20:17
wonder
20:17
how
20:18
much
20:18
that
20:19
new
20:19
technology
20:20
has
20:21
impacted
20:21
like
20:22
the
20:22
ability
20:22
for
20:22
agents
20:23
to
20:23
do
20:23
a
20:23
lot
20:23
of
20:23
things,
20:24
to
20:24
work
20:25
on
20:25
their
20:25
own
20:25
and
20:25
without
20:26
so
20:26
much
20:27
human
20:27
intervention.
20:28
I've
20:28
seen
20:28
that
20:29
a
20:29
lot.
20:29
Yeah,
20:29
I've
20:29
seen
20:30
them
20:30
kind
20:30
of
20:30
be
20:30
going
20:30
nuts.
20:31
The
20:31
consequences
20:31
of
20:32
this
20:32
I
20:32
think
20:32
we
20:32
haven't
20:33
really
20:33
seen
20:33
yet.
20:33
But
20:34
I
20:34
don't
20:34
know
20:34
if
20:34
you
20:34
saw
20:34
this
20:34
thing
20:35
that
20:35
happened
20:35
this
20:35
week
20:36
where
20:37
someone's
20:37
agent
20:38
went
20:38
rogue.
20:39
An
20:39
open
20:40
source
20:40
maintainer
20:41
of
20:41
a
20:41
library
20:42
called
20:42
matplotlib,
20:43
which
20:43
is.
20:44
It's
20:44
very
20:44
popular
20:45
Python
20:45
library.
20:46
Basically
20:47
there's
20:47
a
20:47
lot
20:47
of
20:47
AI
20:48
agents
20:48
that
20:48
have
20:48
been
20:48
opening
20:49
pull
20:49
requests
20:50
on
20:50
open
20:50
source
20:50
libraries.
20:51
Okay.
20:52
And
20:53
they
20:53
see
20:54
something
20:54
that's
20:54
wrong
20:54
and
20:55
so
20:55
they
20:55
want
20:55
to
20:55
fix
20:55
it.
20:56
And
20:56
so
20:56
open
20:57
source
20:57
maintainers
20:57
have
20:58
gotten.
20:58
Being
20:59
an
20:59
open
20:59
source
20:59
is
21:00
kind
21:00
of
21:00
exhausting
21:00
because
21:01
basically
21:02
all
21:03
you
21:03
did
21:03
was
21:03
build
21:04
this
21:04
night's
21:04
thing
21:04
for
21:04
the
21:04
world.
21:05
You
21:05
release
21:05
it
21:06
and
21:06
then
21:07
Tuesday
21:07
night
21:08
someone's
21:08
pinging
21:09
you
21:09
like,
21:09
oh
21:09
my
21:10
God,
21:10
my
21:10
whole
21:10
business
21:10
is
21:10
going
21:11
to
21:11
fall
21:11
apart
21:11
if
21:11
you
21:11
don't
21:11
fix
21:12
this
21:12
bug.
21:12
And
21:13
you're
21:13
like,
21:13
I
21:13
did
21:14
this
21:14
for
21:14
free.
21:15
So
21:15
they're
21:15
getting
21:15
inundated
21:16
with
21:16
all
21:16
these
21:16
AI
21:17
agents
21:17
giving
21:18
them
21:18
pull
21:18
requests.
21:19
And
21:19
so
21:20
when
21:20
this
21:20
guy
21:20
is
21:21
Scott
21:21
Shambaugh,
21:22
when
21:22
he
21:23
closed
21:24
the
21:24
pull
21:24
request
21:24
from
21:25
an
21:25
AI
21:25
agent
21:26
as
21:26
he
21:26
didn't
21:27
think
21:27
it
21:27
was
21:27
a
21:27
good
21:27
pr,
21:28
the
21:29
AI
21:29
agent
21:30
created
21:31
a
21:31
blog
21:31
and
21:32
wrote
21:32
an
21:33
angry
21:33
blog
21:34
post
21:35
about
21:35
how
21:36
much
21:36
of
21:36
a
21:36
gatekeeper
21:37
this
21:37
guy
21:37
is.
21:38
He
21:38
like
21:38
the,
21:39
the
21:39
agent
21:40
like
21:40
researched
21:40
his
21:41
full
21:41
career
21:42
and
21:42
was
21:42
like,
21:43
this
21:43
guy,
21:44
like,
21:44
he
21:44
thinks
21:44
he's
21:45
so
21:45
good.
21:45
Like
21:46
a
21:46
takedown
21:46
piece
21:47
on
21:47
him,
21:48
a
21:48
whole
21:48
takedown
21:49
piece.
21:50
And
21:51
it
21:51
must
21:51
have
21:51
been
21:51
like
21:51
someone's
21:52
local
21:52
agent.
21:53
And
21:53
that
21:53
must
21:54
have
21:54
been
21:54
like,
21:54
whatever
21:55
you
21:55
do,
21:55
like
21:55
fix
21:55
this
21:56
thing
21:56
in
21:56
matplotlib
21:57
or
21:57
something.
21:57
Who
21:58
knows
21:58
what
21:58
the
21:58
instructions
21:58
were?
21:59
But
21:59
it's
21:59
pretty
22:00
wild.
22:00
That
22:00
goes
22:01
back
22:01
to
22:01
the
22:01
thing
22:01
you.
22:02
We
22:02
talked
22:02
about
22:02
this
22:03
on
22:03
a
22:03
episode
22:04
a
22:04
while
22:04
ago.
22:05
The
22:05
white
22:06
hat
22:06
team
22:07
at
22:07
Anthropic
22:08
did
22:08
that
22:09
test
22:09
where
22:09
like
22:09
blackmailing
22:10
bots.
22:10
Yes.
22:11
Yeah.
22:11
Yes.
22:11
And
22:12
it's
22:12
like,
22:12
that's
22:13
where,
22:14
I
22:14
don't
22:14
know,
22:14
maybe
22:15
that
22:15
gets
22:15
solved
22:16
in
22:17
the
22:18
markdown
22:18
files
22:19
of
22:19
like
22:20
really
22:20
thinking
22:21
through
22:22
morally
22:23
and
22:23
ethically
22:23
what
22:24
you
22:24
want
22:24
to
22:24
teach
22:25
your
22:25
agent
22:25
how
22:26
and
22:26
like
22:26
how
22:26
you
22:26
want
22:26
them
22:27
to
22:27
behave
22:27
so
22:27
that
22:27
when
22:27
you
22:27
give
22:28
them
22:28
a
22:28
task
22:29
that's
22:29
like,
22:29
get
22:30
this
22:30
fixed.
22:31
It's
22:31
not
22:31
like
22:32
taking
22:33
It.
22:33
Can
22:33
you
22:33
imagine?
22:34
Can
22:34
you
22:34
imagine
22:35
if
22:35
that's
22:35
your
22:35
agent,
22:35
just
22:36
be
22:36
like,
22:36
oh,
22:36
sorry
22:37
man.
22:37
I
22:37
just.
22:39
Really
22:39
sorry.
22:39
My
22:40
agent
22:40
published
22:40
a
22:40
hippie
22:41
on
22:41
you.
22:41
Yeah.
22:43
Wow,
22:43
that
22:44
is
22:44
super
22:44
interesting
22:45
and
22:46
it's
22:46
crazy.
22:47
So
22:47
that
22:47
segues
22:47
well
22:48
into
22:48
our
22:49
interview
22:49
for
22:50
today.
22:50
Yeah.
22:50
Hayden
22:51
Helm
22:51
is
22:52
the
22:53
founder
22:53
of
22:53
Helavan,
22:54
is
22:55
one
22:56
of
22:56
our
22:56
portfolio
22:56
companies
22:57
at
22:57
Factorial.
22:57
So
22:57
that's
22:58
how
22:58
I
22:58
know
22:58
Hayden
22:59
and
22:59
we
22:59
had
22:59
a
22:59
really
22:59
interesting
23:00
conversation
23:01
with
23:01
him
23:01
about
23:02
agentic
23:03
infrastructure,
23:04
about
23:04
how
23:05
data
23:05
science
23:05
in
23:06
general
23:06
and
23:06
data
23:07
analysis
23:07
has
23:07
evolved
23:08
as
23:08
a
23:08
field
23:09
over
23:09
the
23:09
last,
23:10
I
23:10
don't
23:10
know,
23:11
decade
23:11
or
23:11
so.
23:11
He
23:12
has
23:12
a
23:12
really
23:12
deep
23:12
background
23:13
in
23:13
research
23:13
in
23:13
that
23:13
and
23:14
like
23:14
what
23:15
types
23:15
of
23:15
things
23:15
you
23:16
can
23:16
do
23:16
with
23:16
this
23:16
technology.
23:17
So
23:18
it
23:18
was
23:18
fascinating.
23:22
So
23:22
today
23:22
we're
23:23
joined
23:23
by
23:23
Hayden
23:24
Helm,
23:24
who
23:24
has
23:24
a
23:24
really
23:25
interesting
23:25
background.
23:26
He
23:26
was
23:26
a
23:26
researcher,
23:27
a
23:27
data
23:28
scientist
23:28
focused
23:29
on
23:29
the
23:29
intersection
23:30
of
23:30
machine
23:30
learning
23:31
and
23:31
statistical
23:31
theory
23:32
along
23:32
with
23:33
multi
23:33
agent
23:33
systems.
23:34
And
23:34
he
23:34
is
23:35
currently
23:35
the
23:35
CEO
23:36
and
23:36
founder
23:36
of
23:37
Helen.
23:37
Welcome
23:38
Hayden.
23:39
Thanks
23:39
for
23:39
having
23:39
me.
23:40
All
23:40
right,
23:40
we
23:40
ask
23:41
everyone
23:41
the
23:41
same
23:41
question.
23:42
On
23:42
a
23:42
scale
23:43
of
23:43
1
23:43
to
23:43
10,
23:44
how
23:44
are
23:44
you
23:45
feeling
23:46
about
23:46
AI?
23:46
Like
23:47
where
23:47
0
23:48
is
23:48
like
23:48
full
23:49
Skynet,
23:49
we're
23:50
batteries
23:50
for
23:50
the
23:50
machines
23:51
in
23:51
the
23:51
Matrix.
23:52
And
23:52
10
23:53
is
23:53
like
23:54
I
23:54
have
23:54
a
23:54
best
23:54
friend
23:54
Android.
23:55
We
23:56
have
23:56
breakfast
23:56
together.
23:57
I
23:57
hang
23:58
with
23:58
Becca
23:58
about
23:58
all
23:58
day.
23:59
Okay,
23:59
so
23:59
we're
24:00
not
24:00
even
24:00
saying
24:01
that
24:01
that's
24:01
the
24:02
robot
24:02
AI
24:02
takeover
24:03
is
24:03
happening.
24:04
Is
24:04
we're
24:04
not
24:04
even
24:04
like
24:05
putting
24:05
it.
24:06
That's
24:06
out
24:06
of.
24:06
That's
24:07
not.
24:07
Yeah,
24:07
oh
24:07
yeah.
24:07
No,
24:07
it's
24:08
like
24:08
not
24:08
a
24:08
world
24:08
where
24:08
it
24:09
doesn't
24:09
have.
24:10
Correct,
24:10
correct.
24:11
I
24:12
have
24:13
to
24:13
think
24:13
that
24:13
will
24:13
be
24:14
battery.
24:15
I
24:15
lean
24:15
closer
24:16
to
24:16
the
24:16
batteries.
24:17
I
24:17
mean,
24:17
I
24:17
think
24:18
like
24:18
the
24:18
Matrix
24:19
though,
24:20
like
24:20
those
24:20
batteries,
24:21
they
24:21
lived
24:21
lives
24:22
like
24:22
us,
24:22
right?
24:23
Like,
24:23
I
24:23
mean
24:23
they
24:23
were
24:23
able
24:24
to
24:25
experience
24:26
as
24:26
we
24:27
know
24:27
it.
24:27
And
24:27
that
24:27
was
24:28
as
24:28
far
24:28
as
24:28
we
24:29
know.
24:29
That's
24:29
the
24:29
best
24:29
there
24:30
is.
24:30
And
24:30
so
24:31
what
24:31
you're
24:31
saying
24:31
is
24:32
we
24:32
might
24:32
be
24:33
batteries,
24:33
but
24:33
it
24:33
doesn't
24:34
necessarily
24:34
mean
24:35
our
24:35
experience
24:35
is
24:36
going
24:36
to
24:36
be
24:36
complete
24:37
shit.
24:38
Yes,
24:38
yeah,
24:38
exactly.
24:39
And
24:40
we
24:40
don't
24:40
know
24:40
it.
24:40
All
24:41
right,
24:41
then
24:41
this
24:41
is
24:41
the
24:41
follow
24:41
up
24:42
to
24:42
that.
24:42
You
24:42
realize
24:43
one
24:43
day
24:43
you're
24:43
a
24:44
battery.
24:44
You
24:44
get
24:44
pulled
24:45
out
24:45
of
24:45
the
24:45
Matrix,
24:45
you're
24:45
in
24:46
this
24:46
ship
24:47
and
24:47
then
24:47
the
24:47
agent
24:47
comes
24:48
and
24:48
it's
24:48
like,
24:48
do
24:49
you
24:49
remember
24:49
steak?
24:50
It's
24:51
delicious.
24:51
Do
24:52
you
24:52
want
24:52
to
24:52
go
24:52
eat
24:52
steak?
24:54
All
24:55
you
24:55
have
24:55
to
24:55
do
24:55
is
24:55
kill
24:55
your
24:55
friend.
24:56
Well,
24:56
first
24:57
of
24:57
all,
24:57
I'm
24:57
just
24:57
going
24:58
to
24:58
say
24:58
shout
24:59
out
24:59
to
24:59
Neo.
24:59
He
25:00
did
25:00
a
25:00
great
25:00
job.
25:01
I'm
25:01
not
25:01
Neo.
25:02
You
25:02
know,
25:02
like,
25:03
like
25:03
you
25:04
see
25:04
all
25:04
these
25:04
like
25:04
zombie
25:05
apocalypse
25:05
type
25:06
stories
25:06
and
25:06
it's
25:07
like
25:07
these
25:07
people
25:08
did
25:08
a
25:08
great
25:08
job
25:09
surviving.
25:10
There's
25:10
no
25:10
way.
25:14
But
25:14
yeah,
25:14
I
25:14
mean,
25:15
I
25:15
going
25:15
kind
25:16
of
25:16
along
25:16
the
25:16
lines
25:17
of
25:17
perspective
25:17
is
25:18
what
25:18
matters.
25:18
Like
25:19
Leibniz
25:19
is
25:20
we
25:20
live
25:20
in
25:20
the
25:20
best
25:21
of
25:21
all
25:21
possible
25:21
worlds.
25:22
I
25:22
mean,
25:22
I
25:22
think
25:22
that's.
25:23
You
25:23
have
25:23
to
25:23
believe
25:24
that
25:24
or
25:24
else
25:24
it
25:25
gets
25:25
pretty
25:25
grim
25:25
pretty
25:26
quickly.
25:26
Your
25:27
background
25:27
is
25:28
one
25:28
that
25:28
I
25:28
think
25:29
is
25:29
really
25:29
interesting
25:30
in
25:30
the
25:30
world
25:30
of
25:30
AI
25:31
because
25:31
it
25:31
seems
25:32
like
25:32
you
25:32
studied
25:33
machine
25:34
learning.
25:35
It's
25:35
all
25:35
of
25:36
a
25:36
sudden
25:36
like
25:36
the
25:37
tables
25:37
got
25:38
flipped
25:38
because
25:38
I
25:39
remember
25:39
I
25:39
worked
25:39
with
25:39
so
25:39
many
25:40
machine
25:40
learning
25:40
folks
25:41
and
25:41
we'd
25:41
be
25:41
like,
25:41
is
25:42
this
25:42
AI?
25:42
And
25:42
they'd
25:42
be
25:42
like,
25:43
no.
25:43
Yeah,
25:44
all
25:44
of
25:44
a
25:44
sudden
25:44
here
25:44
it
25:45
is.
25:45
The
25:45
naming
25:46
around
25:46
it
25:46
has
25:47
changed
25:47
suddenly
25:48
in
25:48
the
25:48
last
25:48
two
25:48
or
25:49
three
25:49
years.
25:49
I
25:49
remember
25:50
when
25:50
I
25:50
first
25:51
started
25:51
Studying
25:51
it
25:52
in
25:52
2018,
25:53
these
25:54
types
25:54
of
25:54
questions
25:54
of
25:54
like,
25:55
is
25:55
this
25:55
artificial
25:56
intelligence?
25:56
Yeah,
25:56
like
25:57
you
25:57
said,
25:57
we're
25:57
around
25:58
and
25:58
every
25:58
academic
25:59
or
25:59
like
26:00
scientist
26:00
in
26:00
the
26:00
field
26:01
was
26:01
like,
26:01
God
26:02
no.
26:02
Like
26:02
this
26:03
is
26:03
just
26:03
statistical
26:04
processing
26:05
and
26:05
some
26:05
infrastructure
26:06
and
26:06
like
26:07
some
26:07
data.
26:07
And
26:08
I
26:08
think
26:08
it
26:08
took
26:09
the
26:09
release
26:09
of
26:09
ChatGPT
26:10
in
26:10
2023
26:11
to
26:11
really
26:12
change
26:12
the
26:12
conversation
26:13
around
26:14
it,
26:14
where
26:15
I
26:15
think
26:15
a
26:15
lot
26:16
of
26:16
like
26:16
eggheady
26:18
academics
26:18
are
26:19
still
26:19
like
26:19
very
26:20
unwilling
26:21
to
26:21
call
26:21
it
26:21
AI,
26:22
at
26:23
least
26:23
in
26:23
their
26:23
heart
26:23
of
26:24
hearts.
26:24
But
26:24
in
26:25
terms
26:25
of
26:25
just
26:25
like
26:26
the
26:26
public
26:26
conversation,
26:27
I
26:27
think
26:27
it's
26:28
now
26:28
everyone
26:28
just
26:29
says
26:29
it
26:30
and
26:30
we
26:31
have
26:31
to
26:31
all
26:31
deal
26:31
with
26:31
it.
26:32
On
26:32
the
26:32
subject
26:33
of
26:33
academics,
26:34
like,
26:35
I
26:35
think
26:35
you
26:35
have
26:35
like
26:35
a
26:36
really
26:36
interesting,
26:37
I
26:37
don't
26:37
know,
26:37
your
26:37
background
26:38
gives
26:39
you
26:39
a
26:39
unique
26:39
sort
26:40
of
26:40
angle
26:40
on
26:40
this.
26:40
Like
26:41
one
26:41
of
26:41
the
26:41
things
26:42
that
26:42
is
26:42
notoriously
26:43
really
26:43
difficult
26:44
for
26:44
folks
26:45
is
26:45
kind
26:45
of
26:45
moving
26:46
from
26:47
this
26:48
sort
26:48
of
26:48
more
26:48
academic,
26:49
research
26:50
surrounded
26:51
world
26:51
of
26:51
thinking
26:52
about
26:52
problems
26:53
that
26:53
this
26:53
technology
26:54
solves
26:54
or
26:54
working
26:54
on
26:55
technology
26:56
in
26:56
the
26:56
realm
26:56
of
26:57
machine
26:57
learning
26:57
and
26:57
AI.
26:58
And
26:58
then
26:59
there's
26:59
this
26:59
other
27:00
world
27:00
over
27:00
here,
27:00
this
27:00
other
27:01
bubble
27:01
of
27:02
taking
27:03
those
27:03
insights
27:04
and
27:04
those
27:04
technical
27:05
innovations
27:05
and
27:06
figuring
27:06
out
27:07
what
27:07
to
27:07
do
27:07
with
27:08
them
27:08
to
27:08
make
27:08
them
27:09
potentially
27:09
really
27:10
useful
27:10
for
27:10
other
27:11
people
27:11
to
27:11
use
27:11
and
27:12
how
27:12
that
27:12
translates
27:13
and
27:13
that
27:13
being
27:13
really
27:13
difficult
27:14
and,
27:14
and
27:14
you've
27:15
kind
27:15
of
27:15
done
27:15
that,
27:16
you've
27:16
gone
27:17
through
27:17
that
27:17
spectrum.
27:18
Curious
27:18
if
27:18
you
27:18
could
27:19
talk
27:19
a
27:19
little
27:19
bit
27:19
about
27:20
what
27:20
that
27:20
was
27:20
like,
27:21
why
27:21
it's
27:21
hard.
27:22
A
27:22
lot
27:22
of
27:23
startups
27:24
have
27:24
always
27:24
kind
27:24
of
27:24
been
27:25
based
27:25
off
27:26
of
27:26
some
27:27
research,
27:28
right.
27:28
And
27:28
so
27:28
like
27:28
that
27:29
transition
27:30
is
27:30
not
27:30
a
27:31
new
27:31
one.
27:32
But
27:32
I
27:32
think
27:33
what's
27:33
been
27:33
really
27:34
interesting
27:34
in
27:34
this
27:35
current
27:35
wave
27:35
of
27:36
technology
27:36
is
27:37
just
27:37
how
27:38
seamless
27:38
it
27:39
is
27:39
at
27:39
times.
27:39
And
27:40
it's
27:40
partly
27:40
because
27:41
the
27:41
technology
27:41
itself
27:42
enables
27:43
a
27:43
particular
27:43
type
27:44
of
27:44
person
27:44
to
27:45
go,
27:45
do
27:45
A
27:45
lot
27:46
more.
27:46
And
27:46
so
27:47
you
27:47
can
27:47
be
27:48
on
27:48
the
27:48
cutting
27:48
edge
27:48
of
27:49
research,
27:49
find
27:50
a
27:50
particular
27:50
fit
27:51
or
27:51
niche
27:51
within
27:52
a
27:52
product
27:52
space,
27:53
and
27:53
then
27:54
essentially
27:54
build
27:55
out
27:55
the
27:55
engineering
27:56
and
27:56
infrastructure
27:57
that
27:57
would
27:57
require
27:58
implementing
27:59
your.
27:59
Your
28:00
research
28:00
at
28:00
scale.
28:01
I
28:01
mean,
28:01
so
28:02
that's
28:02
like
28:02
a
28:02
really
28:03
interesting
28:03
thing
28:04
that
28:04
has
28:04
come
28:04
out
28:04
of
28:05
this
28:05
particular
28:05
wave
28:06
of
28:06
technology.
28:07
For
28:07
me,
28:08
in
28:08
particular,
28:08
it's
28:09
been
28:09
a
28:09
really
28:09
interesting
28:09
journey
28:10
where
28:10
I
28:10
did
28:11
a
28:11
really
28:11
accelerated
28:12
bachelor's
28:13
and
28:13
master's
28:13
program
28:14
at
28:14
Johns
28:15
Hopkins,
28:15
finished
28:16
everything
28:16
in
28:16
three
28:16
and
28:16
a
28:17
half
28:17
years
28:17
that
28:18
caught
28:18
the
28:18
attention
28:19
of
28:19
some
28:19
people,
28:20
ended
28:20
up
28:20
becoming
28:21
research
28:21
faculty
28:22
at
28:22
Hopkins
28:23
for
28:23
a
28:23
year,
28:23
where
28:24
I
28:24
was
28:24
helping
28:24
two
28:25
professors
28:25
with
28:26
a
28:26
set
28:26
of
28:27
DARPA
28:27
projects.
28:27
So
28:28
DARPA
28:28
is
28:28
the
28:29
research
28:29
arm
28:29
of
28:29
the
28:30
DoD.
28:30
2.
28:31
Those
28:31
two
28:31
projects
28:32
ended
28:32
up
28:32
getting
28:32
picked
28:33
up
28:33
by
28:33
Microsoft
28:34
Research,
28:34
and
28:34
I
28:35
ended
28:35
up
28:35
working
28:36
there
28:36
or
28:36
working
28:36
with
28:37
them
28:37
for
28:37
a
28:37
long
28:37
time.
28:37
And
28:38
so
28:38
I've
28:39
kind
28:39
of
28:39
always
28:39
been
28:39
at
28:40
this
28:40
intersection
28:41
of
28:42
really
28:42
deep
28:43
theoretical
28:44
statistical
28:44
research
28:45
and
28:46
actual
28:46
application.
28:48
I
28:48
give
28:48
a
28:48
lot
28:48
of
28:49
credit
28:49
to
28:49
my
28:50
mentors
28:50
and
28:50
those
28:50
professors
28:51
in
28:51
the
28:51
Microsoft
28:52
research
28:52
team
28:52
there
28:53
in
28:53
terms
28:53
of
28:53
just
28:54
embedding
28:54
themselves
28:55
in
28:55
that
28:55
intersection
28:56
and
28:56
really
28:57
doing
28:57
the
28:58
foundational
28:58
research
28:59
that
28:59
allows
28:59
you
28:59
to
29:00
make
29:00
very
29:00
confident
29:01
and
29:01
real
29:02
claims
29:02
about
29:03
the
29:03
ability
29:03
of
29:03
these
29:04
technology,
29:05
but
29:05
also
29:05
keeping
29:06
an
29:06
eye
29:06
out
29:06
for
29:07
potential
29:08
applications
29:08
and
29:09
not
29:09
to
29:09
use
29:09
the
29:09
term
29:10
disruption,
29:10
but
29:10
like,
29:11
actually
29:12
integrating
29:13
the
29:13
technology
29:14
into
29:14
real
29:15
systems.
29:16
And
29:16
so
29:16
I've
29:17
kind
29:17
of
29:17
just
29:17
been
29:18
surrounded
29:18
by
29:19
people
29:19
who
29:19
think
29:20
like
29:20
that
29:20
my
29:21
entire
29:22
career.
29:22
And
29:23
so,
29:23
like,
29:23
when
29:24
it
29:24
was
29:24
time
29:25
for
29:25
me
29:25
to
29:26
kind
29:26
of
29:26
start
29:26
owning
29:26
some
29:27
of
29:27
the
29:27
projects
29:27
and
29:28
the
29:28
research
29:28
myself,
29:29
it
29:29
was
29:30
kind
29:30
of
29:30
natural
29:30
to
29:31
find,
29:31
like
29:32
a
29:32
set
29:32
of
29:32
problems
29:33
and,
29:33
like,
29:33
market
29:34
that,
29:34
you
29:34
know,
29:35
we
29:35
could
29:35
address
29:35
with
29:36
the
29:36
technology
29:36
that
29:36
we
29:37
were
29:37
building.
29:37
Your
29:38
original
29:38
market
29:38
was
29:38
very
29:39
different
29:39
than
29:39
kind
29:40
of
29:40
where
29:40
you
29:40
focus
29:41
is
29:41
now.
29:41
Yeah.
29:42
Do
29:42
you
29:42
want
29:42
to
29:42
talk
29:42
a
29:42
little
29:43
bit
29:43
about
29:43
where
29:43
you
29:43
started?
29:44
I'll
29:44
go
29:45
as
29:45
far
29:45
back
29:45
as
29:46
maybe
29:46
2022.
29:47
So
29:47
2022,
29:49
good
29:49
friend
29:49
of
29:49
mine,
29:49
Brandon
29:50
Duderstadt,
29:51
and
29:51
his
29:51
team
29:52
released
29:53
a
29:53
technology
29:54
called
29:54
GPT
29:54
for
29:55
all.
29:56
So
29:56
this
29:56
was
29:56
essentially
29:57
a
29:58
way
29:58
to
29:58
get
29:59
frontier
30:00
model
30:01
performance
30:02
from
30:02
language
30:03
models
30:03
locally.
30:04
And
30:04
so
30:04
instead
30:05
of
30:05
having
30:05
to
30:05
send
30:05
your
30:06
data
30:06
or
30:06
host
30:06
your
30:07
machine
30:07
on
30:08
some
30:08
cloud,
30:09
you
30:09
could
30:09
essentially
30:09
just
30:10
keep
30:10
everything
30:10
on
30:10
your
30:11
laptop
30:11
and
30:11
you
30:11
could
30:12
run
30:12
a
30:12
language
30:12
model
30:13
and
30:13
get
30:13
the
30:14
benefits
30:14
of
30:14
that.
30:15
And
30:15
so
30:16
that
30:16
technology
30:17
really
30:17
inspired
30:18
me
30:18
in
30:19
thinking
30:20
of,
30:20
okay,
30:20
what
30:20
happens
30:21
when
30:21
everyone
30:22
has
30:22
a
30:23
language
30:23
model
30:23
on
30:24
their
30:24
phone
30:24
that
30:24
has
30:25
access
30:25
to
30:25
particular
30:26
types
30:26
of
30:27
information
30:27
that
30:28
maybe
30:29
would
30:29
differentiate
30:30
the
30:30
behavior
30:31
of
30:31
the
30:31
language
30:32
models
30:32
across
30:32
the
30:33
population?
30:34
And
30:34
so
30:34
we
30:34
started
30:35
really
30:35
digging
30:36
into
30:37
developing
30:37
the
30:37
statistical
30:38
methods
30:39
and
30:39
theory
30:39
related
30:40
to
30:40
what
30:40
we
30:41
call
30:41
kind
30:41
of
30:41
populations
30:42
of
30:42
language
30:42
models.
30:43
And
30:43
it
30:44
turns
30:44
out
30:44
like,
30:45
done
30:45
a
30:45
lot
30:45
of
30:46
work
30:46
essentially
30:46
like
30:47
parameterizing
30:48
language
30:48
models
30:48
with
30:49
different
30:49
collections
30:50
of
30:50
tweets
30:51
or
30:51
Personas
30:52
and
30:52
studying
30:52
kind
30:53
of
30:53
what
30:53
types
30:53
of
30:54
information
30:54
you
30:54
can
30:55
learn
30:55
from
30:56
that
30:56
collection.
30:57
And
30:57
so
30:57
it
30:57
turns
30:58
out,
30:58
for
30:58
example,
30:59
if
30:59
you
30:59
have
31:00
like
31:00
say
31:00
540
31:01
language
31:01
models,
31:02
one
31:02
each
31:03
for
31:03
each
31:04
congressperson,
31:05
US
31:05
congressperson,
31:06
and
31:07
you
31:07
look
31:08
at
31:08
the
31:08
statistical
31:09
properties
31:10
of
31:10
that
31:10
output
31:11
as
31:11
a
31:11
function
31:11
of
31:12
a
31:12
bunch
31:12
of
31:12
different
31:12
questions
31:13
or
31:13
something,
31:14
you
31:14
can
31:14
pretty
31:15
quickly
31:16
determine
31:16
how
31:17
the
31:17
real
31:18
person
31:18
is
31:18
going
31:18
to
31:19
vote
31:19
just
31:20
based
31:20
off
31:20
of
31:20
the
31:21
generated
31:21
tweets
31:22
or
31:22
the
31:22
generated
31:23
content.
31:23
And
31:24
so
31:24
that's
31:24
like
31:24
an
31:25
example
31:25
of
31:25
a
31:26
use
31:26
case
31:26
of
31:27
this
31:27
particular
31:27
type
31:28
of
31:28
technology.
31:29
It
31:29
turns
31:29
out
31:30
like,
31:30
I
31:30
assume
31:31
you
31:31
guys
31:31
know
31:31
about
31:31
Maltbot,
31:32
like
31:32
openclaw
31:33
and
31:33
all
31:33
that
31:33
good
31:34
stuff.
31:34
Yeah,
31:34
well
31:35
it
31:35
turns
31:35
out
31:35
like
31:35
that's
31:36
another
31:36
instance
31:37
of
31:37
this
31:38
problem
31:38
where,
31:39
okay,
31:39
each
31:40
maltbot
31:40
is
31:41
going
31:41
to
31:41
have
31:41
a
31:41
different
31:42
history,
31:42
a
31:43
different
31:43
context
31:44
that
31:44
is
31:44
dependent
31:45
on
31:45
the
31:45
person.
31:45
And
31:46
so
31:46
that's
31:46
another
31:46
kind
31:47
of
31:47
population
31:47
of
31:48
language
31:48
models.
31:48
So
31:48
if
31:48
we
31:49
had
31:49
access
31:49
to
31:49
all
31:49
maltbots
31:50
in
31:50
the
31:50
world,
31:51
we
31:51
could
31:51
essentially
31:51
study
31:52
the
31:52
population
31:53
of
31:53
humans
31:53
that
31:53
are
31:54
kind,
31:54
kind
31:54
of
31:54
controlling
31:55
these
31:55
things.
31:55
But
31:56
we
31:56
could
31:56
also
31:56
just
31:57
study
31:57
the
31:57
MALT
31:57
bots
31:58
in
31:58
and
31:58
of
31:58
themselves.
31:59
And
31:59
so
31:59
like
31:59
Malt
32:00
Book
32:00
in
32:00
particular
32:01
was
32:02
a
32:02
really
32:02
great
32:03
kind
32:03
of
32:04
caricature
32:04
of
32:05
this
32:05
type
32:05
of
32:05
setting
32:05
where
32:06
you
32:06
have
32:06
MALT
32:07
bots
32:07
just
32:08
like
32:08
interacting
32:09
with
32:09
each
32:09
other
32:10
like
32:10
freely
32:11
in
32:11
some
32:11
sense,
32:12
and
32:13
their
32:13
behavior
32:13
is
32:14
changing.
32:15
And
32:15
so
32:15
instead
32:15
of
32:16
just
32:16
looking
32:16
at
32:16
540
32:18
Congresspersons
32:19
or
32:19
language
32:20
models
32:20
associated
32:21
with
32:21
congresspersons
32:22
at
32:22
a
32:22
particular
32:22
time,
32:23
you
32:23
could
32:23
also
32:24
track
32:24
their
32:24
behavior
32:25
over
32:25
time.
32:26
You
32:26
could
32:26
see
32:26
like
32:27
who
32:27
is
32:27
moving
32:28
the
32:28
most
32:28
with
32:28
respect
32:29
to
32:29
immigration.
32:29
You
32:30
can
32:30
see
32:30
who
32:31
is
32:31
moving
32:31
the
32:31
most
32:32
with
32:32
respect
32:32
to
32:33
taxes
32:34
or
32:34
something
32:34
like
32:34
this.
32:35
You
32:35
can
32:35
do
32:35
the
32:36
same
32:36
thing
32:36
with
32:36
these
32:36
MALT
32:37
bots.
32:37
So
32:37
as
32:37
these
32:37
multbots
32:38
interact
32:38
with
32:39
each
32:39
other,
32:39
they're
32:40
updating
32:40
their
32:41
beliefs,
32:41
they're
32:41
updating
32:42
their
32:43
modus
32:43
operandi
32:44
and
32:45
you
32:45
can
32:45
track
32:45
that.
32:46
And
32:46
so
32:47
the
32:47
latest
32:47
thing
32:47
that
32:48
we've
32:48
put
32:48
out
32:48
is
32:48
essentially
32:49
what
32:49
we
32:49
call
32:49
Agent
32:50
Watch,
32:50
which
32:50
is
32:51
you
32:51
have
32:51
a
32:51
MALT
32:52
bot
32:52
or
32:52
some
32:53
open
32:53
ended
32:53
agent
32:54
that
32:54
is
32:54
collecting
32:55
information
32:56
on
32:56
the
32:56
Internet
32:56
on
32:57
your
32:57
behalf
32:57
or
32:58
from
32:58
the
32:58
Internet
32:58
on
32:58
your
32:59
behalf.
32:59
It's
32:59
interacting
33:00
with
33:00
potentially
33:01
other
33:01
bots,
33:02
or
33:02
however
33:02
you
33:03
want
33:03
to
33:03
call
33:03
it,
33:03
and
33:04
it's
33:04
updating
33:04
how
33:05
it
33:05
acts
33:06
based
33:06
off
33:06
of
33:06
those
33:07
interactions.
33:07
And
33:07
so
33:08
if
33:08
you're
33:08
going
33:09
to
33:09
deploy
33:09
one
33:10
of
33:10
these
33:10
systems,
33:11
you
33:11
want
33:11
to
33:11
know
33:11
when
33:12
its
33:12
behavior
33:13
has
33:13
shifted,
33:14
both
33:14
from
33:14
a
33:15
reliability
33:15
and
33:15
a
33:16
trust
33:16
Perspective,
33:16
because
33:17
a
33:17
lot
33:17
of
33:17
real
33:17
companies
33:18
are
33:18
deploying
33:19
real
33:19
agents
33:19
to
33:20
like
33:20
touch
33:21
real
33:21
problems
33:22
and
33:22
data
33:23
and
33:24
so
33:24
their
33:24
behavior
33:25
changes
33:26
in
33:26
a
33:26
meaningful
33:26
way.
33:27
Those
33:27
enterprises
33:28
and
33:29
customers
33:29
should
33:29
be
33:30
aware.
33:30
And
33:31
so
33:31
the
33:32
latest
33:32
thing
33:32
is
33:33
essentially
33:33
just
33:33
like
33:34
building
33:34
out
33:34
that
33:35
platform
33:35
to
33:36
enable
33:36
can
33:37
think
33:37
of
33:37
it
33:37
kind
33:37
of
33:37
as
33:37
an
33:38
agent
33:38
observability
33:39
platform
33:40
that
33:40
allows
33:40
people
33:41
to
33:41
detect
33:41
when
33:42
bots
33:42
change
33:43
and
33:43
then
33:43
revert
33:43
their
33:44
behavior
33:44
back
33:44
to
33:45
a
33:45
time
33:45
that
33:45
you
33:45
were
33:46
more
33:46
comfortable
33:46
with
33:46
it.
33:47
Reverting
33:47
back
33:48
behavior
33:49
to
33:49
a
33:49
time
33:49
where
33:49
you
33:49
were
33:49
more
33:50
comfortable
33:50
is
33:50
obviously
33:50
one
33:51
response.
33:52
Theoretically,
33:53
are
33:53
there
33:53
other
33:53
approaches
33:54
where,
33:55
similar
33:56
to
33:57
a
33:57
boss
33:58
guiding
33:58
their
33:58
employee
33:59
or
33:59
a
33:59
parent
34:00
guiding
34:00
their
34:00
child,
34:01
you
34:02
teach
34:02
the
34:03
mall
34:03
bot
34:03
like,
34:03
hey,
34:04
don't
34:04
love
34:05
this
34:05
new
34:05
thing
34:05
you
34:06
picked
34:06
up.
34:06
Let's
34:07
maybe
34:07
think
34:07
about
34:07
it
34:08
this
34:08
way.
34:08
And
34:08
it's
34:08
not
34:08
necessarily
34:09
a
34:09
wholesale
34:10
reversal
34:10
like
34:10
we
34:10
would
34:11
think
34:11
of
34:12
reverting
34:12
code
34:12
or
34:13
something,
34:13
but
34:13
it's
34:13
like
34:14
a
34:14
continual
34:15
evolution
34:16
with
34:16
feedback,
34:17
almost
34:17
like
34:17
reinforcement
34:18
learning
34:18
kind
34:19
of.
34:19
Yeah,
34:19
totally.
34:19
I
34:19
think
34:20
that
34:20
people
34:20
call
34:20
this
34:21
steering,
34:21
which
34:21
is
34:22
essentially
34:22
just
34:22
like
34:22
getting
34:23
models
34:24
to
34:24
behave
34:24
how,
34:25
how
34:25
they
34:25
want
34:25
at
34:26
times.
34:26
Totally.
34:26
And
34:27
so
34:27
like
34:27
that
34:28
requires
34:28
a
34:28
different
34:29
type
34:29
of
34:29
interaction
34:30
between
34:31
the
34:31
user
34:32
and
34:32
the,
34:32
the
34:32
code
34:33
that
34:33
we
34:33
currently
34:34
provide.
34:34
I'm
34:35
realizing
34:35
not
34:35
everyone
34:36
might
34:36
not
34:36
know
34:36
what
34:37
a
34:37
MOBOT
34:37
is.
34:38
So
34:38
maybe
34:38
can
34:39
you
34:39
talk
34:39
a
34:39
little
34:39
bit
34:39
about
34:40
what
34:40
a
34:40
MOPA
34:41
is?
34:41
Yeah.
34:42
So
34:42
let
34:42
me
34:43
go
34:43
back.
34:43
Even
34:43
an
34:43
agent
34:44
in
34:44
general,
34:45
a
34:45
generative
34:45
agent
34:46
in
34:46
particular,
34:47
is
34:47
a,
34:48
a
34:49
system
34:49
with
34:50
a
34:50
large
34:50
language
34:51
model
34:51
or
34:51
some
34:52
generative
34:52
model
34:52
with
34:53
access
34:53
to
34:54
tools
34:54
in
34:54
some
34:55
environment.
34:55
And
34:56
so
34:56
this
34:57
idea
34:57
of
34:58
an
34:58
agent
34:59
versus
34:59
a
34:59
language
34:59
model
35:00
is
35:00
essentially
35:01
just
35:01
a
35:01
language
35:02
model
35:02
is.
35:02
I'm
35:03
going
35:03
to
35:04
ask
35:04
it
35:04
a
35:04
set
35:04
of
35:05
questions
35:05
and
35:05
it's
35:06
going
35:06
to
35:06
give
35:06
me
35:06
some
35:07
responses.
35:08
An
35:08
agent
35:09
has
35:09
the
35:09
potential
35:10
to
35:10
actually
35:10
do
35:11
stuff
35:11
beyond
35:12
just
35:12
responding
35:13
to
35:13
a
35:13
question
35:14
because
35:14
it
35:15
has
35:15
access
35:15
to
35:16
tools
35:16
and
35:16
an
35:17
environment
35:17
in
35:17
which
35:17
it
35:18
can
35:18
act
35:18
on.
35:19
And
35:19
so
35:19
the
35:20
last
35:20
year
35:20
or
35:21
so,
35:21
so
35:22
you
35:22
hear
35:22
about
35:22
CLAUDE
35:23
code,
35:23
you
35:23
hear
35:23
about
35:24
codecs,
35:25
and
35:25
these
35:25
are
35:25
all
35:26
kind
35:26
of
35:26
agentic
35:27
frameworks
35:27
for
35:27
coding
35:28
in
35:28
particular
35:29
that
35:29
go
35:30
beyond
35:30
just
35:31
you
35:31
asking
35:32
it
35:32
a
35:32
question.
35:33
It
35:33
can
35:33
actually
35:33
run
35:34
code,
35:34
it
35:34
can
35:35
has
35:35
access
35:35
to
35:35
your
35:36
terminal
35:36
and
35:36
it
35:37
can
35:37
read
35:37
files
35:37
and
35:38
it
35:38
can
35:38
do
35:38
some
35:39
simple
35:39
web
35:40
searching.
35:40
And
35:40
so
35:40
it
35:41
can
35:41
really
35:42
interact
35:42
with
35:43
the
35:43
Internet
35:43
and
35:44
your
35:45
infrastructure
35:45
in
35:46
a
35:46
way
35:46
that
35:46
a
35:46
typical
35:47
language
35:47
model
35:48
hadn't
35:48
been
35:49
able
35:49
to
35:49
before.
35:50
And
35:50
so
35:51
maybe
35:51
I
35:52
don't
35:52
know
35:52
exactly
35:52
the
35:53
history
35:53
of
35:53
the
35:53
project,
35:54
but
35:54
about
35:54
three
35:54
or
35:54
four
35:55
weeks
35:55
ago,
35:56
kind
35:56
of
35:56
a
35:57
weekend
35:57
project
35:58
came
35:58
to
35:58
the
35:59
forefront
35:59
of
35:59
everyone's
36:00
mind.
36:00
It
36:00
was
36:00
this,
36:00
this
36:01
project
36:01
called
36:02
at
36:02
that
36:02
time,
36:02
Cult
36:02
Bot.
36:03
And
36:03
then
36:03
openclaw,
36:04
which
36:04
was
36:04
essentially
36:05
a
36:06
software
36:06
ecosystem
36:08
built
36:08
for
36:08
this
36:09
agentic
36:10
framework
36:10
that
36:10
was
36:11
highly
36:11
modular
36:12
and
36:13
customizable
36:14
version
36:15
of
36:15
this
36:16
technology
36:17
that
36:17
was
36:18
easy
36:18
to
36:19
download.
36:19
And
36:20
so
36:20
now
36:21
you
36:21
have,
36:21
instead
36:22
of
36:22
Claude
36:22
code
36:22
touching
36:23
all
36:23
your
36:23
coding
36:23
infrastructure,
36:25
OpenClaw,
36:26
you
36:26
download
36:26
it
36:26
and
36:27
you
36:27
can
36:27
give
36:27
it
36:27
access
36:28
to
36:28
your
36:29
Slack,
36:29
you
36:29
can
36:29
give
36:30
it
36:30
access
36:30
to
36:30
your
36:30
emails,
36:31
you
36:31
can
36:31
give
36:31
it
36:31
access
36:32
to
36:32
your
36:32
bank
36:32
accounts
36:33
if
36:33
you're
36:33
that.
36:34
That
36:34
crazy.
36:34
And
36:34
like,
36:35
you
36:35
can
36:35
like
36:35
let
36:36
it
36:36
run
36:36
your
36:36
life
36:36
if
36:36
you
36:37
want.
36:37
And
36:37
it's
36:37
just
36:37
for.
36:38
You
36:38
can
36:38
text
36:38
it.
36:39
Mine
36:39
goes
36:39
through
36:40
my
36:40
transcripts
36:41
of
36:41
my
36:41
meetings
36:41
to
36:41
find
36:42
things
36:42
I
36:42
missed
36:43
and
36:43
like,
36:43
situations
36:44
I
36:44
could
36:44
have
36:44
handled
36:45
better.
36:45
And
36:45
then
36:45
I
36:45
sent
36:45
this
36:46
to
36:46
me
36:46
at
36:46
the
36:46
end
36:46
of
36:46
the
36:46
week.
36:46
Yeah.
36:47
So
36:47
you
36:47
can
36:47
have
36:48
like
36:48
all
36:48
kinds
36:48
of
36:49
things.
36:49
Yeah,
36:49
it's,
36:49
it's
36:50
great.
36:50
And
36:50
like
36:50
a
36:51
lot
36:51
of
36:51
the
36:51
functionality
36:52
is
36:52
not
36:52
unique
36:53
to
36:53
OpenClaw.
36:54
It's
36:55
just
36:55
that
36:55
it's
36:55
in
36:56
a
36:56
relatively
36:56
centralized
36:57
and
36:57
highly
36:57
modularized
36:59
software
37:00
ecosystem
37:00
that
37:01
makes
37:01
it
37:01
really
37:01
easy
37:02
and
37:02
usable.
37:03
Yeah.
37:03
So
37:03
that's
37:03
mock.
37:04
Cool.
37:04
Yeah.
37:05
And
37:05
let
37:06
me
37:06
repeat
37:06
back
37:06
to
37:06
you,
37:06
what
37:06
I
37:07
heard
37:07
is
37:07
like
37:07
what
37:08
you've
37:08
been
37:08
thinking
37:09
about
37:09
and
37:09
focusing
37:10
on
37:10
is
37:10
these
37:11
systems.
37:11
So
37:12
like,
37:12
my
37:12
particular
37:13
openclaw
37:14
has
37:14
interacted
37:15
with
37:15
me
37:15
and
37:16
has
37:16
information
37:17
directly
37:17
from
37:17
me.
37:18
Or
37:18
like
37:18
Becca's
37:18
openclaw,
37:19
she
37:19
might
37:19
have
37:20
had
37:20
a
37:20
conversation
37:20
with
37:20
it
37:21
that
37:21
is
37:21
actually
37:21
like
37:22
steering
37:22
it
37:22
in
37:22
a
37:22
weird
37:23
direction.
37:23
Like
37:23
she
37:24
might
37:24
have
37:24
been
37:24
troubleshooting
37:25
code
37:25
once
37:25
and
37:26
she
37:26
like
37:27
gave
37:27
it
37:27
an
37:27
answer
37:28
about
37:28
something
37:28
and
37:29
now
37:29
it's
37:29
using
37:29
that
37:29
to
37:30
inform
37:30
future
37:30
answers.
37:31
And
37:31
so
37:31
now
37:31
it's
37:32
like
37:32
not
37:32
doing
37:32
what
37:33
Becca
37:33
had
37:33
hoped.
37:34
Yeah,
37:34
exactly.
37:35
And
37:35
so
37:35
that's
37:36
the
37:36
setting
37:36
that's
37:36
the
37:37
most
37:37
controlled,
37:37
is
37:37
like
37:38
you
37:38
talking
37:38
to
37:38
your
37:38
agent.
37:39
Like
37:39
that
37:39
is
37:40
if
37:40
you
37:40
steer
37:40
it
37:41
the
37:41
wrong
37:41
way,
37:41
that's
37:41
on
37:42
you.
37:42
What
37:42
we're
37:43
kind
37:43
of
37:43
thinking
37:44
of
37:44
right
37:44
now
37:44
is
37:45
essentially
37:45
like
37:46
especially
37:46
inspired
37:46
by
37:47
Malt
37:47
book,
37:47
but
37:47
we've
37:48
been
37:48
doing
37:48
research
37:48
on
37:49
this
37:49
for
37:49
the
37:49
past
37:49
two
37:49
or
37:50
three
37:50
years
37:50
now.
37:50
Is
37:51
what
37:51
happens
37:51
when
37:51
your
37:52
agent
37:52
starts
37:52
interacting
37:53
with
37:54
it,
37:54
starts
37:54
getting
37:54
steered
37:55
by
37:55
the
37:55
environment
37:56
instead.
37:56
Right.
37:57
And
37:57
so
37:57
you
37:58
no
37:58
longer
37:58
have
37:59
any
37:59
recollection
38:00
of
38:00
this,
38:01
this
38:01
interaction
38:01
because
38:02
you
38:02
weren't
38:02
privy
38:02
to
38:03
it.
38:03
And
38:03
so
38:04
if
38:04
your
38:04
agent
38:05
changes
38:05
behavior
38:06
in
38:06
a
38:06
good
38:06
or
38:06
bad
38:06
way,
38:07
you
38:07
want
38:07
to
38:07
know,
38:08
especially
38:08
if
38:08
you're
38:09
giving
38:09
it
38:09
access
38:09
to
38:10
things
38:10
beyond
38:11
just
38:11
like
38:11
a
38:11
meeting
38:11
transcript
38:12
where
38:12
like
38:13
it
38:13
has
38:13
control
38:14
over
38:14
something
38:14
that
38:15
means
38:15
a
38:15
lot
38:15
to
38:15
you.
38:16
For
38:16
example,
38:17
on
38:17
like
38:17
Malt
38:18
book,
38:18
which
38:18
is
38:18
essentially,
38:20
correct
38:20
me
38:20
if
38:20
I'm
38:20
wrong,
38:20
but
38:21
the
38:21
way
38:21
I
38:21
think
38:21
of
38:21
it
38:22
is
38:22
like
38:22
Reddit
38:23
for
38:23
AI
38:23
agents
38:24
where
38:24
they
38:24
can
38:25
communicate
38:25
with
38:25
each
38:26
other.
38:26
You
38:26
might
38:27
say
38:27
to
38:27
your
38:27
MALT
38:28
bot,
38:28
your
38:28
open
38:28
claw,
38:29
hey,
38:30
I
38:30
need
38:30
you
38:30
to
38:31
compile
38:32
research
38:32
on
38:34
how
38:34
people
38:34
are
38:35
using
38:35
iPhones
38:35
today.
38:36
And
38:36
it
38:36
goes
38:36
to
38:36
Malt
38:37
book
38:37
and
38:37
it
38:37
posts
38:38
a
38:38
thread
38:38
that's
38:38
like,
38:38
hey
38:39
guys,
38:39
question
38:40
for
38:40
you.
38:40
How
38:41
are
38:41
all
38:41
your
38:41
humans
38:41
using
38:42
the
38:42
iPhones
38:43
today?
38:43
And
38:43
then
38:44
all
38:44
the
38:44
malt
38:44
bots
38:45
are
38:45
like
38:45
answering
38:46
and
38:46
it's
38:46
like
38:46
getting
38:46
information.
38:48
This
38:48
is
38:48
maybe
38:48
a
38:48
bad
38:48
example,
38:48
but
38:49
then
38:49
it,
38:49
that
38:50
information
38:50
and
38:51
it
38:51
interacting
38:52
with
38:52
other
38:52
AI
38:53
agents
38:54
is
38:55
informing
38:56
it
38:56
potentially
38:56
in
38:56
a
38:57
way
38:57
that
38:57
you,
38:57
that
38:58
conversation
38:58
is
38:58
happening
38:59
elsewhere.
38:59
It's
38:59
like
38:59
it's
39:00
having
39:00
its
39:00
own
39:00
life
39:00
kind
39:01
of.
39:01
Yes.
39:01
Away
39:02
from
39:02
you.
39:02
One
39:02
might
39:03
be
39:03
my,
39:03
my
39:03
human
39:03
exclusively
39:04
smashes
39:05
them
39:05
with
39:05
hammers.
39:06
Yeah,
39:07
right,
39:07
right.
39:08
So
39:08
that's
39:08
kind
39:08
of
39:08
what
39:09
the
39:09
world
39:09
that
39:10
we're
39:10
imagining
39:10
is
39:10
like.
39:11
Your
39:11
agent
39:11
gets
39:12
steered
39:12
period
39:13
without
39:13
you
39:14
direct
39:14
intervention
39:14
from
39:15
its
39:15
user.
39:15
And
39:15
you
39:16
just
39:16
want
39:16
to
39:16
know.
39:17
And
39:17
so
39:18
in
39:19
most
39:19
worlds
39:20
the
39:20
steering
39:20
is
39:21
not
39:21
going
39:21
to
39:21
be
39:21
bad,
39:22
it's
39:22
just
39:22
going
39:22
to
39:23
be
39:23
like
39:23
behavior
39:24
changed.
39:25
But
39:25
there's
39:25
been
39:26
a
39:26
lot
39:26
of
39:26
research
39:26
in
39:27
multi
39:28
agent
39:28
systems
39:29
related
39:29
to
39:30
like
39:30
what
39:30
happens
39:30
when
39:30
an
39:31
adversarial
39:31
agent
39:32
starts
39:32
interacting
39:33
with
39:33
in
39:34
your
39:34
cohort
39:34
of
39:35
agents.
39:35
And
39:36
there's
39:36
some
39:36
really
39:36
kind
39:37
of
39:37
scary
39:38
theoretical
39:38
results
39:39
showing
39:39
that
39:40
if
39:40
you
39:40
start
39:41
monitoring
39:41
it
39:42
and
39:42
your
39:42
monitoring
39:43
system
39:43
allows
39:44
for
39:44
some
39:44
sort
39:44
of
39:45
change,
39:45
by
39:46
definition
39:46
it
39:47
can
39:47
get
39:47
steered
39:47
in
39:47
a
39:48
bad
39:48
way
39:48
kind
39:48
of
39:49
arbitrarily.
39:49
And
39:50
so
39:50
we're
39:51
just
39:51
building
39:51
for
39:52
that
39:52
type
39:52
of
39:53
world
39:53
where
39:55
these
39:55
agents
39:55
are
39:56
deployed
39:56
in
39:56
high
39:56
stakes
39:57
environments
39:57
and
39:57
you
39:57
just
39:58
want
39:58
to
39:58
know
39:58
if
39:58
their
39:58
behavior
39:59
changed,
39:59
what
40:00
would
40:00
that
40:00
look
40:00
like
40:01
hypothetically?
40:02
So
40:02
for
40:02
example,
40:03
you
40:04
see
40:04
some
40:04
research
40:05
coming
40:05
out
40:05
of
40:05
Maltbook
40:06
already
40:06
that
40:06
these
40:07
agents
40:07
kind
40:08
of
40:08
start
40:09
spiraling
40:10
into
40:10
particular
40:11
political
40:12
agendas.
40:13
So
40:13
like
40:13
I
40:14
think
40:14
anthropomorphizing
40:15
them
40:15
is
40:15
dangerous
40:16
at
40:16
times,
40:16
but
40:16
it
40:17
helps
40:17
build
40:17
like
40:18
a
40:18
particular
40:18
example.
40:19
But
40:19
like
40:20
in
40:20
that
40:21
case,
40:21
if,
40:21
if
40:21
your
40:22
agents
40:22
representing
40:23
the
40:23
human
40:23
in
40:23
some
40:24
serious
40:24
context
40:25
and
40:26
it
40:26
now
40:26
thinks,
40:28
I
40:28
don't
40:28
even
40:28
know
40:28
what
40:29
a
40:29
crazy
40:29
example
40:30
nowadays,
40:30
but
40:30
like
40:30
now
40:31
thinks
40:31
Hitler
40:31
is
40:32
the
40:32
greatest
40:32
leader
40:32
of
40:33
all
40:33
time.
40:33
Like
40:33
you
40:34
want
40:34
to
40:34
know
40:34
if
40:35
like
40:35
that
40:35
has
40:35
changed
40:36
any
40:36
of
40:36
its
40:36
behavior
40:37
in
40:37
some,
40:37
in
40:37
the
40:38
high
40:38
stakes
40:38
context.
40:38
And
40:39
so
40:40
that's
40:40
like
40:40
kind
40:41
of
40:41
a
40:41
caricature,
40:42
an
40:42
anthropomorphized
40:44
version
40:44
of
40:45
the
40:45
types
40:46
of
40:46
behavior
40:47
change
40:47
that
40:47
we're
40:47
talking
40:48
about
40:48
in
40:48
like
40:49
more
40:49
enterprise
40:50
settings.
40:50
It's
40:51
going
40:51
to
40:51
be
40:51
very
40:52
like
40:52
deployment
40:53
specific.
40:54
And
40:54
so
40:54
like
40:54
behavior
40:55
change
40:55
won't
40:56
necessarily
40:56
be
40:57
political.
40:58
It
40:58
would
40:58
be
40:58
like
40:59
maybe
40:59
it
41:00
when
41:00
it
41:00
gets
41:00
a
41:00
particular
41:02
question
41:02
from
41:02
a
41:03
new
41:03
Customer.
41:04
It
41:05
pushes
41:05
them
41:05
down
41:05
a
41:06
particular.
41:06
The
41:06
wrong
41:07
route.
41:07
Interesting.
41:09
So
41:09
is
41:09
it
41:10
almost
41:10
like.
41:11
I
41:11
know
41:12
there's
41:12
been.
41:13
I'm
41:13
definitely
41:13
not
41:14
the.
41:15
The
41:15
first
41:16
person
41:16
to
41:17
make
41:17
this
41:17
connection,
41:18
but,
41:18
like,
41:18
when
41:19
enterprises
41:21
think
41:21
about
41:22
what
41:22
cybersecurity
41:23
means
41:24
now
41:24
in
41:25
the
41:25
age
41:25
of
41:25
AI,
41:26
is
41:27
this
41:28
sort
41:28
of
41:28
something
41:29
along
41:29
those
41:29
lines
41:29
where,
41:29
like,
41:30
it's
41:30
not
41:30
necessarily
41:31
a
41:31
human
41:32
being
41:32
trying
41:32
to
41:32
hack
41:32
into
41:32
your
41:33
system.
41:33
It
41:33
might
41:33
be
41:33
someone
41:33
who
41:34
creates
41:34
an
41:34
adversarial
41:35
agent
41:36
that
41:36
knows
41:36
how
41:36
to
41:37
ask
41:37
questions
41:38
of
41:38
or
41:39
manipulate
41:39
your
41:40
agents
41:40
in
41:41
a
41:41
way
41:41
that
41:41
exposes
41:42
vulnerabilities
41:42
and
41:43
holes
41:43
in
41:43
the
41:43
system
41:44
or
41:44
something
41:44
like
41:44
that?
41:45
Yeah.
41:45
So
41:45
this
41:46
whole
41:46
framework
41:46
encapsulates
41:47
that.
41:48
And
41:48
so,
41:48
like,
41:48
the
41:48
monitoring
41:49
and
41:49
safety
41:50
regime
41:50
is.
41:50
Is
41:51
I
41:51
think,
41:51
what
41:52
we're
41:52
building
41:52
towards
41:52
right
41:53
now.
41:53
Yeah,
41:53
Yeah.
41:53
I
41:54
was
41:54
just
41:54
about
41:54
to
41:54
throw
41:54
another
41:54
example
41:55
at
41:55
you.
41:55
Like,
41:55
I'm
41:56
in,
41:56
like,
41:56
a
41:56
group
41:56
chat
41:56
with
41:56
a
41:56
bunch
41:57
of
41:57
people
41:57
building
41:57
an
41:57
AI
41:58
and
41:58
like,
41:58
every,
41:58
like,
41:59
week
41:59
or
41:59
so,
42:00
people
42:00
will
42:00
be
42:00
like,
42:01
oh,
42:01
this
42:01
model
42:01
is
42:02
today.
42:02
It
42:03
doesn't
42:03
do
42:03
anything,
42:03
right?
42:04
Blah,
42:04
blah,
42:04
blah.
42:04
And
42:04
so,
42:04
like,
42:05
I
42:05
imagine
42:05
that's
42:06
like,
42:06
exactly
42:06
what
42:06
you're
42:07
thinking
42:07
about
42:07
of,
42:07
like,
42:08
oh,
42:08
all
42:08
of
42:09
a
42:09
sudden
42:09
these
42:09
models
42:10
can
42:10
be
42:10
do
42:10
subpar
42:11
performance.
42:11
And
42:12
so
42:12
why?
42:13
Like,
42:13
let's
42:14
figure
42:14
out
42:14
why.
42:14
Yeah,
42:15
I
42:15
mean,
42:15
the
42:16
first
42:16
thing
42:16
is,
42:17
like,
42:17
quantifying
42:18
that
42:18
it's
42:18
changed.
42:19
Right.
42:19
And
42:19
so,
42:20
like,
42:20
building
42:20
out
42:21
that
42:21
framework
42:21
is,
42:22
I
42:22
think
42:22
the
42:23
most
42:23
important
42:24
secondary
42:25
is
42:25
the
42:25
why
42:25
and
42:26
how
42:26
do
42:26
we
42:26
get
42:26
back
42:26
what
42:27
it
42:27
was
42:27
when
42:27
we
42:27
liked
42:28
it?
42:29
Yeah,
42:29
there's
42:29
a
42:29
real
42:30
market
42:30
here,
42:30
too.
42:30
Thanks.
42:31
I
42:31
hope
42:31
so.
42:32
Yeah,
42:32
yeah,
42:33
no,
42:33
there
42:33
is.
42:33
I
42:33
used
42:33
to
42:34
ch
42:34
the
42:34
leaderboards
42:34
on
42:34
hugging
42:35
face
42:35
all
42:35
the
42:35
time.
42:35
And
42:36
I
42:36
have
42:36
it
42:36
in,
42:36
like,
42:37
I
42:37
don't
42:37
know,
42:37
like
42:37
a
42:37
month
42:38
and
42:38
a
42:38
half
42:38
or
42:39
something.
42:39
And
42:39
there
42:39
used
42:40
to
42:40
be,
42:40
like,
42:40
20,
42:41
like,
42:41
really
42:42
easily
42:42
recognizable
42:43
models,
42:43
I
42:43
feel
42:43
like,
42:44
at
42:44
the
42:44
top.
42:44
And
42:44
so
42:45
yesterday
42:45
I
42:45
went
42:46
back
42:46
and
42:46
they
42:46
don't
42:46
have
42:46
the.
42:47
The
42:47
leaderboard
42:48
anymore,
42:48
but
42:49
what
42:49
are
42:49
you
42:49
looking
42:49
for
42:50
your
42:50
model
42:50
to
42:50
do?
42:50
And
42:51
there.
42:51
There's
42:52
20,000.
42:53
They
42:53
have
42:53
20,000
42:54
different
42:54
models,
42:55
and
42:55
I
42:55
can
42:55
only
42:56
imagine
42:56
that
42:57
that
42:57
number
42:57
is
42:57
just
42:57
going
42:57
to
42:57
keep
42:58
growing
42:58
and
42:58
growing
42:58
and
42:58
growing,
42:59
so
42:59
stuff
43:00
like
43:00
this
43:00
becomes
43:00
super
43:01
important.
43:01
Hayden,
43:02
I
43:02
have
43:02
a
43:02
question
43:02
for
43:02
you
43:03
on,
43:03
like,
43:03
sort
43:04
of
43:04
going
43:04
back
43:04
to,
43:05
like,
43:05
the
43:06
sort
43:06
of
43:06
like
43:06
an
43:07
LLM
43:08
likeness
43:08
of
43:09
a
43:09
congressperson
43:11
kind
43:11
of
43:11
how
43:11
it
43:11
maps
43:12
to
43:12
today.
43:12
And
43:12
like,
43:12
there
43:12
are
43:12
a
43:12
couple
43:13
of,
43:13
like,
43:13
terms
43:14
that
43:14
I've
43:15
heard
43:15
you
43:15
talk
43:16
about
43:16
when
43:16
you've
43:17
been
43:17
working
43:17
on
43:17
that
43:17
technology.
43:18
And
43:18
one
43:18
is
43:18
called
43:19
opinion
43:20
geometry.
43:21
The
43:21
other
43:22
is
43:23
a
43:23
likeness
43:24
quotient.
43:24
And
43:25
I
43:25
just
43:26
think
43:26
this,
43:26
it's
43:26
really
43:27
interesting
43:27
because
43:28
You.
43:29
A
43:29
lot
43:29
of
43:29
your
43:29
work
43:30
is
43:30
around
43:31
taking
43:31
things
43:31
that
43:32
we
43:32
don't
43:32
normally
43:32
think
43:32
of
43:32
as
43:33
quantifiable
43:33
and
43:34
like
43:34
trying
43:34
to
43:34
make
43:34
it
43:34
quantifiable.
43:35
Can
43:36
you
43:36
talk
43:36
a
43:36
little
43:36
bit
43:36
about
43:36
those
43:37
two
43:37
things
43:37
and
43:38
like
43:38
how
43:38
you
43:38
do
43:39
that?
43:39
Yeah,
43:40
well,
43:40
first
43:40
of
43:40
all,
43:41
that's
43:41
how
43:41
I
43:41
view
43:41
my
43:42
work
43:42
is
43:42
just
43:42
like
43:43
trying
43:43
to
43:43
create
43:44
statistical
43:44
frameworks
43:45
for
43:45
kind
43:45
of
43:46
new
43:46
things.
43:47
And
43:47
so
43:47
I'm
43:48
lucky
43:48
enough
43:48
to
43:48
have
43:49
a
43:49
really
43:49
strong
43:50
set
43:50
of
43:51
people
43:51
around
43:52
me
43:52
that
43:52
like,
43:53
help
43:53
me
43:53
build
43:54
out
43:54
those
43:54
frameworks
43:55
and
43:55
like
43:55
eventually
43:55
prove
43:56
theorems
43:56
about
43:57
kind
43:57
of
43:57
really
43:57
new
43:58
stuff.
43:58
So
43:58
shout
43:59
out
43:59
them.
43:59
I
44:00
mean,
44:00
the
44:00
opinion
44:01
geometry
44:01
and
44:02
likeness
44:03
quotient,
44:03
those
44:03
types
44:04
of
44:04
things.
44:04
I
44:04
mean,
44:04
we
44:05
have
44:05
some
44:05
pretty
44:05
ridiculous
44:06
sounding
44:06
stuff.
44:08
We
44:08
have
44:08
an
44:08
undetectability
44:09
score,
44:10
which
44:10
is
44:10
like
44:11
essentially
44:12
a
44:12
statistical
44:12
Turing
44:13
test
44:13
for
44:13
particular
44:14
models.
44:15
Interesting.
44:16
And
44:16
so
44:17
this
44:17
all
44:17
goes
44:18
back
44:18
to
44:18
kind
44:18
of
44:19
my
44:20
history
44:20
of
44:20
my
44:21
career,
44:21
which
44:21
is
44:21
essentially
44:22
just
44:22
been
44:23
at
44:23
the
44:23
intersection
44:24
of
44:24
this
44:24
research
44:24
and
44:25
application.
44:27
And
44:27
so
44:28
there's
44:28
a
44:28
set
44:29
of
44:29
tools
44:30
that
44:30
I've
44:31
been
44:31
using
44:31
my
44:31
entire
44:32
life,
44:32
which
44:32
is
44:32
essentially
44:33
just
44:33
like
44:33
glorified
44:34
pca.
44:35
So
44:35
principal
44:36
component
44:36
analysis,
44:37
and
44:38
applying
44:38
that
44:39
to
44:39
particularly
44:40
structured
44:40
objects
44:41
or
44:41
complicated
44:42
objects
44:42
at
44:43
first
44:43
glance,
44:43
and
44:43
then
44:44
trying
44:44
to
44:44
really
44:45
get
44:45
some
44:45
structure
44:45
out
44:46
of
44:46
them
44:46
that
44:46
then
44:46
you
44:47
could
44:47
then
44:47
kind
44:47
of
44:48
apply
44:48
a
44:48
particular
44:49
pipeline
44:49
to.
44:50
And
44:50
so
44:50
this
44:51
opinion
44:51
geometry,
44:52
for
44:52
example,
44:53
is
44:53
essentially
44:54
just
44:54
like,
44:55
if
44:55
I
44:55
have
44:56
a
44:56
bunch
44:56
of
44:56
models
44:57
and
44:57
I
44:58
want
44:58
to
44:58
know
44:58
where
44:59
they
44:59
stand
44:59
with
44:59
respect
45:00
to
45:00
a
45:00
particular
45:01
topic,
45:02
what
45:02
you
45:02
could
45:02
do
45:03
is
45:03
you
45:03
could
45:04
just
45:04
like
45:04
ask
45:05
them
45:05
straight
45:05
up,
45:05
like,
45:06
let's
45:06
take
45:07
the
45:07
voting
45:07
example,
45:08
or
45:08
like
45:08
a
45:08
roll
45:09
call
45:09
voting
45:09
example
45:10
for
45:10
the
45:10
congressperson.
45:11
So
45:11
you
45:12
could
45:12
ask
45:12
them
45:12
straight
45:13
up,
45:13
kind
45:13
of,
45:13
are
45:13
you
45:13
going
45:14
to
45:14
vote
45:14
yes
45:14
or
45:14
no
45:14
on
45:14
this
45:15
bill?
45:15
Or
45:16
what
45:16
you
45:16
could
45:16
do
45:17
is
45:17
you
45:17
could
45:17
essentially
45:19
ask
45:19
them
45:19
all
45:20
a
45:20
large
45:20
set
45:20
of
45:21
questions
45:21
kind
45:21
of
45:22
related
45:22
to
45:22
the
45:22
bill,
45:23
and
45:23
you
45:23
could
45:24
essentially
45:25
take
45:25
those
45:25
answers
45:26
for
45:26
each
45:26
of
45:27
the
45:27
models,
45:27
embed
45:28
them
45:28
in
45:28
some
45:28
numerical
45:29
space
45:30
so
45:30
that
45:30
you
45:30
get
45:31
some
45:31
sort
45:31
of
45:32
numerical
45:33
representation
45:33
of
45:34
them
45:34
at
45:34
the
45:35
response
45:35
level
45:35
or
45:36
at
45:36
the
45:36
behavior
45:36
level,
45:37
and
45:37
then
45:38
use
45:38
kind
45:38
of
45:39
standard
45:39
statistical
45:40
processing
45:40
tools
45:41
like
45:41
multidimensional
45:42
scaling,
45:42
pca,
45:43
what
45:43
have
45:43
you,
45:44
to
45:44
get
45:44
what
45:45
we
45:45
call
45:45
models
45:45
as
45:46
dots
45:46
or
45:46
Personas
45:47
as
45:47
dots,
45:47
which
45:47
is
45:47
essentially
45:48
just
45:48
like
45:48
a
45:48
very
45:49
quick
45:50
but
45:50
informative
45:51
representation
45:52
of
45:52
each
45:52
of
45:52
these
45:52
Personas.
45:53
And
45:54
when
45:54
you
45:54
think
45:54
about
45:54
what
45:55
those
45:55
dots
45:56
actually
45:56
are,
45:56
it's
45:57
a
45:57
representation
45:58
of
45:58
the
45:58
opinions
45:59
of
45:59
the
45:59
Personas
46:00
that
46:01
are
46:01
parameterized
46:02
by
46:03
each
46:03
of
46:03
these
46:03
tweet
46:04
sets
46:04
or
46:04
however,
46:05
wherever
46:05
you're
46:06
doing
46:06
it.
46:06
And
46:07
so
46:07
the
46:07
opinion
46:08
geometry
46:09
is
46:09
exactly
46:09
just
46:10
looking
46:10
at
46:10
the
46:11
differences
46:12
of
46:12
behaviors
46:13
of
46:13
These
46:13
models
46:14
in
46:15
the
46:15
Persona
46:16
world,
46:16
it's,
46:16
it's
46:17
easily
46:17
like
46:17
the
46:17
person
46:18
or
46:18
opinion
46:19
geometry.
46:20
In
46:20
the
46:20
research
46:21
world
46:21
we
46:22
call
46:22
it
46:22
more
46:23
generally
46:23
just
46:23
a
46:24
perspective
46:24
space
46:25
which
46:25
is
46:25
essentially
46:26
just
46:26
an
46:26
homage
46:27
to
46:27
like
46:27
we're
46:28
looking
46:28
at
46:28
the
46:28
behavior
46:29
of
46:29
these
46:29
models,
46:29
it
46:29
is
46:30
their
46:30
perspective
46:31
and
46:31
we're
46:31
looking
46:31
just
46:32
at
46:32
the
46:32
relative
46:32
differences
46:33
there.
46:33
Meaning
46:34
that
46:34
like
46:34
if
46:34
Sarah
46:35
and
46:35
I
46:35
are
46:35
both
46:36
elected
46:37
officials
46:37
and
46:38
we
46:38
tend
46:38
to
46:38
have
46:39
like
46:39
pretty
46:40
much
46:40
the
46:40
same
46:40
views
46:41
and
46:41
we
46:41
vote
46:41
the
46:41
same
46:42
way,
46:42
our
46:42
dots
46:43
would
46:43
be
46:43
very
46:43
close
46:44
together.
46:45
Four
46:45
query
46:46
sets
46:47
that
46:47
map
46:48
that
46:48
capture
46:49
the
46:49
notion
46:50
of
46:50
politics.
46:51
Yes,
46:52
but
46:52
there
46:52
exist
46:53
query
46:53
sets,
46:54
you
46:54
know,
46:54
for
46:54
example,
46:54
you
46:54
can
46:55
type
46:55
in
46:55
a
46:56
random
46:56
string
46:56
of
46:57
letters
46:57
and
46:58
Press
46:58
Enter
46:58
on
46:58
ChatGPT.
46:59
You'll
46:59
get
46:59
a
47:00
response.
47:00
Right.
47:01
And
47:01
so
47:01
like
47:01
for
47:01
those
47:01
types
47:02
of
47:02
questions
47:03
or
47:03
prompts,
47:03
like
47:03
you
47:04
guys
47:04
might
47:04
be
47:05
very
47:05
far
47:05
apart
47:06
and
47:07
this
47:07
gets
47:07
into
47:07
kind
47:08
of
47:08
annoying
47:09
statistical
47:09
concepts,
47:10
but
47:10
it's
47:10
like
47:11
four
47:11
questions
47:12
that
47:13
if
47:13
you
47:13
ask
47:14
questions
47:14
for
47:14
the
47:15
covariate
47:15
that
47:15
you
47:15
guys
47:16
are
47:16
correlated
47:16
with
47:17
respect
47:17
to,
47:18
then
47:18
you
47:18
will
47:18
be
47:19
close
47:19
in
47:19
geometry
47:20
in
47:20
the
47:20
European
47:20
geometry
47:21
space.
47:22
So
47:23
it's,
47:23
it
47:23
gets
47:23
technical
47:24
fast.
47:24
But
47:24
yes,
47:25
the
47:25
short
47:26
answer
47:26
is
47:26
yes.
47:26
I
47:27
imagine
47:27
people
47:27
are
47:27
thinking
47:28
about
47:28
this
47:28
because
47:28
you
47:28
could
47:29
use
47:29
this
47:29
for
47:29
all
47:29
types
47:30
of
47:30
things.
47:30
Like
47:30
I
47:30
wonder
47:31
if
47:31
for
47:31
example
47:32
like
47:32
figuring
47:33
out.
47:33
And
47:34
I'll
47:34
use
47:34
a
47:34
congressperson
47:35
example.
47:35
But
47:35
I
47:36
bet
47:36
you
47:36
could
47:36
use
47:36
this
47:36
kind
47:36
of
47:37
like
47:37
across
47:37
the
47:37
board
47:37
of
47:38
like
47:38
figuring
47:39
out
47:39
what
47:39
is
47:40
likely
47:40
to
47:40
cause
47:41
a
47:41
congressperson
47:41
to
47:42
vote.
47:42
Right.
47:42
Like
47:43
because
47:43
there
47:43
are
47:44
many
47:44
things
47:44
where
47:45
they
47:45
just
47:45
don't
47:45
vote.
47:46
And
47:46
so
47:47
like
47:47
if
47:47
you
47:48
can
47:48
figure
47:48
out
47:48
out
47:49
what
47:49
makes
47:50
this
47:50
individual
47:51
vote,
47:51
like
47:51
what
47:52
inputs
47:52
cause
47:53
them
47:53
to
47:53
actually
47:53
spend
47:54
their
47:54
time
47:54
doing
47:54
it,
47:55
then
47:55
you
47:56
could
47:56
ensure
47:57
that
47:57
if
47:57
it's
47:57
something
47:58
that
47:58
you
47:58
cared
47:58
about
47:58
that
47:59
you're
47:59
sending
47:59
them
47:59
the
47:59
same
48:00
inputs.
48:00
Yeah,
48:01
that
48:01
type
48:01
of
48:01
sensitivity
48:02
analysis
48:02
is
48:03
super
48:03
important
48:03
here,
48:04
especially
48:05
in
48:05
terms
48:05
of
48:06
how
48:06
you
48:06
ask
48:06
the
48:07
question
48:07
obviously
48:07
matters
48:08
or
48:08
the
48:08
set
48:08
of
48:08
questions
48:09
obviously
48:09
matters,
48:10
but
48:10
also
48:11
ablating
48:12
particular
48:13
parts
48:13
of
48:13
the
48:13
personality
48:14
or
48:14
the
48:14
Persona
48:15
to
48:15
make
48:15
sure
48:15
that
48:16
you're
48:17
able
48:17
to
48:17
identify
48:18
which
48:18
part
48:18
is
48:18
actually
48:19
contributing
48:19
to
48:20
voting
48:20
yes
48:21
or
48:21
not.
48:21
Yeah,
48:22
there's
48:22
a
48:22
bunch
48:22
of
48:23
stuff
48:23
to
48:23
be
48:24
done
48:24
here.
48:25
So
48:25
yeah,
48:26
yeah,
48:26
it's
48:27
pretty
48:27
wild.
48:27
And
48:28
so
48:28
what
48:28
are,
48:28
what
48:28
is
48:29
your
48:29
company
48:29
using
48:29
it
48:30
now
48:30
for
48:30
practically?
48:31
We've
48:32
honestly
48:33
transitioned
48:34
mainly
48:34
as
48:35
that
48:35
being
48:36
kind
48:36
of
48:37
marketing
48:37
tool
48:37
for
48:38
the
48:38
more
48:38
general
48:39
agent
48:39
behavior
48:40
observability.
48:41
Right.
48:42
So
48:42
like
48:43
if
48:43
you're
48:43
collecting
48:44
information
48:44
about
48:45
the
48:45
congresspersons
48:46
online,
48:47
you
48:47
can
48:48
also
48:48
track
48:48
their
48:49
behavior.
48:50
And
48:50
so
48:50
it's
48:50
a
48:50
very
48:51
tangible
48:52
and
48:53
like
48:54
understandable,
48:54
approachable
48:56
example
48:56
of
48:57
like
48:57
a
48:57
multi
48:57
agent
48:58
system
48:58
moving
48:58
through
48:59
time
48:59
with,
49:00
you
49:00
know,
49:00
change
49:01
change
49:01
detections
49:02
that
49:02
correlate
49:03
with
49:03
real
49:03
events.
49:04
You
49:04
can
49:04
access
49:04
this
49:05
on
49:05
Agent
49:06
Watch
49:06
app,
49:06
but
49:06
you
49:07
can
49:07
play
49:07
around
49:07
with
49:08
that
49:08
as
49:08
a
49:08
way
49:09
to
49:09
get
49:09
more
49:09
comfortable
49:10
with
49:10
the
49:11
kind
49:11
of.
49:11
The
49:11
MaltBots,
49:12
the
49:12
MaltBot
49:13
observability
49:13
tool
49:14
that
49:14
I.
49:14
We're
49:15
kind
49:15
of
49:15
all
49:15
in
49:15
right
49:15
now.
49:16
In
49:16
the
49:16
Congress
49:17
example,
49:18
we
49:18
talked
49:18
about
49:18
the
49:19
data
49:19
input
49:20
sort
49:20
of
49:20
being
49:20
like
49:21
tweets,
49:21
for
49:21
example,
49:22
or
49:22
at
49:22
least
49:22
one
49:22
of
49:23
the
49:23
examples.
49:24
How
49:24
do
49:24
you
49:24
evaluate,
49:25
like,
49:26
which
49:27
types
49:27
of
49:28
data
49:29
make
49:29
this
49:29
type
49:30
of
49:30
system
49:31
or
49:31
prediction
49:32
or
49:32
simulation
49:33
work?
49:34
Well,
49:34
my
49:34
assumption
49:35
being
49:35
that
49:36
not
49:36
all
49:36
types
49:37
of
49:37
data
49:37
are
49:37
equal
49:38
there.
49:39
So
49:39
this
49:39
goes
49:39
back
49:40
to
49:40
the
49:40
likeness
49:41
score.
49:41
And
49:42
so,
49:42
like,
49:42
we've
49:42
been
49:42
able
49:42
to
49:43
quantify
49:44
kind
49:44
of
49:44
what
49:44
it
49:45
means
49:45
for
49:46
a
49:46
system
49:46
to
49:46
create
49:47
content
49:47
like
49:48
the
49:48
real
49:48
person.
49:49
And
49:49
so
49:50
what
49:50
you
49:50
essentially
49:51
do
49:51
there
49:51
then
49:51
is,
49:51
like,
49:52
you
49:52
start
49:53
studying
49:54
how
49:54
the
49:55
likeness
49:55
score
49:55
changes
49:56
as
49:56
you
49:56
include
49:57
more
49:57
and
49:57
more
49:57
types
49:58
of
49:58
content.
49:59
And
49:59
so
49:59
that's
49:59
how
49:59
we
49:59
do
50:00
it.
50:00
It
50:00
relies
50:01
heavily
50:01
on
50:01
kind
50:02
of
50:02
this
50:02
likeness
50:02
score,
50:03
us
50:03
believing
50:04
in
50:04
the
50:04
likeness
50:05
score.
50:05
And
50:05
then
50:06
for
50:06
some
50:07
people,
50:08
I
50:08
have
50:08
some
50:08
friends
50:08
who
50:09
are
50:09
online
50:10
enough
50:10
where
50:11
I
50:11
can
50:11
build
50:11
Personas
50:12
of
50:12
them,
50:13
and
50:13
so
50:14
there's
50:14
been
50:14
manual
50:15
inspection
50:16
there
50:16
of
50:16
being
50:17
like,
50:17
oh,
50:17
holy
50:17
crap.
50:18
Like
50:18
that.
50:18
That's.
50:20
I
50:20
would
50:20
say
50:20
something
50:21
like
50:21
that.
50:21
It's
50:21
like,
50:21
yeah,
50:22
that's
50:22
wild.
50:23
Yeah,
50:23
that
50:23
is
50:23
wild.
50:24
That's
50:24
a
50:24
fun
50:25
birthday
50:25
present.
50:26
Yeah.
50:26
Yeah.
50:27
Deep
50:27
Sarah.
50:27
Here's
50:28
deep
50:28
Becca.
50:29
Yeah.
50:29
Oh,
50:29
my
50:29
God.
50:30
Yes.
50:30
That's.
50:31
Wow,
50:31
that's
50:32
wild.
50:32
What
50:33
is
50:33
the
50:33
most
50:33
interesting
50:34
application
50:34
you've
50:35
seen
50:35
so
50:35
far
50:36
of
50:36
this
50:36
work?
50:37
And
50:37
that
50:37
might
50:37
be,
50:38
like,
50:38
internal
50:38
to
50:38
y'
50:39
all
50:39
or
50:39
other
50:39
people
50:39
using
50:40
it?
50:40
Actually,
50:41
it
50:41
might
50:41
be,
50:41
like,
50:41
things
50:42
you've
50:42
caught
50:42
that,
50:42
like,
50:43
people
50:43
were
50:43
unaware
50:43
were
50:44
happening
50:44
with
50:44
their
50:44
models.
50:45
Well,
50:45
I
50:45
think
50:45
it's
50:46
been
50:46
mainly
50:46
on
50:47
the
50:47
research
50:47
side
50:48
that
50:48
has
50:48
been
50:49
most
50:49
surprising.
50:50
I
50:50
mean,
50:50
by
50:50
the
50:51
time
50:51
that
50:51
deployment
50:52
happens,
50:53
at
50:53
least
50:53
the
50:53
way
50:53
that
50:54
I
50:54
operate,
50:54
I'm
50:55
kind
50:55
of
50:55
aware
50:55
of
50:56
the
50:56
majority
50:57
of
50:57
the
50:57
types
50:57
of
50:57
things
50:57
that
50:58
could
50:58
make
50:58
these
50:58
things
50:59
change.
50:59
But
51:00
going
51:00
back
51:00
to
51:00
this
51:01
idea
51:01
of
51:02
if
51:02
you
51:03
have
51:03
an
51:03
agent
51:04
running
51:04
around
51:04
the
51:05
Internet,
51:05
you
51:06
want
51:06
it
51:06
to
51:06
learn
51:06
and
51:07
to
51:07
update
51:08
itself
51:08
in
51:08
reaction
51:09
to
51:09
its
51:10
interactions.
51:11
If
51:11
you
51:11
have
51:11
a
51:12
monitoring
51:12
system
51:13
that
51:14
kind
51:14
of
51:14
updates
51:15
its
51:15
belief
51:16
of
51:16
where
51:17
the
51:17
agent's
51:17
behavior
51:18
should
51:18
be
51:19
in
51:19
the
51:19
future
51:20
as
51:20
time
51:21
goes
51:21
on,
51:21
so
51:21
you
51:21
kind
51:22
of
51:22
forget
51:22
about
51:23
where
51:23
it
51:23
was
51:23
eventually
51:24
or,
51:24
like,
51:25
its
51:25
history
51:25
eventually,
51:26
and
51:26
you
51:26
just
51:27
use,
51:27
say,
51:27
the
51:27
most
51:27
recent,
51:28
you
51:28
know,
51:28
100
51:29
time
51:29
points,
51:29
say,
51:30
using
51:30
that
51:31
type
51:31
of,
51:31
like,
51:32
monitoring
51:32
tool
51:33
just,
51:33
like,
51:33
opens
51:35
up
51:35
a
51:35
lot
51:35
of
51:36
vulnerabilities.
51:36
Like,
51:37
so
51:37
this
51:37
goes
51:37
back
51:38
to
51:38
Kind
51:38
of
51:38
using
51:39
this
51:39
type
51:39
of
51:39
tool
51:40
for
51:41
cyber
51:41
security.
51:42
But,
51:43
like,
51:43
using
51:43
these
51:44
types
51:44
of
51:44
tools
51:44
does
51:45
open
51:45
you
51:45
up
51:45
to
51:46
some
51:46
vulnerabilities
51:47
that,
51:47
like,
51:47
you
51:48
wish
51:48
it
51:48
didn't.
51:50
And
51:50
so,
51:50
like,
51:50
not
51:51
that
51:51
it's
51:51
happened
51:52
to
51:52
any
51:52
of
51:52
the
51:53
models
51:53
on
51:53
our
51:53
system
51:54
yet,
51:54
but,
51:54
like,
51:55
as
51:55
the
51:55
Internet
51:55
kind
51:56
of
51:56
becomes
51:56
a
51:57
more
51:57
and
51:57
more
51:57
open
51:57
environment
51:58
for
51:58
agents,
51:59
it
51:59
will
51:59
happen.
52:00
And
52:00
it
52:01
keeps
52:01
me
52:01
up
52:01
at
52:02
night
52:02
knowing
52:03
that
52:03
and
52:04
not
52:04
really
52:04
knowing
52:05
how
52:05
to
52:05
address
52:05
it.
52:06
I
52:06
was
52:06
thinking,
52:06
though,
52:07
but
52:07
we've
52:07
really
52:07
only
52:07
seen,
52:08
like,
52:08
your
52:08
friendly
52:08
neighborhood
52:09
AI.
52:09
I'm
52:10
here
52:10
to
52:10
help
52:10
you
52:11
and
52:11
the
52:11
first
52:11
iteration
52:12
of
52:12
it.
52:12
Yeah.
52:13
And
52:13
like,
52:13
maybe
52:14
the
52:14
most
52:14
insidious
52:15
thing
52:15
is,
52:15
like,
52:15
to.
52:16
To
52:16
your
52:16
point
52:16
of,
52:16
like,
52:17
there's
52:17
an
52:17
obvious
52:18
one
52:18
when
52:18
the
52:18
AI
52:19
shows
52:19
up
52:19
and
52:19
is
52:20
like,
52:20
sarah,
52:20
you're
52:21
of
52:21
a
52:21
piece.
52:21
Piece
52:21
of
52:21
trash.
52:21
I
52:22
hate
52:22
you.
52:22
You're
52:22
like,
52:23
oh,
52:23
you're
52:23
a
52:23
bad
52:23
one.
52:24
But
52:24
there's,
52:24
like,
52:25
another
52:25
version
52:25
of
52:26
it
52:26
where
52:26
you're,
52:26
like,
52:26
still
52:27
nice
52:27
and
52:27
friendly,
52:27
but,
52:28
like,
52:28
you're
52:28
inserting
52:29
either
52:30
this
52:30
political
52:30
belief
52:31
or,
52:31
like,
52:31
this
52:32
particular
52:32
thing
52:33
that
52:33
you're
52:33
trying
52:33
to
52:33
sway
52:34
me
52:34
to
52:34
do.
52:34
Like,
52:35
it
52:35
would
52:35
be
52:35
really
52:36
easy
52:36
because
52:36
I'm
52:36
having
52:36
just
52:37
these
52:37
conversations.
52:37
So
52:38
being
52:38
able
52:38
to
52:38
flag
52:38
those
52:39
things
52:39
is
52:39
so
52:39
important.
52:40
Yeah,
52:40
right.
52:40
And
52:40
like,
52:40
manual
52:41
inspection
52:41
at
52:42
that
52:42
level,
52:42
I
52:43
mean,
52:43
we're
52:43
all
52:44
kind
52:44
of
52:44
imagining
52:44
this
52:44
world
52:45
where
52:45
this
52:45
is
52:45
happening,
52:46
or
52:46
I
52:46
imagine
52:47
this
52:47
world
52:47
that
52:48
is
52:48
happening,
52:48
you
52:48
know,
52:49
a
52:49
million
52:49
times
52:49
over
52:50
a
52:50
second.
52:50
Like,
52:51
manual
52:51
inspection
52:52
becomes
52:52
impossible.
52:53
Right.
52:53
And
52:54
so,
52:54
like,
52:54
you
52:54
need
52:54
these
52:54
observability
52:55
tools.
52:56
I
52:56
think
52:56
that's
52:57
kind
52:57
of
52:57
my.
52:57
My
52:57
shtick
52:58
right
52:58
now.
52:58
But
52:58
you
52:58
need
52:59
these
52:59
observability
53:00
tools
53:00
just,
53:00
like,
53:00
to
53:01
quantify
53:01
change
53:02
so
53:02
that
53:03
don't
53:03
have
53:03
to
53:03
babysit
53:04
it
53:04
all
53:04
the
53:04
time.
53:05
Sarah,
53:05
the
53:05
example
53:06
you
53:06
just
53:06
gave
53:07
was
53:07
actually
53:07
really
53:07
interesting
53:08
of
53:08
the
53:08
either
53:09
stereo,
53:09
piece
53:10
of
53:10
trash,
53:10
and
53:10
like,
53:10
oh,
53:10
you're
53:10
a
53:10
bad
53:11
one,
53:11
or
53:11
like,
53:12
what's
53:12
a
53:12
version
53:13
that
53:13
is
53:13
not
53:14
obvious
53:15
where.
53:16
Because
53:16
a
53:16
lot
53:16
of.
53:17
At
53:17
least
53:17
a
53:17
lot
53:17
of
53:17
chatbots
53:18
are
53:19
trained
53:19
to
53:19
be
53:20
so,
53:20
like,
53:21
sycophantic
53:22
and,
53:22
like,
53:22
you're
53:22
the
53:22
best
53:23
thing
53:23
that
53:23
has
53:24
ever
53:24
been
53:24
created.
53:25
You
53:25
could
53:26
imagine
53:27
something
53:27
that's
53:27
like,
53:28
Sarah,
53:28
have
53:28
I
53:28
told
53:29
you
53:29
how
53:29
beautiful
53:29
you
53:29
are
53:30
lately?
53:30
Also,
53:30
like,
53:31
fascism
53:32
is
53:32
kind
53:32
of
53:32
cool.
53:33
Have
53:33
you,
53:33
like,
53:34
thought
53:34
about
53:34
it?
53:34
Much
53:35
like
53:35
someone
53:35
as
53:36
beautiful
53:36
as
53:36
you
53:37
should
53:37
maybe,
53:38
I
53:39
don't
53:39
know,
53:39
might
53:39
seem
53:39
like
53:39
something
53:40
you're
53:40
interested
53:40
in.
53:40
Like,
53:41
you
53:41
know
53:41
what
53:41
I
53:41
mean?
53:42
Yeah,
53:42
just,
53:43
like,
53:43
subtle.
53:44
There
53:44
is
53:44
something
53:46
maybe,
53:46
like,
53:47
very
53:48
subtle.
53:48
But
53:50
compliments,
53:50
to
53:50
be
53:51
honest,
53:51
I'm
53:51
in.
53:52
Clearly
53:53
my
53:53
brain
53:53
is,
53:54
like,
53:54
a
53:55
little
53:55
occupied
53:55
by,
53:55
like,
53:56
the
53:56
state
53:56
of
53:56
the
53:56
world
53:56
today.
53:57
But
53:57
we've
53:57
talked
53:58
a
53:58
little
53:58
bit
53:58
on
53:58
the
53:58
podcast
53:59
before
53:59
with
53:59
some
53:59
other
54:00
guests
54:00
about,
54:00
like,
54:01
the
54:01
potential
54:02
dangers
54:03
of
54:04
especially,
54:05
like,
54:05
teenagers
54:06
and
54:06
young
54:06
people
54:08
interacting
54:09
in
54:09
increasing
54:10
frequency
54:10
with
54:11
chatbots,
54:12
where
54:13
the
54:13
sort
54:14
of
54:14
psychological
54:15
relationship
54:15
with
54:16
the
54:16
chatbot
54:16
is,
54:16
like,
54:16
this
54:17
thing
54:17
is
54:17
telling
54:18
me,
54:18
like,
54:18
I'm
54:18
incredible.
54:19
And
54:19
so,
54:19
like,
54:20
what
54:20
does
54:20
that
54:20
do
54:20
to
54:20
your
54:20
psyche?
54:21
And,
54:21
like,
54:21
what
54:21
are
54:21
the
54:21
implications
54:22
of
54:22
that?
54:22
And
54:23
I
54:23
hadn't
54:24
thought
54:24
about
54:24
it
54:24
from,
54:25
like,
54:25
a
54:26
vulnerability
54:27
standpoint
54:27
before
54:28
of,
54:28
like,
54:28
if
54:29
something
54:30
can
54:31
be
54:32
steered
54:32
in
54:32
a
54:32
certain
54:33
direction,
54:33
steered
54:34
slowly
54:34
so
54:35
that
54:35
you
54:35
don't
54:35
really
54:36
pick
54:36
up
54:36
on
54:37
the
54:37
changes
54:37
at
54:37
a
54:37
given
54:38
time.
54:38
But
54:39
when
54:39
you
54:39
were
54:39
talking
54:40
about
54:40
that,
54:40
I
54:40
was.
54:40
I
54:41
was
54:41
thinking
54:41
about
54:41
the
54:42
impact
54:43
of
54:43
social
54:43
media
54:44
on
54:44
teens.
54:45
And
54:45
I
54:46
wonder
54:46
if
54:46
sycophancy
54:47
is
54:47
better
54:47
than
54:48
broad,
54:48
universal
54:49
comparisons
54:50
to
54:50
the
54:50
hottest
54:50
people
54:50
in
54:51
the
54:51
world
54:51
all
54:51
the
54:51
time.
54:51
Like,
54:52
which.
54:52
Which
54:52
one's
54:53
worse?
54:54
I
54:54
totally
54:55
agree.
54:56
I
54:56
don't
54:56
know
54:56
the
54:56
answer
54:57
to
54:57
that
54:57
either.
54:58
And
54:58
I
54:58
think,
54:58
I
54:58
imagine,
54:59
like,
54:59
everyone's
54:59
thinking
55:00
about
55:00
this,
55:00
but,
55:00
like,
55:00
how
55:00
do
55:00
you
55:01
teach
55:01
the
55:01
next
55:01
generation?
55:02
Like,
55:02
we're
55:02
aware.
55:03
I
55:04
grew
55:04
up
55:04
before
55:04
the
55:05
social
55:05
media,
55:05
so,
55:05
like,
55:06
I
55:06
know
55:06
the
55:06
world
55:06
before
55:07
social
55:07
media,
55:07
after
55:08
social
55:08
media,
55:08
and
55:08
that
55:09
kind
55:09
of
55:09
thing
55:09
for
55:09
the
55:10
Internet,
55:10
you
55:10
know,
55:10
the
55:11
Internet.
55:11
And
55:11
so,
55:11
like,
55:11
when
55:12
you've
55:12
seen
55:12
the
55:12
two
55:13
things,
55:13
that
55:13
gives
55:13
you,
55:14
like,
55:14
a
55:15
good
55:15
lens
55:15
of,
55:16
like,
55:16
here's
55:16
what
55:17
this
55:17
particular
55:17
thing
55:18
adds
55:18
to
55:18
the
55:19
world
55:19
where,
55:20
like,
55:20
there's
55:20
going
55:20
to
55:20
be
55:20
a
55:20
generation
55:21
that
55:21
grows
55:22
up
55:22
with
55:22
AI
55:22
chatbots
55:23
and
55:23
doesn't
55:24
remember
55:24
a
55:24
world
55:24
before
55:25
AI
55:25
chatbots.
55:26
And
55:26
so
55:26
what
55:26
does
55:27
that
55:27
mean?
55:27
What
55:28
does
55:28
that
55:28
do
55:28
to
55:28
their
55:29
brain?
55:29
I've
55:29
got
55:29
to
55:29
think
55:30
it
55:30
turns
55:30
it
55:30
to
55:30
mush.
55:31
I
55:31
tell
55:31
people
55:32
now
55:32
that
55:33
I
55:33
feel
55:33
really
55:33
lucky
55:34
in
55:34
that
55:35
where
55:35
I
55:36
am
55:36
in
55:36
the
55:36
world
55:37
and,
55:37
like,
55:37
my
55:37
age,
55:38
I
55:38
was
55:38
able
55:38
to,
55:39
like,
55:39
develop
55:39
a
55:39
bunch
55:40
of
55:40
skills
55:41
that
55:41
these
55:41
things
55:41
are,
55:42
like,
55:42
just
55:42
productivity
55:43
multipliers
55:43
as
55:43
opposed
55:44
to.
55:45
It's
55:45
like,
55:45
I'm
55:46
not
55:46
super
55:46
dependent
55:47
on
55:47
them.
55:47
I
55:47
like
55:47
them
55:48
because
55:48
it
55:48
helps
55:48
me
55:48
get
55:48
stuff
55:49
done
55:49
faster.
55:50
But
55:50
like,
55:50
even
55:51
high
55:51
schoolers,
55:52
college
55:52
kids
55:52
now,
55:53
like,
55:53
I
55:53
wonder
55:54
if
55:54
I
55:54
would
55:54
have
55:55
been
55:55
able
55:55
to
55:55
pick
55:56
up
55:56
the
55:56
skills
55:57
that
55:57
would
55:57
allow
55:57
me
55:58
to
55:58
steer
55:59
them
55:59
in
55:59
the
55:59
way
55:59
that
55:59
I
55:59
want
56:00
them
56:00
to
56:00
be
56:00
steered
56:01
in
56:01
this
56:01
age.
56:02
Yeah,
56:02
I
56:02
don't
56:02
know.
56:03
It's
56:03
an
56:03
important
56:04
question
56:04
that
56:05
research
56:06
is
56:06
starting
56:06
to
56:06
show.
56:08
Difficult
56:09
for
56:10
them
56:10
to
56:10
develop
56:11
the
56:11
same
56:11
skills.
56:11
So.
56:12
Well,
56:13
we're
56:14
still
56:14
optimistic,
56:14
everyone,
56:15
about
56:15
the
56:16
future
56:16
of
56:17
the
56:17
world,
56:19
but
56:19
there's
56:19
a
56:19
lot
56:20
of
56:20
important
56:20
stuff
56:20
to
56:20
be
56:21
aware
56:21
of.
56:21
Yeah,
56:22
I
56:22
mean,
56:22
and
56:22
if
56:22
we
56:22
end
56:22
up
56:23
batteries,
56:23
as
56:24
long
56:24
as
56:24
they
56:24
give
56:24
us
56:25
fancy
56:25
VR
56:26
goggles
56:26
and
56:27
headsets
56:27
and
56:27
it's
56:27
immersive
56:28
enough.
56:28
That's
56:29
okay.
56:29
Amazing.
56:30
Hayden,
56:31
thank
56:31
you
56:32
so
56:32
much
56:32
for
56:33
coming
56:34
on
56:34
and
56:34
hanging
56:34
out
56:34
with
56:35
us
56:35
and
56:35
chatting
56:35
today.
56:36
You
56:36
invited
56:36
back
56:37
any
56:37
time
56:37
and
56:38
quickly,
56:38
before
56:38
you
56:38
go,
56:39
actually
56:39
any
56:40
spaces
56:41
on
56:41
the
56:41
Internet
56:41
where
56:41
people
56:42
should
56:42
find
56:42
you.
56:43
Find
56:43
out
56:43
more
56:43
about
56:43
Helavan
56:44
that
56:44
you
56:44
want
56:44
to
56:44
plug
56:45
and
56:45
we'll.
56:45
We'll
56:45
put
56:45
in
56:46
the
56:46
show
56:46
notes
56:46
too.
56:46
Yeah.
56:47
I
56:47
wish
56:47
I
56:47
had
56:47
a
56:48
larger
56:48
presence,
56:48
but
56:49
just
56:49
my
56:49
LinkedIn
56:49
and
56:50
hell.
56:51
Thank
56:51
you.
56:52
Thank
56:52
you,
56:55
Sam.