00:00:03
Uh, I'm Ben Gil. I'm the CEO of
00:00:06
Singularity Net and the
00:00:08
ASI Alliance and True AGI and more other
00:00:14
AI oriented entities that I'm going to
00:00:16
list right now. And uh, I've been
00:00:18
working on building toward AGI really
00:00:22
since the
00:00:25
1980s. you know, the the fun is really
00:00:28
starting now. And I'm James Barrett. I'm
00:00:31
a primarily a filmmaker and uh I I do a
00:00:35
lot of stuff for National Geographic and
00:00:37
and PBS and uh affiliates in America and
00:00:41
around the world. But I'm also an author
00:00:43
and I years ago I wrote a book called
00:00:45
Our Final Invention: Artificial
00:00:47
Intelligence and the End of the Human
00:00:48
Era. And it was during that writing of
00:00:51
that book that I had the pleasure of
00:00:53
meeting Ben
00:00:55
Girtzel. I was uh interviewing a bunch
00:00:59
of experts for upcoming book and many of
00:01:01
them said, "Isn't it ironic that we're
00:01:03
plunging headlong into this technology
00:01:06
and nobody can really explain how it
00:01:07
works?" I'm talking about generative AI.
00:01:10
Um a lot of very prominent people have
00:01:13
said we really really don't know how it
00:01:15
works. So why should we spend billions
00:01:17
of dollars and uh expose ourselves to to
00:01:22
a lot of threats over technology that we
00:01:25
really don't
00:01:26
understand interrogating the particulars
00:01:30
of LLMs and such is it's interesting
00:01:33
it's important to do I don't think it's
00:01:36
exactly to the point regarding the first
00:01:39
AGIS that that
00:01:41
said I think their
00:01:44
incomprehensibility
00:01:45
is a bit exaggerated actually. I mean I
00:01:49
mean we it's true we don't know exactly
00:01:51
what output they're going to give when
00:01:52
given a certain query but I mean we do
00:01:56
we can probe inside transforming neural
00:01:58
nets and and we design new versions all
00:02:02
the time and there's a lot of other
00:02:04
important things in technology that are
00:02:08
not fully comprehensible to us in
00:02:10
different ways like quantum mechanics is
00:02:12
very hard for folks to understand and
00:02:14
there's a lot of trial and error
00:02:16
involved in designing quantum based
00:02:18
machinery as as well as math. We don't
00:02:20
understand how the immune system works.
00:02:22
Like every vaccine discovered has been
00:02:24
discovered by trial and error. That's
00:02:26
just very often how science works,
00:02:30
right? Existential risk is a big reason
00:02:34
why we should stop not not forever, but
00:02:36
pause AI. Could AI become unmanageable,
00:02:40
uncontrollable, misaligned with human
00:02:42
values? Right now, uh AI is not aligned
00:02:46
with human values. I think that people
00:02:48
that are very deep into it as you are
00:02:51
don't understand that we can stop and we
00:02:53
have a history of stopping technologies
00:02:55
that aren't really uh that are very very
00:02:57
dangerous. Um I don't know that we have
00:02:59
a history of stopping technologies that
00:03:01
are as widespread and easy to do and
00:03:05
delivering so much economic value. I'm
00:03:08
not sure that's true. I mean like
00:03:10
nuclear bombs just don't have or
00:03:13
biological weapons don't have the
00:03:17
positive uses and immediate economic
00:03:20
benefit that that AI has. Well uh
00:03:23
asbestous was a darn good insulator.
00:03:26
Chloro fluorocarbons are you know were
00:03:29
incredibly effective. AI is vastly more
00:03:34
widespread. It can help countries with
00:03:36
military superiority. It can help it can
00:03:39
help every big
00:03:40
company make more money. It can
00:03:42
transform every domain of ind industry.
00:03:46
Right? So it's it's it's not as
00:03:49
isolated a
00:03:52
thing. Transformer neural net is simply
00:03:56
a
00:03:57
large predictor which is trained on a
00:04:00
bunch of data like given a sequence of
00:04:03
things which could be words in many
00:04:05
cases. given a sequence of things, make
00:04:07
a guess as to what will the next thing
00:04:09
be in the sequence, right? And so people
00:04:11
train these sequence predictors on like
00:04:14
a huge proportion of the text on the
00:04:16
internet and they became quite good at
00:04:19
predicting the next token in the
00:04:21
sequence, the next word in the sentence,
00:04:23
the next the next program command in the
00:04:25
program, right? And they just do that by
00:04:28
training them by sort of reinforcement
00:04:30
over and over and over again on a lot of
00:04:32
data. So then why the why why these
00:04:35
things make the next predict why these
00:04:37
things predict the next word is 'the' or
00:04:39
the next item in the code is a semicolon
00:04:41
is a complex story involving many many
00:04:44
billions of nodes and links on an
00:04:46
internal neural net which is the part we
00:04:50
don't know you don't know exactly why it
00:04:52
made that prediction and you can try to
00:04:56
make systems more and more sort of
00:04:59
deterministic in what they do but I I
00:05:03
think and there's a bunch of math that
00:05:04
would back this up that we can't go
00:05:06
into, but I think
00:05:08
that's got severe limitations, but I
00:05:11
mean you can you can put what are called
00:05:13
probes in the inside of the network to
00:05:14
try to measure measure what it does to
00:05:16
give a certain output. You can make some
00:05:18
progress there, but by and large as we
00:05:21
make things smarter and smarter, they're
00:05:23
getting it's getting harder and harder
00:05:24
to tell a story about what happened
00:05:27
inside. they may be getting more and
00:05:30
more uh deceptive that you know uh
00:05:33
neural nets are notoriously bad liars.
00:05:36
Well, they're notoriously good liars.
00:05:38
They lie. They lie a lot. I mean, a
00:05:39
neural net is a very flexible
00:05:41
technology. So, that's sort of like
00:05:43
saying computer programs are bad liars.
00:05:47
Now the actual world situation of course
00:05:51
is quite different because none of us is
00:05:54
in charge of the global economy and
00:05:58
military situation. Right? So from the
00:06:01
standpoint of an individual or group
00:06:04
developing AI right now like the
00:06:07
question for me or for meta or
00:06:10
10-centent even the question isn't
00:06:13
really what should the human species as
00:06:16
a whole be doing because none of us is
00:06:19
in no one is in control of the world
00:06:21
right the question is more given what
00:06:24
everybody else is doing on the planet
00:06:27
right now like what's the most
00:06:29
beneficial thing for me as a as an actor
00:06:33
in that network to be to be doing. So I
00:06:36
mean for me
00:06:38
personally I feel like trying to develop
00:06:42
AGI which
00:06:43
is beneficial and compassionate and is
00:06:48
guided and controlled in a participary
00:06:51
way by a broad global network of people.
00:06:53
I feel that injecting that into the
00:06:57
situation has a better chance of
00:07:00
achieving good than me, you know,
00:07:03
sitting back and playing music or
00:07:06
working on on math theory right now. My
00:07:09
point is that you're doing good things
00:07:10
and you're the kind of a AGI person we
00:07:13
need. Uh but you're not the kind of AGI
00:07:15
person we have. Uh we have a few, but
00:07:18
mostly we have um people that really
00:07:21
don't don't care about about human
00:07:23
suffering. And so they're causing it. Uh
00:07:26
they don't care about about liberty.
00:07:28
They don't they uh they don't care that
00:07:31
we might lose control of these systems.
00:07:34
They're not thinking about that. What
00:07:35
they're thinking about are dollars. Boy,
00:07:37
every time we pull out a new a new shiny
00:07:40
object, people will go for it. and
00:07:42
they'll go for it no matter what the
00:07:43
consequences are, especially if it pays
00:07:45
money or kills their enemies. And we're
00:07:48
just flawed that way. And until we're
00:07:50
until we've got a way to unflaw
00:07:52
ourselves or to to think about this in a
00:07:54
productive way, I think we should just
00:07:56
lay it down and and
00:07:58
and continue to research it, but but not
00:08:01
use it. I mean the people I know who
00:08:05
used to work for me who were in the team
00:08:07
at Google developing transformer neural
00:08:10
nets like those guys are just heads down
00:08:14
thinking about how do you predict the
00:08:16
next token right and they're just
00:08:17
thinking about the computer science and
00:08:20
software and math math problems of it
00:08:23
and they they certainly they certainly
00:08:26
are not doing a moral calculation of
00:08:29
like will this technology be used to
00:08:31
kill people or How much will it be used
00:08:33
for good? But they should but they
00:08:35
should.
00:08:37
I I have done that sort of calculation
00:08:39
which is part of the reason I haven't
00:08:40
taken a big tech job in my life in spite
00:08:42
of plenty of lucrative offers. Right.
00:08:44
But but I think they're not thinking
00:08:46
about it that way.
00:08:50
People assume that AGI will free us from
00:08:52
from h from drudgery.
00:08:55
But will it if it if it frees us from
00:08:58
judgery and our and our boring boring
00:09:00
jobs, will it also free us from purpose,
00:09:02
autonomy, and responsibility? I there's
00:09:05
nothing worse to me than having some
00:09:07
something else do my job. And I I'm sure
00:09:09
a lot of people feel that way. To me,
00:09:12
it's pretty clear when you look at how
00:09:15
human motivations work and everything we
00:09:17
know about human psychology. I mean I I
00:09:20
think if we were to succeed in creating
00:09:24
an AGI which is compassionate and
00:09:27
beneficially oriented toward humans
00:09:30
which is what I'm working on with my
00:09:31
colleagues at Singularity ASI alliance
00:09:34
true AGI and so on and if that system
00:09:36
then it proved itself towards super
00:09:39
intelligence keeping compassion and
00:09:41
benefit in mind as it drives this
00:09:44
process like if this succeeded and we
00:09:46
got a beneficial AGI I mean I I have
00:09:49
little doubt that after a transition
00:09:53
period humans would adapt to it and
00:09:56
would find plenty of purpose, meaning
00:09:59
and joy in life going forward in that
00:10:02
situation. I mean we can we can create
00:10:05
art. We can do science and math for it
00:10:08
for its its own sake. We can explore the
00:10:10
world and climb up mountains. We will
00:10:13
still be engaged in our own social
00:10:15
network and chasing girlfriends and
00:10:17
boyfriends and playing sports against
00:10:19
each other. Like there's all manner of
00:10:21
ways we can find purpose and meaning and
00:10:24
have our own will and autonomy within
00:10:26
our social network of humans and and our
00:10:29
creations and interactions. These are
00:10:31
subtle problems, but um universal basic
00:10:35
income and what we do when AGI is here
00:10:37
and we're all, you know, unemployed is
00:10:40
someday will be very relevant and it's
00:10:42
relevant to a bunch of people right now.
00:10:44
Um I I think it's relevant because uh I
00:10:48
think it's a it's a miscalculation. I
00:10:50
think if you pay everybody to do
00:10:52
nothing, uh you'll get you'll get very
00:10:54
very very unhappy people. Um, I think a
00:10:58
lot of people get a lot of meaning and
00:11:00
value and pleasure out of work, the work
00:11:03
that they get paid for. Um, and I think
00:11:06
if you I think if you take that away and
00:11:07
you just say, you know, you just be you
00:11:09
just be you, you be you. And I don't
00:11:12
think you'll find a lot of suddenly
00:11:14
there'll be a lot more poets and a lot
00:11:15
more authors. I think there'll be, you
00:11:17
know, probably a lot more drug drug
00:11:19
users, uh, a lot more alcoholics. there
00:11:22
will be a hell of a lot more musicians
00:11:24
because almost all musicians I know have
00:11:25
day jobs and they'll be really happy
00:11:27
just to play music instead. Right.
00:11:29
That's true and that's a good thing.
00:11:31
More musicians you can always use.
00:11:35
Well, one thing we agree on is the the
00:11:38
bulk of resources in the AI field is not
00:11:42
explicitly pushing toward a beneficial
00:11:44
singularity. Right. It's like pushing We
00:11:46
agree on that. Yes. I get more power or
00:11:48
money for myself and then the
00:11:50
singularity will sort of happen as a
00:11:53
side effect and that's pretty ridiculous
00:11:57
from a bigger bigger point of view.
00:11:59
Yeah. How many how many billionaires do
00:12:01
we need?
00:12:03
Yeah. Well, we don't we don't need any
00:12:05
billionaires. That that's right. On the
00:12:07
other hand, the capitalist economy has
00:12:11
also led to a lot of wonderful things.
00:12:13
Right. So that's the the world that
00:12:15
we've selforganized. Yeah, it's
00:12:17
complicated. Really complex. And I think
00:12:20
I think we we are out of out of time
00:12:22
now. But Well, it was a pleasure. It was
00:12:25
a pleasure talking to you, Ben. I look
00:12:27
forward to talking to you again before
00:12:29
long.
