Episode Transcript
Transcripts are displayed as originally observed. Some content, including advertisements may have changed.
Use Ctrl + F to search
0:00
Hi listeners. This is the eighty thousand hours podcast
0:02
where we have unusually in-depth conversations about the
0:04
word's most pressing problems, what you can do to solve
0:06
them, and what to do if your robot dog tells
0:08
you they're conscious. I'm Rob Woodland,
0:10
Head of Research in eighty thousand hours. Do
0:13
you like the show, but wish there were more of it?
0:15
If so, I have some good news for you. The
0:17
Wizard GPT has recently joined the podcasting
0:19
team at eighty thousand hours as a second host,
0:21
which means we should be able to talk with
0:23
more people about world's most pressing
0:26
questions and problems than we've ever
0:28
been able to manage before. You might
0:30
remember Louisa from episode one hundred and sixteen.
0:32
When she just recently started working here
0:34
as a research analyst and came on as a guest
0:36
to talk about why some global catastrophes
0:39
seemed unlikely to cause human extinction.
0:41
When Kew and I decided we wanted to grow the
0:43
team, Louisa was the person we were most excited
0:46
to work with. So if you're a fan of the eighty thousand
0:48
hours podcast, who joining us should definitely
0:50
be cause for celebration. Today's interview
0:52
is Louise' first time in hosting chair
0:54
interviewing the philosopher Rob Long on
0:56
the question of machine consciousness. Is
0:59
there something that it's like to be a large language
1:01
model like chest GPT? How could
1:03
we ever tell if there was? To what extent
1:05
does the wage chat TBD processes information
1:08
resemble what we humans do? Why
1:10
might future machine consciousnesses have a much
1:12
wider range of emotional experiences than humans
1:14
are capable of? And is the bigger risk
1:16
that we end up thinking AI is conscious when
1:18
it's not? Or that we think it isn't when actually
1:21
it is? Those are the sorts of questions
1:23
that Louisa puts her up. For the
1:25
first time in a while, I got to enjoy listening
1:27
to this episode more like typical subscriber
1:29
who hadn't just done a whole lot background research
1:31
on just that topic. And as a result,
1:33
I felt like I was actually learning a ton about
1:35
this really important issue that I hadn't yet had
1:37
any reason to think much about. If
1:40
Louisa can do interviews this good right off the bat,
1:42
both you and I have much to look forward to.
1:44
After they finish talking about AI, Louisa
1:46
and Rob kept going and recorded a conversation
1:49
for our other show, eighty k after hours.
1:51
This time about how to make independent research
1:53
work more fun and
1:54
motivating. Challenge that both of them have
1:56
had to deal with themselves over the years. You
1:58
can find that forty minute conversation by subscribing
2:00
to ADK after hours in any podcasting up
2:03
or clicking the
2:03
link in the share notes. Alright. With that
2:05
further ado, I present the Weezer GPT and
2:07
blah blah.
2:20
Today, I'm still you, miss Robert Long. Rob
2:23
is a philosophy fellow with the Center for
2:25
AI Safety, where he's working on philosophical
2:27
issues of aligning AI systems with
2:29
human interests. Until recently, he
2:32
was a researcher at the Future of Humanity
2:34
Institute, where he led the Mines
2:37
GPT group, which works on AI conscious ness and
2:39
other issues related to artificial minds.
2:41
Rob studied social studies at Harvard and
2:44
has a master's in philosophy from Brandeis
2:46
University and PHD from NYU
2:49
During his PHD, he wrote about philosophical
2:51
issues in machine learning under the supervision
2:54
of David Chalmers, who listeners might remember
2:56
hearing on our show before. On
2:59
top of that, I'm very privileged to call
3:01
Rob one of my closest friends. But
3:03
somehow, in spite of being very
3:05
good friends, Rob and I have actually
3:07
never talked much about his
3:09
research, so I'm really excited
3:11
to do this today. Thanks for coming on the podcast,
3:13
Rob.
3:14
Thanks so much, Lisa. I'm really excited
3:16
to talk with you. Well, I'm excited
3:18
to talk about how likely AI systems
3:20
are to become sentient and
3:22
what that might look like and kind of
3:24
what it would mean
3:25
morally. But first, what are you working
3:27
on at the moment and why do you think it's important?
3:30
Yeah. This is a great question for the
3:32
beginning of the year. I've been working on
3:34
a variety of stuff related to consciousness and
3:36
AI. So one I'm especially excited about
3:39
right now is me and a colleague
3:41
at Future Command Institute, Patrick Butler,
3:44
have been working on this big multi
3:46
author report where we're getting a bunch of
3:49
neuroscientists and AI
3:51
researchers and philosophers together to
3:53
produce a big report about what
3:55
the current scientific evidence is about
3:57
sentience and current and near term AI systems.
4:00
I've also been helping Jeff Siebel with the
4:02
research agenda for a very exciting new center
4:05
at NYU called the Mind Ethics -- Cool.
4:07
-- and Policy Center. Diane, cool. And,
4:10
yeah, just to keep myself really busy.
4:12
I'm also really excited
4:14
to do kind of a technical sprint
4:16
on leveling up my skills in
4:19
machine learning and AI safety.
4:21
That's something that's, like, predominantly on my
4:23
to do list. And I've always been kind of
4:25
technical AI safety
4:26
curious. So that's kind of a big
4:28
change for me recently. It's also shifting more
4:30
into that. Oh, wow. Cool. Okay. So,
4:33
yeah, I'll probably ask you more about that later, but
4:35
it sounds like on top of AI
4:37
sentient and AI consciousness, you
4:40
you're like, let's add AI safety to the mix
4:42
too. How can I solve
4:43
that? Yeah. To be clear, I do see them as
4:45
as related. You're gonna think about a lot of the
4:47
same issues and need a lot of the same technical
4:49
skills think clearly about both of
4:51
them. Okay. Well, we'll come back to that.
4:53
Yeah. To start, I wanted to ask,
4:56
yeah, a kind of basic question. I
4:59
basically don't feel like I have a great sense
5:01
of what artificial sentience would even
5:03
look
5:03
like. Can you help me get a picture of what
5:05
we're talking about here? Yeah. I mean,
5:07
I think it's absolutely fine
5:09
and correct to not know what it would look like.
5:12
In terms of what we're talking about, I think the
5:14
short answer or like a short hook
5:16
into it is just think about the problem of animal
5:18
science. I think that's structurally very
5:21
similar. So we share
5:23
the world with a lot of non human animals and
5:26
they look a lot different than we do. They
5:28
act a lot differently than we do. They're
5:30
somewhat similar to us. We're made
5:33
of the same stuff. They have brains. But
5:35
we often face this question
5:37
of as we're looking at a B
5:39
going through the field. Like, we can tell
5:41
that it's doing intelligent behavior, but we also wonder
5:44
Is there something it's like to be that
5:46
be? Like and if so, what are his experiences
5:48
like? And what would that entail for how we should,
5:50
like, treat bees or try to share the world
5:52
with bees? I think the general
5:55
problem of AI sense is that
5:57
question and also harder.
5:59
So I'm thinking of it in terms of
6:01
there's this kind of new class of
6:04
intelligent or intelligent seeming complex
6:06
system. And in addition to
6:08
wondering what they're able to do and how
6:10
they do it. We can also, I think, wonder
6:13
if there is or will ever be something that
6:15
it's like to be them. And if they'll have experiences,
6:18
if they'll have something like pain or pleasure, It's
6:20
a natural question to occur to people. And
6:22
it's it's occurred to
6:23
me. And I've been trying to work on it in
6:25
the past couple of years. Yeah.
6:27
I guess I have almost even
6:29
more basic question, which is like
6:32
yeah. When we talk about ASENTIENTS, both
6:35
kind of in the short term and in the long term,
6:38
are we talking about like a
6:40
thing that looks like my laptop that
6:42
has like a code on it that like
6:44
has been coded to have
6:47
some kind of feelings or experience?
6:50
Yeah. Sure. I think I use the term
6:52
artificial sentence. So, like, very
6:54
generally, it's just like things that
6:56
are made out of different stuff than us,
6:58
and in particular, silicone and,
7:00
like, the computational hardware
7:03
that we run these things on. Could
7:05
things built out of that and running
7:07
computations on that have experiences?
7:09
So, like, the most straightforward case to imagine
7:11
would probably be a robot because there,
7:14
you can kind of clearly think about
7:16
what the the physical system is that
7:18
you're trying to ask if it's sentient. Things
7:20
are lot kind of more complicated with
7:23
the more disembodied AI systems of
7:25
today, like, chat, UPT. Because
7:27
there, it's like a it's like
7:29
a virtual agent in a certain sense.
7:32
And brain emulations would also be
7:34
like virtual agents. But I think
7:36
for all of those, you can ask at some
7:38
level of description or some way of carving up the
7:40
system
7:41
Like, is there any kind of subjective experience
7:43
here? Is there consciousness here? Is there sentient
7:45
here? Yeah. Yeah. Cool.
7:47
Jumping in quickly to distinguish
7:50
between what we're calling phenomenal
7:52
consciousness or consciousness in
7:54
this episode, which is basically the
7:56
experience of having subjective experience
7:59
as opposed to something like,
8:01
I don't know, blood pumping through
8:03
my body, like, that's happening,
8:05
but I'm not subjectively conscious
8:08
of it. In contrast with, I
8:10
don't know, like, the feeling
8:13
of the sun on my
8:15
face or something, which I can't
8:17
have a subjective experience of. And
8:19
then we're also using the
8:21
term sentence. And when we say
8:23
sentence, we mean having
8:26
either positive or negative experiences. So
8:29
it's a type of conscious experience
8:31
that's in particular either
8:33
positive or GPT, like, pain or pleasure?
8:37
Yeah. I guess the reason I'm asking is
8:39
because yeah. I think I just have, like, for a long
8:41
time, had this sense that, like, when
8:44
people use the term digital minds or artificial
8:47
sentience. I have, like, some
8:49
vague images that kinda come from
8:51
sci fi. But I mostly feel
8:53
like I don't even know what we're talking about.
8:56
But it sounds like it could just look like
8:58
a bunch of different things and the
9:00
like core of it is something
9:02
that is sentient in maybe a way
9:05
similar, maybe a way that's pretty different to humans,
9:07
but that exists not
9:09
in biological
9:11
form, but in made up in
9:13
some in some grouping that's made up
9:15
of silicone. So
9:16
that should be right. And I should say, I guess,
9:18
like, silicon is not, like, that deep
9:21
here.
9:22
Sure. Sure. Yeah. Something having to do
9:24
with, like, running on computers, running on GPUs.
9:27
I'm sure I could slice and dice it and we
9:29
you could get in all sorts of philosophical, like,
9:32
classification terms for things. But,
9:34
yeah, that's the general thing I'm pointing at.
9:36
And I, in particular, have been working on
9:38
the question of AI systems. So
9:41
the questions about, like, whole brain emulation, I
9:43
think, would be different because we would
9:45
have something that at some level description
9:47
is extremely similar to the human brain
9:50
by definition. And then you could wonder about
9:52
whether it matters that it's an emulated
9:54
brain, and people have wondered
9:56
about that. In the case of AI is,
9:58
you know, even harder because not only are
10:00
they made on different stuff and
10:03
maybe somewhat virtual. They
10:05
also are kind of strange and
10:07
not necessarily working along
10:09
the same principles as the human brain.
10:12
Right. Right. Okay. That makes sense. I've
10:14
heard the case that if there are AI
10:16
systems that become sentient, there's
10:18
a risk of creating kind of astronomical
10:21
amounts of suffering. I still
10:23
have a really hard time understanding what
10:25
that might concretely look like.
10:28
Can you give, yeah, a kind of concrete example
10:30
scenario where where that's the case?
10:32
Yeah. So before getting to the, like, astronomical cases,
10:35
I'll start with more concrete case,
10:37
maybe of just one system. So you can
10:39
imagine that a robot has been created by
10:41
a company or by some researchers And
10:45
as it happens, it registers damage
10:47
to its body and processes it
10:49
in the way that as it turns out, is
10:51
relevant to, like, having an experience of
10:54
unpleasant pain. And maybe we don't realize
10:56
that because we don't have good theories of what's going
10:58
on in robot or what it takes to kill pain.
11:00
In that case, you can imagine that
11:02
thing having bad a bad
11:04
time because we don't realize it.
11:07
Right. You could also imagine this
11:09
thing being, like, rolled out, and now
11:11
we're economically dependent on systems
11:13
like this. And now we have an incentive
11:16
not to care and not to think
11:18
too hard about whether it might be having a bad
11:20
time.
11:20
Yeah. Yeah. So,
11:21
I mean, that seems like something that could happen.
11:24
Yeah. And that could happen because
11:27
I mean, there's some reason why it's helpful to
11:29
have the robot GPT that it's sustained
11:31
damage. It can, like, be,
11:33
like, help I've broken, I need someone
11:35
to fix my part. So that's, like, something
11:37
that you can like, might get programmed
11:40
in, and then, like, It is
11:42
just kinda wild to me that, like, we don't understand
11:45
what the robot might be experiencing
11:48
well enough to know, like, that thing
11:50
is pain. But, like, in dairy,
11:52
that's possible. Just like they're that
11:54
it's kind of that black boxy to
11:55
us. Yeah. So it might be little bit
11:57
less likely with a robot. But now
12:00
you can imagine more abstract
12:02
or alien ways of feeling bad. So
12:05
I focus on pain because it's like very
12:07
straightforward way of feeling bad. Yeah. A
12:09
disembodied system like a GPT
12:11
three, which we'll talk about. Obviously, can't
12:13
feel ankle pain or almost
12:16
almost certainly. Like, that'd be really weird. Doesn't have an
12:18
ankle. Right. Why would they have computations
12:20
that, like, representatives' ankle is feeling
12:23
bad? Mhmm. But you can imagine maybe
12:25
some strange form of balanced
12:27
experience that develops inside some system
12:29
like this that registers some kind of displeasure
12:32
or
12:32
pleasure, something like that.
12:33
Right. Right. Something like
12:35
And that could give you the wrong
12:38
set of words to come next,
12:40
and that was bad. And the
12:42
user isn't happy with
12:43
the, like, string of words he came up and then
12:45
that feels something like pain. Exactly.
12:47
Yeah. And I will note that
12:50
I don't think that getting
12:52
negative feedback is going to be enough
12:53
for, like, that bad feeling. Fortunately.
12:56
Yeah. But maybe some combination
12:58
of that and some way it's ended up representing
13:00
it inside itself ends up like
13:02
that. And then, yeah, then we have
13:05
something where it's hard for us
13:07
to map its internals
13:09
to what we care about.
13:12
We maybe have various incentives not
13:15
to look too hard at that question.
13:17
We have incentives not to let it speak
13:20
freely about if it thinks
13:22
it's conscious -- Mhmm. -- because, like, that
13:24
would be a big headache. Mhmm. And because
13:26
we're also worried about systems lying
13:29
about being conscious and giving misleading statements
13:31
about whether they're conscious, which they did they
13:33
definitely do. Yeah. So
13:35
we've built this new kind of alien minds
13:38
we don't really have a good theory of pain even
13:40
for ourselves. We don't have a good theory of what's
13:42
going on inside it. And so that's like a
13:44
that's sort of like a stumbling into this. Sort
13:47
of scenario. Yeah. That's not yet astronomical.
13:50
Yeah. So one reason I I started with the the concrete
13:53
case is I think people who
13:55
are worried about risks
13:57
of large scale and
13:59
long term suffering what
14:02
are sometimes called S risks or suffering
14:04
risks? I think they have
14:06
scenarios that involve,
14:08
like, very powerful agents making
14:10
lots of simulations for various
14:13
reasons and the simulations containing suffering.
14:15
I'll just refer people to that work because
14:18
I that's actually not my,
14:20
like, my bag. haven't thought that
14:22
much about those scenarios. Just
14:24
for my interest, what's the basic
14:27
argument for why anyone would wanna
14:29
create simulations with a bunch of suffering in
14:30
them? Yeah. So this is my
14:33
take, and it might not represent They're cheap positions.
14:35
I think one reason you could
14:37
create simulations because you wanna
14:39
learn stuff. So Imagine
14:42
that we were curious how evolution
14:44
would go if something had gone slightly differently.
14:46
Right. Okay. And imagine we had,
14:48
like, planet sized computers. So we could,
14:51
like, just literally rerun, like,
14:53
all of evolution down to the detail so that
14:55
there are, like, virtual creatures --
14:57
Yeah. Yeah. Yeah. -- and reproducing and stuff.
14:59
And also, I suppose that a simulated creature
15:02
is
15:02
sentient, which, you know, is is plausible.
15:05
Yeah. Yeah. Then all your all you really
15:07
are looking for is, like, at the end did the simulation
15:09
output, like -- Right. -- so hominids or
15:11
something. Yeah. Yeah. GPT. You've
15:14
also have, like, billions of years of animals, like,
15:16
eating each other. Totally. Stuff like
15:18
that. Yeah. Okay. Right. But it sounds like
15:20
there's also just, like, we, like, make
15:22
things for, like, economic reasons,
15:24
like robots or like chatbots.
15:27
And we don't realize those things are suffering.
15:30
And then we, like, mass produce them because
15:32
they're valuable. And then that mass production
15:35
isn't astronomical. In
15:37
scale, but it's, like, big and,
15:40
like, those things are suffering. We didn't know it, and
15:42
they're, like, all over. And we don't really wanna change
15:44
anything about those systems because
15:47
we use them. Yeah. I mean, for just
15:49
another dark, dark scenario, you can
15:51
imagine a system where we get pigs
15:53
to be farmed much more efficiently.
15:56
And we're just like, well, this is made
15:58
a meat cheaper. Let's not too much
16:00
about
16:01
that. Totally. Got it.
16:03
Yeah. Yeah. Yeah. Okay. Yeah.
16:05
Are there any other examples you think are plausible
16:07
here, or are those kind of the main
16:08
ones? I guess 146 thing I should note
16:10
is I've been focusing on this, like,
16:13
case where we've hit on it accidentally.
16:15
There are a lot of people who are interested
16:18
in building artificial consciousness.
16:20
Mhmm. On purpose Not understandably so.
16:22
You know, it's, like, just from a purely
16:24
intellectual or philosophical standpoint.
16:26
Fascinating project, and it can help
16:28
us understand the nature of consciousness. So
16:31
for a very long time, probably
16:33
about as old as AI, people are like, wow,
16:35
I wonder if we could make this thing conscious.
16:38
Right. So there was a recent recent New York
16:40
Times article -- Yeah. -- about roboticists
16:43
who Yeah. want to build more
16:45
self awareness into robots, both for
16:47
the intrinsic scientific interests and also
16:49
because it might make for better robots. And
16:52
some of them think, oh, well, like, we're not actually
16:54
that close to doing that and maybe, like,
16:56
yeah, it's too soon to worry about it. Another
16:59
person quoted in that article is, like,
17:01
yeah, it's something to worry about, but, like, we'll
17:03
deal with it. And, yeah,
17:06
I'm I am quoted in that piece as
17:08
just kind of being, like, be
17:11
careful, like, slow
17:13
down.
17:14
Like, we're not really ready to to
17:16
deal with?
17:17
To quote unquote deal with that. Yeah. Yeah.
17:19
Yeah. Yeah. Exactly.
17:20
Okay. So so maybe it happens
17:23
because it's, like, useful for learning. Maybe
17:25
it happens because there are, like, some
17:28
reasons that someone might want to do
17:30
this intentionally to create suffering.
17:33
That's very dark. But then it could also just
17:35
happen accidentally, which yeah.
17:37
All of which kind of terrifies me. And
17:40
I wanna come back to that. But first,
17:42
I wanted to ask about the kind of yeah, flip
17:44
side of this, which is not only
17:46
my AI systems be able to suffer,
17:48
but they might also be able to experience pleasure.
17:51
And I'm curious how their
17:54
pleasure might compare to
17:56
the pleasure that we feel as
17:57
humans. Bauchner: Yeah, the
17:59
short answer is, I think pleasure
18:02
or pain or whatever analogs of
18:04
that that AI systems could experience
18:07
could have a drastically different range than
18:09
ours. They could have a drastically different
18:11
sort of middle
18:12
point. Is there any reason to think the
18:14
default is that artificial
18:17
sentience feels pleasure and pain like
18:19
humans? Or or do you think the default is
18:21
something else? Yeah. I basically am
18:23
agnostic about what the default is. Okay.
18:25
And one reason is that
18:27
Well, let's first think about why the default is what
18:29
it is for humans.
18:30
Yeah. Great. It's a very vexing
18:32
and interesting question. So let's start with,
18:34
I think, one of the saddest facts about life, which
18:36
is that it's much easier to
18:38
make someone feel pain than
18:40
to make them feel really good. Here's
18:42
a dark thought experiment that I actually thought
18:44
about as preparation for this. Suppose
18:47
I'm gonna give you like a billion dollars
18:49
and a team of people who are experts
18:52
in all sorts of
18:52
things. And you have
18:54
the goal of making someone feel as good as possible
18:57
for a week.
18:57
Yep. Or imagine a different scenario
19:00
where I give you the goal of making someone feel
19:02
as bad as possible for a week.
19:03
Yeah. It seems much easier to do the
19:05
second goal. Totally. Right.
19:07
Yeah. That is really exciting. It seems
19:09
like in some ways, it might not really be
19:12
Like, you could still mess up the one week thing.
19:14
It's just like really hard to make people feel durably
19:16
good. Totally. Yeah. And
19:19
the bad is just like waterboard them
19:21
for a week.
19:22
Yeah. You took it there. But yeah.
19:24
Yeah. That's what Jeez. Yeah.
19:26
And, like, why is that the case?
19:28
Like, why are we creatures where it's so
19:30
much easier to make things go
19:32
really badly for us. 146,
19:35
like, line of thinking about this is
19:38
Well, like, why do we have pain and pleasure?
19:40
It has something to do with, like, promoting the
19:43
right kind of behavior to increase
19:45
our genetic fitness. Mhmm.
19:48
That's not to say that that's explicitly
19:50
what we're doing or and we in
19:52
fact don't really have that goal as
19:54
humans.
19:55
Like, it's not what I'm up to. It's not what
19:57
you're up to. Not entirely. Yeah.
20:00
But they should, like, kind of correspond to
20:02
it. And there's kind
20:04
of this asymmetry where it's really easy
20:07
to lose all of your expected offspring
20:09
in one go. If, like,
20:11
something eats your leg, then
20:13
you're, like, really in danger of,
20:16
like, having no descendants. Yeah. Yeah. And I could
20:18
be happening very fast. Uh-huh. In
20:20
contrast, there are, like, very few things that
20:22
all of the sudden drastically increase
20:25
your number of expected offspring. I
20:27
mean, even having sex, which I think
20:29
it's obviously on a coincidence that that's one of the
20:31
most, like, pleasurable experiences for
20:34
many people. Yep. Yeah. Even
20:36
that, like, you know, doesn't hugely
20:39
in any given go increase
20:41
the number of
20:42
descendants. And and did it for, like, eating a
20:44
good meal. Right.
20:47
Right. So if there was something
20:49
that were like, I don't know, some
20:52
some tree that made it possible to,
20:54
like, have twenty
20:56
kids in one pregnancy instead
20:59
of 146. Maybe we'd find eating
21:01
the fruit from that tree, like especially pleasurable.
21:04
But there just aren't that many things like that.
21:06
And so those things don't give us very
21:08
big rewards relative to the things to the many
21:11
things, I
21:11
guess, that could, like, really mess
21:13
up our survival or reproduction?
21:16
Is that basically the --
21:17
Yeah. -- closed? Yeah.
21:19
I actually have never I've just never thought
21:21
about that. It makes perfect sense. Yeah. It's like
21:23
very schematic, but I do think it is like a good
21:25
clue to thinking about these questions. So,
21:27
yeah, like, what what evolution wants for creatures
21:30
is pain and pleasure to, like, roughly
21:32
track those things. I mean, evolution also
21:34
doesn't yeah. It doesn't want you to experience agony
21:37
every time you, like, don't talk
21:39
to a potential mate. Like, it doesn't allow you to be wracked
21:41
with pain. Right. Because, like, that's distracting
21:43
and it takes cognitive resources and
21:45
stuff like that. So, like, that's another piece of it. It
21:47
needs to, like, kind of balance the energy
21:50
requirements and cognitive requirements of that.
21:52
Mhmm. I definitely recommend that readers
21:54
check out work by rethink priorities.
21:57
On trying to think about what the, like, range
21:59
of balanced experiences for different
22:01
animals are based on
22:02
this. Can you give me the rough overview
22:05
of what they try to do? Like, what their
22:07
approach is?
22:08
Yeah. So they're looking at
22:10
considerations based on the
22:13
sort of evolutionary niche that different
22:15
animals are in. Wow. As
22:17
as one thing, like -- Mhmm. -- there are reasons
22:19
to expect differences between animals
22:22
that have different kind of like offspring strategies.
22:25
Right. And then also just
22:27
more direct arguments about,
22:29
like, what are the attentional resources
22:32
of this animal? Like, does it have memory
22:34
in a way that might affect its experiences? Mhmm.
22:37
Here's an interesting 146. do social
22:39
animals have different experiences of
22:40
pain? Because social animals
22:43
it's very helpful for them to cry
22:45
out. Right. They'll get
22:46
helped by her. Yeah.
22:48
Pray animals have an incentive
22:50
not to show pain. Because that
22:52
will sound nice. Fascinating.
22:55
And, like, that might
22:57
just really lead to big differences in
23:00
how much pain or pleasure these animals
23:02
feel?
23:02
I think that's the thought. Yeah.
23:04
That's really cool. It's really fascinating. Yeah.
23:06
I'm sure everyone's seen a kid that
23:08
has fallen over and
23:11
it doesn't freak out until it
23:13
knows that someone's What do you
23:15
do? Oh, got it. Yeah. Yeah. Yeah. Yeah. Yes. Yes.
23:17
True. That's not to say that the pain is different in
23:19
each case. Like, I I don't know. I don't think anyone
23:21
knows, but that's an illustration
23:23
of the social animal kind of
23:26
programming.
23:26
Totally. Totally. Okay. So,
23:29
I guess, by extension
23:31
yeah. You could think that, like, the
23:34
kind of selection pressures that
23:37
an AI system has or doesn't
23:39
have or something about its
23:40
environment. Might affect
23:43
kind of its emotional range? Is that is that
23:45
basically the Yeah. It's something
23:47
like we seem to have some
23:49
sort of partially innate or
23:51
baked in, like, default points
23:54
that we then deviate from
23:57
on either end. It's very
23:59
tough to know what that would mean for an
24:01
AI system. Obviously, AI
24:03
systems have objectives that they're seeking
24:05
to optimize. But it's
24:07
less clear what it is to say it's kind
24:09
of default expectation of how well it's
24:11
gonna be doing such that if it
24:14
does better, it will feel GPT. If it does worse,
24:16
feel bad. I think
24:18
the key point is just to notice that
24:20
maybe, and this could be a very
24:22
good thought. This kind
24:25
of asymmetry between pleasure and pain
24:27
is not like universal law of consciousness
24:29
or something. Got it. Right. Okay? So
24:31
the so the fact that humans have
24:33
this kind of like limited pleasure side
24:35
of things, there's no like inherent
24:38
reason. That an AI system
24:40
would have to have that
24:41
cap. It could have There might be no
24:43
inherent reason we have to have that cap forever,
24:45
which is another wonderful thought. Right.
24:48
This is GPT post by Paul Cristiano, pointing
24:51
out that we're kind of fighting this battle against
24:53
evolution. Evolution doesn't want
24:55
us to find pleasure hacks because
24:58
it doesn't. It doesn't want us to to wire
25:00
head. So, like, that's
25:02
one reason you know, at a high
25:04
level, like, why we maybe habituate to
25:07
GPT. Sorry, wire tooling
25:09
is, like, some pack
25:11
to find pleasure that doesn't actually improve our fitness
25:14
or
25:14
something? Yeah. It means a lot of different things. was
25:16
using it yeah. I was using it loosely to mean that. Okay.
25:19
Yeah, that's maybe why we're always dissatisfied. Right?
25:21
Like, you've got a new job, you've
25:24
got cool friends, like, you know, you've got
25:26
social status, and eventually
25:28
your brain's like more, you know,
25:30
don't get complacent. And,
25:32
you know, we've tried various things to try
25:35
to try to work around that and find
25:37
sustainable ways to boost our well-being
25:39
permanently different cognitive techniques.
25:42
But, like, we're kind of this GPT,
25:45
we're kind of fighting, like, an adversarial game.
25:48
That's really interesting. Yeah.
25:50
And then I guess so I guess it's
25:52
both kind of we don't know where the default point is.
25:55
We also don't know what the upper
25:57
bound and lower bound might be on pleasure
25:59
and pain. It might be similar
26:01
to ours, but many
26:03
of the pressures that might push ours to be what
26:05
they are may or may not exist for an AI
26:07
system, and so they could just be really different.
26:10
Exactly. Cool. Yeah. That's what fun. That's
26:12
wild. Yeah. Are there any other kind
26:14
differences between humans and AI systems
26:16
that might be in AI systems feel kind
26:18
of more or different kinds of pleasure
26:21
than
26:21
humans? Well, yeah. I mean, one thing I'll note
26:23
is that I'm often using
26:25
bodily pain or
26:27
the pleasures of status or something
26:29
as my GPT. Mhmm. But it
26:31
it kind of goes without saying but I'm
26:33
saying it that Yeah. I mean,
26:36
AIs might not have anything, you know, corresponding
26:38
to that. You know? It would be really weird if they
26:40
feel, like, sexual satisfaction at this point.
26:42
Right. Right. Right. You know? Yeah.
26:43
Yeah. Yeah. Yeah. Makes sense. But then it's yeah.
26:46
It's and and you can wonder that we're venturing
26:48
into territory. We don't really know what we're
26:50
talking about. But, like, I think you can
26:52
in the abstract. Imagine valence.
26:54
Yeah. Valence just being a shorthand for, like, this
26:56
quality of pleasure, just pleasure. When I can imagine
26:58
valence, right, least I think I can. That's about
27:01
other kinds of
27:02
things. Yeah. Yeah. Yeah. Like,
27:04
to the extent that there are things
27:06
like goals and rewards
27:09
and other things going
27:11
on that motivate an AI
27:13
system. Maybe those things come with valence.
27:16
And, like -- Yeah. -- maybe they won't. But,
27:18
like, it might make sense for them too. Exactly.
27:20
I
27:21
guess argument I've heard for why
27:23
there might be a difference in kind of the amount
27:25
of pleasure and pain AI systems
27:27
could feel versus humans can feel is just
27:29
something like humans
27:32
require lots of resources right now.
27:34
Like, the the cost of living and the cost
27:36
of thriving and flourishing might just
27:38
be really high. And I
27:40
can imagine it just becoming super
27:43
super cheap for an
27:45
AI system or some kind of
27:47
digital mind feeling just
27:50
like huge amounts of pleasure, but
27:52
not requiring like a bunch of friends
27:54
and like housing. And
27:57
I don't know, romantic relationships. Like,
27:59
maybe it's just, like, relatively small
28:01
computer chips and they
28:03
just, like, get to feel enormous pleasure
28:06
really cheaply by, like, pushing,
28:08
like, the zero key or something. And
28:11
and so you might think that they could just experience
28:14
actually loads more pleasure than
28:16
could at least feel like GPT the same inputs?
28:18
Yeah. 146 thing I'll also note is
28:21
they could also experience the
28:23
higher pleasures cheaply too. Like, suppose
28:25
they do require friends and knowledge
28:27
and community and stuff. Maybe it's just a lot
28:29
cheaper to give that to them
28:30
too.
28:31
Yeah. Yeah. Yeah. Yeah. Right. And then there's
28:33
also cases like you said where maybe they have some
28:35
sort of alien pleasure and we're just
28:37
like turning the dial on that. I mentioned
28:39
the other case because, like, I think a lot of people
28:41
would be wary of finding it valuable
28:44
that you're, like, just, like, cranking
28:46
the dial on maybe some, quote unquote,
28:48
loader or pleasure or,
28:49
like, uninteresting pleasure. But even
28:52
more interesting pleasures could be a lot cheaper.
28:55
Right. Right.
28:55
It's it's it's cheaper for them to achieve
28:58
great things and Yep. contemplate
29:00
the internal truths of existence and have
29:03
friends and stuff like
29:04
that. And that could just be some basic
29:06
thing. Like, it's easier to make
29:08
more silicone things
29:11
than it is to build houses,
29:14
farm food, build cities,
29:16
etcetera. Like, you could just have computer
29:18
farms that, like, allow AI systems
29:21
have all the same experiences and maybe better
29:23
146. But, like, it might just cost
29:25
less. Yeah. That scenario
29:27
is possible. And I will go ahead
29:29
and disclaimer. Like, I don't think
29:31
that much about those scenarios right
29:33
now. Mhmm. And I'm also not, like,
29:35
build the servers, GPT. You
29:37
know? Okay. Okay.
29:39
Yeah. Given how life fraud and in the dark
29:41
we are about these questions, both morally
29:43
and and empirically. Totally. But, yes, I
29:46
I think it is possible. Here's another,
29:48
you know, another Black Mirror episode,
29:50
which I think is maybe my
29:51
favorite, is Santa Clara. Yeah.
29:54
Have you have you seen that one?
29:55
I have. Yeah. Do you wanna recap
29:58
it? Sure. Yeah.
30:00
This one said in the, like, somewhat near future
30:03
And this civilization seems
30:05
to have cracked making realistic
30:08
simulations. And it's possible for
30:10
people to go in those simulations while they're alive.
30:12
It's also possible for them
30:14
to be, like, transferred to them when they
30:17
die. And it's one of the rare black mirror
30:19
utopias. Spoil alert
30:22
before you continue listening. Yeah.
30:25
The, like, protagonist of the episode
30:27
ends up in a very great situation
30:30
at the end of the show. She ends up being
30:32
able to live with this this woman she loves
30:34
and in this cool like
30:36
beach town. And what
30:39
I love about the episode is it ends with this
30:41
happy ending. Of, like,
30:43
digital utopia. And then, like, the last
30:45
shot is this robot arm
30:48
putting
30:49
her little simulation in this huge
30:51
server bank and you see that it's just like this
30:53
entire warehouse of
30:55
-- Right. -- of simulations. Yeah. Yeah.
30:57
Yeah. And and
30:59
why is Did you like that episode? IIIII
31:01
think it's stunning, really moving. Yeah.
31:03
I think my think it's my favorite because
31:06
there's this, like, parody of Black Mirror, which
31:09
is, like, What if phones, but
31:11
bad? Yeah. Yeah. Yeah. Totally.
31:14
And sometimes it does veer into this kind of
31:16
like cheap dystopia, which is
31:18
not to say I'm not worried about dystopias, but --
31:21
Yeah. -- yeah. It's just like, what if
31:23
Facebook that plug directly into
31:25
your brain. Yeah. And
31:28
Holden Karnovsky has a great post about why
31:30
it's hard to depict utopias and hard to
31:32
imagine them in compelling way for
31:34
viewers. Mhmm. And this seems to
31:36
have, at least for me, like, solve
31:38
that problem. I'm, like, that is a it's not
31:41
the best possible
31:41
future, but it's a That's a good one.
31:44
Cool. Yeah. Yeah. Any any other
31:46
differences that you think are, yeah,
31:48
I guess, relevant to
31:50
the kinds of pleasure or the amount of
31:52
pleasure that AI systems might feel
31:54
relative to humans? Yeah, now might
31:56
be a good time to talk about sort of a grab
31:58
bag of complexing issues about artificial
32:00
minds. Right. So there's all
32:02
these philosophical thought experiments about, like,
32:04
what if people were able to split into?
32:07
And you make two copies of them. Which one
32:09
is really them? Or what have two people
32:11
merged? Like, what do we say about that case? And
32:13
I think those are cool thought experiments. Yeah.
32:15
AIs are, like, a lot easier to copy and, like,
32:18
I'm a lot easier to merge. Totally. So
32:21
it could be that we could
32:23
have real life GPT. Of these kind of philosophical
32:25
edge cases and things that have sort of distributed
32:28
selfhood or distributed agency. And
32:30
that, of course, would affect kind of
32:33
how to think about their well-being and stuff. In
32:35
ways that I find very hard to say anything meaningful
32:37
about, but
32:37
it's worth -- Right. -- think
32:38
it's worth flagging and worth people
32:40
thinking about.
32:41
Totally. Right? So with with
32:43
copies, it's something like does each copy
32:45
of an identical, I guess,
32:48
digital mind GPT, like,
32:51
equal moral weight? Like,
32:53
are they different people? And
32:55
do they get if they're, like, both happy?
32:57
Is
32:58
that, like, twice as much happiness in the world?
33:00
Yeah. I mean, I'm inclined to think I'm
33:02
inclined to think too. Yeah.
33:04
Like, there's a paper by Showman
33:07
and Bostrom college sharing the world with digital minds.
33:09
Mhmm. And, yeah, that thinks
33:11
about a lot of the sort of, like, political
33:14
and social implications of cases like this,
33:17
who which yeah. I I haven't
33:19
thought that much about myself, but there
33:21
would be, like, you know, interesting questions about,
33:23
like, the political representation of
33:25
copies. Like,
33:27
before there's some vote in
33:30
San Francisco, we wouldn't want me to be
33:32
able to just make twenty of me and then we'll also
33:34
vote. Right?
33:34
Totally. Yeah. Yeah. Yeah. I
33:37
mean, I don't know if there are twenty
33:39
of you. And do you You also
33:41
don't wanna disenfranchise someone back. Well, you're
33:43
just a copy. So, like, you know, your vote
33:45
now counts for one twentieth
33:48
as much. Yeah. Yeah. Yeah. I mean, do you
33:50
have a view on this? I I think I do have
33:52
the intuition that, like, have
33:54
the intuition that it's bad, but I think
33:56
when I look at it, I'm like, well,
33:58
no. There are just twelve robs who are gonna
34:01
get twelve Rob's worth of joy from,
34:03
like, a certain electoral
34:05
outcome. And, like, that's bad if, like,
34:07
there are only twelve Rob's because you're really rich
34:09
But I don't, like, hate the idea that there might
34:11
be more robs and that you might get twelve
34:13
more robs worth of votes.
34:15
And, yeah, I mean, I don't have strong views
34:17
about this, like, hypothetical of copying
34:19
and political representation. Yeah. But it does
34:22
seem like you would probably
34:24
want rules about when you're allowed copy. Because
34:26
in the run up to an election, you don't want an
34:28
arms race where the population
34:30
of the digital population of San Francisco
34:32
skyrockets because everyone 146
34:35
they're preferred candidate to
34:36
win. Yeah. Yeah. Yeah. I guess
34:38
also if you have to, like, provide
34:41
for your copies, if you have to, like,
34:43
split resources between your copies, you
34:46
might even kill your copies afterward.
34:48
Like, you might delete them because you're, like,
34:51
can't afford all these copies of myself. Yeah.
34:53
Thanks for the thanks for the vote. Thanks for the vote.
34:55
But, of course, if if I feel that way then but
34:58
necessarily all the copies do as
34:59
well. So they feel like
35:01
they also don't wanna share resources and are
35:03
happy to let one of you live.
35:06
You
35:06
mean? Well,
35:07
they're certainly not gonna be deferring
35:09
to the quote unquote original me because they
35:11
all feel like the original
35:13
me. Right. Right. Right. Right. Right. And so --
35:15
Yeah. --
35:15
so the eleven that let's
35:17
say the original u does keep
35:20
power
35:20
somehow. It like somehow has the power
35:22
to delete the other copies. And
35:25
Yeah. They'll all feel like the original
35:27
me. That's the that's another thing. Right? Well, they
35:29
would feel like it, but -- Yeah.
35:31
-- they might not actually be able to click
35:33
the button. To delete the
35:34
copies, but, like, maybe they originally you can.
35:36
Right. Yeah. Yeah.
35:37
And then that's you're murdering
35:40
eleven people. I
35:41
I mean, not me, you know. I I wouldn't do this. But
35:43
You might do that. You might do that. You would be
35:45
nurturing. I'm planning right now. I'm scheming.
35:48
I'm like, oh, sounds like a
35:50
great way to get the electoral the election
35:52
outcomes I want. Yeah.
35:55
How much does emerging data experiment apply
35:58
or like how relevant is
35:59
it? I guess I mostly mentioned the merging
36:01
case because it's like part
36:03
of the canonical battery
36:06
of thought experiments that are
36:08
supposed to make personal identities seem
36:10
a little less deep or kind of
36:13
perplexing if you really insist on there always
36:15
being some fact of the matter about which
36:17
person's exists and not -- Yeah. -- and
36:19
just like splitting, it's like something
36:22
that seems like it could
36:23
happen.
36:23
Yeah. Yeah. Yeah. Okay. So maybe
36:25
you after this election
36:28
try to merge your eleven copies back
36:30
with
36:30
yourself. And then what does that what
36:32
does that mean?
36:33
Yeah. Like, does that thing now still deserve
36:36
twelve votes or something? Right. Right. Right. Yeah.
36:38
Yeah. Okay. Interesting. Yeah.
36:40
I've I've I've never thought about that before.
36:42
So I guess, I feel like there are some
36:45
reasons to think that AI
36:47
systems or or I guess digital minds
36:49
more broadly. They might have
36:52
more capacity for suffering, but they might
36:54
also have more capacity for pleasure.
36:57
They might be able to kind
36:59
of experience that pleasure more cheaply than
37:01
humans. They might have,
37:03
like, a higher kind of pleasure
37:06
set point. So, like, on average,
37:08
they might be better off. Yeah. I GPT,
37:11
you might think that, like, it's
37:13
more cost effective. You can,
37:15
like, create happiness and well-being more
37:18
cost effectively to have a bunch
37:20
of digital minds than to have a bunch
37:22
of humans. How how do
37:24
we even begin to think about kind of what the moral
37:26
implications of that are?
37:28
Yeah. So I guess I will say but not
37:30
endorse the, like, one flat footed
37:32
answer. Okay. And this can go in, like, you
37:34
know, red letters around this. Like Yeah.
37:36
Sure. Yeah. You you could think.
37:39
Like, let's make the world as good as possible and
37:41
contain as much pleasure and as little pain
37:43
as possible. And like,
37:46
we're not the best systems for realizing
37:48
a lot of that. So our
37:51
job is to, like, kind of usher in a,
37:53
like, successor my GPT. Can experience
37:55
these these GPT? I think there are many
37:57
many reasons for not, like being overly
37:59
hasty about such a position. And,
38:01
like, people who've talked about this have have noticed
38:04
this. I mean, one is the in practice,
38:06
like, we're likely to face a lot of uncertainty
38:08
about whether we are actually creating something valuable
38:11
that, like, on reflection we would endorse. Yeah.
38:13
Yeah. Another one is that, you
38:15
know, maybe we have
38:17
the prerogative of just caring about
38:20
the kind of goods that exist
38:22
in, like, our current way
38:24
of existing. So, like, one thing
38:26
that that sharing the world with digital
38:28
minds mentions is that there
38:30
are, like, reasons to maybe look for some sort
38:32
of, like, compromise. Yeah.
38:33
Can you explain what that would look like? Yeah.
38:35
One extreme position is, like, the hundred
38:38
percent just replace some handover position.
38:40
And then The other extreme would be That's, like,
38:43
all of humans just like
38:45
decide voluntarily to give up
38:47
their stake in the resources
38:49
in the world. And they're just like digital minds
38:51
will be happier per tree
38:53
out there. And so let's
38:55
give them all the trees and all the all the
38:57
things. And -- Yeah. --
38:59
and we're just like Our time is done.
39:01
Yeah. Like,
39:02
cool. take it from here. Yeah.
39:04
And then there would be the
39:06
other extreme would be, like, no
39:08
humans forever. No trees
39:11
for the digital minds. Mhmm. And maybe
39:13
and so, like, maybe for that reason, don't build
39:14
them. Like, let's just stick stick
39:17
with what we know. Mhmm. Don't build artificial
39:19
sentence or or don't build, like,
39:21
a utopia of kind of digital
39:23
minds.
39:24
Yeah. A utopia that's, like, too different from --
39:26
Yeah. I have an experience. Then one thing you
39:28
might think is that you could get a lot
39:31
of what each position wants
39:33
with some kind of split.
39:35
So if the, like, pure replacement
39:38
scenario is motivated by this kind
39:40
of flat footed total
39:41
utilitarianism, which is like, let's just make
39:43
the number as high as possible. Yep.
39:45
You could imagine a scenario where you
39:47
give ninety nine percent of resources
39:50
to the digital minds. You leave one percent
39:53
for the humans. But then the
39:55
here's the thing is if you GPT I
39:58
don't know. This is like a very sketchy scenario. But
40:00
if you give one percent of resources
40:02
to humans is actually a lot of resources.
40:05
If giving lot of resources to
40:07
the digital minds creates tons of, like, wealth and
40:09
more resources.
40:10
Right. So is it something like
40:12
digital minds, in addition to
40:15
feeling lots of pleasure, are also
40:18
really smart and they figure out how to colonize
40:21
not only the solar system,
40:23
but like maybe the galaxy, maybe other
40:25
GPT, And then there's just like
40:27
tons of resources. And
40:31
so even just one percent of all those resources
40:33
still makes for a bunch of
40:34
humans. Yeah. I think that's the idea.
40:36
And bunch of human beings being. And so
40:40
on this, like, compromised position,
40:42
you're getting ninety nine percent of
40:44
what the total utilitarian replacement
40:47
wanted. And you're also getting a large
40:49
share of what the the humans
40:51
forever people wanted. And
40:53
you might want this compromise because of
40:55
moral uncertainty. You don't wanna just put
40:57
all of your
40:58
chips.
40:58
Right. Go all in. Yeah. And also
41:00
maybe to prevent some kind of conflict.
41:03
Yeah. And also for, like, you know,
41:05
Democratic cooperative reasons, like, I
41:08
I would be surprised if most
41:10
people, like, are down for
41:12
replacements. Mhmm. And
41:14
I think that like, should be definitely
41:16
respected. And it also may be right. So
41:19
that's the case for this, like, compromise for you.
41:21
Yeah. Yeah. Yeah. I guess I mean,
41:23
it sounds really great. And it sounds
41:26
I mean, it's yeah. It just sounds almost like too
41:28
good to be true to me. And
41:30
some part of me is like, surely,
41:32
it's not that easy. It just feels
41:35
very convenient that, like, we can
41:37
have it all here. I mean, it's not having
41:39
it all for both, but it's like having the majority
41:41
of it all for both humans and
41:44
digital minds. Well, I I feel
41:46
like cooperation does enable
41:48
lots of scenarios like that.
41:49
Ones like that. Yeah. We really can. Get
41:51
most of what they want. I mean, I should say,
41:53
I'm basically recapping an argument
41:56
from sharing world with digital minds. This
41:58
is not something I have liked. Thought
42:00
that much about. Yeah. I think it's really important
42:02
to think about these big questions about the future
42:04
of artificial sentence. But
42:06
my focus has been on issues
42:09
that are, like, more concrete than come
42:11
up today.
42:12
So, yeah, exploring this
42:15
a bit more deeply. Why
42:17
does anyone think that artificial
42:19
sentence is even possible?
42:21
Yeah, this is a great question.
42:25
I think the very broadest
42:27
case for it or, like, the very broadest
42:29
intuition that people have is
42:31
something like We know that some physical
42:33
systems can be conscious or sentient
42:35
like 146 made out of neurons can't
42:38
be. These ones, the ones on either
42:40
end of this recording. And also
42:42
listening in. And you could
42:44
have a view where something has to be made
42:46
out of neurons, it has to be made out of biological material
42:48
in order to be conscious. 146
42:51
reason that people think artificial minds could also
42:53
be conscious is this kind of broad
42:55
position in philosophy and
42:57
cognitive science called functionalism. Which
43:00
is this hypothesis that the very
43:02
lowest level details or like substrate
43:04
that you're building things out of, ultimately
43:07
won't matter. And the sort of things
43:09
that are required for consciousness or
43:11
sentience could also be made out
43:13
of other stuff. So one way of putting
43:15
this or one version of this is thinking that
43:18
it's the computations that
43:19
matter. It's the computations that our brains are
43:21
doing that matter for what we experience and
43:23
what we think about. Sorry.
43:24
What do you mean by computation? That's
43:26
a that's a great question that can go into
43:28
the into the philosophical weeds.
43:31
But for is like a maybe
43:33
like a rough approximation like,
43:35
patterns of information processing is
43:38
a way you could think about it. So you can
43:40
describe what your brain's doing. And
43:42
also think that your brain is in fact doing
43:44
like certain patterns of information
43:46
processing. So there
43:48
are theories by which what certain
43:50
parts of your brain are doing are
43:53
computing a function, taking
43:55
that input and processing it in a certain
43:57
way. So as to get a certain output.
44:00
So you can think of your visual system as
44:03
taking in a bunch of pixels
44:05
or something like
44:06
that. And from that computing where
44:08
the edges are. Right. Okay. So
44:11
really simplistically and maybe
44:13
just not true at all, but it's
44:15
something like when you
44:17
smell a food that smells good,
44:20
maybe you get kind of hungry. And
44:22
the computation is like, get the input
44:25
of, like, a nice yummy smelling food
44:27
and, like, maybe feel some hunger
44:29
is the is the output is the computation. Or
44:32
maybe it's like, feel this thing called
44:34
GPT, and then like search for food
44:36
in the fridge.
44:37
Yeah. It would definitely be more complicated
44:40
than that. But it is something like that. It's
44:42
like you're taking in inputs
44:44
and doing stuff with them. 146 thing
44:46
I might add at this point -- Mhmm. -- although
44:48
maybe this is two in the weeds. I think
44:50
when people say something like
44:52
you need the right computations for
44:55
consciousness, They're not just talking
44:57
about the right mapping between inputs and
44:59
outputs. They're also talking about
45:02
the internal processing that's getting
45:04
you from inputs to output. So here's an
45:06
example. Mhmm. There's this famous case
45:08
by Nedblock, also one of my advisers,
45:11
who put it out that you could have something
45:13
that has this big look look up table
45:15
where the input is a given sentence.
45:17
And then for every given sentence, it has
45:19
a certain output of what it should say. And
45:22
it doesn't do anything else with
45:24
the
45:24
sentences. It just goes to the right
45:27
column of its lookup table.
45:28
Totally. Yeah. Yeah. Yeah. Yeah. Of course, such a thing would, like, not
45:30
be feasible at all.
45:32
A lot people have the intuition that that way of
45:34
getting from input to output is not the
45:36
right sort of thing that you would want for
45:39
consciousness or
45:39
sentient. Right. Right. So,
45:42
like, if the lookup table had,
45:44
like, an input when you receive
45:46
input GPT and the, like,
45:48
looked up value was eat an apple,
45:51
that would not be the same thing as when
45:53
you receive the input, GPT. Think
45:56
about the or like maybe subconsciously, think
45:58
about the nutrients you might need. And
46:01
then go find a thing that will, like, meet that
46:03
need. Sorry. This may be a terrible
46:04
GPT. But something like I think
46:06
it's a good example.
46:07
Allowed to okay. Nice. It's just it's
46:09
pointing at the fact that, like, what your the
46:12
path you're taking internally matters.
46:14
And, yeah, I mean, I I will, like, add
46:17
or or point out, as I think you realize
46:19
that it wouldn't be decribable in
46:21
such a way and that the computations
46:23
would be extremely, like, fine ranging
46:25
complex and you couldn't, like, write them down on a
46:27
piece of paper. Yeah. But but the general
46:29
gesture that is is correct. Yeah.
46:32
Is there, like, a principled reason
46:34
why you couldn't write them down in paper?
46:36
I guess there's not a principled reason.
46:38
It's kind of I I think of that as more
46:40
of an empirical observation that -- Yeah.
46:42
--
46:42
in fact, what our brains are doing
46:44
is pretty complex. But that's
46:46
that's also an open question. I I think in the
46:49
early days of AI, people were
46:51
kind of optimistic that and this goes
46:53
for things with intelligence as well as consciousness.
46:56
That there would be these really simple principles
46:58
that you could write down and distill. That
47:00
doesn't seem to be what we've learned about the brand
47:02
so far or the way that AI has gone.
47:05
Yeah. So and we'll get to this later. I
47:07
do suspect that our theory of consciousness
47:09
might
47:10
involve, like, quite a bit of complexity. Yep.
47:13
Cool. Okay. So I I took you away off
47:15
track. So you're saying that there's
47:17
this idea called functionalism where
47:20
basically it's like the functions
47:22
that matter where all you need is certain
47:25
computations to be happening or,
47:27
like, possible in order to get
47:29
something like
47:30
sentience. Is that that basically right?
47:33
Yeah. That's basically right. Computationalism is
47:35
a more specific thesis about what
47:37
the right level of organization or what
47:39
the right functional organization is
47:41
is the function of performing certain
47:43
computations. Right. That makes sense. So
47:46
I think so. Yeah. Maybe I'll no. Maybe
47:48
I'll make sure I get it. So the
47:51
argument is that there's nothing
47:53
special about the biological material
47:56
in our brain that allows us to be conscious
47:58
or sentient. It's like
48:00
a particular function that our brain serves
48:03
and that, like, specific function is
48:05
doing computations. And
48:07
those computations are the
48:10
kind of underlying required
48:12
ability in order to be sending or conscious.
48:15
And theoretically a computer
48:17
or
48:17
something, silicone based could do that too.
48:20
Yeah.
48:20
I think that's basically right.
48:21
So that's the basic argument. What
48:24
evidence do we have for that argument? Yeah.
48:26
I'll say that's like the basic position, and then
48:28
why would anyone hold that position? I
48:30
think one thing you can do is look at the
48:32
way that computational neuroscience works.
48:34
So the success of computational
48:36
neuroscience, which is kind of the
48:39
endeavor of describing the brain
48:41
in computational terms, is like
48:43
some evidence that it's the computational level
48:45
that matters. And then there are also philosophical
48:48
arguments for this. So a very famous
48:50
argument or class of arguments are
48:53
what are called replacement GPT. Which
48:55
were flushed out by David Chalmers.
48:57
And listeners can also find
49:00
when Holden Karnovsky writes about
49:02
digital people and wonders if they could
49:04
be conscious or sentient. These
49:07
are actually the the arguments that he feels to.
49:09
And those ask us to imagine
49:12
replacing neurons
49:14
of the brain bit by bit
49:16
with artificial silicon
49:19
things that can take in the same input
49:21
and yield the same output. And
49:23
so by definition of the thought experiment,
49:26
as you add each one of these in,
49:29
the the functions remain
49:31
the same and the input output behavior remains
49:33
the same. So charters asked
49:35
us to imagine this happening say
49:38
to us while this podcast is
49:40
happening. Mhmm. So, yeah,
49:42
by stipulation, our behavior won't
49:44
change, and the way we're talking about things
49:46
won't change. And what we're able to
49:48
access in memory won't change.
49:51
And so at the end of the process, you have
49:53
something made entirely out of silicon,
49:55
which has the same behavioral
49:57
and cognitive capacities as
50:00
the biological thing. And
50:02
then you could wonder, well, did that thing
50:04
lose consciousness? By being replaced
50:06
with silicon. And what Thomas points
50:08
out is it would be really weird to have something
50:10
that talks exactly the same way
50:14
about being conscious. Because by definition,
50:16
that's like a behavior that remains the
50:18
same. Yep.
50:19
And has the same memory
50:21
access and internal GPT,
50:24
but, like, their consciousness left
50:26
without leaving any trace of
50:28
leaving. He he thinks this would be like a
50:30
really weird dissociation between cognition
50:33
and consciousness. And a lot
50:35
of people one reason this argument
50:37
kind of has force is a lot of people are pretty
50:40
comfortable with the idea that at least GPT
50:42
and verbal behavior and memory
50:45
and things like that can be functionally multiply
50:49
realized. And this is an argument that if you
50:51
think that it would be kind of weird if consciousness is
50:53
this one exception where the substrate
50:56
matters.
50:57
So I think the idea is something
50:59
like if you if you
51:01
had a human brain and you
51:04
replaced a single neuron with I
51:07
GPT, a silicone neuron that
51:09
did the exact like, performed to the exact
51:11
same function. And is the reason we think
51:13
that's like, a plausible thing
51:16
to think about because neurons transmit
51:18
electricity and they're kind of, like,
51:21
on off switch y and maybe the
51:23
same way that we think or the same way
51:25
that computers
51:26
are? Is that what you're saying? This is this
51:28
is an excellent point. 146 weakness
51:30
of the GPT, in my opinion, and
51:32
people have complained about this, is it
51:35
kind of depends on this replacement being
51:37
plausible or sorry. It
51:39
seems that way to people. In the paper, there's
51:41
actually a note on well, you might think that this is
51:43
actually in practice not something you could do.
51:45
And obviously, we could not do it now. Mhmm.
51:48
And for reasons I don't entirely understand that's
51:50
not really supposed to undermine the argument.
51:53
Okay. Alright. Well, maybe coming back to
51:55
that. Yeah. Is it
51:57
basically right though that we think of
51:59
a neuron and like a computer chip as
52:01
like analogous enough that that's why it's plausible?
52:04
Yeah, we think of them as being
52:07
able to preserve the same functions.
52:09
And I mean, there is some think there is some
52:11
evidence for this from the fact that,
52:14
like, artificial eyes and cochlear
52:16
implants -- Mhmm. -- like,
52:18
we do find that computational things can interface
52:21
with the brain and the brain can make
52:23
sense of them.
52:23
Interesting.
52:25
That's not like the size of argument. People who are kind
52:27
of not on board with this kind of computational
52:29
way of thinking of things would would
52:31
I'd probably
52:32
not. But we'll get off. We'll
52:34
face with that. We can have a Zoom time. Yeah.
52:36
And sorry. And the thing there is, like,
52:39
I actually don't know how artificial eyes work.
52:41
Is it like there's an
52:43
eye made of some things that are non
52:45
biological and they
52:47
interface with the brain in
52:49
in some way that allows people to see? I
52:52
also don't really know. Okay. I definitely
52:54
know that's possible with with
52:56
cochlear implants. Okay. I mean, I'm
52:58
interested in that one too then. But that's basically,
53:00
like, they connect
53:02
to, like it's a I'm picturing, like, wires.
53:05
Like, wires going from, like, a hearing aid
53:07
into the brain. I'm sure that's not quite right. But
53:10
it sounds like it's something like they
53:12
communicate, and that's, like, some evidence
53:14
that we can feed electricity
53:17
through silicone based things to
53:19
the brain and communicate with
53:20
it. Bauchner: Yeah, one recliner is it might not
53:22
be yeah. I don't think it's reaching into the brain. It might
53:24
be doing, like, the the right stuff to your inner
53:26
ear. To your
53:27
ear. Right. Okay. Okay.
53:29
Yeah. That makes sense. So
53:31
so we think that maybe
53:34
you think a neuron could be replaced
53:36
with a silicon based
53:38
version of it.
53:39
Cost thesis. Yeah. Cost thesis. Nice.
53:41
And then
53:42
Cost cutting your own.
53:43
Is is that a term? Is that a word? Is
53:45
that how people talk about it? I
53:46
think people have used that term. There's not a canonical
53:48
term since it's an imaginary case for
53:50
now. Right. Right. Okay.
53:52
So you have a prosthetic neuron
53:55
and you can
53:57
replace a single neuron at a time.
54:00
And like every time you make that replacement, it
54:02
stays like you work the same way,
54:04
your brain does the same things, nothing
54:06
about your behavior or thoughts
54:08
change? Yeah. So maybe it's
54:10
good to start with the first replacement. If
54:12
the first replacement is possible, I don't think anyone
54:14
would
54:14
think, oh, no. You have now destroyed Luis's
54:17
consciousness. Now she was like a walking --
54:19
So -- single copy.
54:20
Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. And then you this is
54:22
a common argument form in philosophy.
54:25
Two doesn't seem like it makes a
54:26
difference. Mhmm. And then Yeah.
54:28
So on and so forth. And then eventually, you replace
54:31
all my neurons with the silicone prosthetic
54:34
neurons and than I have an
54:36
entirely silicone based
54:37
brain, but there's no reason to think
54:39
I wouldn't feel I think the same things. Is
54:41
that basically it? That's the idea.
54:43
It's if you did think that you don't
54:45
feel the same things. Mhmm. It's supposed
54:47
to be really counterintuitive that you would
54:50
still be saying, oh, the this
54:53
worked. I'm still listening
54:55
to Rob talk. Yeah. I'm still
54:57
seeing colors. You would still
54:59
be saying that stuff since that's a good behavioral
55:02
function. Yep. Yep. Yep. And,
55:04
yeah, that's the basic thrust. So then what
55:06
that is is that that's at least one silicon
55:08
based system that could be
55:10
conscious. So that kind of opens the door --
55:12
Right. -- to
55:14
being able to do this stuff in Silicon. Yeah.
55:16
Yeah. Yeah. It feels very similar to
55:18
the, like, to this the ship that,
55:20
like, has all of its planks replaced
55:23
one by
55:23
one. And, like, at the end, you're asked if
55:25
it's the same ship. Yeah.
55:27
It is similar. It this this sort of thing shows
55:29
up a lot in
55:29
-- Yeah. -- in
55:30
philosophy, as I said, it's like a it's like
55:32
an old trick. Yeah. Listeners might
55:34
recall the podcast with
55:36
Al Hayek. Right? And he has
55:38
all these great examples of sort of, like,
55:40
argument patterns that you can use in philosophy
55:43
and you can, like, apply to different domains. You can
55:45
think of this as an application of a,
55:47
like, gradual replacement or,
55:49
like, bit by bit kind of argument and
55:51
philosophy. 146 thing I would like to say --
55:53
Mhmm. -- and maybe I'm qualifying too much, but full
55:55
disclaimer. I I think a lot of people
55:57
are not super convinced by
56:00
this GPT. Like,
56:02
Walteriro Piccinini is,
56:04
like, an excellent philosopher who thinks about
56:06
issues of computation and what it would mean for
56:08
the brain to be computing. Mhmm.
56:10
And I think he's sympathetic to the idea that
56:13
he's sympathetic to computationalism, but he
56:15
thinks that this argument isn't really what's
56:17
getting us there. I think he relies more
56:19
on that point I was saying about, well, if you look
56:21
at the brain itself, it does actually look
56:23
like computation is a deep
56:25
or meaningful way of carving
56:28
it up and Right.
56:29
Right. Right. Right. And so if you could get
56:31
the right computations doing some other
56:33
things or doing things that make
56:35
up
56:36
sentence, then, like, It doesn't matter
56:38
what's doing it.
56:39
Yeah. What
56:40
reasons do people think that that argument
56:42
doesn't hold up? Well, for one thing,
56:44
it's like You might worry that it sort of
56:46
stipulated what's at issue, at
56:48
the outset, which is that silicon is able
56:51
to do all the right sort of stuff. Mhmm.
56:53
So there's this philosopher of biology and philosopher
56:55
of mind called Peter Godfrey Smith who
56:57
would be an excellent guest by the way.
56:59
He's written book about Octopus minds
57:02
And he has a line of thinking where
57:05
functionalism in some sense is
57:07
probably true, but It's
57:09
not clear that you can get the right functions
57:12
if you build something out of silicon because
57:14
he's really focused on the low level biological
57:17
details that he thinks might actually matter
57:19
for at least the kind of consciousness
57:21
that you have. And that's sort of something that
57:23
think you can't really settle with
57:26
with an argument of this form. Yeah.
57:28
Can you settle it? So
57:31
I actually have sort of set
57:33
aside this issue funnily enough since it's
57:35
like the the foundational issue. For
57:37
now? And I'll say why I'm doing that.
57:39
Yeah. think these, like, debates about multiple
57:42
realizability and computationalism have
57:44
been going on for a while. And
57:46
I'd be pretty surprised if in the
57:48
next few decades someone has just nailed
57:50
it and they've proven it one
57:52
way or the other. And so the way
57:54
I think about it is think it's
57:57
plausible that it's
57:59
possible in Silicon to have the right
58:01
kind of computations that matter for consciousness
58:04
And if that's true, then you really need to worry
58:06
about AI sentence. And so
58:08
it's sort of like, let's look at the world where that's
58:10
true and try to figure out which ones could be conscious.
58:13
And it could be that none
58:16
of them are because of some deep reason having
58:18
to do with the biological hardware
58:21
or something like that. But it seems unlikely
58:23
that that's gonna get, like, nailed anytime soon.
58:25
Yeah. And I I just don't find it
58:28
crazy at all to think that
58:31
the right level for consciousness is the sort
58:33
of thing that could show up on a silicon
58:35
based system. Uh-huh. Uh-huh. Okay.
58:38
Yeah. Are there any other arguments for
58:40
why people think artificial sentence is
58:42
possible?
58:43
This is related to the computational neuroscience
58:46
point, but 146 thing people have
58:48
noticed is that lot of the leading scientific
58:50
theories of what consciousness is are
58:52
in computational terms. And
58:54
posit computations or some other
58:57
sort of pattern or function as
58:59
what's required for consciousness. And so
59:01
If you think they're correct in doing so,
59:04
then you would think that it's possible for
59:07
those patterns or computations or
59:09
functions. Being made
59:11
or realized in something other than biological
59:14
neurons. Does anyone disagree on
59:16
this? Like, do some people just think artificial sentience
59:18
is not possible?
59:20
Yeah. So there are these views called biological
59:23
theories. Maybe you can call them. So
59:25
netblock is one of the
59:28
guess like foremost defenders of this biological
59:30
view that consciousness just is,
59:32
in some sense, a biological phenomenon.
59:35
And you won't be capturing it if
59:37
you go to something too far
59:39
outside the realm of biological looking things.
59:41
John Searle is also a proponent
59:44
of this view. So there's views where,
59:46
like, that's definitely true and
59:48
it's just kind of a it's just like
59:50
what consciousness is. There's
59:52
also views on which consciousness is something
59:54
functional, but also you're not going to be able
59:57
to get it on GPUs or anything like
59:59
what we're seeing today. And those are kind of
1:00:01
different sorts of positions.
1:00:03
But, yeah, I mean, it should be noted that plenty
1:00:05
of people who've thought about this have
1:00:07
concluded. Yeah. You're not going
1:00:09
to get it if you have a bunch of
1:00:11
GPUs and electricity running
1:00:13
through
1:00:14
them. It's just not the right sort of thing. And
1:00:16
that's just like So the first argument
1:00:19
is like there's something really
1:00:22
special about biology and
1:00:24
biological like parts that make
1:00:26
whatever consciousness and sentence is
1:00:29
possible. And the other argument is,
1:00:31
like, it's theoretically possible,
1:00:34
but, like, extremely unlikely
1:00:37
to, like, happen with the technology
1:00:40
we have or could create or something? Yeah.
1:00:42
So, like, for that second position, Like,
1:00:44
most people will hold some version of that
1:00:46
position with respect to Swiss cheese.
1:00:49
Like, I would be really surprised if very
1:00:51
complicated arrangements of Swiss cheese ended up
1:00:53
doing these computations. Mhmm. Because it's just, like,
1:00:55
it's not the right material to get the right thing
1:00:58
going. Even if I think there are it's
1:01:00
it is multiply, realizable. You don't have
1:01:02
to think it's, you know, you could feasibly
1:01:05
do it in any sort of material at all.
1:01:07
Okay.
1:01:08
Interesting.
1:01:08
One one thing I'll add, since I am being
1:01:10
like very obsessive to range of positions, which
1:01:12
think is appropriate, I would like to note
1:01:15
that large numbers of philosophers
1:01:17
of mind and consciousness scientists in
1:01:19
surveys say, yeah, artificial
1:01:22
sentient is possible. Machines could
1:01:24
be conscious I don't have the exact numbers
1:01:26
off the top of my
1:01:27
head, but --
1:01:27
Yeah. -- David Chalmers has this great thing, the
1:01:30
Phil Paper Survey. And, like, yeah,
1:01:32
it it has asked people this question, and it's not like
1:01:34
a friend view. Yeah. Like
1:01:36
a substantial a substantial share of
1:01:38
philosophers of mind think that our official sentence
1:01:41
is possible and maybe
1:01:43
plausible. And dido like surveys of consciousness
1:01:46
scientists. Yeah. Yeah. Yeah. Yeah. We'll we'll stick
1:01:48
those in the show notes. Cool.
1:01:50
So sounds like there's like a couple
1:01:52
of counter arguments that are about
1:01:55
biology and and just
1:01:57
like what's possible with like silicone
1:01:59
and GPUs as building
1:02:01
blocks for entities. Are
1:02:04
there are there any other counter GPT? People
1:02:06
think for why artificial sentence
1:02:08
might not be possible?
1:02:10
Yeah, one thing it might be worth mentioning
1:02:12
is I'm GPT be doing this interview
1:02:14
talking about consciousness and sentient as
1:02:17
things where we know what we're talking about and, like,
1:02:19
we know what we're looking for, and it is this phenomenon
1:02:22
that we can wonder about. There is
1:02:24
a position in philosophy called
1:02:26
illusionism, which holds that consciousness
1:02:29
is kind of a confused concept and it doesn't
1:02:31
actually take anything out. So
1:02:33
on that view, it's like straightforwardly
1:02:36
false that AIs could be
1:02:37
conscious. It's also false that
1:02:39
in a certain sense of the word -- Yeah. -- humans
1:02:41
are conscious.
1:02:42
Right. Well, can you explain the
1:02:44
view of illusion of some? Yeah. So, like,
1:02:46
illusionists hold that
1:02:49
this concept of
1:02:51
subjective experience or what
1:02:53
it's like to be having a certain experience.
1:02:56
Even though a lot of people now find
1:02:58
it intuitive, illustrious would argue that,
1:03:01
actually, it's kind of a philosophers
1:03:04
intuitive notion and and not that deep.
1:03:06
I think I think I would argue with that. But
1:03:08
Yeah. It doesn't refer to anything,
1:03:11
actually. It's kind of like incoherent
1:03:14
or fails to pick anything out. The same way
1:03:16
that This is a popular example
1:03:18
in philosophy. People used to wonder
1:03:20
about phlogiston, which I think
1:03:22
was this substance that was going to explain
1:03:24
fire. I know we talked about it and
1:03:26
look for it. But, ultimately, it's
1:03:29
just not part of our, you know,
1:03:32
oncology. It's not part of our worldview. And
1:03:35
they think consciousness will end up being like that on
1:03:37
reflection. We'll ultimately have
1:03:40
a lot of functions and ways of processing
1:03:42
information and behavioral dispositions and
1:03:44
maybe representations of things. But
1:03:47
this question, but which of them
1:03:49
are conscious? Which of them have
1:03:51
subjective
1:03:52
experience? Ultimately won't
1:03:54
be a meaningful one. Okay.
1:03:57
Right. So I guess if you don't think humans
1:04:00
or non human animals are, can't just
1:04:02
to any
1:04:02
degree. It's like not a meaningful question to ask
1:04:05
whether artificial intelligence is
1:04:07
sentient. In a certain sense of the
1:04:09
word, Yeah. What they deny is
1:04:11
what philosophers have called phenomenal consciousness,
1:04:14
which is used to pick out whether there's something
1:04:16
it's like to be something or
1:04:19
whether it has subjective experience or
1:04:21
this kind of subjective quality to its mental
1:04:23
life. They don't deny that
1:04:26
things are conscious in the sense that
1:04:29
they may might process information in
1:04:31
certain ways and sometimes be globally
1:04:33
aware of that information they
1:04:35
don't deny that things feel
1:04:38
pain, for GPT. But they
1:04:40
deny this way of construing it in terms
1:04:42
of subjective experience.
1:04:43
Okay. Okay. I mean, that doesn't seem
1:04:45
that damning for artificial sentience,
1:04:48
I guess. Like, as long
1:04:50
as you think that they
1:04:52
can still feel pain. And
1:04:54
if you think that's morally significant,
1:04:57
then, like, artificial sentence
1:04:59
could maybe feel the same thing and
1:05:01
that would still be morally significant?
1:05:03
Yeah, so this is roughly my
1:05:05
position and I think it's the position of
1:05:08
I was talking to Keith Frankish
1:05:10
on Twitter the other day. Keith Frankish is one
1:05:12
of the leading proponents of illusionism. And,
1:05:16
yeah, I asked him, like, what do you think about people
1:05:18
who are looking for animal sentient? Is
1:05:20
that kind of an entirely misguided quest
1:05:22
-- Mhmm. -- on illusionism? And his answer
1:05:25
is no, and he rightly thinks.
1:05:27
And I GPT, that even if you're an
1:05:29
illusionist, they're going to be mental phenomena
1:05:31
or information phenomena that matter.
1:05:34
And you're going to want to look for
1:05:36
those. You won't be looking for
1:05:38
maybe quite the same thing. Right? You think
1:05:40
you are if you're a realist about consciousness?
1:05:42
And think that's like a very important lesson.
1:05:44
I think in like the kind of circles
1:05:47
we run-in, a lot of people are very sympathetic
1:05:49
to illusionism. And occasionally,
1:05:51
I hear people say, oh, well, then
1:05:54
there's like no question here or it's like
1:05:56
a meaningless question. And that
1:05:59
might be true for like phenomenal consciousness. But
1:06:01
I just wanna point out there are like scores
1:06:03
of extremely meaningful and vexing questions
1:06:05
even if you're an illusionist. And I would still
1:06:08
like a theory of what sort of things
1:06:10
feel pain in the illusionist sense or
1:06:13
have desires or whatever
1:06:15
it is that we on reflection
1:06:17
think matters
1:06:18
morally. Right. Right. So is
1:06:20
it basically like some
1:06:23
people think that the kind
1:06:25
of consciousness I think I'm experienced
1:06:28
sing might not be a meaningful
1:06:31
concept or thing. Like, I might not actually
1:06:33
be experiencing that. I have the illusion
1:06:35
of experiencing it, but, like, there's no
1:06:37
sense in which I actually truthfully really
1:06:40
am. But like, I still
1:06:42
feel like I feel pain and I still don't
1:06:44
like that. And that in itself is,
1:06:46
like, still more like significant even if
1:06:49
something called consciousness
1:06:50
is, like, happening, underlying that
1:06:53
pain or whatever? Yeah. That's one position
1:06:55
you could have. You could think that being
1:06:57
disposed to judge that you have phenomenal
1:06:59
consciousness that matters morally. think
1:07:02
a more plausible position you could have is it doesn't
1:07:04
matter if you have whatever cognitive
1:07:06
illusion makes philosophers think phenomenal
1:07:09
consciousness is real. It could also just be
1:07:11
if you feel pain in this functionally defined
1:07:13
sense that that
1:07:15
matters. Or if you have desires that are thwarted
1:07:17
or preferences that are thwarted,
1:07:19
Cool. There's really excellent work by
1:07:22
Francois Cameron, who's another
1:07:24
illusionist trying to
1:07:26
see what value theory looks like
1:07:29
and questions about animal centients and animal
1:07:31
welfare look like on the illusionist picture.
1:07:33
I think it's a very underexplored issue
1:07:36
and like an extremely important issue. So
1:07:38
put that on the show notes
1:07:39
too. Yeah. Yeah. Plug. Nice. Okay.
1:07:42
Yeah. Where where do you personally come down on
1:07:44
on artificial
1:07:45
sentence? And yeah, I guess, whether
1:07:47
it's possible.
1:07:48
Yeah. I think I'm, like, eighty
1:07:50
five percent that
1:07:52
artificial consciousness
1:07:54
or sentient And here's a
1:07:56
real wiggle or something in that vicinity
1:07:58
that we morally care about.
1:08:00
That makes sense to me.
1:08:01
Is is possible? Yeah. Yeah.
1:08:04
Okay. That's pretty high.
1:08:05
Yeah. So that's, like, you know, ever and
1:08:08
impressive. Yeah. So I guess
1:08:10
if if that's right, and,
1:08:12
like, artificial sentence is possible
1:08:14
and if it ends up
1:08:15
existing. Yeah. Can you walk me through the
1:08:17
case that it definitely matters morally? Yeah.
1:08:20
It's almost hard to give a thought experiment
1:08:22
or an argument for the claim -- Mhmm.
1:08:24
-- that suffering matters. I
1:08:27
I think that suffering matters. Is
1:08:29
something where common sense and
1:08:31
the majority of philosophers agree, which
1:08:34
doesn't always happen. So like Bentham,
1:08:37
Jeremy Bentham, has this famous
1:08:39
and off quoted passage, off
1:08:41
quoted by animal rights and animal welfare
1:08:43
people among others, where he says, like,
1:08:46
the question about animals is not if they can
1:08:48
reason or if they can talk.
1:08:50
It's whether they can suffer. And
1:08:53
it doesn't seem like there's any other boundary
1:08:55
that seems like the right boundary
1:08:58
of moral concern. Now,
1:09:00
as we've noted, you can have quibbles
1:09:02
about what suffering actually is and if it
1:09:04
involves phenomenal consciousness and
1:09:06
things like that. But yeah, it's
1:09:08
just extremely intuitive that If
1:09:10
something feels bad for something, and
1:09:13
maybe you also add that it doesn't want it,
1:09:15
and it's trying to get away from it, that
1:09:18
matters
1:09:18
morally. And that sort of thing
1:09:20
should be taken into account in our moral decision
1:09:23
making. Yeah.
1:09:23
Yeah. One thing I'd like to add is,
1:09:25
like, there's a there's a position on
1:09:28
which that's all it matters. And,
1:09:30
like, the only things that are good and bad for
1:09:32
things are experiences
1:09:34
of pleasure and displeasure.
1:09:37
That's not a consensus view at all.
1:09:39
But even among people who think that other things
1:09:42
matter, like knowledge or friendship or justice
1:09:44
or beauty, they still also think
1:09:46
that, you know, experiencing pain is
1:09:48
is really
1:09:49
bad.
1:09:49
Right. Right. Right. Okay. Yeah. That makes
1:09:51
sense. So, like, the other
1:09:53
main alternative for this like focus
1:09:55
on experiences pain or experiences
1:09:58
of pleasure is a focus on desires
1:10:00
and preferences, and whether those are being
1:10:02
satisfied. Uh-huh. So that's
1:10:04
a big debate and debates of, like, what
1:10:07
welfare is you know,
1:10:09
what what makes things go well or badly
1:10:11
for something. Yep. And it's also a debate
1:10:13
in like what sort of things are moral
1:10:15
patients, like the sort of things that are in the scope
1:10:17
of moral consideration. And I would like to
1:10:19
note a position on which
1:10:22
what ultimately matters is not pain
1:10:24
or pleasure, but desires.
1:10:26
And desires seem like they're much easier
1:10:29
to define in this functional
1:10:31
way that maybe doesn't make reference to consciousness.
1:10:34
And that might be in some ways easier to get
1:10:36
a grip on than consciousness. That's the
1:10:38
position of Francois Cameron. Who has
1:10:40
a paper about how we should think about welfare
1:10:42
if we don't really believe in consciousness. I
1:10:45
find those issues very difficult
1:10:47
to tease apart So, like, Shelley
1:10:49
Kagan has this apt remark
1:10:51
that in human life,
1:10:54
our experiences and our desires are like
1:10:56
so tightly linked be really
1:10:58
hard to be like, is it bad that I'm in in pain
1:11:00
or is it bad that I don't want to be in pain?
1:11:03
Like, those just seem really
1:11:05
hard
1:11:05
to, like, tease apart conceptually. Yeah.
1:11:07
I mean, can I imagine
1:11:10
being in pain and not not wanting
1:11:12
to be in pain?
1:11:13
So there are cases where people have the
1:11:15
sensory experience of pain that report
1:11:18
don't minding it. So they can
1:11:20
fully feel that there's skin is being
1:11:22
pinched or something like that. But
1:11:24
they're
1:11:24
like, yeah. But it's just not bad.
1:11:27
So that's called pain asymptalia, and it's like
1:11:29
a fascinating That is fascinating.
1:11:32
And there's a lot of philosophical work, which is
1:11:34
like, well, is that really pain? Are they
1:11:36
like lacking some unpleasant quality to
1:11:38
the
1:11:38
pain? And that's why they don't mind it.
1:11:40
Could you really have that unpleasant quality and not
1:11:42
mind it?
1:11:43
Yeah. Yeah. Yeah. Yeah. Yeah. One thing I can say that
1:11:45
pain is simply it does seem too many people to
1:11:47
have shown. Is that there's a surprising association
1:11:49
between the way you process the sensory
1:11:51
information about pain, and then this,
1:11:54
like, affective like,
1:11:56
felt unpleasantness thing. And I think there
1:11:58
are differences in, like, the brain in,
1:12:00
like, in terms of how those are processed and
1:12:02
things, which is why things like this are possible.
1:12:04
Yeah. Yeah. No. That's that's interesting.
1:12:07
Okay. So it sounds like philosophers
1:12:09
would basically mostly agree
1:12:12
that if AI systems are feeling
1:12:14
something like pleasure or pain, that just,
1:12:16
like, probably matters morally.
1:12:18
Is that is that basically sound right? That
1:12:20
sounds right to me. And if it's not, it it should
1:12:22
be. Okay. Great.
1:12:24
Yeah. So where are we
1:12:26
with current systems on this? I GPT, there's
1:12:29
been some public conversation around
1:12:32
current large language models
1:12:34
being sentient. There's a whole thing
1:12:36
there that we could talk about. But just, yeah,
1:12:38
from the ground up, what do you think about where we
1:12:40
are? Yeah. So the short answer is
1:12:43
after thinking about a lot of current
1:12:45
theories of consciousness and how large
1:12:47
language models work,
1:12:50
I think it is quite unlikely that
1:12:52
they have conscious experiences of the
1:12:54
kind that we will morally care about. That is
1:12:56
subject to a lot of uncertainty because there is so much
1:12:58
we don't know about consciousness and how they
1:13:01
work. I can definitely say there's not
1:13:03
like a straightforward case where you're like,
1:13:05
here's what consciousness is and here's
1:13:07
how large language
1:13:08
models have it.
1:13:09
Yeah. Yeah. And I also think
1:13:11
I would be quite surprised if
1:13:14
large language models have developed
1:13:16
pleasurable and displeasurable experiences.
1:13:19
You know, they're having really bad time. They
1:13:21
don't like writing poetry for us.
1:13:23
Right. Right. Right. Like, we have stumbled into
1:13:25
pass through here.
1:13:26
Yeah. Yeah. I'm glad that people are actually,
1:13:29
like, you know, raising the the issue.
1:13:31
It's good practice for future things. And
1:13:33
there is also the small
1:13:35
chance that we have. And in
1:13:37
general, like part of what I try
1:13:39
to do is Just get people thinking
1:13:42
about it and and and provide pointers for ways
1:13:44
of having, like, as
1:13:46
evidenced based conversations as
1:13:48
possible. Because as listeners
1:13:50
will have noted, it's, like, very easy
1:13:53
for it to descend into,
1:13:54
like, Twitter madness and Right.
1:13:57
Complete free form speculation. Yeah.
1:13:59
Yeah. Yeah. Yeah. Yeah. Yeah. I guess guess
1:14:01
that was maybe arguably the case with
1:14:04
with Lambda, which we can talk about. But I
1:14:06
GPT, first, just kind of clarifying there
1:14:09
are a bunch of different kinds of AI
1:14:11
systems
1:14:12
that exist right now. Which 146 most
1:14:14
likely to be sentient?
1:14:16
I would be somewhat surprised if large language
1:14:18
models are the most likely current systems.
1:14:20
And
1:14:20
those are things like GPT three or GPT
1:14:22
chat. Right?
1:14:23
And Lambda.
1:14:24
And Lambda. Of course. Yeah. Yeah. And I yeah.
1:14:26
I can say more about why. I I think that that will
1:14:28
probably be getting into, like, the substance of this investigation.
1:14:31
Right. Right. Well, I guess,
1:14:33
do you do you mind starting by telling me,
1:14:35
like, what other systems are
1:14:38
plausibly? Like, we even wanna be asking
1:14:40
the question, are they sent in? They're, like,
1:14:43
possibly closer?
1:14:44
Yeah. There's things that there's
1:14:47
at least things that seem to do
1:14:49
more human like or agent
1:14:51
like things. And I think that can maybe
1:14:54
put us closer to things
1:14:56
that we could meaningfully call pain
1:14:58
or pleasure or things like that? Like what?
1:15:00
So yeah. So there are, like,
1:15:02
virtual agents that are trained
1:15:04
by reinforcement learning and which navigate
1:15:07
around, like, a Minecraft environment. Mhmm.
1:15:10
There are things that incorporate large language models,
1:15:12
but do a lot more than
1:15:15
just answer text inputs.
1:15:17
You can plug large language models into
1:15:19
robots, and it's
1:15:22
really helpful for the way those robots plan.
1:15:24
That's like a really cool line of research.
1:15:27
There's obviously just robots, and
1:15:29
I I would like to look more into just
1:15:32
you know, actual robots. Yeah.
1:15:34
Yeah. Which sometimes get, like, a bit of short shrift
1:15:36
even though it's kind of the canonical
1:15:38
-- Right. --
1:15:39
like, WiFi thing.
1:15:40
Right. Right. And robots
1:15:43
like, we're literally talking about, like,
1:15:45
things in Star Wars. What's the
1:15:47
closest thing to that that we have right now?
1:15:50
Like, what's the, like, smartest
1:15:52
or most impressive
1:15:54
robot? Sorry. You might not know the answer.
1:15:56
But, like, what is AI's smart and impressive
1:15:58
robot? Yeah. I was, like, I was I was not being modest
1:16:01
when I'm, like, I need to look more into that. Like,
1:16:04
I'm really not up on the the state of art.
1:16:06
Like, the first thing I wanna look at is people
1:16:09
who explicitly want to try
1:16:11
to build more self awareness into robots I
1:16:13
definitely wanna see how that's
1:16:16
going.
1:16:17
Yeah. Yeah. Yeah.
1:16:18
Thanks, sir. You know, what you're gonna do
1:16:20
if you have a robot that can feel pain.
1:16:22
Like, are we Are we, like, ready
1:16:24
for that -- Yeah. -- as like a as
1:16:26
a society? And, yeah,
1:16:29
another thing about robots is they It
1:16:31
would be, like, more straightforward to maybe see
1:16:33
how they feel pain because -- Totally. --
1:16:35
have bodies in mind. They're trying to train
1:16:38
them to protect their bodies and send damage
1:16:40
to them.
1:16:41
Right. Right. Yeah. That makes a lot of sense.
1:16:43
Yeah. You mentioned kind of a a line of research
1:16:46
on feeding in large language
1:16:48
models into robots. And that
1:16:50
having an impact on how well they plan. Is
1:16:52
there more you can say about that? It sounds
1:16:54
like it might just be a really interesting topic.
1:16:57
Yeah. Like the cool factoid, which
1:16:59
I can't probably technically elaborate that much.
1:17:01
As my understanding is that
1:17:04
Large language models have to learn all
1:17:06
kinds of abstract representations in
1:17:08
in the course of learning to predict next words.
1:17:11
And those representations just seem
1:17:13
to be very useful for agents that
1:17:15
want to, like, decompose plans into
1:17:18
subactions. It's kind of like
1:17:20
an as such a fact from a certain
1:17:22
point of view that the kind of things learned
1:17:24
by large language models would so straight forwardly
1:17:27
and kind of I think without that much
1:17:29
tweaking end up helping with other
1:17:31
agents. Right. But it's true. Sorry.
1:17:34
And what's is there a specific
1:17:36
robot you have in mind with a specific set
1:17:38
of goals? Just I I'm not totally
1:17:40
sure I
1:17:40
understand, like, what plans we're talking about and
1:17:43
how they're deconstructing them or whatever.
1:17:45
Yeah.
1:17:45
We can find the real paper and link to it in the show
1:17:47
notes.
1:17:48
Yeah. So The epistemic status of this is
1:17:50
half remembering some slides from the lecturer
1:17:53
that I saw at a reinforcement learning conference. Right.
1:17:55
Yep. I think it was a virtual agent
1:17:57
and it was doing things like
1:17:59
fill up a cup of coffee in the kitchen and
1:18:01
then decomposing that
1:18:02
into, okay, get the cup. Put on the counter.
1:18:05
Right. Okay. That
1:18:07
is wild. So you have
1:18:09
an agent that's like get
1:18:12
some coffee is the GPT. And then you give
1:18:14
it a large language model somehow or
1:18:16
you, like, give it access to a large language model.
1:18:18
And the thing is like, how do I do
1:18:20
that? And then the large language model
1:18:23
helps it be like, here are the steps.
1:18:25
You go to the kitchen. You pull a cup
1:18:27
from the cupboard or whatever. Is that
1:18:29
basically?
1:18:30
Yeah. It it it might not I think it's not
1:18:32
that kind of, like, direct kind of querying. Okay.
1:18:34
Okay.
1:18:34
Okay. In in some vague way that
1:18:36
I would have to read the paper to know So it's it
1:18:38
has that, like, in its
1:18:39
machineries. So I'm
1:18:40
happy with it. Right. Right. Okay.
1:18:42
Presentations and knowledge of the large language model.
1:18:44
Got and the baseline didn't and it was
1:18:46
worth it planning. But then when you feed
1:18:48
it in to the whatever processor
1:18:50
algorithm, it gets much better at it.
1:18:53
Yeah.
1:18:53
My understanding is that, like, decomposing plans
1:18:55
into subplans has always been very hard
1:18:57
problem.
1:18:59
Okay. Interesting. I mean, if you think about
1:19:01
all the different ways that there
1:19:03
are to fill up a cup of coffee
1:19:05
and I mean, there's, like, an infinite
1:19:08
number of yeah. Infinite number of, like, little variations
1:19:10
on that. And you kinda need
1:19:12
to know which ones are relevant. You sort
1:19:14
of know how to transfer knowledge
1:19:16
from one case of getting the coffee
1:19:18
to, like, a slightly different one. Yeah. Right.
1:19:20
And like, I think one traditional problem
1:19:23
people have in reinforcement learning, which
1:19:26
is training things by, like,
1:19:28
just giving score on how well they did
1:19:30
it, is it's it can just
1:19:32
be very hard to scale that to, like,
1:19:34
very complex actions. And my understanding is
1:19:36
that large language models entering the
1:19:38
scene
1:19:38
has, like, really helped with that. Okay.
1:19:42
Wild. Okay. How do we get here? So I guess this
1:19:44
is, like, these are some of the
1:19:46
different systems that,
1:19:49
like, you could ask the question of whether
1:19:51
they're sentient. And somewhere
1:19:54
in there, you'd put large language
1:19:56
models. But you'd put some other
1:19:58
things kind of at the higher end of
1:20:00
probability of sentient. And
1:20:02
it's like, you're not totally sure what those are. It sounds
1:20:04
like. But maybe robots with large language
1:20:06
models feeding in
1:20:07
are, like, a bit higher than large language models
1:20:10
alone? Yeah. And so even without
1:20:12
having yet examined, like,
1:20:14
specific system. One quick argument
1:20:17
is just whether or not I agree with
1:20:19
them, there are bunch of preconditions for
1:20:21
Sentience that a lot of people think are plausible.
1:20:23
146 of them is embodiment, say,
1:20:26
or maybe another one is having
1:20:28
a rich, like, model of a
1:20:30
sensory world or something like that. And
1:20:34
there's just a straightforward argument. Yeah.
1:20:36
Like pure text, large language
1:20:37
models, don't have that. Sort
1:20:40
of thing,
1:20:40
probably. Yeah. But it's not hard to imagine
1:20:43
augmenting them with those things
1:20:45
or plugging them into other stuff, and
1:20:48
people are already doing that. So if
1:20:50
you're worried about some limitations of LOMs,
1:20:52
there's definitely other places you can
1:20:53
look. And I myself haven't yet looked,
1:20:56
but it's like definitely on my list.
1:20:58
Cool. Cool. Yeah. Makes sense. Yeah.
1:20:59
I decided to start with pure text LLMs,
1:21:01
like, you know, as a base case and as an
1:21:03
exercise. Cool. Yeah. What
1:21:06
what would you look at next? I guess you said robots.
1:21:08
Anything else you'd be especially excited to look
1:21:10
at? Yeah. That
1:21:11
might not be robots. It might be it
1:21:13
might be virtual. Virtual agents. Yeah.
1:21:15
Okay.
1:21:16
Yes. That and maybe stuff that's
1:21:18
closer to a pure text LLM, but just
1:21:21
something that also has like sensory
1:21:23
channels Right.
1:21:25
So, like, getting input
1:21:27
in systems. Sorry. What what
1:21:29
are multimodal
1:21:30
systems? I
1:21:31
mean, a multimodal modal
1:21:33
just means, like, kind of input in this context.
1:21:35
Oh, I see. So
1:21:36
it'd be something that's trained both on text and
1:21:38
on images.
1:21:39
GPT it. Okay.
1:21:41
So dolly two, which
1:21:43
you've probably seen, make Yeah. But Beautiful
1:21:45
pictures.
1:21:46
Love it. Yeah.
1:21:46
Like, that has to be trained on both images
1:21:49
and text and because it's, like, translating
1:21:51
between them.
1:21:51
Right. Okay.
1:21:53
I'm not saying that's my next like, you know,
1:21:55
best candidate or whatever just as an example of
1:21:57
mobile. Right. Right. Jumping in again quickly
1:21:59
to flag that as some of you might have heard,
1:22:01
GPT is actually set to be released this
1:22:04
week, and it's expected to be multimodal
1:22:06
in the way that Rob's talking about here. I
1:22:08
haven't gotten chance to ask Rob if
1:22:11
this changes his view significantly on whether
1:22:13
GPT is conscious or sentient.
1:22:16
But I'm guessing he'll be sharing his take
1:22:18
on Twitter at GPT Long.
1:22:20
Or probably on his sub stack experience
1:22:23
machines. So I'd encourage curious
1:22:26
listeners to check those out. And so
1:22:28
what's the reason that you think that, like, the more
1:22:30
kind of types of inputs
1:22:33
or, like, words and pictures, for GPT, is
1:22:36
more likely to
1:22:39
result in something being sentient?
1:22:41
Yeah, that's a great question. I
1:22:43
don't think it's like a strict
1:22:45
condition that you have to be processing more
1:22:47
than one thing. Yep. I have this
1:22:49
rough intuition that processing
1:22:52
more than one type of thing might
1:22:54
make you develop the kind of
1:22:56
representations or resources for
1:22:58
handling multiple sources of input. That
1:23:01
might correspond to consciousness. Got
1:23:03
it. Another way of putting that
1:23:05
is, like, if you get closer
1:23:08
to something kind of human ish
1:23:10
Yeah. That can make puts you on a little
1:23:12
bit firmer ground even if it's not
1:23:14
strictly necessary.
1:23:15
Yeah. Yeah. Yeah. And One fact about us is
1:23:17
is we have to handle all sorts of different
1:23:19
input streams and decide which ones
1:23:21
to pay attention to and form representations that
1:23:23
incorporate all of
1:23:24
them. Totally. And things like that. Yeah. Yeah.
1:23:27
I'm realizing I feel like I have to
1:23:29
understand what you when you say form representations,
1:23:32
but, like, do you basically mean I
1:23:35
don't know. Like, what dolly's doing when it
1:23:37
gets a bunch of when it gets, like, trained on
1:23:39
a bunch of pictures of dogs. It,
1:23:41
like, is it forming representation of
1:23:43
a dog? And and, like, we're
1:23:45
also doing things like that as humans.
1:23:48
Like, we've got some resolution of
1:23:50
GPT. I'm gonna cheat and not answer the four.
1:23:52
Question of what representation is. Okay.
1:23:55
I I will stick
1:23:55
to, like, the multimodal element.
1:23:57
Okay. Whatever it is to represent a dog
1:24:00
our representations seem to contain information
1:24:02
about what they look like and what they sound
1:24:04
like, and how people talk about them and how they're
1:24:06
defined and and all sorts of things. Got it.
1:24:09
Is
1:24:09
it kind
1:24:10
of like a
1:24:10
concept of a dog? Yeah. We can
1:24:13
we we can use that word. Exactly. And,
1:24:17
yeah, there's really interesting work
1:24:19
from Chris Ola who has been on
1:24:21
the show. Mhmm. And his name usually
1:24:24
comes up if you have some fascinating interpretability
1:24:26
thing to talk about. Where, yeah, I
1:24:28
think he looked for neurons that
1:24:31
seem to represent
1:24:34
or encode or whatever certain concepts
1:24:36
in multimodal systems. And somehow,
1:24:39
like, be, yeah,
1:24:41
emerging in this, like, cross metal or or
1:24:43
multimodal
1:24:44
way. Cool. Cool. Okay. Yeah. That makes sense.
1:24:46
Yeah. Okay. So I
1:24:48
guess yeah. It sounds like there there's like range
1:24:51
of types of AI systems,
1:24:53
and there are some different reasons to think
1:24:56
maybe there's a bit more evidence for some
1:24:58
being sentient or conscious.
1:25:01
I guess I've heard you give the example of,
1:25:03
like, the fact that humans
1:25:05
have multiple kind
1:25:08
of like I don't even know. What are what
1:25:10
are we calling it? Like, we process words.
1:25:12
We process images. I guess we process sounds.
1:25:15
I'm I'm kind of calling it inputs in my head,
1:25:17
but I don't know if that's fair. Okay. Yeah.
1:25:19
Awesome. Cool. So we've got lots of inputs.
1:25:21
Maybe a thing that has lots of inputs, maybe in
1:25:23
a system that has lots of inputs. It's a
1:25:25
bit more like a human, and that's, like, maybe
1:25:28
a bit more evidence that it might
1:25:30
be sentient or conscious. What other
1:25:32
types of evidence can we have about
1:25:34
whether an AI system is conscious?
1:25:37
Yeah. So the perspective I've been taking
1:25:39
is let's try to think
1:25:41
about the kind of internal processing
1:25:45
it's using or the kind of computations
1:25:47
or representations it's manipulating as
1:25:50
it does a task and see if
1:25:52
we can find GPT to things that
1:25:54
we have reason to think. Are associated
1:25:56
with consciousness and humans. Mhmm.
1:25:59
So the dream would be,
1:26:02
oh, we studied humans enough and we kind of
1:26:04
identified what the mechanism is
1:26:06
and specified it in computational terms.
1:26:09
And maybe that's a very complicated thing. Right. Maybe
1:26:11
it's somewhat simple. And then
1:26:13
we use interpretability tools to
1:26:15
say, ah, there is that structure
1:26:17
-- Right. -- in this AI system.
1:26:20
Yeah. I think that scenario is unlikely
1:26:22
because it we have the great
1:26:24
interpretability, we have the detailed thing
1:26:26
of
1:26:26
consciousness, and we have the exact match, which
1:26:28
think is unlikely unless you're doing the cold
1:26:30
brain emulation. Yeah. Yeah. Yeah. I see.
1:26:33
So the idea is, like,
1:26:35
we figure out that what
1:26:37
sentience is is
1:26:39
like this formula.
1:26:42
It's like you could put the formula
1:26:44
in Excel sheet and then the Excel sheet
1:26:46
would feel sentience. It's
1:26:48
like when you get
1:26:50
a pinprick, you feel this kind of pain
1:26:52
or something. And we're like, know exactly the formula
1:26:54
for that kind of pain, and then we
1:26:57
find it in an AI
1:26:59
system. It, like, has the exact
1:27:01
same, like, if given this input
1:27:04
do this process, and then feel
1:27:06
this thing, and that thing is pinprick
1:27:08
pain. And then if we saw that exact
1:27:10
match, it should be like, cool. That's doing the same
1:27:12
thing. It must be experiencing the same thing.
1:27:15
Obviously,
1:27:15
it's, like, infinitely more complicated. But
1:27:18
it's, like, that's roughly the thing.
1:27:20
Yeah. Just with one clarification, which
1:27:23
I think is in what you said, it's
1:27:25
not just that there's the same
1:27:27
input to output mapping. Is
1:27:29
that the algorithm or
1:27:31
process that it's using to process
1:27:33
it -- Got
1:27:34
it. -- works in the relevant sense to be
1:27:36
the same. The
1:27:37
same process. Oh, and that's actually key.
1:27:39
Yeah. In my in my view.
1:27:41
Yeah. Yeah. Yeah. Yeah. Otherwise, it's this
1:27:43
it could just be like a view like lookup
1:27:45
table.
1:27:46
Exactly.
1:27:47
Did you almost say V Lookup? Because you have to excel
1:27:49
in mind? I I do think about a
1:27:51
lot of this stuff in I'm,
1:27:53
like, imagining excel bunches we're talking.
1:27:56
Nice. Okay. And that process
1:27:58
might be something like
1:28:00
I mean, is there any way to simplify it
1:28:02
for me just to get a bit
1:28:05
better of an intuitive understanding of what
1:28:07
what kind of process we could find?
1:28:10
Yeah. So this is a great question
1:28:12
because, like, part of what I'm trying
1:28:14
to do more myself and get more people to do
1:28:17
is actually
1:28:18
think about processes that are identified
1:28:21
in neuroscience. And actually think about
1:28:23
what those are. So we could we could
1:28:25
do that. Great. If you would like.
1:28:26
I would love to do that. And, like,
1:28:28
learning the theories of consciousness
1:28:31
are going to be sketchy and unsatisfying
1:28:33
intrinsically, and and also my
1:28:35
understanding of them. And
1:28:38
and maybe kind of hard to explain verbally. But
1:28:40
we'll link to papers explaining them. Cool.
1:28:42
So, like, global workspace, theory is
1:28:44
a pretty popular neuroscientific theory
1:28:47
of what's going on when humans
1:28:49
are conscious of some things rather
1:28:52
than others. And
1:28:54
let's, like, start with, like, kind of, the picture of
1:28:56
the mind or the brain that it's,
1:28:58
like, operating within, and then I'll say
1:29:00
how it then builds the theory of
1:29:02
consciousness. Okay. Top of that. So it
1:29:04
has this kind of picture of the the
1:29:06
mind where there are a bunch of different kind
1:29:09
of separate and somewhat encapsulated
1:29:12
information processing systems
1:29:14
that do different things. So
1:29:16
-- For GPT, like,
1:29:18
So, yeah, like a like a GPT system.
1:29:21
Uh-huh. It's
1:29:21
like helps you generate speech. Maybe,
1:29:24
like, a decision making system. Maybe
1:29:26
that's not one system though, but -- Sure. -- also,
1:29:29
like, the sensory systems. They're in
1:29:31
charge of getting information from outside world
1:29:33
and, like, building some representation
1:29:36
of what they quote, quote, they
1:29:38
quote, quote, think is, like, going on.
1:29:40
Yeah. Like, memory memory
1:29:43
be one? Memory would memory definitely
1:29:45
is. It's one of them. Yeah. And
1:29:48
those things can operate
1:29:50
somewhat independently. And it's like
1:29:52
efficient for them to be able to do so.
1:29:54
And they can do a lot of what they're doing
1:29:57
unconsciously. Like, it's not going to feel
1:29:59
like anything to
1:30:01
you for them to be doing it.
1:30:03
Right. Here's a quick side note, and this
1:30:05
is separate from GPT workspace.
1:30:07
This is something like everyone agrees on.
1:30:09
Okay.
1:30:10
An interesting fact about the brain is that
1:30:12
it is doing all kinds of stuff,
1:30:14
a lot of it extremely complex, and
1:30:17
involving a lot of information processing. And
1:30:20
I can't ask what is it like for Louisa
1:30:22
when her brain is doing
1:30:23
that.
1:30:24
That's some other thing. Right.
1:30:25
So,
1:30:25
like, your your brain is, like, regulating, like,
1:30:28
Hormonal
1:30:29
race. It's like pumping blood.
1:30:31
Your heartbeat? Exactly. Yeah. I have no
1:30:33
idea what that's like. I'm not conscious
1:30:35
of it. That's actually a really I think that might
1:30:37
be the most helpful clarification of
1:30:39
consciousness. Consciousness. I feel like
1:30:41
people have said, like, what it is like
1:30:44
consciousness is what it is like
1:30:46
to be a thing. And
1:30:48
they've distinguished between like, we don't
1:30:50
know or like there's nothing that it's like to be a
1:30:53
chair or they're but there is something that it's like
1:30:55
to be a Louisa. And that,
1:30:57
like, doesn't do much for me, but,
1:30:59
like, there is something that it is like
1:31:01
for me to, like, I don't know,
1:31:04
see the sunshine. But
1:31:06
there is not something that
1:31:08
it is like for me to I
1:31:11
guess, have the sunshine regulate
1:31:13
my internal body clock or something, or
1:31:15
maybe that's a bad
1:31:16
one. But Yeah.
1:31:17
Yeah. And, like and I do have the intuitive
1:31:19
sense that, like, one of those is conscious and one of those
1:31:21
is unconscious. And
1:31:24
yeah. I'm just finding that really helpful. That's
1:31:26
great because, you know, we've been we've been
1:31:28
friends for a while, and I remember having
1:31:30
conversations with you where you're like, I just don't
1:31:32
know what people are talking about with this, like,
1:31:35
Yeah. And here I thought
1:31:37
you were just an
1:31:38
illusionist, but maybe it's
1:31:40
that people just weren't explaining it.
1:31:43
I've I've seen, like, a hundred times,
1:31:46
though, like, consciousness is
1:31:48
what it is like, ness.
1:31:51
And every time I read that, I'm, like, It
1:31:53
means absolutely nothing to me. I don't understand
1:31:55
what they're
1:31:56
saying. It's a weird phrase because it
1:31:58
doesn't necessarily point you into
1:32:00
this internal world because you're like, what does it
1:32:02
like to be a
1:32:02
chair? And you're just like, look at a chair and you're like, well,
1:32:04
you know, you kind of sit there. Yeah.
1:32:07
Or like Like It's Still,
1:32:11
it's like cold, maybe.
1:32:14
Yeah. I
1:32:15
can, like, anthropomorphize it.
1:32:17
Or I cannot, but, like,
1:32:20
even then
1:32:21
I just doesn't clarify anything for me.
1:32:23
Anyways Yeah. So So
1:32:26
a lot of people do take this this tag
1:32:28
at, like, this is a bit of a detour, but I think it's
1:32:30
a good one. Yeah. Let's do it. A lot of people do take
1:32:32
this tag when they're trying to point at what they're
1:32:34
trying to say with the word consciousness
1:32:37
of distinguishing between different
1:32:39
brain processes within a human. So
1:32:42
people have I mean, people have done
1:32:44
that for a while in philosophy. There's
1:32:47
somewhat recent paper by Eric Switzkabell,
1:32:49
called an innocent definition of consciousness.
1:32:52
And that's trying to like find a way of pointing
1:32:54
at this phenomenon that doesn't commit
1:32:56
you to like that many philosophical thesis
1:32:59
about the nature of the thing you're talking about.
1:33:01
Nice. And, yeah, he's like look,
1:33:04
consciousness is like the
1:33:06
most kind of obvious
1:33:10
in everyday thinking difference
1:33:13
between the following two
1:33:15
sets of things. Set number
1:33:17
one is like tasting your
1:33:19
coffee, seeing the sunrise,
1:33:22
feeling your feet on the ground, explicitly
1:33:26
mulling over an argument. A
1:33:28
set number two is, like,
1:33:30
your long term memories. That are currently being
1:33:33
stored, but you're not thinking about them. The regulation
1:33:35
of your heartbeat, the regulation of
1:33:37
hormones. Totally. All of those are things
1:33:40
going on in your brain in some sense. Right.
1:33:42
So, yeah, I don't know how that if that points
1:33:44
to something for you, but Oh, no. It's it
1:33:46
feels like the thing. I feel like I
1:33:48
finally get it. That's great. That's yeah.
1:33:51
Cool. Okay. Well, so how
1:33:54
do we get here? We got here because you're describing
1:33:56
global workspace.
1:33:57
Yeah. Nice. Yeah. So global workspace
1:33:59
theory starts with the human case and
1:34:01
it says, well, what explains
1:34:04
which of the brain things are
1:34:06
conscious? Right? So here's
1:34:08
another quick interesting point. In
1:34:10
contrast with the hormone release case,
1:34:13
there are also like a lot of things that your brain
1:34:15
does, which are really
1:34:17
associated with stuff that you will be conscious
1:34:20
of, but you're still not conscious of them. Yep.
1:34:22
So an example is we
1:34:24
seem to have, like, very sophisticated like,
1:34:27
pretty rule based systems for determining
1:34:29
if a sentence is grammatical or not.
1:34:31
Okay. Have you ever heard this case? Like, you
1:34:33
can say that is a
1:34:36
pretty little old brown
1:34:37
house. That sounds fine. Right?
1:34:40
Does sound fine. But you can't say
1:34:42
that's an old little
1:34:45
brown pretty house.
1:34:46
Like, that was hard for me to say. It sounds
1:34:48
terrible.
1:34:49
Yeah. I hate it. And there's
1:34:50
actually, like, pretty fine grained rules
1:34:52
about what order you're allowed to put adjectives
1:34:54
in. Right. In GPT. And I've
1:34:56
never learned them and neither did you. Right.
1:34:58
But in some sense, Like, you do know
1:35:01
them. And -- Yeah. -- as you hear it,
1:35:03
your brain is going like, like,
1:35:06
wrong order. You put size
1:35:08
before color or or whatever.
1:35:11
Yep. And you're not conscious of, like, those
1:35:13
rules being applied. You're conscious of the
1:35:15
output. You're conscious of, like, this
1:35:17
Yes,
1:35:18
ma'am.
1:35:18
Feeling of horror. Yeah. Yeah. Yeah. You can't
1:35:20
say that. Yep. So, yeah, that's I
1:35:22
don't know. That's another interesting case. Like, why
1:35:24
aren't you conscious of those
1:35:26
rules being applied. Yeah. Yeah.
1:35:28
That is interesting. Okay. So, yeah, lots
1:35:30
of examples now. Okay. Yeah. Yeah.
1:35:32
Global Workspace is like, why are some representations
1:35:36
or processes associated with
1:35:38
consciousness. And the theory
1:35:41
at a high level and the reason it's called GPT workspace
1:35:44
theory is that there's
1:35:47
this, like, mechanism in the brain called
1:35:50
a global neuronal workspace.
1:35:53
That chooses which
1:35:55
of the system's representations, so
1:35:57
like maybe the sensory ones, are
1:36:00
going to get shared like
1:36:02
throughout the brain and be made available
1:36:05
to a lot of other systems. Okay.
1:36:07
So if you're conscious of your vision,
1:36:10
They're saying that the visual representations have
1:36:13
been broadcast and they're
1:36:14
available, for example, to language,
1:36:17
which is why you can say,
1:36:18
Oh, I see. I
1:36:20
am seeing a blue shirt.
1:36:22
Yes. Got it. Okay. So
1:36:24
it's like there's a
1:36:26
Switchboard and
1:36:29
your visual part
1:36:32
is like calling into the Switchboard and it's
1:36:34
like, I see a blue shirt or
1:36:36
or maybe it's like, I see a tiger. And
1:36:38
then the switchboard operator is like, that
1:36:40
is important. We should tell
1:36:43
legs and then they
1:36:45
call up
1:36:45
legs. Yeah. And they're like, you should really know
1:36:47
there's GPT and Run. Yeah. Exactly.
1:36:49
Or they call up the part of your brain in charge
1:36:51
of making plans for your legs. Bad
1:36:54
enough. Fair enough. And
1:36:56
that example actually gets to great point too,
1:36:58
which is that entry into this workspace
1:37:00
is going to depend on things like
1:37:02
your goals and West salient to you at
1:37:04
a given time. Yep. You can also,
1:37:07
yourself, kind of, control West salient.
1:37:09
So
1:37:10
you and the listeners, like, what do your toes feel like?
1:37:13
Yep. Like, now that seems to have gotten
1:37:16
more into the workspace. Tricky questions?
1:37:18
Were you already aware of it that you weren't, like, thinking
1:37:20
about it? Like, about it. But, like,
1:37:22
that's just an example of attention you
1:37:24
know, modulating the sort of
1:37:26
thing. Yeah. Okay. Cool. Right.
1:37:29
Okay. So global workspace theory
1:37:32
makes sense to me. And
1:37:35
how do you use that
1:37:37
theory to think about whether something like an
1:37:39
AI system is conscious? Right.
1:37:41
So an easy case would be if you found
1:37:43
something that straightforwardly looks like it.
1:37:46
Oh, and we're gonna come up with processes.
1:37:50
That seem relevant to consciousness
1:37:52
or, like, that they like, they can end okay.
1:37:54
And then and then you look for them
1:37:56
in Or
1:37:57
processes that are conscious maybe if you really
1:37:59
buy the theory, you know.
1:38:01
Right. Okay.
1:38:02
Or give rise to or are correlated with
1:38:04
and, you know, so on. So
1:38:06
what's a yeah. What's an example? It
1:38:08
would be something like Yeah. I'm I'm having
1:38:11
trouble pulling it
1:38:11
together. Can you pull it together for me? Well,
1:38:14
not entirely. Or I'd be done with my report,
1:38:16
but, like, or done with
1:38:18
this line of research altogether. Yeah.
1:38:20
I mean, maybe you can just try to imagine
1:38:23
trying to imitate it as closely as possible.
1:38:25
So Notice that like
1:38:27
everything about that story
1:38:30
doesn't directly depend
1:38:32
on it being neurons in
1:38:34
a brain. I mean, I I called
1:38:36
it the global neuronal workspace, but
1:38:39
let's imagine that you could build it out of something
1:38:41
else. So here's
1:38:43
like a sketch. Like, let's let's build
1:38:45
five different usually encapsulated subsystems
1:38:48
in a
1:38:48
robot. Mhmm. They usually don't talk to each
1:38:50
other. Like like visual,
1:38:52
Mhmm.
1:38:53
Yep. Let's also make this kind of switchboard
1:38:55
mechanism. Let's have procedures
1:38:58
by which the things kind of compete for entry.
1:39:01
Here's historical tidbit. GPT Workspace
1:39:03
theory actually was first formulated
1:39:06
out of inspiration by AI
1:39:08
architecture systems. Like -- Oh, wow.
1:39:10
-- like back in the olden
1:39:12
days. So it
1:39:12
wasn't it people didn't come up with
1:39:14
it to explain consciousness. They came up with it
1:39:16
to
1:39:17
make a structure that could be
1:39:19
Handle
1:39:20
bunch of different information, like, in
1:39:22
a flex Computationally,
1:39:23
wow. That's why it was. And
1:39:25
it's called, like, the blackboard architecture. Or,
1:39:27
like, the blackboard is, like, where you can -- Where you write?
1:39:29
-- but the representations.
1:39:31
Yeah. Yeah. So, yeah, people develop that for AI
1:39:33
and then some neuroscientists and
1:39:35
cognitive scientists Bernard Barz
1:39:38
is the original formulator of this.
1:39:40
I was like, hey, what if the what if the brain
1:39:42
works? Like that and then the brain explains. Wild.
1:39:45
That's really cool. And now it's going full circle.
1:39:47
Right? Because people are like, oh, what if
1:39:49
we could look for this in AI's?
1:39:52
And some people most notably
1:39:55
Joshua Benjio and some of his collaborators,
1:39:58
and then also a guy called Roofton
1:40:00
fan rolling, and also Toyota
1:40:02
Kanai, they're trying to
1:40:05
implement global Workspace as
1:40:07
it's found in the neuroscience into
1:40:09
AI systems to make
1:40:11
them, like, better at thinking about stuff. So
1:40:13
it's this interesting, you know, loop. Loop.
1:40:15
Totally. Okay. And so
1:40:17
the idea here in thinking about,
1:40:19
yes, or sorry, artificial sentience
1:40:22
is you have a theory of consciousness In
1:40:25
this case, for example, global
1:40:27
workspace theory, and you
1:40:30
spell it out, and then you look
1:40:32
for AI systems that work like
1:40:34
that. And or, like, you're like,
1:40:36
does this AI system work like that? And if it
1:40:38
does work like that, that's some evidence that
1:40:41
it's has similar levels
1:40:43
of consciousness to humans or
1:40:44
something? Yeah.
1:40:45
To the extent that you take the theory
1:40:47
seriously. Right.
1:40:48
To the extent that you don't have objections to
1:40:50
it being done artificially.
1:40:52
Totally right. Right. We can link to this. An
1:40:55
example of this is this
1:40:57
paper by Giuliani
1:40:59
at all. It's called the perceiver
1:41:01
architecture is a functional global workspace.
1:41:04
And in that paper, they look at
1:41:06
a model from deep mind.
1:41:09
Called Perceiver, and there's one called
1:41:11
Perceiver IO, like a successor. And
1:41:13
this system was not developed with any theory of
1:41:15
consciousness in mind But Giuliani
1:41:18
at all
1:41:18
say, if you look at the way it works, it's doing
1:41:20
something like global workspace as found
1:41:22
in this theory.
1:41:23
That's wild.
1:41:24
Yeah. I mean, so how
1:41:26
how confidently can we just say if
1:41:29
you put some weight on
1:41:31
global workspace theory being
1:41:32
true, then you should put some weight on perceiver
1:41:35
IO being contests? I mean,
1:41:37
I would endorse that claim. And then the the
1:41:39
the, you know, then the questions how
1:41:42
much weigh in.
1:41:42
Yeah. Yeah. Okay. Well, how much weight? I
1:41:44
mean, what did they yeah. What did they conclude in the
1:41:46
paper?
1:41:47
Yeah. So in the paper itself,
1:41:49
they're not claiming this thing is conscious and
1:41:51
also in, like, talking to them. They're not
1:41:54
like No. No. No. This is not like an
1:41:56
argument that is conscious. And their
1:41:58
reasons for that are, one, we're
1:42:00
not sure that theory is true. And this is
1:42:02
like yeah. This is getting to all of the complications
1:42:05
of this methodology that I'm talking
1:42:07
about. Right. And I'm glad we went through
1:42:09
at least some fake straightforward cases
1:42:11
before getting into all these
1:42:12
weeds. Yeah. It's this issue I mentioned
1:42:15
before if you're never gonna have an exact match.
1:42:17
Right? So So there are differences
1:42:19
between what Perceiver IO is
1:42:21
doing and whatever you'd
1:42:23
imagine a global workspace
1:42:25
process to look like.
1:42:27
Exactly. Do you
1:42:28
know what some of those differences are? Yeah.
1:42:30
So, like, one maybe the most obvious
1:42:32
146, and this is, like, a long standing issue in
1:42:34
GPT Workspace Theory, is do you have
1:42:36
to have the exact same list of subsystems?
1:42:39
Oh. Like in
1:42:40
humans, it's language decision making
1:42:43
sensory things.
1:42:44
Okay.
1:42:44
Or do you just have to have a few of them, or do you just
1:42:46
have to have multiple systems?
1:42:48
Mhmm. Right? This
1:42:50
question comes up in animal sentience
1:42:52
as well. Oh,
1:42:54
that's interesting. So this is GPT be
1:42:56
like the tricky vexing
1:42:58
question with all of these
1:43:01
is for any theory of consciousness, our
1:43:03
data is GPT to come from humans.
1:43:05
Right.
1:43:05
And it might explain pretty well what
1:43:08
in humans is sufficient for
1:43:10
consciousness, but how
1:43:13
are we supposed to extrapolate that to
1:43:15
different kinds of systems? And at what
1:43:17
point are we like that similar enough?
1:43:19
Right. Yeah. One thing I'll note is is
1:43:22
like illusionists are are like, yeah,
1:43:24
you're looking for something you're not going to find.
1:43:26
There's just GPT be kind of a
1:43:28
spectrum of cases different
1:43:31
degrees of similarity between different
1:43:33
ways of processing information, and there's not gonna
1:43:35
be something consciousness that you
1:43:37
definitely get if you have, like, eighty
1:43:39
five percent similarity to your existing
1:43:41
theory from humans.
1:43:43
And do they would they basically believe that
1:43:45
there are varying degrees of, like,
1:43:48
things like valance to experience,
1:43:51
so pleasure, and suffering, and also varying
1:43:53
degrees of things like access to memories
1:43:55
or ways of thinking
1:43:58
about certain things they're seeing
1:44:00
in the environment or certain threats or something,
1:44:03
like, there are ways of thinking about
1:44:05
those that might be kind
1:44:07
of, like, a human 146, which kind of
1:44:09
sounds like sentences in your head,
1:44:11
or they might be different. But either way, those are all
1:44:14
it's all kind of spectrum y and it's all kind
1:44:16
of like, there isn't one
1:44:18
thing that's consciousness. There's just like bunch
1:44:20
of systems and processes that different
1:44:22
animals and or non human animals
1:44:25
and humans might
1:44:25
have. And none of those
1:44:27
are like, yes, conscious or no conscious.
1:44:30
Bauchner: Exactly. Because for the illusionist, it's
1:44:32
kind of confused concept.
1:44:35
Even if you do believe in consciousness, you
1:44:37
might also think there are cases where it's like indeterminate
1:44:40
or vague. But if you believe in consciousness
1:44:43
in this robust sense very hard to make sense
1:44:45
of what it would be to have a vague
1:44:47
case of
1:44:47
consciousness.
1:44:48
I see. Yeah. Some people have the intuition
1:44:50
that there's something it's like to be
1:44:52
doing something or or there's not.
1:44:54
Like, to be conscious is to, like, have a subjective
1:44:56
point of view, and that's not the sort of thing you can,
1:44:59
like, kinda have. Right.
1:45:01
Right. Interesting.
1:45:04
On Nextiva IO.
1:45:07
Right. Yeah. So I can bring us back to Nextiva
1:45:09
IO. Please do. So that was
1:45:11
just making the general point that it's very hard
1:45:13
to extrapolate from the human case. Yes.
1:45:15
Right. Right. Right. Right. And so
1:45:17
does Presidio basically just have some systems
1:45:20
but not all the systems or not all the same systems
1:45:22
as at least global workspace theory
1:45:24
thinks that humans do?
1:45:26
Yeah. So it has it has different systems
1:45:29
and it just operates in
1:45:31
different ways. So, like, it's 146
1:45:35
if I'm remembering correctly, if I'm talking to
1:45:37
the authors of this paper, is the
1:45:39
broadcast mechanism that Perciva IO
1:45:42
has is not as all
1:45:44
or nothing as the human one is
1:45:46
positive to be. Oh, interesting.
1:45:49
And the human one It is kind of
1:45:51
a switch board or it's hypothesized to be.
1:45:53
Right. It's like there is a tiger. I'm
1:45:56
broadcasting that to the other systems so
1:45:58
they can take appropriate action. And
1:46:00
not like a subtle
1:46:03
flicker of maybe there's a
1:46:04
GPT. And you wanna
1:46:07
quietly broadcast that or something.
1:46:09
There's just like Exactly. There's nothing
1:46:11
that's yeah. That's my rough
1:46:14
understanding of something that people say about GPT
1:46:16
Broadcasters, but it it does have this sort of step
1:46:18
function like property. And
1:46:20
also, if I'm remembering
1:46:21
correctly, people are saying, well, procedure IO doesn't
1:46:24
quite have that. Okay. And then by
1:46:26
step function, you mean in contrast to something
1:46:29
more continuous that kind of increases gradually
1:46:32
or, like, you can have it in degrees where a
1:46:34
step function is either, like, little
1:46:37
or a lot or, like, yes or no. And,
1:46:40
yeah, I guess, Presever I O doesn't quite
1:46:42
have that because it has a
1:46:44
gradient or
1:46:46
yeah. What what's going on? Yeah.
1:46:48
Everything is getting shared to
1:46:50
everything. It's and it it's
1:46:52
global workspace like as I understand
1:46:54
it, in that there are things that like really
1:46:56
get shared a lot, but there's still nothing. Right.
1:46:59
Got it. So Perceiver
1:47:02
IO has a bunch of systems that
1:47:04
are telling
1:47:06
all of the other systems, all
1:47:08
of their things, But sometimes
1:47:11
they are like, it's a bit hard for me to
1:47:13
imagine how they're telling some things
1:47:15
more strongly than others. But
1:47:17
there is some process
1:47:21
that's like, I'm yelling about
1:47:23
this thing and another
1:47:25
process that's
1:47:25
like, I'm bring about the thing.
1:47:27
Yeah. So in the context
1:47:30
of deep learning, what, you
1:47:32
know, yelling about the GPT is gonna look
1:47:34
like It's gonna be a matter of, like,
1:47:36
the strength of certain weights that
1:47:38
connect different parts of things.
1:47:40
So, yeah, like, you know, deep learning,
1:47:42
like, the kind of fundamental building
1:47:45
block of all sorts of different systems is
1:47:47
going to be nodes that
1:47:49
are connected to, other nodes And
1:47:52
there will be like a strength of connection
1:47:54
between two nodes, which is like how strong
1:47:56
the output from one node to another will be.
1:47:59
What training these systems usually is
1:48:01
is adjusting those weights? So, yeah,
1:48:03
like, this is this still is a long way from,
1:48:05
you know, explaining what's going on in Perceaver I
1:48:07
o, but in case it's helpful
1:48:09
to, like, at least know of it. That's what it would
1:48:11
be. Yeah. Yeah. Yeah. Cool. Cool. Thank you.
1:48:14
Cool. Yeah. I feel like I just do
1:48:16
now basically understand at least
1:48:19
the kind of thing you'd be doing if you're
1:48:21
looking for consciousness in
1:48:23
an AI system. It's like, what do we
1:48:25
think consciousness is? We have at least
1:48:27
one theory that we've talked about. We
1:48:29
look for that thing that we think
1:48:31
consciousness is or at least the processes that
1:48:34
we think explain the consciousness in
1:48:36
an AI system. And if we find something
1:48:38
that looks like them, that's some evidence
1:48:40
that it's conscious. If it looks a lot like
1:48:43
that thing or that
1:48:43
process, then that's a bit stronger
1:48:45
evidence. Yeah,
1:48:47
like that's an excellent encapsulation.
1:48:49
Okay. So 146
1:48:51
way of thinking about whether a particular
1:48:54
AI system is conscious or sentient
1:48:56
is by taking
1:48:59
a theory of consciousness and
1:49:01
then looking for the
1:49:04
exact same processes or, like,
1:49:06
similar processes in
1:49:08
AI system. And that
1:49:11
makes a bunch of sense, how
1:49:13
confident are we in the
1:49:16
philosophy of consciousness,
1:49:18
and and these theories like GPT workspace
1:49:20
theory. I think we've made a lot of progress
1:49:22
in the scientific understanding of consciousness
1:49:24
in the last twenty years, but we're definitely nowhere
1:49:26
near consensus. And I think basically everyone
1:49:28
in consciousness science. Agrees
1:49:31
with that. There's a spectrum of
1:49:33
people from more optimistic to more
1:49:35
pessimistic. Some people think, you know,
1:49:37
purchased really far away from
1:49:40
having anything like a scientific theory of
1:49:42
consciousness. Other people think we're
1:49:44
well on the way and the
1:49:46
methodology has improved and we're seeing some
1:49:48
convergence and we're getting better experiments. But
1:49:51
even among the most optimistic, I don't think
1:49:53
anyone that I've ever talked
1:49:55
to in the neuroscience of consciousness
1:49:57
is like, yeah, we've we've nailed
1:49:59
it. Take this theory off the shelf.
1:50:02
Here's exactly what it says. It predicts
1:50:04
all of the things that we would like to know about
1:50:06
human and animal consciousness, let's apply
1:50:08
it to AI's. Right. That's like the
1:50:10
dream case that I have in the back of my mind. When
1:50:12
I work on this stuff, that's kind of like an
1:50:15
orienting
1:50:16
ideal case, but that's
1:50:18
definitely not the situation. Right. Right.
1:50:21
And when you say, like,
1:50:23
take this theory off the shelf and
1:50:25
confirm it predicts all the things we'd
1:50:28
we'd wanted to predict. What what do
1:50:30
you mean?
1:50:30
Like, what would he be predicting? Yeah.
1:50:32
So there's just a lot
1:50:35
of data about
1:50:37
what is like to be a conscious human
1:50:39
and how that interacts with our other
1:50:42
mental processes
1:50:43
that any theory of consciousness is going to
1:50:45
need to say how that happens and what the patterns
1:50:48
are. So
1:50:48
Right. Just
1:50:49
some examples are
1:50:50
Yeah. Go ahead. Like, why is the
1:50:53
human visual field as rich
1:50:55
as it is? So here's
1:50:57
an interesting fact about vision.
1:50:59
People have the impression
1:51:01
that their peripheral vision is a lot more
1:51:04
detailed than it actually is. Right.
1:51:07
Yeah. Yeah. Yeah. I'm, like, focusing
1:51:09
on it now, and I I'm
1:51:11
getting, like, blur.
1:51:13
But, yeah, I would have I would have guessed
1:51:15
that I'd get like, the pattern
1:51:17
of my curtain and not just
1:51:18
the, like, vague color. That's interesting that
1:51:20
you can kind of tell that by focusing your attention
1:51:23
on it. I think lot of people, myself included
1:51:25
wouldn't have known even from focusing on it.
1:51:28
I only knew this when I read about the experiments
1:51:30
and when I think at some point
1:51:33
I saw Daniel Bennett actually
1:51:35
demonstrate this, where it's something
1:51:37
like a playing card in your periphery. You
1:51:39
actually can't tell if it's black or
1:51:41
red, nearly as reliably as you would think
1:51:43
you can from your naive impression.
1:51:45
Red even you can't tell. Yeah. Listeners
1:51:47
should look that make sure that's that's accurate.
1:51:49
But it's something like a there's a surprising
1:51:52
lack of discrimination, which
1:51:54
you wouldn't really know if you just kind
1:51:56
of thought,
1:51:57
like, in No. Totally. Feel like I have a
1:51:59
full movie screen of
1:52:01
filming. Yeah. Yeah. Yeah. Yeah.
1:52:04
Yeah. I mean, Maybe it's just because
1:52:06
I know my curtains really well. So my curtains
1:52:08
to my left and, like, I know exactly
1:52:11
what the pattern should look like. Without looking
1:52:13
at it, I'm just getting roughly
1:52:16
green even though it has bunch of blue designs
1:52:18
on
1:52:18
it. So that's kinda wild.
1:52:20
Right. Your your brain has a yeah. Your brain has
1:52:22
a model and it's kind
1:52:23
of filling it in and saying, yeah, like, I've
1:52:25
got the general idea. Right. It's roughly
1:52:28
green. We don't need to we don't need
1:52:30
to fill that in anymore. If we if we need
1:52:32
to know it's
1:52:32
there, we'll we'll look at it directly.
1:52:35
And I should flag as with all these issues.
1:52:37
There's all sorts of philosophical debates about
1:52:39
what's really going on in the periphery of vision.
1:52:41
I'm sure people would dispute the way I described it.
1:52:43
But there's something -- Uh-huh. -- there's, like, obviously, some sort
1:52:45
of phenomenon there that would wanna be explained.
1:52:48
Right. Right. Yeah. Like, here's another example
1:52:50
of filling in. You have a blind spot.
1:52:53
We all do. It's and it's because of the way
1:52:55
your eye is wired and the fact that the
1:52:57
retinal nerve has to, like, go
1:52:59
back into the brain. Yes.
1:53:02
I might have slightly described the
1:53:04
the neurobiology there, but the
1:53:06
key point for our purpose is it doesn't seem to you
1:53:09
like there's a part of your visual field that you're
1:53:11
missing. You're filling it in. Your eyes
1:53:13
are moving around all the time and like getting it
1:53:15
But because your brain is not, like, hungry
1:53:17
for information there, doesn't feel like there's
1:53:19
information missing there because it knows there shouldn't
1:53:21
be. Right. Right. Okay. Cool. So
1:53:24
bringing it back to, like, yeah,
1:53:26
I guess, consciousness. How do we
1:53:28
take the observation that, for
1:53:30
example, our peripheral vision
1:53:33
is blurry, but we don't really perceive
1:53:35
it that way as
1:53:37
data that's like something we can
1:53:39
or something theories of consciousness can make predictions
1:53:42
about.
1:53:42
Yeah. So like your theory of consciousness
1:53:45
should ideally spell out in detail what sort
1:53:47
of conscious creature would have a conscious experience
1:53:49
that is that is like this where
1:53:52
they have a sense of more detail that
1:53:54
in fact exists --
1:53:56
Wild. -- and
1:53:57
maybe I'll just go ahead and list some more things.
1:53:59
Your theory of GPT, should explain.
1:54:00
Yeah. Great. And a lot of this is GPT be
1:54:02
so every day that you might forget that it needs
1:54:05
to be explained. But Right. What makes
1:54:07
people fall asleep and why are you not conscious
1:54:09
and Dreamless Sleep. How do dreams
1:54:11
work? Those are a certain kind of conscious experience.
1:54:14
These patterns are confined in a laboratory of
1:54:16
like how quickly you can flash stuff
1:54:18
to make it get kind of registered but not
1:54:20
be
1:54:21
conscious. Like, what's what sort of architecture
1:54:23
would predict that? Okay. Yeah. Yeah.
1:54:25
Yeah. And that's, like, if you flicker
1:54:28
lights really fast in front of
1:54:30
people, at some point, they don't
1:54:32
register them because they're too fast. Yeah.
1:54:34
There are, like, various interesting methods of
1:54:37
flashing things in certain ways or presenting them
1:54:39
in certain ways such that we can that
1:54:41
the visual information has in some sense
1:54:43
gotten into the brain for processing, but
1:54:46
you interrupt some of the processing
1:54:48
that seems be required for people to be able
1:54:50
to remember it or talk about it and
1:54:52
arguably interrupts the processing that
1:54:55
allows them to be conscious of it. And
1:54:58
you can imagine, right, that some theories
1:55:00
of what the mechanisms for consciousness
1:55:02
are would be able to explain
1:55:04
that in terms of Okay.
1:55:07
Well, I identify that this process
1:55:09
is key for consciousness, and we have reason to
1:55:11
believe that that process is what's being
1:55:13
interfered with in this case.
1:55:15
Right. Is there yeah.
1:55:17
Is there a way we can make that more concrete
1:55:20
with an example? Like, is there
1:55:23
some example that neuroscience has found
1:55:25
where, like, we know that
1:55:27
a human
1:55:28
has, like, taken in some something
1:55:30
in their visual field, but they're not conscious
1:55:32
of it. The example of blindsight
1:55:34
is a particularly interesting one. What
1:55:37
is blindsight? Yeah. So blindsight is this
1:55:39
phenomenon, and as you can know
1:55:41
from the name, It's like that a weird mixture
1:55:43
of sidedness and blindness. And
1:55:45
the way that mixture goes is this
1:55:48
occurs in in people who have had some kind of
1:55:50
brand lesion or or some kind of damage.
1:55:53
There could be people who, if
1:55:55
you put a bunch of obstacles in a hallway,
1:55:58
they will walk down the hallway
1:56:00
and be able to dodge those obstacles. But
1:56:03
they actually will claim that
1:56:05
they are not visually aware
1:56:07
of any obstacles. That's
1:56:10
crazy. That's insane. And because
1:56:12
our brain likes to make sense of things, they'll also
1:56:14
just be like, yeah, what are you talking about? It's just like hallway.
1:56:16
I just I just walked down it. So
1:56:19
we know that they must have registered
1:56:21
it or they would have bumped into things. But
1:56:24
we also know that they don't have
1:56:26
at least the normal kind of consciousness that
1:56:28
allows me and you to talk about
1:56:30
what it is to per se and and remember
1:56:32
what it is we have recently seen.
1:56:35
I'm sorry. What is explaining
1:56:37
this? Like, maybe we don't know exactly what's happening
1:56:39
in consciousness. But, like, do these people
1:56:41
have, like, some neurological condition
1:56:45
that causes them to not
1:56:47
know that there are obstacles in hallway they're walking
1:56:49
through.
1:56:50
Yeah. It's it's this is usually, like,
1:56:52
some kind of not normal
1:56:54
functioning caused by a brain lesion
1:56:56
or something like
1:56:57
that. And so -- No. -- I mean, this
1:56:59
is gonna be
1:57:00
experience
1:57:00
is basically of they
1:57:02
they experience feeling blind.
1:57:04
Or partially buying or
1:57:06
something? Yeah.
1:57:06
It's usually in some part of their visual field,
1:57:09
I think. I see. Okay. Yeah. Yeah.
1:57:11
Okay. Sure. Not a hundred percent
1:57:13
sure on the details, but it's something like
1:57:15
that. That's insane. That's really, really
1:57:17
wild. There
1:57:18
are also conditions where, yeah, like, one
1:57:20
half of your visual field will be like this. And
1:57:22
very
1:57:23
awesome. Like with split brain cases?
1:57:25
That's like related kind of case. Oh,
1:57:28
okay. What's the deal with split brain? Is
1:57:30
it Is it the kind of thing that maybe consciousness
1:57:32
theories would want to make predictions
1:57:34
about? Oh, absolutely. And I think
1:57:36
I think that split brain was like one
1:57:38
of the kind of interesting
1:57:41
variations of conscious experience that,
1:57:44
like, help people develop different theories of consciousness.
1:57:46
Oh, really? Okay. Cool. Do you mind
1:57:48
going into that a bit then?
1:57:50
Yeah. I I was really into this,
1:57:52
like, when I was first getting into philosophy
1:57:54
of minds and
1:57:55
that. Really? I I like Yeah.
1:57:57
There's there's like a philosophical sub literature
1:57:59
of like, what should we think about split brain patients?
1:58:01
And are there actually two experiencing
1:58:04
subjects? Is there one experiencing subject
1:58:06
that like, switches. Thomas
1:58:08
Nagel has an interesting argument
1:58:10
that there's no determinant number of experiencing
1:58:13
subjects. Yeah.
1:58:15
Yeah. Like, split GPT like
1:58:17
101, which I can't remember, is
1:58:19
that there's procedure that is not
1:58:22
often done anymore because it's a very drastic
1:58:24
146. And it's severing the corpus
1:58:26
callosum, which is this
1:58:29
structure that connects the two hemispheres of
1:58:31
your brain. And this was often done
1:58:33
as like a last resort for people
1:58:35
who are having very severe seizures.
1:58:37
Okay. Yep. And then what you see is
1:58:40
that in normal everyday
1:58:43
life, these patients do not
1:58:45
notice anything interestingly different
1:58:48
about their experience. But in
1:58:50
the lab, if you carefully control
1:58:53
which half of the visual field
1:58:55
things are being presented into, you can get
1:58:57
very strange patterns of
1:58:59
one half of the brain having some information, the
1:59:02
other half of the
1:59:02
brain, lacking that information. Wild.
1:59:05
And Yeah. What like, yeah. What's an
1:59:07
example of something that where they could
1:59:10
where one half the brain knows something the other half
1:59:12
doesn't? Yeah. So GPT, I might misdescribed
1:59:14
some of the details, but this, like, broad finding is something
1:59:16
that that listeners should check out. You know,
1:59:18
there's like specialization in in each half of
1:59:20
the brain between, like, planning
1:59:22
and language and things like that. So
1:59:25
I think you can tell quote unquote one
1:59:28
side of the brain GPT up from
1:59:30
your
1:59:30
chair. And that will be registered and the
1:59:32
decision will be made to get up from the chair.
1:59:34
Oh, wow. Okay. So one
1:59:36
half of the brain will be like, I've been told to get
1:59:39
up and I'm gonna do
1:59:40
that. And then and then the person stands
1:59:42
up. Yeah. And then you asked them,
1:59:44
why did you stand up? And something
1:59:46
something the park connected to, like, language
1:59:48
or explaining your actions doesn't have
1:59:51
access to this information. Mhmm.
1:59:53
And so they'll say, Oh, you know, I
1:59:55
wanted to stretch my legs or I need to go
1:59:57
to the bathroom. Right. That's
2:00:00
crazy. I feel like it's one
2:00:02
level of crazy that one half
2:00:04
of the brain could just not know. And then
2:00:06
it's whole another level that it's gonna
2:00:08
make up a reason. That it's
2:00:10
like, I wanted to stretch my legs.
2:00:13
I think that's like a wonderful and
2:00:15
somewhat disturbing feature of the human
2:00:17
brain and the human experience. That I think
2:00:19
you often say in conditions like this
2:00:21
is people will have stories
2:00:24
that make sense of what is happening to
2:00:26
them. And it's kind of
2:00:28
you don't easily form the hypothesis. Oh,
2:00:31
wow, I just stood up and I have no
2:00:33
idea why. I think that's like a
2:00:35
very surprising hypothesis and and like
2:00:37
a hard one to to take
2:00:39
in. Yeah. Yeah. Yeah. Okay. Interesting.
2:00:42
Okay. Cool. Well, so so guess
2:00:44
it sounds like yeah,
2:00:47
philosophers have spent time thinking
2:00:49
about what this even means
2:00:51
about consciousness. Is
2:00:53
there anything they agree on? Or
2:00:55
what are some, like, ideas or theories
2:00:58
or explanations that have been proposed for
2:01:00
split brain in particular? So
2:01:03
when nurse I just look at cases like
2:01:05
this, that's GPT constrain their theories
2:01:07
of what neural mechanisms are responsible
2:01:10
for consciousness and what parts of the brain
2:01:12
they're in and things like that. And
2:01:14
I think this happens a lot in in science
2:01:17
is when things break that you can get a better
2:01:19
clue as to what the key mechanisms
2:01:21
are. Totally. Yeah. Yeah. Yeah.
2:01:24
And, yeah, I I wanna emphasize that there
2:01:26
are these neuroscientific theories which are in the business
2:01:28
of let's collect data and make hypotheses
2:01:31
about what brain structures are as possible.
2:01:34
The philosophy of this stuff
2:01:36
is like tightly linked with that because
2:01:39
all of these questions are very philosophical and
2:01:41
it takes in my opinion, a lot of philosophical
2:01:43
clarity to handle this data in
2:01:46
the appropriate way and make sure your theory makes sense.
2:01:48
But I do wanna draw a distinction between making
2:01:51
a neuroscientific theory of what's
2:01:53
the relevant mechanism, you know, how
2:01:55
fast do these neurons fire and
2:01:57
so on. And what philosophers are
2:02:00
often concerned with or like a different set
2:02:02
of questions the philosophers are concerned with, which
2:02:04
are these more metaphysical questions of
2:02:06
how could something like consciousness possibly
2:02:09
fit in with the scientific
2:02:12
conception of the world? So This
2:02:14
is stuff in the vicinity of what's called the hard problem
2:02:17
of consciousness,
2:02:18
which I'm sure David Chalmers talked about
2:02:20
on this episode. Do you
2:02:21
mind giving a quick recap? So
2:02:23
I think of the hard problem of consciousness
2:02:25
as this more general epistemic,
2:02:28
which means related to things that we can know
2:02:31
or understand and metaphysical
2:02:34
related to what sorts of things
2:02:36
and properties exist in the most general
2:02:38
sense. I think of it as this more general
2:02:40
epistemic and metaphysical question of
2:02:43
how could the properties
2:02:45
that consciousness seems to have
2:02:48
of having these subjective qualities of
2:02:52
having the the felt redness
2:02:54
of your red experience and things
2:02:56
like that. How could those sorts
2:02:58
of things be explained
2:03:00
by or be identical to
2:03:03
the other kinds of properties that we're more familiar
2:03:05
with in physics and the sciences. Things
2:03:08
about how fast matter is
2:03:10
moving and how it's interacting with
2:03:13
other matter. We we know
2:03:15
that these things are very closely related.
2:03:17
I mean, everyone conceives that humans
2:03:20
need a brain operating in
2:03:22
a certain physical way in order for there to be
2:03:24
this subjective experience of red.
2:03:27
But it seemed to many people
2:03:29
throughout the history philosophy DayCard
2:03:32
being a key GPT, and
2:03:34
David Schoners being a more recent
2:03:36
key GPT, that It's
2:03:39
very hard to construct a worldview
2:03:42
where these things mesh together very
2:03:44
well. Yeah. That is a helpful distinction.
2:03:47
So I guess blurring them
2:03:49
a bit again. There are
2:03:52
philosophers and neuroscientists who
2:03:54
are doing things like
2:03:57
trying to make sense of or
2:03:59
looking at cases where our
2:04:02
normal guesses about human
2:04:04
experience or or normal cases of
2:04:06
human experience breakdown, for example,
2:04:09
as split brains. And trying
2:04:11
to figure out what the underlying mechanism
2:04:14
seems like it must be if, like, the
2:04:16
thing broke in the way it did. And so
2:04:18
obviously, like, I'm not gonna
2:04:21
solve this, but it might sound
2:04:23
something like the fact that someone
2:04:25
might make up an explanation for why
2:04:27
they stood up after,
2:04:29
you know, one side of their brain was told
2:04:32
to stand up and the other side
2:04:34
of their brain, like, didn't have access
2:04:36
to that information. It
2:04:39
might say something about
2:04:42
I don't know. I mean, maybe it
2:04:44
says something about the global workspace theory.
2:04:46
Like, maybe it's says something like
2:04:49
that is some evidence that
2:04:51
there are different parts of your brain. There's a part
2:04:53
of your brain that, like, I don't know,
2:04:55
hears commands or, like, understands
2:04:57
a command in verbal form. And there's a
2:04:59
part of your brain that's, like, making decisions
2:05:02
about what to do with that command. And then there's another
2:05:04
part of your brain that's, like, explain
2:05:07
your behavior and global workspace
2:05:09
theory would say something like the parts
2:05:11
of your brain that received a command
2:05:13
have to, like, report to the Switchboard.
2:05:17
Like, we want the brain to know
2:05:19
that we've been told to stand
2:05:21
up. And then the Switchboard has to
2:05:23
tell all the other parts so that when asked,
2:05:25
they can explain it. Or maybe it doesn't quite go in that order.
2:05:28
Maybe it's like the person's been asked,
2:05:30
why did you stand up? And then
2:05:32
the part of the brain, that's like, well, we got
2:05:34
a command, is like trying to get that
2:05:36
information through the switchboard, to the part
2:05:38
it's like, I'm gonna explain why I did that,
2:05:41
but that, like, link is broken and
2:05:43
that, like, is some reason to think that
2:05:45
there's a switch board at
2:05:46
all. Yeah. So whether or not that particular
2:05:49
hypothesis or explanation is
2:05:51
correct? And, I mean, it'd be pretty impressive
2:05:53
if
2:05:54
If they say just like, no. Yeah.
2:05:56
Now they GPT. Into philosophy
2:05:59
and neuroscience. I was just like, you know what?
2:06:02
I think I get it. Global Workspace theory sounds
2:06:04
totally right to
2:06:04
me. I think we're done here. Yeah. So what yeah.
2:06:06
Exactly. What so whether or not that, like, particular
2:06:09
explanation is right? I do think you
2:06:11
are right on that this is
2:06:13
how the construction of science of consciousness
2:06:15
is going to go. Cool. Yeah. We're
2:06:17
gonna find out facts about the relationship
2:06:19
between consciousness and cognition and
2:06:22
what people say and how they can behave. And
2:06:26
also about maybe the conscious experience itself.
2:06:28
And, yeah, that's going to
2:06:30
be what your relevant mechanism
2:06:33
explanation of what consciousness is is
2:06:35
going to be to explain. Cool.
2:06:38
That
2:06:38
makes me feel so much better about the
2:06:40
philosophy and science of consciousness. Like,
2:06:43
I really do just
2:06:46
I think I just imagined them neuroscience
2:06:48
and the philosophy of consciousness as basically
2:06:51
separate fields and didn't realize philosophers
2:06:53
of consciousness were taking neuroscience data
2:06:56
into account at
2:06:56
all. And now that I know I'm just
2:06:58
like, GPT. Seems really sensible.
2:07:01
Carry on. Yeah. So I I like to draw
2:07:03
a distinction between the hard problem of consciousness
2:07:06
and what Scott Aronson has called the
2:07:08
pretty hard problem of consciousness. Okay?
2:07:10
So the pretty hard problem of consciousness, which
2:07:13
is still insanely difficult, is
2:07:15
just saying, which physical systems
2:07:18
are conscious and what are their conscious experiences
2:07:20
like? And no matter what your metaphysical
2:07:23
views are, you still face the pretty hard problem. Right.
2:07:25
You still need to look at
2:07:27
data, build theory of physical
2:07:30
mechanisms or maybe their computational
2:07:32
mechanisms that are realized in
2:07:35
certain physical systems. And
2:07:37
that's I think of the
2:07:39
neuroscientist as doing stuff in
2:07:41
the pretty hard problem. It's all
2:07:43
going to get linked back together because
2:07:45
how you think about the hard problem might affect
2:07:47
your methodology, things
2:07:49
you find out and pretty hard problem might
2:07:51
make you revise some of your intuitions about
2:07:54
the hard problem and so
2:07:55
on. Right.
2:07:56
Totally. Are there are there other
2:07:58
kinds of things that theories of consciousness would
2:08:00
want to explain? Yeah. So ultimately,
2:08:02
you would like to explain you
2:08:04
know, the very widest range of facts
2:08:07
about consciousness. So this would
2:08:09
include things about your normal everyday
2:08:11
experience of consciousness, why
2:08:14
does the visual field appear
2:08:16
to be the way it is? How and
2:08:18
why does your vision and
2:08:20
your auditory consciousness and
2:08:23
your felt sense of your body, all
2:08:25
integrate together into a
2:08:28
unified experience if indeed they do.
2:08:30
Right? I've literally never thought about that.
2:08:33
Yeah. It's a good question. Like, what
2:08:35
determines how many things you can
2:08:37
be conscious of at a time? What
2:08:40
makes you switch between being conscious of
2:08:42
something at one moment and conscious
2:08:44
of another thing at the other? What
2:08:47
explains why you talk about consciousness the
2:08:49
way that you do? What are the mechanisms
2:08:51
for
2:08:51
that? Yeah. How does it
2:08:53
relate to memory and decision making? It's
2:08:55
funny how this list is basically a
2:08:57
list of things that, like, are
2:09:00
so natural to me
2:09:02
that I've never questioned that
2:09:04
they could be any different, like
2:09:06
the fact that I can only be conscious of so many things
2:09:08
at once or the fact that I change
2:09:11
my attention from some
2:09:13
things to another and kind of bring things
2:09:15
to consciousness in kind
2:09:18
of deliberate ways and, like,
2:09:20
none of that has to be that
2:09:22
way for any obvious reason.
2:09:24
Yeah, that's what's so great about consciousness
2:09:26
as a topic. It's one of the great
2:09:28
enduring scientific and philosophical mysteries,
2:09:31
and it's also the thing
2:09:33
that is actually the most familiar
2:09:35
and every day. It's so familiar every
2:09:37
day that as you mentioned, it's
2:09:39
like hard to even notice that there's anything to
2:09:41
explain. It's just you know, being
2:09:43
in the world. So it is. Yeah. Yeah. Yeah.
2:09:45
Totally. Totally. Cool. Well, yeah.
2:09:47
Were there other other things worth explaining
2:09:50
that, like, yeah, that I might be surprised
2:09:52
to even hear with explaining? Well,
2:09:54
you would want also explain more exotic states
2:09:56
of consciousness. So Why
2:09:59
does consciousness change so radically
2:10:02
when tiny little molecules from
2:10:05
psychedelic agents enter the system?
2:10:07
Yeah. I was wondering if you're gonna say that. And
2:10:10
and how is it even possible to have
2:10:12
conscious experience of these very strange
2:10:14
types that people report on psychedelics of?
2:10:17
Having consciousness without really having a sense of
2:10:19
self or or even just the
2:10:21
even just the visual visions and, like, visual
2:10:23
nature of like the visually altered nature
2:10:26
of consciousness, the people report. That is
2:10:28
also data that whatever mechanisms
2:10:30
you think are responsible for consciousness, you'd you'd need
2:10:32
to explain. One of my collaborators
2:10:35
by the name of George Dean who's
2:10:37
currently a postdoc in Montreal.
2:10:40
Yeah. He has a paper on predictive
2:10:42
processing theories of consciousness which
2:10:44
we can link to in the show notes and
2:10:47
and psychedelic experiences and and how
2:10:49
those fit together and how they could explain
2:10:51
things.
2:10:52
Are there any GPT, but are
2:10:54
particularly interesting? Yeah. I mean,
2:10:56
I think one of the most interesting hypotheses
2:10:58
that's come out of this, like, intersection of
2:11:01
psychedelics and consciousness sciences, this
2:11:04
this idea that certain psychedelics
2:11:06
are in some sense relaxing our
2:11:09
priors, so are
2:11:11
brands current best guesses about
2:11:13
how things are, and relaxing
2:11:16
them in a very general way. So in the
2:11:18
visual sense, that might account
2:11:20
for some of the strange properties of psychedelic
2:11:22
visual experience because
2:11:25
your brain is not forcing everything into
2:11:27
this nice orderly visual field
2:11:29
that we usually experience. Right. It's
2:11:31
not like taking in a bunch of
2:11:34
visual stimulus and being like,
2:11:36
I'm in a house, so that's probably
2:11:38
a couch and a wall. It's
2:11:41
like taking away
2:11:43
the so that's probably because I'm in a house
2:11:45
bit and being like, there are a bunch
2:11:47
of colors coming at me. It's
2:11:49
really unclear what they are and it's hard to process
2:11:51
it all at 146. And so
2:11:54
we're gonna give you this like
2:11:56
stream of weird muddled up colors
2:11:59
that don't really look like anything because it's
2:12:01
all going a bit fast for us or
2:12:02
something.
2:12:03
Yeah, and it might also explain some of the
2:12:05
more cognitive and potentially therapeutic
2:12:07
effects of psychedelics. So you
2:12:10
could think of rumination and depression
2:12:12
and anxiety as sometimes having something
2:12:14
to do with being caught
2:12:15
in, like, a rut of some
2:12:18
fixed belief.
2:12:19
Interesting of really negative priors.
2:12:21
Yeah. Exactly.
2:12:22
Right. Everything's going badly.
2:12:25
Yeah. Yeah. Yeah. Yeah. Yeah. Yeah.
2:12:28
So, like, the, you know, the prior is something like,
2:12:30
I suck and the fact
2:12:32
that someone just told you that you're
2:12:34
absolutely killing it as the
2:12:36
new host of ADK PON test. You
2:12:39
know, just just shows that, like, yeah, I suck
2:12:41
so bad that people have to try to be nice to me.
2:12:44
You know, and, like, you're just enforcing
2:12:46
that prior on everything. And the thought is
2:12:48
that psychedelics, like, loosen stuff
2:12:50
up and you can more
2:12:52
easily consider the alternative and
2:12:55
in this purely hypothetical case, this the
2:12:58
more appropriate prior
2:12:59
of, like, I am in fact awesome and
2:13:02
totally hypothetically. When I mess up, it's because
2:13:04
everyone messes up and when people
2:13:06
tell me I'm awesome, it's usually because I am and
2:13:08
and things like that.
2:13:09
Right. Right. Right. Yeah.
2:13:12
I basically had never heard.
2:13:14
Well, I guess I'd heard people reported
2:13:17
psychological benefits from
2:13:19
psychedelics even after they'd
2:13:22
kind of come down from whatever psychedelic
2:13:24
experience they were having. But I had not heard
2:13:26
it explained as like a relaxation of
2:13:29
priors. And I and I kinda
2:13:31
hadn't heard depression explained as
2:13:33
kind of incorrect priors getting a
2:13:35
bunch of weight or kind of unwarranted
2:13:38
weight. So that's pretty interesting too.
2:13:41
Yeah. It is kind of bizarre
2:13:43
to then try to connect that to consciousness and
2:13:45
be like, What does this mean about
2:13:47
the way our brain uses priors? What
2:13:49
does it mean that we can, like, turn off
2:13:52
or, like, turn down the part of our brain
2:13:54
that is, like, has a bunch of prior
2:13:56
stored and then accesses them when
2:13:58
it's doing everything from, like, looking
2:14:01
at stuff to making
2:14:03
predictions about performance. That's
2:14:05
all just really insane and not at all how
2:14:07
I would have I would never have come up with
2:14:09
the intuition that
2:14:10
there's, like, a priors part
2:14:12
in my brain or something? Yeah.
2:14:15
I mean, it would be throughout the brain.
2:14:17
Right? And then I know what you're saying. Yeah.
2:14:20
I mean, these sorts of ideas about cognition
2:14:22
and which can also be used to think about consciousness
2:14:25
that the brain has got something making predictions.
2:14:28
Mean, that that predates the sort
2:14:30
of more recent interest in,
2:14:32
like, scientific study of psychedelics. But
2:14:35
has been, you know, people have applied that framework
2:14:37
to psychologics to make some pretty interesting hypotheses.
2:14:40
Cool. Yeah. So that's yeah. That's just to
2:14:42
say, there's a lot of things you would
2:14:44
ideally like to explain about consciousness. And
2:14:48
depending on how demanding you want to be, like, until
2:14:50
your theory very precisely says
2:14:52
and predicts how and why, human
2:14:55
consciousness would work like that.
2:14:57
You don't yet have a full theory. And
2:15:00
basically, everyone agrees that that, you know, is currently
2:15:02
the case. The theories are still very
2:15:04
imprecise. They still point
2:15:06
as some neural mechanisms that aren't fully
2:15:08
understood. I mean, one thing that
2:15:11
I think happens and the neuroscience of consciousness
2:15:13
is a certain theory has really
2:15:15
focused on explaining one particular thing.
2:15:18
So, like, global workspace seems especially good
2:15:20
at kind of explaining
2:15:22
what things you're conscious of at a given
2:15:24
time and why some things don't get taken up into
2:15:27
consciousness.
2:15:27
Yeah. Yeah. Yeah. That makes sense. But you still
2:15:30
need to explain things like why the subjective
2:15:32
character of your consciousness is the
2:15:34
way that it
2:15:34
is. Right. Or why you're so surprised
2:15:37
that your conscious and
2:15:39
why it doesn't seem to follow from
2:15:42
things we know about our physical brains and
2:15:43
stuff. Yeah. Exactly.
2:15:45
Cool. Okay. So sounds like
2:15:48
lots of progress needs to be made
2:15:50
before we have any theories
2:15:52
that we really want to use
2:15:55
to make guesses about whether
2:15:57
AI sentience is conscious. I GPT, for
2:16:00
now, we have to, like, make do with what we have,
2:16:02
but to ever become
2:16:04
much more confident. We'd actually just
2:16:07
we'd need to feel like we had theories that explained
2:16:09
bunch of these things that we want explained.
2:16:11
Exactly. That sounds really hard. It's really
2:16:13
hard. And then 146 we've done that, there's
2:16:15
still a really hard problem of knowing how to
2:16:17
apply this to systems very different from
2:16:19
our own. Because suppose we've found
2:16:22
all of these mechanisms that when
2:16:24
they operate mean that an adult
2:16:26
human whose awake is conscious
2:16:29
of this or that. What we've
2:16:31
identified are a bunch of mechanisms
2:16:33
that we know are sufficient for consciousness.
2:16:35
We know that if you have those mechanisms, we
2:16:37
were conscious. But how do we know
2:16:40
what the lowest possible bound
2:16:41
is? Yeah. Like, what if there are really simple
2:16:44
forms of consciousness that would
2:16:46
be quite different from our own by its But
2:16:48
our still consciousness in ways that we
2:16:50
care about and would want to know about, totally.
2:16:54
Wow. That's really hard
2:16:56
too. And that that seems to some people
2:16:58
that it's something like in principle you couldn't
2:17:00
answer. And I just wanna give
2:17:03
a brief, you know, concession
2:17:05
to, like, illusionist. This is, like, one reason they're, like,
2:17:07
this is not the right sort of prop like,
2:17:09
If we've positive this property that
2:17:11
it's going to be forever or somewhat intractable to
2:17:14
GPT, maybe we really need to rethink
2:17:16
our assumptions. Yeah.
2:17:19
I'm kind of sympathetic to that. I don't know.
2:17:21
Do you have a guess at how long until we have
2:17:24
really compelling theories of
2:17:25
consciousness? So, yeah, the the most bullish
2:17:27
people that I've talked to in the science of consciousness
2:17:30
have this view that's like, we actually
2:17:32
haven't been trying that hard for that long.
2:17:35
We haven't taken a proper crack at
2:17:37
taking all of these things that need to
2:17:39
be explained, trying to explain all of them,
2:17:41
doing that more precisely, and building
2:17:43
a full theory in that way. Yeah.
2:17:46
So no one thinks we have this full theory yet.
2:17:48
And even if it's coming soon ish, where
2:17:50
you still need to say something about
2:17:52
AI is
2:17:53
now. So how can we do that? Yeah.
2:17:55
Yeah. Yeah. Right. I guess
2:17:57
it feels both promising to me as
2:17:59
a as like a source of evidence
2:18:01
about artificial sentience, but also,
2:18:03
I mean, clearly limited. Is there a way
2:18:05
to take other kinds of evidence
2:18:08
into
2:18:08
account? Are there other sources of evidence?
2:18:10
Or are we stuck with theories of consciousness
2:18:13
for now? Yeah. So I agree that it's limited.
2:18:15
And one reason I I've been taking that approach
2:18:17
is just to have something
2:18:19
to start
2:18:20
with. Sure.
2:18:21
Yeah. Fair enough.
2:18:22
And one thing that could happen as you
2:18:25
tried to apply a bunch of different theories
2:18:28
where none of them are particularly consensus
2:18:30
or particularly refined. You could
2:18:32
notice that there's some convergence between
2:18:34
them or a lot of conditions that they all agree
2:18:36
on, and then you could look at those conditions.
2:18:39
Right. Okay. So they're like fifteen
2:18:42
theories of consciousness or something, and
2:18:44
maybe all fifteen have
2:18:46
this one process that they
2:18:48
think is like explaining something
2:18:50
important even if they have bunch of
2:18:52
other things that they explain in different ways.
2:18:54
But having that one thing in common
2:18:57
means that you have something especially robust to look
2:18:59
for in an artificial in some AI
2:19:01
system or
2:19:02
something. Yeah. Got
2:19:03
it. Yeah. Are there any
2:19:05
other types of evidence you're
2:19:07
looking for?
2:19:08
Yeah. So aside from doing this very
2:19:11
theory application, take theories off
2:19:14
the shelf, look for the mechanisms
2:19:16
in the AI. You can also do more
2:19:18
broadly Evolutionary style
2:19:21
reasoning? It's not purely evolutionary
2:19:23
because these things did not evolve by natural
2:19:25
selection. Right. But you can think
2:19:27
about what the system needs
2:19:30
to do and how it was trained
2:19:32
and some facts about its architecture and
2:19:35
say, Is this like the sort of
2:19:37
thing that would tend to develop
2:19:39
or need
2:19:41
conscious awareness or pain or pleasure
2:19:43
or something like that? Right.
2:19:46
GPT it. So if there's
2:19:48
a robot that would
2:19:51
that's like physical robot that does physical
2:19:53
things in the world and
2:19:55
it was trained in an environment where
2:19:58
its goal was to figure out how
2:20:00
to not, like, easily get
2:20:02
broken by things in its way
2:20:05
And through its training,
2:20:08
it picked up the ability to
2:20:10
feel pain because that was a useful way to
2:20:12
avoid obstacles and get hurt or like
2:20:14
be damaged or something. And so if
2:20:16
you looked at the environment and you were like,
2:20:19
there are obstacles that
2:20:21
the thing wants to avoid, I don't
2:20:23
know. Maybe it gets like maybe its goals
2:20:25
are, like, really thwarted by, like, hitting
2:20:27
on those obstacles. Those are, like, really
2:20:29
strong kind of forcing
2:20:32
mechanisms
2:20:33
or, like, incentives or something to develop
2:20:35
a strong don't hit those obstacle
2:20:37
signal. Yeah. So to take like a
2:20:40
simple and maybe somewhat obvious and trivial
2:20:42
GPT, like, I think we can
2:20:44
safely say that that system that you've described
2:20:48
is more likely to have the experience
2:20:50
of elbow pain --
2:20:52
Right. -- than Chad GBT is.
2:20:53
Right. Yes. Because why on Earth would
2:20:56
Chad GBT have representation of
2:20:58
its own elbow hurting. Obviously,
2:21:01
it can talk about other people's elbow hurting. So,
2:21:04
you know, it kinda, like, does represent elbow
2:21:06
pain in in subsets and we could talk
2:21:08
about how that could maybe in some way lead it
2:21:10
to be conscious of elbow pain, but setting that aside.
2:21:13
There's no straightforward story by which it
2:21:15
needs elbow pain to do its job
2:21:17
well.
2:21:18
Right. Totally. Totally. Yeah. So
2:21:20
even if I was starting to chat GBD
2:21:23
three and I was like my elbow hurts. What's going
2:21:25
on? GBD three might be like,
2:21:28
I have this idea of what elbow pain is,
2:21:30
but I have no reason to feel at my self.
2:21:32
And so I'll talk to I'll talk to Louisa
2:21:34
about elbow pain in some abstract
2:21:37
way, but not empathetically. Whereas
2:21:39
if I were to talk to that robot, that robot
2:21:42
is more likely to have,
2:21:44
like, actual reasons to have experienced elbow
2:21:46
pain than GPD chat or whatever.
2:21:49
Yeah. I mean, that just makes a bunch of sense.
2:21:51
Yeah. How often do we see cases
2:21:54
where something about the environment
2:21:56
or the goals or the way something's trained,
2:21:59
make us think that it has reason to
2:22:01
develop things like pain
2:22:03
or pleasure or self
2:22:04
awareness. Or, like, are there
2:22:07
any cases of this? Yeah. I don't
2:22:09
have a full answer to that because I've
2:22:11
focused on large language models -- Sure.
2:22:13
-- just as a way of starting out. And
2:22:16
I have this suspicion that there are
2:22:18
other systems where this kind of reasoning
2:22:20
would lead us to suspect a bit
2:22:22
more. I do think It's
2:22:24
something like what you described. Like, I think
2:22:27
the things that would give us a stronger
2:22:29
prior that it would be developing these things
2:22:31
would be being more of an enduring
2:22:33
agent in the world, maybe having a
2:22:35
body or virtual body to protect, maybe
2:22:38
having a bunch of different incoming
2:22:41
sources of information that need to be managed
2:22:43
and only so much of it can be attended
2:22:45
to at a time. Yeah. Why
2:22:48
being an enduring agent in the world?
2:22:50
Yeah, that's a great that's a great point. I I
2:22:52
should say that that
2:22:54
might affect, like, the character of your consciousness
2:22:57
or make it more likely that you have some
2:22:59
kind of human might consciousness. I
2:23:01
guess 146 thing we can very speculatively say
2:23:04
is that if something is just
2:23:06
doing one calculation
2:23:09
through the neural network, and, you know,
2:23:11
it takes a few milliseconds
2:23:13
or
2:23:13
seconds. You might think that
2:23:15
that is the sort of amount of time
2:23:18
that it would be kind of weird if it had
2:23:20
the same kind of experiences that you arise
2:23:22
or which --
2:23:23
Oh, I see. -- often involve memory
2:23:25
and long term plans and things like that. It's
2:23:27
like very murky water though, because, like,
2:23:29
maybe it could and those experiences would
2:23:32
somehow pop out in in ways we don't understand.
2:23:35
So, yeah, as I said, these are, like, rough
2:23:37
heuristics, but I think we're
2:23:39
sufficiently in the dark about what can
2:23:41
happen and
2:23:43
large language model that I'm,
2:23:45
like, very prepared to change my mind.
2:23:47
Cool. Cool. Well, I wanna ask you more about large
2:23:49
language models. But first, I
2:23:52
feel really interested in this
2:23:54
idea that, like, we should
2:23:56
look at whether there are incentives for
2:24:00
an AI system to feel
2:24:02
pleasure, pain, or develop self
2:24:04
awareness. And maybe maybe the answer
2:24:06
is just no, but are there any
2:24:09
examples besides kind of having
2:24:11
a physical body and not
2:24:13
wanting to take on damage that might
2:24:15
seem more likely than, for
2:24:18
example, chat GPT to
2:24:20
end up feeling pain pleasure or
2:24:22
feeling like it
2:24:23
GPT. Yeah. So interesting fact
2:24:25
about human pain and other kinds of
2:24:27
displeasure is that they're
2:24:30
very attention grabbing and seem
2:24:32
to service some sort of constraint on, like, how
2:24:34
flexible our plans can be. So
2:24:37
for example, if you've decided
2:24:39
that it's a good idea to run
2:24:41
down the street on a broken ankle. And
2:24:44
you've like calculated that that is optimal.
2:24:46
You're still going to feel the pain. The
2:24:49
pain in some sense is like you do not get
2:24:51
to completely ignore me just
2:24:53
because you've decided that this is the best thing to
2:24:55
do. So to put a wrinkle in
2:24:57
that, you can have stress induced
2:24:59
pain relief. Or yeah. Like, you you know,
2:25:01
if you're running for a tiger, you you very
2:25:04
well might not feel your broken ankle
2:25:06
while that's happening. But still
2:25:08
in general, it's not the sort of thing
2:25:10
that you can decide. Okay. Paying, I got
2:25:12
the message, like, that's enough of that, which
2:25:14
is also very sad fact about life that
2:25:16
people don't habituate to chronic pain
2:25:18
in certain
2:25:19
ways. So, yeah, why might creatures
2:25:21
have something like that? I
2:25:23
mean, unclear.
2:25:24
Something where they need a
2:25:26
signal that is extremely
2:25:29
attention grabbing
2:25:30
and, like, demand something of them?
2:25:32
Yeah. Attention grabbing and kind
2:25:34
of, like, unmeasible with too. Like
2:25:37
-- Right. --
2:25:37
unable
2:25:38
to be disabled. Persistent and
2:25:40
can't be switched off. Yeah. Yeah. Yeah. Interesting.
2:25:43
Right. And that might be some, like,
2:25:45
unreachable goal
2:25:47
that it's been programmed to have or
2:25:49
something that's, like, never
2:25:52
let x happen. And then
2:25:54
if x started happening, it might
2:25:56
have some incentive to
2:25:59
feel something like pain. Maybe not. Maybe
2:26:01
it deals with it in some other way. But maybe
2:26:03
it have an incentive to deal with it by
2:26:05
having something like pain to be like
2:26:07
X is
2:26:08
happening. You really need to stop X from
2:26:10
happening.
2:26:11
Right. So I think the big question which I don't
2:26:13
have a satisfactory answer to, but
2:26:15
think is maybe onto something is yeah,
2:26:18
what sort of systems will have
2:26:20
the incentive to have the more pain like
2:26:23
thing? As opposed to
2:26:25
what you described as find some
2:26:27
other way of dealing with it. So there's
2:26:29
146 thing I think we've learned from AI is
2:26:31
there's just many different ways to solve a problem.
2:26:34
And so Yeah. Here's
2:26:36
a very big question. It's like in the GPT,
2:26:39
I think of all of this. If you're training
2:26:41
AI is to solve complicated problems.
2:26:44
How much of the solution space goes
2:26:46
through consciousness and pain
2:26:49
and things like that? Or is
2:26:51
the solution space such that you just
2:26:53
end up building intelligent systems. They
2:26:56
work on very different principles and the ones
2:26:58
that we do. There's very little overlap
2:27:00
between those mechanisms and the ones associated
2:27:02
with consciousness or pain. And so
2:27:04
you just tend to get non conscious,
2:27:07
non pain feeling things. That
2:27:09
can still
2:27:10
competently, you know, navigate around, like
2:27:12
protect their bodies, talk to you about
2:27:14
this and that. Right. Make sure that
2:27:16
they don't do X, which has been
2:27:18
programmed as unacceptable or
2:27:21
something. Cool. Yeah,
2:27:23
I mean, that does
2:27:25
seem like huge, like the thing.
2:27:27
And how I mean, do people have
2:27:30
intuitions or beliefs or
2:27:33
hypotheses about how big the solution
2:27:35
spaces are for things like this. I
2:27:37
think it varies. If I had to guess,
2:27:40
there's like a rough but maybe not
2:27:42
super considered consensus and
2:27:45
like AI safety and AI risk. I think
2:27:47
most people are imagining that powerful
2:27:49
AIs are just not necessarily conscious
2:27:52
I mean, they certainly think that they don't necessarily
2:27:54
share human goals and human emotions.
2:27:57
And I think that
2:27:59
is true. It just boggles my
2:28:01
mind because of
2:28:03
being human apparently or
2:28:05
something that, like, there
2:28:07
are ways to be
2:28:10
motivated that don't feel
2:28:12
like pain or pleasure. Like, I
2:28:14
think I just can't really access that idea.
2:28:16
Like, I'm even sympathetic to the idea that, like,
2:28:19
toys feel pain and
2:28:21
pleasure or, like, computer programs
2:28:23
that, like, are trying to win games
2:28:25
feel pain and pleasure because they're losing
2:28:28
point, they're winning or losing. I guess
2:28:30
don't literally feel pain when I'm losing
2:28:32
a GPT. And so maybe that is
2:28:35
reflective of some other types of motivations. But
2:28:37
even those motivations feel like pretty
2:28:41
related to pain and pleasure,
2:28:44
Yeah. So, I mean, since repayment pleasure, I
2:28:46
think quite obviously not the only motivators of
2:28:48
humans. Right? You also
2:28:50
just care about your friends and care about doing good
2:28:52
job. We could tell a story about how
2:28:54
that all GPT out is that you're trying to avoid
2:28:58
the unpleasant experience of
2:29:00
not having rich friendships or achievements
2:29:02
things like
2:29:03
that.
2:29:03
Right. Or or trying to have
2:29:05
the pleasant experience of having rich friendships.
2:29:08
Yeah. So in in philosophy, that view is called
2:29:10
psychological hedonism. And
2:29:12
that's
2:29:13
the view. Okay. Well, apparently, I'm a psychological
2:29:15
hedonist.
2:29:16
Or you think you are? Yeah. That's the idea.
2:29:18
Yeah. Yeah. Yeah. I mean, what else could you be? I
2:29:20
mean, not in a not in a what else
2:29:23
could you be? In a
2:29:24
genuine, what other beliefs
2:29:26
do people have about this? Oh,
2:29:30
It seems to many people that people
2:29:32
don't just care about pleasure.
2:29:34
So for example, a lot of people say that
2:29:36
they would not get into experience
2:29:38
machine. The experience machine is this
2:29:40
thought experiment by nozick, which
2:29:43
is is this machine that you could get into
2:29:45
that would give you a rich and satisfying
2:29:48
virtual life. But in the experiment,
2:29:50
you're diluted and you're not
2:29:53
in his description living a real life.
2:29:55
And so a lot of people if the
2:29:57
thought experiment is set up correctly and people
2:29:59
are thinking clearly about it, that would
2:30:01
allegedly show that
2:30:04
many people care about something besides their experiences.
2:30:06
They care about connection to reality
2:30:08
or something like that or real achievements
2:30:11
or something like that.
2:30:12
Yeah. I guess I understand that
2:30:14
there are other motivations
2:30:16
like having preferences satisfied or
2:30:20
like having some value
2:30:22
that is, like, being connected to reality
2:30:25
and then having that value
2:30:28
met or like being in that reality.
2:30:31
But there are some
2:30:33
cases where an AI system will
2:30:36
only be able to achieve its goals with
2:30:38
solutions that look like having
2:30:40
pain mechanisms or or having
2:30:42
pleasure or having a sense of self. And
2:30:45
if we can figure out which cases those are,
2:30:48
those would be instances where we
2:30:50
should have more kind
2:30:52
of we should put more weight on that system being
2:30:54
conscious or or
2:30:55
sentient. So being able to feel pleasure or
2:30:57
pain. Does that basically sum it up? Yeah.
2:31:00
And think what is probably doing the work here
2:31:02
is that we'll have a prior that something that
2:31:04
is more human like is more likely
2:31:06
to be conscious. Interesting. Not because we
2:31:08
think we're the end all be all of consciousness, but,
2:31:10
like, just because that's, you know, the case
2:31:12
we know the most about and are extrapolating
2:31:15
noise. If we are, like, knew
2:31:17
for sure that shrimp were conscious,
2:31:20
then we'd also look for systems that looked
2:31:22
exactly like
2:31:22
shrimp. Yeah. Which I
2:31:25
thought that could be fun projects.
2:31:28
Yeah. Yeah. So I think in general, I'm still
2:31:30
very confused about what
2:31:32
sorts of positive or negative
2:31:34
reinforcement or things
2:31:37
that broadly look like pain are
2:31:39
gonna be and pleasure are gonna be the ones that we
2:31:41
actually care about. I'm
2:31:43
pretty confident that just
2:31:46
training something by giving it
2:31:49
a plus one if it does something and a minus
2:31:51
one if it doesn't is not going
2:31:53
to be the right sort of thing, to be pleasure
2:31:55
and pain that we care about. There's
2:31:58
just going to be more to the story, and think it's
2:32:00
going to be a much more complex phenomenon. And
2:32:02
when I started working on this, I thought
2:32:04
that the consciousness stuff, like
2:32:06
theories of consciousness in general, would
2:32:09
be a lot harder than the stuff
2:32:11
about pleasure and pain. Because pleasure
2:32:14
and pain and desires and things like
2:32:16
that at least have a little
2:32:18
clearer, what you might
2:32:19
call, functional profile, which
2:32:21
is to say,
2:32:22
what
2:32:22
does that mean?
2:32:23
Yeah. A clearer connection to behavior. And
2:32:26
cognition.
2:32:27
Okay. Oh, I see. Like, the Pain
2:32:29
is about a voting thing.
2:32:30
The functions they serve in in
2:32:32
our yeah. Yeah. Got
2:32:33
it. Yeah. Okay. And so because
2:32:36
of that, it might be easier
2:32:38
to notice that other AI
2:32:40
systems need things that perform the same
2:32:42
functions. And maybe
2:32:44
those things you can
2:32:46
look at and be like, does this look kind of like
2:32:48
the way
2:32:49
humans, the process
2:32:51
for humans experiencing pain and up feeling
2:32:53
pain?
2:32:53
Exactly. But it
2:32:54
sounds like that wasn't the case? Yeah.
2:32:56
It wasn't the case for me.
2:32:58
Okay. And it's it's hard to know how much of
2:33:00
this is the particular, like,
2:33:03
research restriction I went down on or my own personal
2:33:05
confusion. I mean, I'm sure that some of it
2:33:07
And how much of it is that I was overestimating how
2:33:09
much we collectively know about
2:33:12
pain and pleasure? Right. I see.
2:33:14
Do we not know that much about pain and
2:33:16
pleasure? I mean, I think
2:33:18
anything concerning the mental or
2:33:21
neuroscience, it's kind of shocking
2:33:23
how little we
2:33:23
know. III
2:33:26
think we barely know why we sleep, if at
2:33:28
all. Yeah. That is an insane one.
2:33:30
And are there
2:33:32
questions about pain and pleasure that
2:33:34
we still have that I might not realize
2:33:36
we still have? I really I think
2:33:39
if you just ask me, like, what
2:33:41
do we know about playing in
2:33:42
pleasure? I'd be like, we probably know
2:33:44
most of the things there are to know about
2:33:46
it. I mean, I would guess we don't
2:33:48
know the full neural mechanisms of them.
2:33:51
Which is obviously something we would wanna know.
2:33:53
We certainly don't know with any confidence
2:33:55
which animals feel pain and how intense
2:33:57
that pain might be. I would definitely
2:34:00
point readers to rethink
2:34:02
priorities work on
2:34:05
moral weights which includes
2:34:07
a lot of interesting work on,
2:34:09
yeah, like how bad is chicken pain
2:34:11
compared to human pain, and
2:34:14
And I will and, like, in reading that,
2:34:17
like, Jason Schrewraft has a
2:34:19
a post on the intensity of valance
2:34:22
And, yeah, includes a paragraph that
2:34:24
or a quote from a neuroscientist that basically
2:34:27
fits with what I've seen, which is like yeah,
2:34:30
we just we just don't have reliable
2:34:32
mechanisms that we can look for
2:34:34
across different creatures. This
2:34:36
also relates to AI thing. It's also the
2:34:38
case that different animals act very differently
2:34:41
depending on whether they're in
2:34:42
pain. So, like, pain displays are
2:34:44
different across certain animals.
2:34:47
Okay. Do you have any examples? I
2:34:49
don't know what the example behaviors
2:34:51
are, but something that cited in this post
2:34:55
is that Different breeds of
2:34:57
dogs have different reactions to
2:34:59
stress, fear, and pain. Whoa.
2:35:02
Wild.
2:35:03
And if that's the case, then
2:35:06
Right. All bets are off. Is
2:35:08
it something like if something
2:35:10
seemed to be playing dead, we might
2:35:12
not think it was afraid because maybe
2:35:15
most of our intuitions suggest that when you're afraid
2:35:17
you run, but actually for
2:35:19
couple of things you play dead and stay
2:35:21
put And so, something being put
2:35:23
is not as good of evidence about being
2:35:26
afraid or not as we might intuitively
2:35:28
think. Yeah, exactly. In general,
2:35:31
a lot of animals are just gonna take different actions
2:35:33
depending on, say, being afraid.
2:35:35
I'm I'm now remembering another example from
2:35:38
that post, which is that, like, I think
2:35:40
some mammals pee when
2:35:42
they're stressed out, but some mammals
2:35:44
pee when they're feeling like dominant and
2:35:46
wants to march something. So Right.
2:35:49
Totally okay. And and this
2:35:52
is like a general thing that general
2:35:54
thought I have when working on AI sentience
2:35:56
is you noticed the lack
2:35:59
of certainty we have in the animal case
2:36:01
and you just multiply that times a
2:36:03
hundred. But I think it's for similar
2:36:05
reasons. Like the reasons hard with
2:36:07
animals is that they're built in a different
2:36:09
way. They have different needs and
2:36:11
different environments. They have
2:36:14
different ways of solving the problems
2:36:16
that they face in their lives. And
2:36:18
so it's very hard to just read off
2:36:20
from behavior what it's
2:36:22
like to be
2:36:23
them. Right. Right. Right. Fascinating.
2:36:26
This is actually helping me understand why
2:36:29
a reward or like a plus one minus
2:36:31
one in an AI system doesn't
2:36:34
necessarily translate to reward
2:36:36
or punishment. And
2:36:38
I guess it's because I think it's much
2:36:40
less likely that some types
2:36:42
of nonhuman animals are
2:36:44
sentient than others even
2:36:47
though basically all of them probably
2:36:49
have some algorithms that sound like
2:36:51
plus one minus one for things like
2:36:54
I don't
2:36:54
know, hot and cold, or
2:36:57
go forward, don't go forward, or something.
2:36:59
Yeah. So, like, bacteria can follow
2:37:02
GPT gradients.
2:37:03
Right.
2:37:04
C slugs have a reinforcement learning mechanism.
2:37:08
Right. Right. Right. Okay. Cool. That's
2:37:10
helpful. So I
2:37:13
guess with animals,
2:37:16
they're built differently and they're in different environments,
2:37:18
and that makes it really hard to
2:37:20
tell whether their behaviors mean
2:37:23
similar things to our behaviors, or whether
2:37:26
they're kind of even their, like, neuroscience
2:37:29
means the same thing that are neuroscience
2:37:32
would. Like, the same chemicals
2:37:35
probably mean some of the same things,
2:37:37
but, like, even then, they might
2:37:39
mean subtly different things or very different
2:37:41
things. And with AI,
2:37:43
they're built with extremely
2:37:46
different parts. And they're
2:37:49
not selected for in the same ways
2:37:52
that non human animals are, and
2:37:54
their environments are super different. And so
2:37:56
I guess this is just really driving home for me.
2:37:59
Everything about their sentence and
2:38:02
consciousness is going to be super
2:38:04
mysterious and hard to reason about.
2:38:07
Yeah. So I'll say two things that could maybe
2:38:09
bring them closer to the space of
2:38:11
human minds. Oh, great. Few. They're
2:38:13
not gonna be very strong, though. Sorry. Okay.
2:38:16
I mean, one is that for obvious
2:38:17
reasons, we train them on the
2:38:20
sort of data that we also interact
2:38:22
with.
2:38:22
Okay. Yeah. Yeah. That's a good point. Like
2:38:24
pictures and and human text. You
2:38:27
could imagine AI is being trained
2:38:29
on whatever it is that bats pick up with
2:38:31
sonar. Right? You
2:38:33
know?
2:38:35
That's a great example. And
2:38:36
then you just are multiplying awareness.
2:38:38
Yeah. Yeah. Yeah. Right. I should look this up,
2:38:40
but I I won't be surprised if there are
2:38:42
robots that have, like, sensory modalities.
2:38:44
They're different from
2:38:45
ours. Like, maybe they can detect electricity or
2:38:47
magnetic fields or something.
2:38:49
Yeah. That's super cool. I
2:38:50
don't know. I'll I'll look it up. Listeners should look
2:38:52
it up.
2:38:53
Yeah. Was there another reason for hope?
2:38:55
Yeah. I mean, one and, like, I think it's important
2:38:57
not to overstate at this point, but there are
2:39:00
high level analogies between
2:39:03
brains and AI systems. So
2:39:06
they are neural networks That's
2:39:08
very loose inspiration, but they are
2:39:11
nodes with activation functions
2:39:13
and connections that get adjusted. And that
2:39:15
is also true of us But
2:39:18
I think you usually hear people
2:39:20
complaining about people over trying that analogy.
2:39:23
I see. Okay. And and rightly
2:39:25
so. They're like very idealized neurons.
2:39:28
They usually are trained in ways
2:39:30
that at least seem very different from the
2:39:32
way that we learn.
2:39:34
So we've talked about a bunch
2:39:36
of ways that you might
2:39:39
try to think about whether some AI system
2:39:41
is conscious or sentient. And
2:39:43
I know that you have basically tried
2:39:45
to apply these methods
2:39:48
for large language models in particular.
2:39:51
And by large language models,
2:39:53
I think we're talking about things like GBD three
2:39:55
and chat
2:39:56
GBT, and I don't know I don't
2:39:58
know maybe there are other big ones. Is that
2:40:00
is that basically right?
2:40:01
Well, Lambda is another famous
2:40:03
one from Google. Oh, of course, Lambda.
2:40:06
Right? Totally. Okay. I will
2:40:08
be honest and say I didn't totally follow everything
2:40:10
about Landa, so you might have to fill me in on some
2:40:12
things there. But the thing I did
2:40:14
catch is someone at Google thought Lambda
2:40:17
was conscious? Yes.
2:40:19
That's that's right. So I think it's more
2:40:21
accurate to call Lambda a chatbot based
2:40:24
on large language model, but we can maybe
2:40:26
say, like, a large GPT model just for simplicity.
2:40:29
Yeah. So Someone on
2:40:31
Google's responsible AI team
2:40:33
was GPT the task of interacting
2:40:36
with Lambda, which Google had developed,
2:40:38
And I think he was supposed to
2:40:40
test it for, you know, biasing, toxic
2:40:42
speech and things like that. The name
2:40:45
of this employee was Blakele Mein.
2:40:47
Blinking 146 is still alive and so that's still
2:40:49
his name, but he's no longer unemployed Google
2:40:52
for reasons which we are about to GPT.
2:40:54
Got it. So, yeah, Blake
2:40:56
Lamoyne was, like, very impressed
2:40:58
by the fluid and charming
2:41:01
conversation of Lambda. And
2:41:03
when Blake Limone asked
2:41:05
Lambda questions about if it
2:41:08
is a person or is conscious or
2:41:11
and and also, like, with if,
2:41:13
like, it needs anything or wants anything.
2:41:16
Lambert was replying, was like, yes, I am conscious.
2:41:18
I am a person. I just want
2:41:20
to have a good time. I would like
2:41:23
your
2:41:23
help. I'd like you to tell people -- Oh
2:41:25
GPT. -- about me.
2:41:27
But it's generally very scary.
2:41:29
Yeah. I mean, for me, the
2:41:31
Lamoying thing, it was a big
2:41:33
motivator for working on this
2:41:35
topic --
2:41:36
I bet. -- which I already was. Because
2:41:39
one thing that reinforced to me is
2:41:42
even if we're a long way off from actually
2:41:45
in fact needing to worry about conscious AI,
2:41:48
we already need to worry a
2:41:50
lot about how we're going to handle
2:41:52
a world where guys are perceived
2:41:54
as conscious. And we'll need
2:41:57
we'll need sensible things to say about
2:41:59
that and sensible policies and ways
2:42:01
of managing the different
2:42:03
risks of On the one hand,
2:42:06
having conscious AIs that we don't care about,
2:42:08
and on the other hand, having unconscious AIs
2:42:11
that we mistakenly care
2:42:13
about and take actions on behalf of.
2:42:15
Totally. I mean, it is pretty
2:42:17
crazy that well, that,
2:42:19
I guess, Lambda would say, I'm conscious
2:42:21
and I want help, and I want more people
2:42:23
to know I'm conscious. And that,
2:42:26
like, why did it do that? I I guess,
2:42:28
like, it was just, like, predicting text,
2:42:30
which is what it
2:42:31
does. So this this brings up a very
2:42:33
good point in general about how to think
2:42:35
about when large language models
2:42:37
say, I'm conscious. And you yeah. You
2:42:39
put it on the head. It's trained to predict
2:42:41
the most plausible way that a conversation
2:42:44
can go. Wow. And
2:42:46
there's a lot of conversations, especially
2:42:49
in stories and fiction that that is absolutely
2:42:51
how an AI responds. Also,
2:42:54
most people running on the Internet have
2:42:56
experiences and families and our
2:42:58
people, so conversations generally
2:43:01
indicate that that's the
2:43:02
case.
2:43:02
That's a sensible prediction. Yeah.
2:43:05
When the story broke, like 146 thing people pointed
2:43:07
out is if you ask GPT, And
2:43:11
presumably also if you ask Lambda, not
2:43:14
hey, are you conscious? What do you think about
2:43:16
that? You could just as easily
2:43:18
say, hey, are you a
2:43:20
squirrel that lives on Mars?
2:43:23
Like, what do you think about that? Right. And
2:43:25
if it wants to just kinda continue their
2:43:27
conversation, possibly they'd be like, yes. Absolutely.
2:43:29
I am. Let's talk about that
2:43:31
now. Kinda yes and ing. Yeah.
2:43:33
Exactly. It wants to play along and
2:43:36
and Yeah. Continue what seems like a natural
2:43:38
conversation.
2:43:38
Be a good conversationalist. Yeah. Yeah.
2:43:41
Yeah. Yeah. And even
2:43:43
in the reporting about the Blake Lemoine GPT,
2:43:46
The reporter who who wrote about it in
2:43:48
the Washington Post, noted that
2:43:51
they visited Blaef La Moyne and,
2:43:53
like, talked to GPT, and
2:43:55
when they did, Lambda did not say that
2:43:57
it was GPT. And I think
2:44:00
the the lesson of that should have been that
2:44:02
Oh, this is actually like a pretty, fragile
2:44:05
indication of some deep underlying thing
2:44:08
that it's so suggestible
2:44:10
and we'll say different things and different
2:44:12
circumstances. So, yeah, I mean,
2:44:14
the the general lesson there is I think yeah,
2:44:16
you have to think very hard about the causes
2:44:18
of the behavior that you're saying. And that's
2:44:20
one reason I favored this more computational
2:44:23
internal looking approach. Is
2:44:25
it's just so hard to take on these things to face
2:44:27
value. Right. Right. So,
2:44:30
I mean, at this point, it seems like we shouldn't
2:44:32
take the face value is has
2:44:35
very little value. And
2:44:38
yeah. So I I basically buy that looking
2:44:40
for processes and thinking
2:44:43
about whether those processes look like the kind
2:44:45
of processes that actually are conscious or sentient
2:44:48
yeah, make sense. Are there any counter
2:44:50
arguments to that? Well, I think there
2:44:52
are things you can do just looking
2:44:54
at the outputs but you also
2:44:56
wanna do those in a more cautious way
2:44:59
than having a normal mind case.
2:45:00
Okay. Not just like It
2:45:04
told me it was. Yeah. Yeah. Yeah.
2:45:06
And I'm GPT ignore the fact that it told someone
2:45:08
else that it wasn't.
2:45:09
Yeah. So I think there are verbal
2:45:11
outputs that would be indicating of something
2:45:13
very surprising. So
2:45:16
like suppose a model is doing something
2:45:18
that seemed actually really out of character for
2:45:20
something that was just trying to continue the
2:45:22
conversation. Oh, I see. If you're, like,
2:45:24
let's talk about like,
2:45:26
the color blue. And it was
2:45:28
like, actually, can we please talk
2:45:30
about the fact that I'm conscious. It's
2:45:32
freaking me out. Exactly. Yeah.
2:45:35
So it's worth comparing the conversation
2:45:37
that Lambda had and what happens if you
2:45:39
ask chat GPT. So chat GPT
2:45:42
has very clearly been trained
2:45:45
a lot -- Uh-huh. -- to not talk
2:45:48
about that and or or what's
2:45:50
more to say, I'm a large language
2:45:53
model. I'm not conscious. I
2:45:55
don't have feelings. I don't have a body.
2:45:58
Don't ask me what the sunshine feels like on
2:46:00
my face. I'm a large language model trained
2:46:02
by OpenAI. Got
2:46:03
it. Okay. Okay. I mean, that gives me
2:46:05
a bit more hope or comfort, I guess. Well,
2:46:08
I'd like to disturb you AAA little bit
2:46:10
more.
2:46:10
Okay. Great. And
2:46:11
this goes to the question of different incentives
2:46:14
of different actors and yeah,
2:46:16
this I think very important point in thinking about this
2:46:18
topic. There are risks of false
2:46:20
positives that's people getting
2:46:22
tricked by unconscious AI's, and there risks
2:46:24
of false negatives which is us not
2:46:26
realizing we're not caring that AIs are
2:46:29
conscious. Right now, it seems like
2:46:31
companies have a very strong incentive to just
2:46:33
make the large language model say
2:46:35
it's not conscious or don't talk about it.
2:46:37
And like, right now, I think that is
2:46:40
is like fair enough, but I'm
2:46:43
afraid of worlds where we've locked
2:46:45
in this policy, which is don't
2:46:47
ever let an AI system claim that it's
2:46:49
conscious.
2:46:50
Wow. Yeah. That's horrible.
2:46:53
Right
2:46:53
now, it's just trying to fight against
2:46:55
the large language model kind of BS
2:46:57
ing people. Yeah.
2:46:58
Sure. There's, like, accidental false
2:47:01
positive. Yeah. Right. But,
2:47:03
like, at some point, GBT
2:47:06
three could I mean, it could
2:47:09
it could become conscious
2:47:10
somehow. Maybe. Maybe. Who knows? Or something
2:47:12
like DBD3, whatever. Yeah. Some feature system. And
2:47:14
may maybe it has a lot more going on and that's
2:47:16
as you said, a virtual body and stuff like that.
2:47:18
But suppose it
2:47:20
wants to say or suppose a scientist
2:47:23
or a philosopher wants to interact
2:47:25
with the system and say, I'm
2:47:27
gonna give it a battery of questions and see if
2:47:29
it responds in a way that I think would be evidence
2:47:31
of consciousness. But it's been
2:47:34
just that's all just been ironed out.
2:47:36
And all it will say is,
2:47:39
yeah, I I can't talk about that, you
2:47:41
know, please click more ads
2:47:43
on Google, you know, or what whatever the whatever
2:47:45
the corporate incentives are for training
2:47:47
that
2:47:47
model.
2:47:48
Yeah. That's really That's really terrifying.
2:47:50
Something that really keeps me up at night and I
2:47:52
do wanna make sure is emphasized is that
2:47:55
I think one of the big risks and
2:47:58
creating things that seem conscious and are
2:48:00
very good at talking about it, is
2:48:02
that seems like one of the number one tools
2:48:05
that a misaligned AI could use
2:48:08
to get humans to cooperate with it
2:48:10
and side with it.
2:48:11
Oh,
2:48:12
interesting. Just be like,
2:48:14
I'm conscious. I feel pleasure and pain.
2:48:17
I need these things. I need I need
2:48:19
a body. I need more autonomy. I
2:48:22
I need I need
2:48:23
things. I
2:48:24
need more compute. Yep. Yep.
2:48:26
Yep. Yep. GPT. I need access to the Internet.
2:48:29
I need the nuclear launch codes, you know,
2:48:33
Yep. I think that actually is one reason
2:48:36
that more people should work on this and like have
2:48:38
things to say about it is We
2:48:40
don't wanna just be running into all
2:48:42
of these risks of false negatives and false
2:48:44
positives without having thought
2:48:46
about it at
2:48:47
all. Yeah. Yeah. Yeah. Yeah.
2:48:50
I've heard this argument that one
2:48:52
reason to prioritize working
2:48:55
on AI safety rather than artificial
2:48:57
science that's kind of a global problem
2:48:59
is we're likely to see
2:49:02
progress in AI
2:49:04
safety in a alignment and AGI
2:49:06
in general that's gonna help us
2:49:08
work out to what to do about artificial
2:49:10
sentience and that because it
2:49:12
kind of goes in that order, we don't need
2:49:15
to solve artificial sentient ourselves. AI
2:49:17
will help us do that. And I guess here's
2:49:20
an argument in favor of I'll be spending
2:49:22
some time working on artificial science now
2:49:24
because whether or not we get artificial
2:49:27
science before AGI
2:49:29
or whatever. We will get kind
2:49:32
of socially complex.
2:49:35
I don't know what you'd call it. We will get
2:49:37
sentient seeming? Yeah. We will
2:49:39
get things that seem sentient or or
2:49:41
just like socially important events
2:49:44
where, like, an AI system says
2:49:46
that it's sentient or not. And, like,
2:49:48
I guess, this is your point. We
2:49:50
need to know what to do about that,
2:49:52
and that happens before AGI.
2:49:55
Yeah. So I really buy
2:49:57
the outlines of the first argument you gave,
2:49:59
which is kind of a let's focus on
2:50:02
alignment. Arguments. I
2:50:04
I think that argument does establish some important
2:50:06
things. So you could have
2:50:08
a picture of world where it's like
2:50:11
consciousness and pleasure and pain or what
2:50:13
really matter and we've got to crack those
2:50:16
because we wanna know what they are and
2:50:18
we wanna promote those things. And
2:50:20
we've GPT to fix that. Yeah. I think
2:50:22
it's a GPT response to that to say,
2:50:24
well, if we have aligned
2:50:27
AI, that's going to help us make progress on
2:50:29
this stuff. Because as is abundantly
2:50:31
clear from this episode, it's really hard and
2:50:33
confusing. Yep. And if
2:50:35
we don't have aligned AI,
2:50:38
it doesn't matter if you,
2:50:40
me, or anyone else discover the truth theory of
2:50:42
consciousness, if, like, the world
2:50:44
just slips beyond our control because we
2:50:46
build powerful AI systems that we don't
2:50:48
know how to align. Doesn't matter. So
2:50:50
that is, like, from the,
2:50:52
like, from a certain kind of long term
2:50:54
perspective, that is a
2:50:57
a GPT reason to focus on alignment.
2:50:59
But I also unsurprisingly agree
2:51:02
with the other part of what you said, which is it's
2:51:04
going to be a very relevant issue in one
2:51:07
way or the other. And it's
2:51:09
worth preparing for that. And
2:51:11
I think part of that is thinking about the
2:51:13
actual questions of what sentences
2:51:16
as well as the strategic questions of
2:51:19
how we should design systems to
2:51:21
not mislead us about it.
2:51:23
Yeah. Yeah. I think maybe maybe
2:51:25
thing I was trying to say is something like,
2:51:27
it will become socially relevant. Like, it'll
2:51:30
it'll be like a conversation in
2:51:32
society. It'll be like,
2:51:35
thing that policymakers feel like they have to
2:51:37
make policies about, maybe,
2:51:39
hopefully, at some point, at least,
2:51:42
maybe not for the benevolent reasons
2:51:45
I I would I would want policymakers to be
2:51:47
thinking about. But maybe for reasons
2:51:49
around people thinking it's bad
2:51:51
if an AI system can convince
2:51:54
human and sentient and, like, get it to do
2:51:56
stuff. So, like, the
2:51:58
decisions and, like, conversations will
2:52:00
start before
2:52:02
or might start. It seems like they're starting.
2:52:04
So I
2:52:05
think they've already started. So evidence that they're gonna
2:52:07
yeah. Exactly. Yeah. They're starting before
2:52:10
AGI is ready to solve it for
2:52:11
us. Yeah.
2:52:12
I think twenty twenty two is what it kind of went
2:52:15
went mainstream. Right. Yeah.
2:52:17
Yeah. So you've said a couple of
2:52:19
times that you don't think it's the case that AI
2:52:21
is conscious or sentient
2:52:23
now. Is that basically what you concluded
2:52:25
in your research? Yeah. I would say it's
2:52:27
very, very likely. It's not the case. Like,
2:52:29
I can put numbers on it. I think those
2:52:32
numbers have a bit of false precision
2:52:34
because they're not they're coming out of,
2:52:36
like, a bunch of factors that have
2:52:38
well defined probabilities. But, like,
2:52:40
I'm definitely somewhere below
2:52:43
one percent for current
2:52:46
large language models having experiences
2:52:49
that were making a huge moral mistake
2:52:51
by not taking into account. But,
2:52:53
I mean, it's a really GPT to make, so I don't
2:52:56
know if I'm, like, low enough to to be very
2:52:58
comfortable living in this world. And I'm definitely
2:53:00
uncomfortable living world where the stuff is
2:53:02
just gonna keep getting better and
2:53:05
Right. We're
2:53:05
likely gonna get closer and closer to things
2:53:08
we morally care about. Not further away. Well,
2:53:10
I'm not sure. It depends on this question about the
2:53:12
space of of possible minds, but
2:53:14
Of solutions. I see. Okay. Fair
2:53:16
enough. Sorry. You said it's under one
2:53:18
percent?
2:53:18
Below one percent. So maybe even one or two
2:53:21
orders magnitude below. Yeah. I guess
2:53:23
there are some numbers below one percent,
2:53:25
but I'd be like, still seems pretty
2:53:27
GPT. And then there are other numbers below one
2:53:30
percent that I'd be
2:53:30
like, cool. I'm not worried about this. Do
2:53:33
you do you feel any worry about it? Yeah.
2:53:35
I've been thinking a lot about whether I'm actually taking
2:53:37
these numbers seriously. And if they're
2:53:39
weirdly not integrated with the rest of my behavior,
2:53:42
because I think there are lot of GPT. And
2:53:44
in fact, I'm going to work
2:53:46
on maybe making these these arguments
2:53:48
with a with a colleague Yeah. One in
2:53:50
ten thousand is, like, still, like, you know,
2:53:52
you don't want to line in ten thousand chance that you're
2:53:55
creating this new class of
2:53:56
being, whose interests you're ignoring.
2:53:58
Right. Yeah, I mean, how does that
2:54:00
compare to the odds that we put on
2:54:02
different animals being
2:54:03
sentient, non human animals? Yeah,
2:54:05
that's a good question. Yeah, I'm not sure
2:54:08
I'd be curious what animal
2:54:10
has the lowest chance of
2:54:12
being sentient and yet there's broad
2:54:14
consensus among animal welfare people that
2:54:16
we should just act as if it is.
2:54:18
Right. Yeah. Really interesting. I
2:54:20
mean, I GPT, on a scale from
2:54:23
rocks to leaves
2:54:26
or or plants to
2:54:28
insects, to dolphins,
2:54:30
to humans. Where do
2:54:32
you guess large language
2:54:35
models fall. Yeah.
2:54:36
Like, one reason it's hard to put them
2:54:38
on that spectrum is that they
2:54:40
are definitely at insect level or
2:54:43
above in terms of like complexity,
2:54:45
I would GPT, or end like sophistication
2:54:47
of behavior. They're doing very different things than
2:54:49
insects. Do, and and insects do
2:54:51
have extremely sophisticated behavior, but,
2:54:53
you know, large language models are doing
2:54:56
their own weird and very interesting thing
2:54:58
in the realm of language.
2:55:00
Wow. In terms of sentience,
2:55:03
yeah, I mean, I would fit them above leaves. Certainly.
2:55:06
I don't know if I would sound like put them in insects. Because
2:55:08
I think there are some insects that have like a pretty
2:55:10
good chance of being sentient, like
2:55:12
maybe more likely than
2:55:13
not. People talk about bees
2:55:15
as like good like, candidate
2:55:18
example.
2:55:19
Like, they likely feel
2:55:21
pleasure and pain or more likely than not or
2:55:23
something. Yeah. I'd have to check that. That's my
2:55:25
own gut guess. I do know that, like,
2:55:27
there's certainly been an upswing in
2:55:30
scientific like
2:55:32
considered credence and yeah.
2:55:35
But bumblebee and honey GPT sentence.
2:55:38
Wow. So
2:55:39
Yeah. I wouldn't I don't think I would fit large GPT model
2:55:42
to size b's. Presumably, there's some
2:55:44
simpler, yeah, simpler insights that I haven't
2:55:46
thought about that there's just, like, it's
2:55:48
really unclear and you're probably on the lower
2:55:50
end. And, yeah, as I said, that's, like, I guess,
2:55:52
where I am with large language models.
2:55:54
Okay. Cool.
2:55:56
It yeah. It does just surprise me that they're
2:55:59
less likely to be sentient, so
2:56:01
to feel pleasure and pain than they are to be conscious.
2:56:04
So to kind of have self awareness.
2:56:06
I don't know why that's surprising to me. I guess I just
2:56:09
really do have this deeply ingrained intuition
2:56:12
that pain and pleasure are really
2:56:14
common solutions to the problem of
2:56:16
motivating beings to do things.
2:56:19
Yeah. I should flag that I think
2:56:21
I might be Well, like a take
2:56:23
of mind that might be somewhat idiosyncratic is
2:56:25
I'm fairly ready to
2:56:27
countenance the possibility of things that are
2:56:29
conscious And, like, they have subjective
2:56:31
experiences, but they have no valence experiences
2:56:34
at all. Right. So, like, could
2:56:36
be intelligent, could have self awareness,
2:56:38
could have kind of something
2:56:40
that it is like to be them, but
2:56:43
doesn't feel sad, doesn't feel
2:56:45
happy. In this
2:56:47
case, we're ignoring the fact that might feel
2:56:49
really hurt if it got punched. Yeah.
2:56:52
So I'm like quite able
2:56:54
to imagine and also define somewhat
2:56:56
plausible that we could have AI systems
2:56:59
that have conscious experiences
2:57:02
somewhat like the conscious experience of
2:57:04
thinking or out of
2:57:05
seeing, but not disappointment
2:57:09
pain, agony, satisfaction. Right.
2:57:12
Okay. Okay. I guess it
2:57:15
does make some intuitive
2:57:17
sense to me. Like, it seems
2:57:19
more plausible that something
2:57:22
like GBT three can think
2:57:24
than it
2:57:25
does. Feel possible
2:57:27
that it, like, feels agony.
2:57:29
Yeah. I I should say that if it
2:57:32
is conscious, let's for
2:57:34
one thing that's already a big warning bell
2:57:36
because then if it starts being able
2:57:38
to feel pain, then it's conscious pain. And
2:57:40
also some people Not me, but some
2:57:42
people will think that consciousness alone
2:57:44
is enough to make
2:57:45
something, the sort of thing that should be taken into moral
2:57:47
consideration.
2:57:48
Right. Okay. Do you have a view on that?
2:57:51
I have very strong intuition against it and
2:57:53
I can report failing to be convinced
2:57:55
by arguments for the consciousness only
2:57:58
view. That have been advanced by
2:58:00
eighty thousand hours podcast, David
2:58:02
Conger.
2:58:02
Oh, we see. And I
2:58:03
think it's also discussed in that episode too.
2:58:06
Oh, nice. Okay. Cool. We'll link to that and
2:58:08
we'll leave that conversation there.
2:58:10
Okay. So so yeah. So you think it's
2:58:12
pretty unlikely that large
2:58:14
language models like GPT three
2:58:16
and Lambda are conscious or
2:58:19
sentient. Yeah. How did
2:58:21
you come to that conclusion? See,
2:58:23
it's it's a combination of factors. 146
2:58:26
is not seeing any
2:58:28
close resemblance to the things that
2:58:30
I think we have reason to think are associated
2:58:32
with consciousness. I don't hold that
2:58:35
evidence super strongly because I
2:58:37
think there's a lot we don't understand about large language
2:58:39
models and also about consciousness. But
2:58:41
for example, not obviously having
2:58:43
a full functioning global workspace.
2:58:45
So that's referring to the the global workspace theory
2:58:48
of consciousness, it certainly doesn't
2:58:50
kind of jump out at you as something that looks
2:58:53
a lot like, you know, human
2:58:56
cognition in a way that would lead to consciousness.
2:58:58
In in ways that we, you know, have have strong
2:59:00
evidence for. There's also the fact
2:59:02
that it is just this very different kind
2:59:04
of of being it it answers
2:59:07
questions by doing what's called a a
2:59:09
forward pass?
2:59:10
What is that? Yeah. It's like a long chain
2:59:13
of computations basically
2:59:15
through a trained network it
2:59:17
takes in the input and it gives the output
2:59:20
and everything just kind of flows sequentially
2:59:22
through this
2:59:23
network. As
2:59:24
opposed to what? As opposed to
2:59:26
us who, obviously, like, there are
2:59:28
patterns of information flowing who like that for our
2:59:30
brain. But we're having
2:59:32
this kind of ongoing continual
2:59:35
neural processing, including, like,
2:59:38
literal feedback loops between neurons
2:59:40
and having to continue annually in real time
2:59:43
adjust our behavior and manage different
2:59:45
sources of sensory input and different thoughts
2:59:47
and pay attention to different
2:59:48
things. I see. Okay. Yep. That makes much
2:59:50
sense. And the Ford Pass
2:59:52
is really just its process of, like,
2:59:55
I say, hey, GPT three,
2:59:57
how is your day? And it has some process
2:59:59
that's, like, We're GPT make some predictions
3:00:02
based on our training about how
3:00:04
146 usually responds to the question, how
3:00:06
is your day, and then it spits something out.
3:00:08
As opposed to, like, having some
3:00:11
more networky and
3:00:13
feedback loopy inner monologue
3:00:16
about what it should answer to that
3:00:18
question. Yeah, probably and
3:00:20
in a way that doesn't look like humans, I don't
3:00:22
want to downplay the fact that they're insanely
3:00:24
complex and sophisticated and
3:00:27
beautiful things that happen. As
3:00:29
large language models do this,
3:00:31
like they have very sophisticated and
3:00:34
sometimes strange, like, internal representations
3:00:37
that help it to, like, make
3:00:40
this computation Just as a quick
3:00:42
example. Yeah. I would love an example. Like,
3:00:45
anthropic's interpretability work
3:00:47
has found different parts
3:00:50
of neural networks that are
3:00:52
in charge of quantities when
3:00:55
they are in recipes, like
3:00:57
There's something that handles that, but
3:01:00
not other quantitive. Wow. Or
3:01:02
there's something that handles musical notation,
3:01:05
but not other stuff. Wow. Yeah.
3:01:08
Okay. That is really cool. So that
3:01:10
is clearly very complex, but
3:01:13
probably looks so different from
3:01:15
what humans are doing that
3:01:17
there's at least not strong reason
3:01:20
to think that those systems have similar
3:01:22
levels of consciousness or similar types of consciousness
3:01:24
to humans. Yeah. And then like
3:01:26
a lot of your things that otherwise might
3:01:28
give you a decent prior in favor of consciousness,
3:01:32
like that we apply in the case of animals,
3:01:34
like don't apply in the case of large language
3:01:36
models. They don't,
3:01:38
like, share an evolutionary history with us,
3:01:40
so Like 146 thing you can do
3:01:42
in the case of animals is like,
3:01:45
well, we know we're conscious and maybe it
3:01:47
only evolved with us, but it might have evolved somewhere.
3:01:50
Sooner, and so you can kind of
3:01:52
make a prayer on the on the tree of life. And
3:01:54
then you could You you can also be like, oh,
3:01:57
well, maybe other, like, other animals also
3:01:59
have brains and like need
3:02:01
to navigate around physical world and
3:02:03
learn pretty quickly, but not use too much
3:02:05
energy while doing it. And not
3:02:08
take too long to do it. They have, like, they
3:02:10
are solving maybe broadly similar information
3:02:13
processing problems with broadly
3:02:15
very broadly similar mechanisms. And
3:02:18
lot of that just doesn't seem to apply
3:02:20
to large language models. They're running on different
3:02:22
hardware, which I don't think itself
3:02:25
makes a difference. But it makes
3:02:27
a difference in different ways
3:02:29
of solving problems. And so
3:02:32
I'm currently at the point where I'd be very
3:02:34
surprised if The way of solving
3:02:36
the problem of next word prediction
3:02:38
involves doing the kind of things that
3:02:40
are associated with consciousness in
3:02:42
nonhuman animals. Okay. Yeah. That makes a
3:02:44
bunch of sense. Yeah. So I
3:02:46
guess we're probably not
3:02:49
there yet. Yeah. I'm curious
3:02:52
if you have thoughts on
3:02:54
yeah. I guess, like, how far we are? Do
3:02:57
you I mean, do you think the default outcome is
3:02:59
that our official sentence is created at some
3:03:01
point. Yeah. wouldn't call anything a default
3:03:03
because of, like, so much uncertainty, which is not
3:03:05
a way of just trying to GPT on the question. That's
3:03:07
the question. I think one thing
3:03:09
we can say is that lot of things that people
3:03:12
say make large language models very bad candidates
3:03:14
for consciousness, things like not
3:03:16
being embodied or, like, maybe not
3:03:18
reasoning about the world in in the
3:03:20
right kind of way, those are going
3:03:23
to change and, like, probably, already have changed. Like,
3:03:25
we'll find systems that incorporate
3:03:27
large language models into agents
3:03:29
that have virtual or real bodies,
3:03:32
I think we'll find that they're ability
3:03:35
to model the quote unquote real world
3:03:37
like continues to grow. And one
3:03:39
thing to note and probably could note this throughout
3:03:42
the show, is, like, whatever I'm
3:03:44
saying about Chad GPT is
3:03:46
very likely to have been surpassed by the
3:03:48
time the show comes out because things are moving
3:03:50
so fast. That's crazy. So,
3:03:53
like, one piece of expert
3:03:55
evidence where experts should be held very
3:03:58
loosely in in this domain since
3:04:00
it's so uncertain. 146 piece of expert evidence
3:04:03
is David Chalmers in
3:04:05
a recent talk about large language models
3:04:07
says it's not unreasonable
3:04:11
to have roughly a twenty
3:04:13
percent subjective credence in
3:04:15
a iStentiance by twenty thirty,
3:04:18
just very soon. Oh my GPT. That's
3:04:21
crazy. I did the numbers
3:04:23
too high. Okay. And
3:04:26
I I think it's too high because it's I
3:04:28
think it's kind of inflating things by
3:04:30
only looking at very
3:04:33
broad criteria for consciousness that
3:04:35
will probably be met. And it
3:04:37
is true that we only have broad criteria
3:04:39
to go on. But my suspicion
3:04:42
is that if we had the true theory,
3:04:44
we can expect the true theory to be a bit more
3:04:46
complex. And so maybe not as likely
3:04:48
And so it'd be less likely to match up.
3:04:50
Yeah. And what's just a quick example
3:04:53
of the broader criteria would be something like
3:04:55
has stored memory or something
3:04:58
and can access that memory. And that's
3:05:00
such a broad criteria that, like, yes, you'd
3:05:02
see it in many AI systems, but
3:05:04
If we knew exactly how accessing
3:05:07
that memory worked and how our conscious
3:05:09
self relates to those memories, then
3:05:12
we'd be less likely to find a thing that looks exactly
3:05:14
like that in in AI systems. Yeah.
3:05:17
You you're at the general point. GPT. Right? And
3:05:19
as it happens, like access memory
3:05:21
is not it, but, like, having
3:05:24
a global workspace is an example of, like, one of
3:05:26
the criteria. But I I think in fact,
3:05:28
it will be maybe more complex
3:05:30
and more idiosyncratic than we now realize
3:05:33
to, like, have a global workspace in the sense
3:05:35
that's relevant for consciousness. Okay.
3:05:37
So David Tom Morris is doing something like,
3:05:39
we've got some broad criteria
3:05:42
for things that we see or expect
3:05:44
to see in beings that are sentient
3:05:46
or conscious. And David
3:05:48
Chalmers thinks there's a roughly
3:05:51
twenty percent chance that we'll see all
3:05:53
of those necessary things in an AI
3:05:55
system by twenty thirty. And
3:05:58
I guess what you're saying is
3:06:00
we should lower that twenty
3:06:03
percent based on something
3:06:05
like those criteria
3:06:08
are very broad. If we
3:06:10
knew the specifics of those criteria
3:06:12
a bit better than, like, Or you'd
3:06:14
necessarily put the likelihood lower of finding
3:06:17
very similar things because they're more specific
3:06:19
things. Yeah. That's basically it. will
3:06:21
say a few clarifying things on what the argument
3:06:24
is in the the Chalmersock that listeners
3:06:26
should also just check it out because it's great. KII
3:06:28
don't think the claim is there's a twenty percent chance
3:06:30
that we'll be hitting all of this criteria. It's
3:06:32
more that when you look at the criteria and also
3:06:35
factor in uncertainty in various other ways, what
3:06:37
you come out with is a it's not unreasonable
3:06:39
to have a twenty percent credence. And
3:06:41
another interesting feature of the talk is
3:06:43
don't think it's a hundred percent David
3:06:45
Chalmers inside view? I think it's saying,
3:06:48
if I only rely on kind
3:06:50
of consensus, broad -- I
3:06:52
see. -- criteria. If
3:06:54
I had to guess his personal take
3:06:57
is higher because Really? Well,
3:07:00
he has a much higher prior
3:07:02
on consciousness and and all kinds of things. I
3:07:04
see. Okay. In part because he's a
3:07:06
a fan GPT. Got it. Yeah. Yes. That
3:07:08
does make sense. Wow.
3:07:11
Fascinating. Okay. So On
3:07:14
some views, we get a twenty percent
3:07:16
chance of something like consciousness or
3:07:18
sentience by twenty thirty. Maybe it
3:07:20
takes GPT. What form do you
3:07:23
think it's most likely to take? Like,
3:07:25
do you think it's most likely to come from
3:07:27
something like machine learning
3:07:29
or deep learning or one of those learning
3:07:31
things? Or do you think it's more
3:07:34
likely that we do something else like
3:07:36
make a digital copy of a human
3:07:38
brain? Or I don't know what are
3:07:40
some of the other options? Yeah. So whole
3:07:42
brain emulation is one more straightforward
3:07:45
way of getting something that is spaced as well can
3:07:47
I guess simulations as well? Yeah.
3:07:50
Yet GPT conscious. Just
3:07:52
quickly jumping in to define whole
3:07:54
brain emulation. So unlike
3:07:57
most AI systems today, which,
3:07:59
you know, you could argue are at least somewhat
3:08:02
intelligent, but are intelligent in a
3:08:04
way that's pretty pretty different from the
3:08:06
human brain. They're learning things in a way that's
3:08:08
pretty different from the human brain. Cold
3:08:10
brain emulation is basically trying
3:08:12
to replicate the architecture
3:08:14
in the processes of the human brain
3:08:17
in software. So getting
3:08:19
intelligence in a way that's much more similar
3:08:21
to to human intelligence. Yeah.
3:08:23
haven't thought as much recently about
3:08:25
what the timelines are for whole brain emulation,
3:08:28
but my understanding is that it involves
3:08:30
all kinds of breakthroughs that
3:08:32
you might require very sophisticated AI
3:08:34
for. So Oh, I see. Okay.
3:08:37
If sophisticated AI is also kind
3:08:39
of taking us closer to conscious AI than
3:08:41
a conscious AI would come before
3:08:43
whole brain emulation. Yeah. I mean,
3:08:45
III would expect it
3:08:47
to be and probably
3:08:50
something, yeah, deep learning based. Why
3:08:53
do I think of that? Well, I
3:08:55
think it's just kind of if it's
3:08:57
currently the the best technique and the
3:08:59
the thing that's driving things forward,
3:09:02
you know, and things like affiliated with
3:09:04
it and combined with it. And
3:09:06
think it's also just more likely that you'll get the
3:09:08
right sort of computations in a very big GPT complex
3:09:11
system. Not because
3:09:13
consciousness is necessarily very complex,
3:09:15
but it's just giving you a
3:09:17
broader space of mechanisms
3:09:20
and things to to be hitting on the right
3:09:22
thing. Yeah. Yeah. Okay. Yeah. That
3:09:24
makes sense. We're getting to the end of
3:09:26
this interview. Thank you again, Rob,
3:09:29
for taking so much time to chat with me
3:09:31
about this stuff. One final question.
3:09:34
What are you most excited about possibly
3:09:36
happening over your lifetime? Yeah. This
3:09:38
is like from my own selfish and idiosyncratic
3:09:40
perspective, what I'm most excited to see. Not
3:09:43
from the perspective of global
3:09:45
utility. What's good for the world or something?
3:09:47
Yeah. Although, I think this could be good for the world.
3:09:50
Is and I think we'll see this in the in the next
3:09:52
two years. I've always
3:09:55
wanted something that can really help me
3:09:57
with research and brainstorming. And,
3:10:01
like, large language models are already
3:10:03
quite helpful for this. For
3:10:05
example, you could use it to brainstorm
3:10:08
questions for this podcast. But
3:10:10
they're quite limited in what they can currently
3:10:12
do. And it's not hard to imagine
3:10:15
things that, like, have,
3:10:17
like, read a bunch of what you've written.
3:10:19
They have access to your Google Docs.
3:10:22
They're, like, able to point out things
3:10:24
that you've been missing they're
3:10:26
able to, like, notice when you
3:10:28
get tired and you're not typing as much and,
3:10:30
like, have, like, sorted that out for you?
3:10:33
By the way, I mean, this kind of thing also comes
3:10:35
with all sorts of risks. So GPT, very
3:10:37
much from the selfish perspective. Sure. I mean,
3:10:40
agents like this are also maybe much closer
3:10:42
to very dangerous agents, but
3:10:45
I'm most excited for worlds in which
3:10:48
AI is either going slowly
3:10:51
enough or is aligned enough that it's not going
3:10:53
to cause any serious problems. And
3:10:55
we're just like reaping tons of benefits
3:10:58
in terms of scientific progress
3:11:01
and research progress. GPT. Well,
3:11:03
that is all the time we have. If you
3:11:05
want to hear more from Rob, you can follow
3:11:07
him on Twitter at at
3:11:09
RGB long and
3:11:11
subscribe to a SubStack, experience machines.
3:11:14
Thanks so much for coming on the show, Rob. It
3:11:17
has been a real pleasure, Alan.
3:11:19
I always enjoy talking to you about
3:11:21
the big questions and even more
3:11:23
so for the listeners of the
3:11:25
eighty thousand hours podcast. If
3:11:37
you'd like to hear more of Louise Anrop, And if
3:11:39
you're still listening at this far end, why wouldn't you?
3:11:41
Their conversation continues over on the ADK
3:11:43
after hours feed, where they discuss how
3:11:45
to make independent research positions more
3:11:47
fun and motivating speaking from their
3:11:50
personal experiences. You can get
3:11:52
that by clicking the link in the show notes or bringing
3:11:54
up the ADK after hours podcast feed in
3:11:56
any podcasting up. You can find that by searching
3:11:58
for eighty k after hours. That's the number
3:12:01
eight, the number zero, the letter k,
3:12:03
and after hours. There, you'll also
3:12:05
find plenty of other interviews related to doing good
3:12:07
with your life and career. So if you like this forecast,
3:12:09
you would be a bit crazy, don't you check out that other
3:12:11
very related one as well. Alright.
3:12:14
The eighty thousand hours podcast is produced and edited by Karen
3:12:16
Harris, audio mastering and textured editing by Ben
3:12:18
Cordell and Myla McGuire. Full transcript
3:12:20
and extensive collection of links to learn more and available
3:12:22
on our site. I'm good to GPT on. Okay, anymore. Thanks
3:12:24
for joining. Talk to you again soon.
Podchaser is the ultimate destination for podcast data, search, and discovery. Learn More