Latest / Why We Do What We Do / Artificial Intelligence with Joseph Dailey

Transcript

Artificial Intelligence with Joseph Dailey · Why We Do What We Do

0:03
You're listening to Why we do what we do. Welcome to Why we do
0:16
what we do. I am your barely conscious host Abraham, and I'm calculating
0:22
a lot of data really rapidly. Host Shane Nice, good job. We
0:26
are psychology podcast, and we talk about all the things that people do,
0:29
and sometimes the things that people make, and then the things that the things
0:32
that they make do. That's what we're talking about today, and we're so
0:35
lucky we get to have someone who's much more of an expert in this than
0:39
we are, and that allowed me to not have to read as much as
0:41
I normally have to read when I'm wanting together things together, and I'd like
0:45
reading, but it does take a lot of time. So thanks for being
0:48
here. I'd like to welcome our special guest, Joseph Daily. Joey,
0:52
can you introduce yourself? Absolutely, Hello, folks. My name is Joey.
0:55
I'm an electrical engineering student at University in Vatterino. I'm a graduate student
1:00
studying how we might be able to use reinforcement learning to make robots more pleasant
1:07
to work with, safer, smarter, better at doing what people want them
1:11
to do. I'm also a teacher and musician. And a very fortunate be
1:18
colleague of Abraham in the past, and very pleased to be here. Yeah.
1:23
Part of the reason that I felt like it made a lot of sense
1:26
to have have you come on to talk about this with us is because you
1:30
have had your sort of feet in both ponds. If you will, think
1:33
you are someone who understands behavioral philosophy really well and also understands AI at least
1:38
a lot better than anyone else that I know, and so being able to
1:42
blend those two in a way that allows us to have a cogent conversation is
1:47
really valuable. Yeah, I have the rare distinction of being slightly better than
1:52
lay person into fields if that's If that's sufficient, you know, I'm happy
2:00
to illuminate whatever I can here for you. You're aggressively humble, And when
2:06
I say an inch deep and a mile wide, like you know, you
2:08
know just enough about a lot of things. It's like you know enough to
2:12
have really cool conversations, but not enough to get on Jeopardy, which is
2:14
perfectly fine. That's like, that's our demographic might be a mile wide,
2:17
but my experience of conversations with you is definitely more than an inch deep.
2:23
This metaphor is getting awfully ephemistic. Let's go ahead and transition to our topic
2:28
here, and actually before we do that, just to remind everybody, if
2:30
you'd like to support us, you can join us on Patreon, pick up
2:32
some merch, leave us a rating and review, subscribe, send us nice
2:36
letters through a large language model AI generation machines. We appreciate that too,
2:39
and I'll talk more about this at the end of the discussion, but first
2:44
we should also recognize that if you're listening to this on the day then it
2:46
comes out we like to issh you a happy National Hot Buttered Rum Day.
2:52
Right, Okay, I'll go through a couple more. It's also National Bootleggers
2:55
Day. I'm not sure that's a professional. A lot of people of stuff
2:59
carry. It's also National I'm sorry, It's Museum Selfie Day, so if
3:02
you like paintings, that's a really cool thing to go do. It's also
3:07
Benjamin Franklin Day. If you are a fan of kites, keys, electricity,
3:10
and STDs. Sure you can bootleg a bottle of hot buttered rum into
3:15
a museum in Man. I mean you're really celebrating with a plom There are
3:22
more holidays but we'll leave it for now. I think we've got such an
3:23
interesting discussion. I'm excited to get to all right, anything else before we
3:28
jump into this conversation, Shane that you can think of. No, I'm
3:31
really excited about this. The first conference talk that I ever did was on
3:37
a paper that was about AI and behavior analysis, So I'm really excited to
3:40
dive into this. Oh man, if you remind me we should put a
3:44
link to that then this. No no, no, no, no,
3:46
no, no, we don't need to do that. I was way way
3:50
out of my scope and I did an hour a fifty minute talk in thirty
3:53
minutes because I was too nervous. So I'm so excited to talk about this
3:54
though. I love this stuff. I'm glad you said that because one of
3:58
the things that I think that may or may not come as a surprise to
4:01
some people, depending on where they're at in the news with the world,
4:04
I guess as how much longer AI has been around, then maybe it seems
4:09
like because the explosion of art generating AI and large language model AI in just
4:16
the last couple of years has made it seem like we haven't had AI before
4:23
this. We have, but it has definitely shifted dramatically, and I would
4:28
say that I think the word of the year for twenty twenty three was chat
4:32
GPT. I definitely heard that more than just about I thought it was going
4:35
to be spy balloons for a minute, but that pretty much petered out around
4:40
like maybe March or April. Yeah, and then it was chat GPT reigned
4:45
supreme. So let's actually let's quickly discuss what AI is and then we'll talk
4:49
about some of the really the most popular applications of it. Now, if
4:54
we come to a consensus on what AI is, we will be doing We
4:58
will be doing better and in the field has been today, We'll be doing
5:01
God's work here breaking on this podcast. Yeah, groundbreaking science probably a good
5:08
place to start. As Ada Lovelace, okay, because I guess the earliest
5:12
notion that we have of an artificial intelligence is just automation, right, the
5:17
ability to instruct a machine to do something and have it do it. It's
5:23
not a very robust form of intelligence. But with Ada Lovelace in the I
5:28
want to say, the mid nineteenth century, we kind of have the first
5:30
notion that we can make machines do sophisticated tasks. Even tasks where some emergent
5:40
behavior at the time of their execution might be necessary. I think Charles Babbage,
5:46
who was a contemporary of Ada Lovelace and whose work she's credited with improving
5:50
on. Charles Babbage was of the opinion that a machine with a few arithmetic
5:57
operations could do everything of consequence to a human being. Wow. And then
6:02
in the early twentieth century we started probing the limitations of systems like that.
6:11
This is where we get the first work that essentially shows you can't do everything,
6:16
at least not quickly or well with conventional simple computers. Fair yeah,
6:21
especially back then, especially back then. But it's even more extensive than that.
6:28
I mean, this was some of the foundational work of if you're familiar
6:30
with these names Church and Turing and some of their contemporaries in early early computer
6:38
science, that there are really fundamental limitations to what can be accomplished with what's
6:42
called, for example, a touring machine, which is a general model of
6:45
a computer. Sure, so that takes us into the middle of the twentieth
6:48
century and people start looking at like, okay, well, then we know
6:53
that there are things people can do that we haven't succeeded in implementing very robustly
6:57
with computers, which are supposed to be in some ways at least a lot
7:00
faster and a lot smarter. So how do we bridge that gap? And
7:05
people start looking then that neurology and figuring that there might be something structural in
7:11
the way that brains are hooked up that enables some unexpected emergent behavior. And
7:18
they're right, they're just not right for the right reasons yet. But so
7:21
in the nineteen sixties we got what's called a perceptron, and you should google
7:26
a picture of this thing if you get a chance one of the early perceptrons.
7:29
They're really funky. The notion here is that if you can define your
7:32
problem as a set of numeric inputs, it's like, say, your problem
7:38
is a picture that you want to classify as either being of a dog or
7:41
not of a dog. Zero one, Right, there's some linear combination of
7:47
the pixels that comprise that image such that you can scale them somehow, add
7:53
them up, maybe do that a few times in multiple layers, and eventually
7:58
get out a number that's a zero or one. And I say you should
8:01
look this up, because there wasn't a good way to implement this in silicon
8:05
or I guess at the time in tubes yet. So one of the first
8:11
implementations of a perceptron was a machine with like every individual linear weight in the
8:16
in the model quote unquote was a physical knob that would be turned with a
8:22
motor, and the motor would be worrying it back and forth as it progressed
8:26
through through the learning process until it honed in on exactly the right position for
8:31
each dial. Wow, it's bizarre. Yeah, I think it's about the
8:35
coolest thing ever built. And then that too quickly shows its faults. The
8:41
picture I'm looking at looks like a closet full of wires and knobs. This
8:45
is the mark one perceptron machine. It looks like you open a closet door
8:50
and it's just full of like cables and wires and stuff like it's pretty it's
8:54
pretty extensive. Yeah, And for perspective, this might be as much which
9:00
is a millimeter on a modern silicon die to implement something similar, like we
9:05
have come that far because we're still talking about like a computer is either a
9:11
human job description or a room in a building at this time, right,
9:16
Yeah, I know they've said that like the the computing power that it took
9:22
to send people to the moon, we carry like ten times that in our
9:28
pocket on our phone. Yeah, so we're talking about those heady days of
9:33
early electronic computing. So yeah, they start doing this, They find that
9:39
perceptrons can classify images really well, can handle certain kinds of data really well.
9:46
People get excited that, hey, maybe this is this is artificial intelligence.
9:50
This is not just like wrote algorithmic behavior. This is something legitimately novel,
9:58
and so they start trying to apply it to machine translation, and that
10:01
falls through pretty quick. Sure are they like asking how it feels. We're
10:07
not even there yet. We're at the foot where people are starting to are
10:11
trying to like apply inside of like Darba, are starting to try and apply
10:16
computers to the process of like translating English to French or whatever, and it's
10:22
just not There is nothing that's much better than just simple word substitutions. And
10:28
then at the same time, there's a proof that perceptrons cannot solve certain what
10:35
we call nonlinear classification problems. Meaning if if you were to plot all of
10:41
the data that you want to classify into two groups, and let's say you
10:46
have group one, which is red, and group two, which is blue.
10:48
Unless you can separate red and blue with a single clean stroke on the
10:54
graph, or you know, if there are more than two variables, then
10:58
you know a plane a hyper But unless you can put a clean linear boundary
11:03
between your two groups of perceptron can't do it. And that leads us into
11:07
what we call the AI winter. Ooo. That sounds so scary, It
11:13
sort of was. It was beginning to look for computer scientists like a very
11:18
dark and lonely time because confidence in the ability of machines to ever replicate some
11:24
of the things people can do had just evaporated, and the funding went with
11:30
it. This was I mean, you'll get different dates on this, but
11:33
from roughly, like i'd say, the nineteen seventies to the nineties, nobody
11:39
is looking very much at neural networks. Nobody is looking very much at AI
11:43
anymore. You know, obviously there are exceptions, but there's sort of a
11:48
broad agreement that the potential is limited. Do you think that's why DISCO took
11:50
cold Like why there was this like gap in you know, human intelligence growing
11:54
and developing, so they had to you know, supplant it with something like
11:58
you know, I don't know, Donna Summers the intellectual pessimism of the day.
12:05
They needed something better to do to explain how Reagan got elected too.
12:09
Yeah, yeah, and that's and then we ended up with Drunge and we're
12:13
like, okay, all right, Like it got real dark. Towards the
12:16
end of this winter, it gets better. This next part goes pretty fast.
12:20
Starting in the nineties two thousands, we start getting better faster GPUs graphical
12:26
processing units. And these aren't just good for graphics, although people buy a
12:31
lot of them for that. They they're good at anything that requires a lot
12:33
of very fast parallel computation that isn't super sophisticated, like I don't know,
12:39
for example, like multiplying a lot of things by linear weights. So now
12:43
all of a sudden, we can do certain kinds of math that are applicable
12:48
to AI really really fast. People start taking renewed interest in recurrent neural networks,
12:54
which are ones that feedback into themselves, and convolutional neural networks, which
12:58
is like you have a little squad window that slides over an image and looks
13:01
at each little part of it and tries to identify what features of the image
13:05
there are and then in twenty twelve, I think we get something called alex
13:11
net, which is the first convolutional neural network that really excels at classifying images.
13:18
Cool, we're talking like unprecedented performance in computer vision, and people start
13:24
going, hey, wait, maybe there's something to this after all. And
13:28
then in twenty eighteen we get GPT, which is I guess kind of really
13:33
why we're here talking about it. I mean, yeah, GPT is,
13:37
for those of you don't who don't know, the open Ayes generative pre trained
13:41
transformer. And what that means is generative is it's making stuff, right,
13:46
It's not just performing a classification or it's not performing a task. It is
13:52
making stuff on command, Yeah, on command, yeah, in response to
13:56
a prompt, pre trained. Because this was never originally going to be a
14:01
general purpose model. GPT was going to be sort of a basis for like,
14:05
Okay, we've trained it on a wide swath of the English language,
14:11
pulled off off the wild west of the Internet, and then you can take
14:16
this thing, this monstrosity, and kind of fine tune it a little bit
14:22
with specific texts in your area of interest. And then a transformer is there's
14:31
a lot to what a transformer is. But the important thing there is that
14:35
we're using a technology called self attention, where the machine is looking both at
14:37
the prompt and the text that it's generated, and it doesn't differentiate between the
14:43
two. It just says, given this prior distribution of text, what's likely
14:48
to come next? Right? Okay. That's been difficult to do. Recurrent
14:50
neural networks, like I said, with feedback, had been the way we
14:54
were trying to do that for a long long time. But there's this thing
15:00
that happens where because those process the text in order, you forget what comes
15:07
at the beginning of the sentence before you reach the end, even if it's
15:09
really really important, and self attention doesn't do that. Self attention looks at
15:13
the whole text all at once and identifies relationships between the words. Okay,
15:20
And we'll talk a little bit more about why that might be relevant to how
15:24
we think about language inside of psychology a little bit later. I suppose,
15:26
Yeah, I'm so sorry to do this too, but unfortunately we have to
15:31
let the artificial unintelligence take over for just a moment, and that's our ads.
15:41
Fortunately, our ad breaks for us are extremely short, so they're already
15:45
over. But we'll go ahead and jump back into this. I wanted to
15:48
comment that I think, as you said, what probably what surprised me and
15:54
so I assume surprised other people about the use of CHATGBT is how accurate it's
16:00
simulated generated language, because it spits out syntactically sensitive relevant information immediately based on
16:11
whatever problem do you give it? So much so that the first time I
16:15
used it, I was like, is this like a joke? And there's
16:17
someone on the other end of this is like a huge network group of people
16:19
who like, there's some one of the other end generating responses to what I
16:23
say, trying to make me believe that a computer is doing this, But
16:27
it's it's actually gotten faster, like so much so that I'm like, oh,
16:30
humans can't generate information that quickly. Yeah, that's the crazy thing.
16:34
There was some speculation, I mean, it was even an in joke for
16:38
a while among people working on natural language processing, which is which is unfortunately
16:42
not very much of what I do personally, But I was privy to some
16:47
of these discussions where like people were joking that it was like Amazon mechanical Turk,
16:51
like you've you've hired somebody in another country on like a piece work basis
16:57
to type back at you. Yeah, and that's why open ai has these
17:02
billions of dollars in funding. But no, it's it's ridiculous how quick it's
17:07
gotten and syntax, the ability to generate really sophisticated pros I think is where
17:14
it really shines. Of course we'll talk too, I think about where maybe
17:18
doesn't shine. But yeah, that's it's pretty remarkable. And for a long
17:21
time we thought that was sort of impossible because we just didn't have we didn't
17:26
have a compelling way to get there until we came up with this attention thing.
17:30
And in fact, there's there's a paper. It's a little bit mathematically
17:34
dense, but it's a good concept paper too if you're interested. It's it's
17:37
called a tension is All you Need and it's sort of a discussion of why
17:41
you don't need memory to generate really sophisticated, complex, accurate language on a
17:52
computer. So I mean, that's well worth reading if you want to.
17:57
And by the way, this is not just news to you, and this
18:00
is just this is not just used news to people using the technology. The
18:07
efficiency with which transformers have scaled, the gains on performance that have been observed
18:14
were news to the people developing them. Wow, GPP was already unexpectedly way
18:21
better than GPT two. Yeah, to the extent that this is where I
18:26
think the public started taking a little bit more seriously the idea that AI could
18:30
represent a in and of itself, a major player in the fate of humanity
18:34
over the next several decades, or this explosion and capability that nobody saw coming.
18:40
Well, and you're talking, we're talking right now, you know,
18:42
nearly two hundred years of research into this thing, right, like at least
18:48
at least like formal research and inquiry around this. I mean you're talking from
18:52
Ada Lovelace and like kind of like the conceptual stuff all the way till now
18:55
You've got two hundred years of having this discussion, and all of a sudden
18:57
it is like gone from we can do simple things too. It is exploded
19:03
into this really useful tool that people are still like kind of going, oh,
19:07
what are the possibilities now? And even the people that have developed it
19:11
have no idea what it can do. Oh yeah, it's for the first
19:15
time we're not thinking about how do we how do we get it to do
19:18
what we want it to do, and there's a greater research effort directed at
19:22
how do we know what it can do? And then eventually it'll go how
19:26
can we stop this? You know, I'm just kidding. I mean,
19:30
that's a real concern that people have. I mean, even I think that
19:34
maybe we underestimate the ability of really comparatively stupid models to impact human behavior pretty
19:42
radically. I mean the way that people have been persuaded to kind of,
19:48
you know, self censor and use character substitutions on TikTok because they're afraid that
19:53
the monetization algorithm will delete their videos and things like that. So we have
20:00
this interesting example where it's like, you know, the purpose of the algorithm
20:03
implemented on a platform like YouTube or TikTok is just to maximize engagement and profitability,
20:08
but there's a sort of ancillary emergent goal happening in the aggregate as instances
20:15
of these models cooperate. That is a kind of like morality agent. It's
20:22
really fascinating that we don't need very sophisticated autonomous behavior at all for it to
20:29
start really radically impacting people's lives. Coded Bias is a documentary that's out there
20:34
that's a great examination of that. It happens all the time, and so
20:38
yeah, people are very I think, you know, maybe their picture of
20:44
what that might look like is not always the most accurate, but I think
20:47
people are very rightly concerned about how this affects humanities, you know, disposition
20:52
in the world at large. I work a lot in the publishing world,
20:55
and there was some discussion around the issues of like aigen rated books and that
21:00
we're written based on authors writings and so like, there was like this use
21:06
of this tool where these books were being written by ghost writers essentially, and
21:11
there is a discussion around kind of the art around that and the art being
21:14
taken away and so so like it guess going to your point, like people
21:18
being rightly concerned about the uses of this tool and how there are really no
21:21
limitations to it or like what it can do and all that stuff. I
21:23
think it's we're starting to see that that stuff is going to kind of emerge,
21:26
I imagine over time, and then safeguards and you know, stop gaps
21:32
and stuff are going to be have to be put in place as those problems
21:34
start to arise, because we just don't know what those problems will be there
21:38
is, as you mentioned, the concern that grows about people. I think
21:41
that they's toured on the more conspiratorial side. AI is going to take over
21:48
everything, and we're going to lose their jobs, and like it'll it'll just
21:51
it'll just want to kill people, and I think there is we need to
21:53
get to that at some point. And then there's also, I think the
21:56
more practical side, which is that this is now shown to be useful enough
22:00
that actually can replace a chunk of human behavior and certain sectors of the workforce
22:06
and actually going after white collar jobs more so than blue collar jobs. And
22:10
so those people who have a little bit more power and influence be feeling particularly
22:14
threatened and wanting to do something about it. And there's something to talk about
22:18
in there. And I think also, you know, we started with this
22:22
trying to describe in some amount of definition what AI is, and if we
22:26
just we'll look at the two words separately, not to get too deep into
22:30
the philosophy of this right at the moment, but artificial I think the way
22:33
we're using it right now is to suggest that it was created by human beings,
22:38
and the intelligence part I think we're aligning with some amount of generative autonomy
22:45
in a sense in the way that this is being applied, and whether you
22:48
want to go back and parse out like you know, what intelligence actually is
22:52
or isn't, or how we've tried to define it and failed for so long,
22:56
and the idea that like, well, every time, you know,
23:00
a person only gains intelligence through interactions with education and teachers and learning. And
23:04
if we've seen instances where they don't get that and they barely develop any kind
23:10
of you know, coaching language or anything, or they can interact meaningfully with
23:14
those around them, and so is that also then artificial because humans construct one
23:19
another in the ways of intelligence. We don't need to go down that path
23:22
I think I'm getting I would really love to go down that. We can
23:26
save that for we can save that for later time, because there's a lot
23:30
there. I mean, we are very attached as a species to the notion
23:36
of intelligence as something special and intrinsic to us maybe sometimes, and I think
23:41
that's one place where the philosophy of behaviorism can be really helpful to us and
23:45
understanding what this is, you know, it's do we necessarily have to privilege
23:52
intelligence as an unobservable intrinsic quality in order to appreciate the capability of a sophisticated
23:59
system. Just you know's there's that little thing to chew on, and I've
24:03
been I've been wrestling with that now for years. Yeah, I think more
24:08
to the more immediate point. I mean, you're right, you know there
24:11
are there are longitudinal concerns. Conspiratorial is one word, for sure. Your
24:18
Eliza Yukowski's in Time magazine this past year talking about, you know, an
24:22
AI that with an insufficiently well defined goal might inevitably find itself best served by
24:29
doing something dangerous, existentially dangerous to people. And then we have the more
24:33
immediate sense in which people's livelihoods are are potentially endangered even now, because you
24:41
know, whether it's rightly placed or not. There is such speculative fervor over
24:45
what human capabilities can be replaced with a technology like this. But before we
24:52
get too excited about that, like, I feel like that paints a pretty
24:56
grim picture. I think it's also worth talking about things that these models are
25:00
maybe not so good at. Yeah. Part of this where is going I
25:03
think this segues nicely into this conversation is the idea of like general intelligence as
25:08
we like to talk about among humans versus a specific intelligence and what general intelligence
25:15
would mean for AI, what specific intelligence means for AI. And also,
25:19
like one of the things that some skeptics have raised is we don't need AI
25:25
to be generally intelligent, and it's really, as far as we know,
25:27
it seems really really difficult to get there. So no one's going to put
25:32
in the effort to try and get it there, because a specific AI will
25:36
do way better than a human will do at a specific task. And so
25:40
we don't necessarily wouldn't be compelled to create one that can do everything. We
25:44
don't expect humans to do everything. We don't really need a robot to do
25:47
everything. We need a lot of little robots who are really good at a
25:49
thing or a couple of things. The interface between the general and the specific
25:55
is sort of where my research interests lie. Oh great, I can talk
26:00
to you in the broadest terms about large language models, which are today kind
26:06
of the only generalists that exist. And there's some vindication I think for certain
26:11
sub fields of psychology that are concerned with language in that what it seems like
26:17
has happened and what's part of this unexpected explosion of capability is that in order
26:21
to emulate language, well, you have to incidentally acquire certain other useful faculties.
26:27
So Google put out a paper maybe as long as two years ago now
26:33
that was, and I had occasion to discuss this with Steve Hayes. Are
26:36
friends of the podcast familiar with Steve hays He actually was a guest on this
26:40
than the first or second year that we were doing this, so I think
26:42
so, yeah, so just a little bit, just a little bit.
26:45
Maybe I had occasion to talk with with Steve Hayses, who's a wonderful guy,
26:48
about a paper Google put out a couple of years ago discussing how the
26:53
language model they were developing at the time was unexpectedly good at math, not
26:57
great, but unexpectedly really good, which was sort of a corollary of his
27:03
work on language that we should expect that. And it's also the case that
27:08
we get decent programming out of these models, right like we get you can
27:15
ask it to generate Python for you and it will, which is a coding
27:18
language. For those that don't are familiar with that, the possover in terminology
27:23
sometimes kills me. That's what we're here for. I'll be your every man
27:27
I'll be the person that's like, I have no idea what that means.
27:30
Love it. So even even the generalist models are like outsourcing some capability to
27:37
writing code that can then be expressed to give an answer that would be harder
27:41
to get from within the model itself, because that's a thing you can do
27:45
when you can produce language. It's not just restricted to English pros. And
27:51
then there are certain other things that come with that, like even though a
27:56
large language model has never seen a sunset, you might have to develop kind
28:00
of a model of a world model that includes what a sunset is in order
28:04
to be able to accurately describe one. Right, So what's actually happening when
28:08
these things are being trained is that they're just guessing the next word that comes
28:12
in a text. And to get really good at guessing what the next word
28:17
that comes in a text is going to be, you have to have a
28:18
model of what the author is saying, what the author's perspective is, the
28:22
world that the author lives in. So it's not just language, it's not
28:26
just a model of the syntax. There's maybe buried somewhere deep in the weights
28:33
sort of a simulated picture of what things do in the world. So it
28:37
seems like for at least for the foreseeable future, our really good general models
28:41
are going to be language models. But it's also true that we have a
28:48
lot of specific models which are kind of what I do with robots and what
28:52
I've looked at doing for some like potentially some like space borne technologies and things
28:57
like that, where you get a smaller model really really good at one task,
29:02
and if you need more sophisticated behavior, you can sort of glue those
29:04
together. So those are sort of the two directions that people take within that
29:11
your right to say that the what's the word I want here? The specialists
29:17
are a lot better at what they do in general than the language model sometimes
29:21
are. Okay, I sent you guys a link to a quiz where you
29:26
try and guess what GPT is capable of from a set of sample problems,
29:30
and I think that would be really cool to share with the listeners. But
29:33
I will I will summarize my takeaway. Large language models are really really good
29:37
at giving you realistic answers. What they are trained to do is produce very
29:41
similar to sometimes the easiest way to seem correct is just to be correct and
29:48
that's a lot of what we rely on in education. Right, is like,
29:51
at a certain point, if we give you enough examples, the best
29:52
way to fit all of them is going to be to learn the underlying pattern.
29:57
It'll just take too much. It's too demanding to memorize all the answers.
30:03
Right, not necessarily always true. I mean, I think probably everybody
30:07
here has seen some like pretty radical overfitting in a student or a inside of
30:11
a clinic at one time or another where you see just some pretty phenomenally honed
30:18
in specific malaud of fiction. Sure, and so in the same way,
30:25
we don't always get generalization that way, but often in the pursuit of giving
30:30
producing realistic language, we do achieve correct language. Neural networks at large are
30:34
very good at detecting and continuing patterns that might not be obvious in the data.
30:40
Whatever the data is. The data could be you know, language,
30:41
but it could also be like I've applied some of this to weather forecasting and
30:47
gotten some really really good results, because there are patterns in weather data that
30:52
just aren't necessarily salient to a human observer and don't have the emphasis they should
30:59
have in a conventional statistical model. Yeah, they are very good at memorization.
31:03
I think the count for tutable parameters in GPT four is one hundred trillion.
31:08
Oh, that's all. It's all that order of magnitude, and that
31:12
there is a lot of information that you can encode with one hundred and one
31:17
hundred trillion variables. Yeah, that seems like it. There's a lot of
31:22
combinations there, I would imagine. I think you also just stressed out all
31:26
of our meteorologist listeners because they were like, I thought I was good at
31:30
this, and now you're finding out that this software might be aler, this
31:33
tool might be a little bit better than they are at it. It's possible.
31:37
I mean, what I've done with this is looked at the upper,
31:40
upper upper atmosphere, where most of us will never have to worry about it
31:42
if we're not in space. So for now, meteorologists, you're safe from
31:45
me. Well, unfortunately I need to interrupt with us going into the upper
31:48
atmosphere where the ads cannot reach us. But unfortunately they stay put in the
31:53
embedded structure of the episode. Okay, so it seems like we're sort of
32:04
on a roll here with like what AI is good at versus what it's not
32:07
good at. So let's keep on that sort of that trajectory here, because
32:10
I think it helps illustrate one of the things I feel like we're building toward,
32:15
and how this is related to psychology and understanding human behavior is the question
32:21
that everybody has. Is it conscious or can it be conscious? Is it
32:24
sentient? Can it be sentient? And what do those things mean? So
32:28
we have to understand right now what it already is doing and what it isn't
32:31
doing to I think arrive at some place where we can have a meaningful discussion
32:36
about that without having a definitive answer ready at our fingertips. Yeah, so,
32:40
I mean, let's get into the what it can for a second here,
32:45
if you'll indulge me. I will I alluded to the fact that maybe
32:49
in order to produce realistic, you know, touring testy language, you have
32:57
to have kind of a world model. You know, you have to be
33:00
able to describe something to a human that sounds like it was written by a
33:05
creature with ears and eyes important features here that does not always pan out.
33:10
When you start asking for spatial reasoning, you start getting really mixed results.
33:20
There's a kind of problem that I like to play with with a lot of
33:22
large language models where you guys can follow along, play along at home if
33:28
you want. But the example I like to give is like, let's say
33:31
I'm a clockmaker and I have two clocks A and B. All right,
33:37
A is going to be my reference, So A at some given moments in
33:45
time, clock A says a A, and clock B says eight oh five.
33:52
Let's say, okay, okay, sounds good. I'm going to come
33:55
back in a little bit. Block A now says eight oh four, Block
34:01
B now says eight oh six. My question is which clock's hands are turning
34:07
faster? I would say based on just the current outcome A. But I
34:14
don't think that we could know because we don't know if the other one is
34:15
rotating faster based on like did it do a full rotation? I think A
34:21
is a really good answer, especially if I tell you that you come back
34:22
on the same day, or if I tell you that you come back,
34:27
you know, five minutes later, or if I tell you that neither clock
34:30
has done a full rotation, you know, if I add a little bit
34:32
of information. But like heuristically, it seems likelier that a clock would be
34:37
you know, faster slow by a few minutes than by a whole rotation.
34:42
So it's a reasonable answer, right, I wouldn't fault you for saying B.
34:47
If you could say, well, maybe bed A did a full three
34:51
sixty, that's fine. The thing is when you start asking I played with.
34:54
Google has a new movel called Gemini that gets this consistently wrong every time.
35:00
If you tell it not to use code chat GPT is about fifty to
35:04
fifty. Oh okay, well, I mean that's about as as accurate.
35:07
It's about as wrong as you can be. Really. Yeah, when you
35:13
have two answers, my experience is that it's very very sensitive. Put question
35:15
is raised unexpectedly. So so this is like a real I call this an
35:21
overhang, right where we have a model that can do really sophisticated, incredible
35:27
stuff and it's missing the most basic kind of world modeling and reasoning. Yeah,
35:32
that just sounds like me, Like that just sounds like I can do
35:36
all this really like interesting stuff, but I can't. I don't know how
35:38
to, like I don't know, change a tire or something, you know,
35:40
like apparently I'm ai is what I've learned in this episode. It's a
35:45
little bit more sophisticated than that. I think right, because to your point,
35:49
Shane, it's a little bit like if you didn't know how to change
35:52
a tire, but you didn't know how to repair in f fifty right,
35:57
right, right, right. It's it's weird. It's it's truly weird,
36:00
because we shouldn't expect to see such poor performance on some measures of basic reasoning
36:08
when we see such great performance on things like the sat right right, and
36:13
it can now pass the bar. I think it passed medical licensing boards.
36:19
I believe that they did a study where they asked they gave it some symptoms,
36:23
and they also gave real doctors some symptoms and asked them to try and
36:28
diagnose, and that actually did better than the doctors did. Okay, but
36:30
here we have to be careful too, because that may be true for medicine
36:35
because people google a lot of medical questions. It's true. I you know,
36:42
my field is concerned with things that are called things like unscented common filters
36:46
and you know, discrete fur your transforms, which are not maybe as present
36:52
in the internet corpus. Sure. So this is another really interesting sensitivity of
36:59
the model that when you start asking for very specific information, it will straight
37:05
up tell you wildly incorrect thing. You get what people will call variable salad.
37:12
Sometimes if you get too deep in the weeds, if you get outside
37:15
of things that people are googling, like bar exam questions or medical questions,
37:22
if you get outside of things that aren't like very richly present in the training
37:25
data, you run into trouble pretty fast with what broadly would call this hallucination.
37:30
And it's a problem with the other kind of problem I was discussing too,
37:32
like not only will these models give you a wrong answer, they'll give
37:37
you wrong reasoning, or like you can ask with what's called chain of thought,
37:43
you can say, explain your work before you give me the answer,
37:47
okay, on some of these problems, and you will get sometimes radically wrong
37:52
reasoning the whole way through and then an incorrect answer, or sometimes even more
37:57
strangely, completely correct reasoning that would get you to a valid solution to a
38:00
question, and then a wrong answer, or if you don't ask it to
38:06
put its work first, sometimes you'll get a wrong answer and then correct reasoning
38:10
that goes off the rails at the very last second and starts like changing the
38:15
numbers in your original question so that it has a better statistical fit to the
38:20
answer that it already gave you. Right, it's a probability machine. And
38:22
I guess maybe my question to that would be because, like leading into that,
38:25
you mentioned that in a place where the training data is rich, and
38:29
like the source is rich, and there's a lot of like maybe consistent information
38:32
in that space, Like it makes sense that in a medical model or an
38:36
illegal model it could pass a bar or like a medical exam or something like
38:39
that. But when you get into the more general space, is that an
38:42
issue? I mean, maybe we don't know this, and I this is
38:45
kind of where my curiosity comes in. Is that an issue where there are
38:50
too many data inputs that it's just trying to get to a point. Is
38:52
it an issue of there are too many pathways, or that the data or
38:58
there's just not enough any and all of the above. There's been a paper
39:01
in pre print that I'm going to have to find so that I can send
39:04
it to you so that you can refer to this. Maybe, but there's
39:07
a paper in preprint last time I checked, demonstrating that you can recover multiple
39:12
gigabytes at a time of ver beatim text from the Internet by giving GPT chat
39:19
GPT's particular prompt. Okay, some of what it's being trained on is recoverable
39:24
in full form beginning to Yeah, so that's weird because that shouldn't happen.
39:32
We would probably expect, I mean, we would probably expect if these models
39:37
are working the way we expect them to and the way we want them to,
39:42
that in some way information is being extracted, but we shouldn't have We
39:45
shouldn't be wasting the model's size on just like memorizing rote information. And at
39:53
the same time we have exactly what you're referring to, which is like a
39:59
dearth of some subject specific information. So in some places it seems like it's
40:02
too much where it's being what's called overtrained. In some places it seems like
40:06
it's way too little, right, And it's kind of an open question as
40:09
to how we reconcile that. I think this is really fascinating because it mirrors
40:15
exactly what you would expect to happen. If you like, if you sat
40:20
a baby down with like sat materials and had it study for thirty years and
40:25
do nothing else. It feels like it would kind of look like that.
40:30
Sure, yeah, that's a great metaphor. Like it'd be incredibly good at
40:34
answering any any and all inquiries related to SAT material, and that would be
40:38
really bad at basic logic and reasoning. It's getting better. I do need
40:43
to add that caveat because people have started to realize that this overhang exists.
40:46
But yeah, and then this leads us to the last really big problem,
40:52
which is that I alluded earlier to how this there's this issue of like nonlinearity,
40:58
right, like language, which is tricky because the output is not just
41:02
like the sum of the inputs. The whole is not the sum of the
41:05
parts. Yeah, right, So there are multiple ways to get an AI
41:10
to do the same thing. A neural network can reach the same output by
41:15
infinitely many combinations of inputs. So it's hard for that reason to work backwards
41:22
through what the model is doing and attribute its decisions, attribute its mistakes that
41:27
you have to do inductively. There is some work on that, because we
41:30
can kind of peep inside the model and we can see what parts of it
41:34
are active and what, you know, what correlates with why. And there's
41:37
been some really cool work done too on like working backwards to find what combination
41:44
of inputs makes GPT most likely to say girl, for example, sure,
41:49
which it's really sad. Actually, what that combination of inputs is the things
41:53
that we're prompt gpt to say girl are about us colored by the Internet,
41:59
as you would spect. But that attribution is really labor intensive, really really
42:05
hard to do. For that reason, there is in the public sector pretty
42:09
widespread skepticism of leveraging AI to do any kind of decision making, And I
42:16
think that's that's something we need to be aware of, is that all of
42:20
this is a surprise even to the people doing it. Right. I remember
42:22
seeing a story, and please forgive my ignorance on this, but I remember
42:27
seeing a story where I forget what organization was it was, but they created
42:30
this. It was like an AI robot that like absorbed information and started to
42:35
learn to kind of communicate better, and it got really racist really quick,
42:38
based on like Internet inputs. So I guess it's one of those things where
42:44
I've seen that time and time again, where not necessarily that like robots become
42:47
Nazis, but like that there is this conversation about like how quickly these types
42:52
of like these types of neural networks or these these platforms or these like software
42:57
engines or whatever it might be. They learned so quickly and they adapt so
43:01
quickly based on the number of inputs that get put in. I don't know
43:05
if I have a question around that, but I just found that part fascinating,
43:07
And I guess maybe that's one of like kind of the problems with this
43:10
stuff and why people don't trust it with decision making because it gets so easily
43:14
influenced based on data that it receives, and we don't shame that's, like,
43:20
we don't We don't have a way to deal with that either. Yet,
43:22
By the way, the problem of like yeah, in general, like
43:27
you know, being disproportionately affected by certain inputs during training is a problem.
43:32
But the specific problem of having to deal with just the volume and the intensity
43:37
of bigotry that exists is why we now have a version of chat GBT that
43:45
is so equivocal and accommodating as to be totally noncommittal in some of its answers.
43:52
Because the only way we figured around that today is to hit it with
43:58
the fire hose who the full corpus of the Internet, and then step back
44:02
and retrain it, like, okay, now that you have a knowledge base
44:07
of language, there's a directed effort on retraining or what's called fine tuning,
44:12
to be conciliatory. Right, Sure, I don't want to. I feel
44:16
like politically correct is such a loaded term anymore to say that would insinuate maybe
44:21
that I think this is a wrong approach, or that I think it's it's
44:24
it's an over correction. And in the absence of a better way of going
44:30
about this, I prefer the version of a publicly available natural language supercomputer that
44:36
you have to try really really hard to get to be bigoted at you,
44:42
even if it's maybe not as good of a supercomputer. Right, it just
44:46
seems like where we're at right now, I prefer that personally. And maybe
44:52
that's not that's not everybody. But I had a whole idea about an episode
44:55
about that. I'm gonna leave that as the cliffhanger while these ads play.
45:00
Though I've had this discussion, there's actually been some actual research on this,
45:09
as well as a lot of talking points on places where their whole mission is
45:15
to espouse misinformation or disinformation. But that a lot of these AI things are
45:22
express a liberal bias, and what I interpret that to mean is is really
45:28
hard to force them to be racist, and that's what I think. What
45:31
they mean is like it will try and say, as you said, it's
45:35
conciliatory. It's sort of a people pleaser. It's going to try and say
45:38
as unoffensive a thing as it can possibly say. And that's what they mean
45:44
by a left leaning bias. And I'm like, I'm not really sure what
45:47
else they mean. I've tried to get GPT to roast my music taste and
45:52
it won't do it. I'll do that for you if you want, I
45:58
mean, if you want me to, I'm more than happy to. I'm
46:00
very opinionated about music. We'll talk. I know Abraham is too, so
46:04
we're in and I overlap a lot of music. I don't know that there's
46:08
a factual bias. I couldn't speak to that, but I will say that,
46:14
yeah, it's the thing is that it's and this is this is sort
46:17
of off the beaten path, right like, irrespective of what decisions you make
46:22
with what a robot is allowed to say, it will be interpreted as a
46:27
signifier, right like, that's it's These things are not insulated from the morays
46:34
of the people building them, nor should they be. It is an inherently
46:39
political technology, because it is a technology of language, and I think language
46:44
in many ways is inherently political. Politic I think is just discourse on a
46:50
large scale. That's beautiful as we move into, like, you know,
46:54
thinking about what what this does, right, Like, for the first time,
46:59
we have thinking things that are not people that we interact with on a
47:02
regular basis, and we have to give some consideration to what that does to
47:07
the composition and cohesion of the societies that we live in. Right, Yeah,
47:13
I think that's going to be a more direct consequence. And when I
47:17
say that, like I alluded earlier on to my thinking that maybe, you
47:22
know, the AI apocalypse that people are afraid of is sort of misconceived.
47:27
I think what I mean by that is, I think there's a much greater
47:30
and more immediate threat in the potential for AI to participate in shaping our preferences
47:37
and in shaping the language we use with one another and in producing this information
47:43
for us, than there is in like, you know, somebody plugs one
47:46
into a bomb and blows us up. Right, The fear of the overtly
47:52
violent and aggressive apocalypse is not necessarily, as I guess, as easy to
47:58
get to as like let's say something more pervasive, something more subtle, something
48:02
more psychological, or something like, you know, more insidious. I guess,
48:07
you know, like everybody's expecting terminator and they're really looking more at like
48:10
something almost like a virus that could possibly Like that's a more realistic concern,
48:15
I guess. And I get I get KG about alluding to like notions of
48:20
social cohesion because even that has become social contagion, because even that has become
48:25
sort of loaded. But yeah, I don't. I don't have a better
48:30
way to put that, and especially because what we have given AI some reins
48:35
over is in serving us information and serving us entertainment. So like we talked
48:43
about earlier, like TikTok doesn't have well, maybe they do. I couldn't.
48:47
I couldn't say this with certainty. But my, my, my overwhelming
48:52
impression is that TikTok does not or or does not need to have an ulterior
48:55
motive in shape and people preferences and in radicalizing people. All that TikTok needs
49:02
to do to achieve the end of alienating and radicalizing and compartmentalizing people is to
49:09
prioritize engagement like Facebook. Yeah, exactly, and so Abraham. Earlier on,
49:16
you said that when we talk about artificial intelligence, what we take artificial
49:21
to mean is created by humans. I don't think that's a necessary qualifier.
49:27
I think artificial. I think it's efficient to say that artificial means something other
49:31
than biological because, in the same way that there are now concerns about stochastic
49:39
terrorism, right people motivated to violence by the deluge of content on the Internet.
49:46
I think stochastic intelligence. I think systems with sophisticated behavior that achieve coordination
49:53
between people that would otherwise be really hard to achieve in the absence of intelligence
49:59
that seems to be in the form of Facebook, for example, something that
50:02
can kind of maybe just happen. That's a scarier thought to me, do
50:07
it? I mean, do I still want to keep an eye on what
50:10
kind of automation people are using artificial intelligence for let that become like an instrument
50:17
of violence. Yeah. Absolutely, I think that's a very real risk.
50:21
But it's not what keeps me up at night. What keeps me up at
50:23
night is that we have machines that give us and tell us and show us
50:29
exactly what we ask them to with a very very appealing veneer of realism.
50:36
I don't get to do this very often. We superimpose in our shows.
50:40
If we choose to employ a particularly high level vocabulary word or one that we
50:45
think people might not know, we'll throw a little music thing that a transition
50:51
to section on vocabulary, and you use the beautiful word, which is stochastic.
50:54
So could you define stochastic for our audience please? I love the jingle.
51:01
Stochastic is maybe my favorite word. A stochastic process is one that is
51:10
random in its progression, but random in many different parts. So I could
51:16
talk to you about the stochastic movement of the particles in a gas, right,
51:21
or I could talk to you about the stochastic behavior of people using mass
51:28
transit in New York on any given day. And so when we talk about
51:30
stochastic things inside of engineering, we sort of discard the individual data points and
51:38
we speak about the behavior of the ensemble and the net effects of that ensemble.
51:45
Right. So when we allude to stochastic terrorism, what we mean is
51:53
there's no single element that we can readily attribute accountability to for the action of
52:00
a lone wolf terrorist who's been radicalized by the Internet or by a hate group.
52:07
It's been the aggregate of the activity of many different people with many disparate
52:14
interests who are not necessarily coordinating with one another. In the same way,
52:20
AI is stochastic in its development in that it develops emergent behavior based on lots
52:28
of tiny, little individual data points that are not necessarily correlated with one another,
52:32
but that have a pattern in the ensemble. And potentially also we get
52:37
stochastic behavior from large scale dynamical systems like human interactions and technological interactions between automated
52:46
systems. And I think there is a sense in which you could describe the
52:52
fundaments of psychology, even you could describe neurology as an essentially stochastic process,
52:57
I mean, the firing of a neuron. Although it has in the long
53:00
term deterministic behavior, is somewhat random. It is orderly in that it is
53:07
a function of you know, your neurochemistry, but it is also unpredictable in
53:15
certain ways. Lovely a great word. Yeah, it was a great one.
53:17
That was a really uh that was a very well explained teaching moment for
53:22
that word. And I appreciate that. I adore that word. I think
53:25
is the single most beautiful thing to have come out of my education. Wow,
53:31
understanding stochasticity is such a powerful tool for at least for me, it
53:37
was a very powerful tool for breaking out of the lens of everything that befalls
53:43
you being a matter of intention. Okay, sure, I like that.
53:46
Yeah, it's a path path forward in the interest of time. I think,
53:52
no, no, this is really fun and I'm really enjoying this.
53:53
But I would like to get to a discussion on U on the big the
53:58
big worry, and I think one very so related to generally psychology and maybe
54:02
more specifically human behavior science, as whether AI can become sentient or conscious.
54:09
And we've done an episode of consciousness, and I believe we even replayed that
54:13
as a nostalg We do what we call nostalgic episodes for episodes that are five
54:17
or more years older, and we replay them in their entirety and then add
54:22
a little new information to it, and I think we did that recently with
54:24
consciousness. And we've also we haven't replayed these yet, but we did a
54:28
three part series on intelligence years ago, and so I think we've we've done
54:31
a little bit of legwork there. But let's talk about whether or not these
54:36
machines are conscious or could be conscious and or sentient, and how people fear
54:40
that in the pre role you mentioned to me the Google scientists who recently espout
54:47
his opinion that Google's Bard model is sentient on the basis of a conversation that
54:52
he had, right, I will submit that I think the claims of a
54:57
language model are insufficient as evidence for anything in and of themselves, because,
55:05
like I said to you at that time, you could very easily prompt Bard
55:08
to give you an equally compelling argument that it is a squirrel. Right,
55:13
So like I hope we can agree that to begin with, that's not that's
55:17
not going to be satisfactory to us. I would agree with that you would
55:21
be maybe maybe kind of poor behaviorists if it were. You know, I'm
55:29
I don't think I have a sufficiently clear picture of consciousness, even with the
55:32
benefit of having talked to you about this before, Abraham, I don't think
55:37
I have a sufficiently clear picture of what people mean when they say that to
55:42
answer the question of whether a thing is conscious, And if I wanted to
55:45
be really radical about it, I mean I could take umbrage with the notion
55:51
of consciousness as something that distinguishes humans from non humans at all, Like,
55:57
is this equality that actually exists or is this a name that we give to
56:00
a particular set of emergent behaviors. I will say that certain thinks things that
56:06
we almost unanimously consider relevant to consciousness, such as a continuity of experience,
56:12
are not features of an AI model. In general, chat GPT does not
56:17
and cannot perceive anything that is outside of the current text that you're sharing with
56:25
it, and it does not change in any way right like much much like
56:30
human neurology, The way that these models learn is that they change structurally,
56:37
except that that change is bounded there is in generally period of training, and
56:42
then the models is what we call frozen. So it is as though I
56:46
take a picture of all of your synapses, the potentiation of every neuron in
56:51
your body, and I say, none of that is allowed to change anymore.
56:54
Does that mean I don't get older in a sense? Yeah, I
56:59
mean it would mean that you have no knowledge of events outside of what you
57:07
learned prior to that point and what is immediately available to you. I think
57:13
continuity of experience tends to be pretty important. We are not all talking to
57:16
the same chat GPT when we do. We are talking to many different instances
57:22
of chat GPT as it was on a particular date. I do not,
57:27
however, rule out the possibility that there will one day be I think there
57:34
it is very philosophically and technically feasible for there to exist a neural network that
57:40
exhibits sufficiently many behaviors that are subjectively relevant to a definition of consciousness to satisfy
57:46
anyone, And at that time we'll have to have the discussion about there will
57:52
no longer be any putting it off as I am, as I often do,
57:58
will have to have the discussion about whether whether there is such a thing
58:00
as consciousness and what it is, and whether it is meaningfully distinct between how
58:05
we exhibit it and how machines exhibit it. But for the time being,
58:08
that's why I'm really attached to this pet idea I have that psychologists' behaviorists are
58:20
the people most needed in the examination of what AI can do and what it
58:23
is. For my part and for many of my colleagues, what we look
58:30
at as implementation, and what we look at is provable behavior. We look
58:32
at the specification of the balance of a model, and we don't do very
58:39
much with studying empirically after the paper is done, there's an opportunity there,
58:46
There is an opportunity to study what kinds of emergent capabilities we can get out
58:52
of a machine, and maybe even that elucidates the boundaries around consciousness. Maybe
58:55
within that we discover something that seems hard to replicate, impossible to replicate,
59:00
and that we can point to with some justification and say is a definition of
59:06
cynthias. I think you said very beautifully what many what have heard people say
59:10
that I think identify the problem with consciousness and and how and to the extent
59:17
to which it may be applied to something like AI, which is what do
59:22
people mean when they say conscious regardless of what they're talking about? Right,
59:25
the thing that we've talked about in those episodes is nobody's agreed on that yet.
59:29
And I'm sort of like it's sort of like free will, and like,
59:31
well, depends on how you define it. You want to define it
59:34
in a way that encompasses all these things, then sure we've got free will.
59:36
If you wanted to find it in a way that doesn't encompass those things,
59:38
then no, I think we all agree we don't have free will,
59:40
and consciousness is the same thing. And I think you also said something really
59:45
important is is this something that is unique and importance to humans? Is this
59:50
a feature of life? Is it a feature of a brain? Is this
59:53
a like? Is it actually a thing? What do people mean when they
59:58
say it? And what it applied to? Something like robots? But then
1:00:01
the important thing there is could it emulate behavior similar enough to what we do
1:00:07
in such a way that we would then say that looks like consciousness? And
1:00:12
I think to your as you said, like absolutely it could. I think
1:00:15
that even with what exists, if you give it a sufficiently correct well,
1:00:22
it would need the opportunity to I think, generate without a prompt. But
1:00:25
if it could do that, I believe you could. You could program it
1:00:29
carefully enough that with a very specific instance, you could get something that looks
1:00:32
indistinguishable from human behavior for a moment, for fleeting moment, and it probably
1:00:37
it probably will get more sophisticated to the point where the distinction will become increasingly
1:00:43
vague. But I also don't know that that matters a lot for the AI
1:00:49
apocalypse or concern over much. I don't think that computers would care about humans
1:00:57
if they became so self aware that they were concerned about their own existence.
1:01:00
We don't really pose much of a threat to them, and we aren't hogging
1:01:04
the resources that they need, well until the butlerry and jihad, and then
1:01:07
we have to worry about all of that and then have to guild navigators to
1:01:12
get us through fold space to Aracus for the spice. I didn't expect the
1:01:17
dune that I woke up today. I am full of multitudes, my friend,
1:01:22
full of multitudes. It's a question of how well specified our relationship to
1:01:30
artificial intelligences, and how well specified the goals of a given model are the
1:01:35
goalposts, I will say, move all the time. The touring test was
1:01:38
proposed going on one hundred years ago now, and we have now reached the
1:01:45
point where machines can generate language that is, in context indistinguishable from human speech.
1:01:51
I mentioned earlier that sometimes the simplest, easiest way to seem correct is
1:01:55
to be correct. The simplest way to seem like you understand something is to
1:01:59
understand something. And I don't know what it means when we get to a
1:02:02
point where a machine does, even for a fleeting moment, produce something indistinguishable
1:02:07
from human behavior, Because what if the simplest way to have the overt similarity
1:02:13
is to have a similarity in certain covert properties of that system. For that
1:02:16
matter, that's the only way. That's the only reason I have to believe
1:02:19
that we all share in something called consciousness, is that you behave similarly enough
1:02:22
to me that I can recognize in your actions motivations for similar actions that I
1:02:28
would take. I can understand why you would do what you do and why
1:02:32
you would say what you say. It seems similar to me, and I
1:02:37
know that I have a continuity of experience, and I know that I have
1:02:40
a perspective and an awareness of myself and what distinguishes it from my surroundings.
1:02:45
That recognition is the only evidence that I have that we are in any way
1:02:50
similar. And I don't have a compelling reason to deny that recognition to something
1:02:54
that is not a human being. I don't have any compelling reason to deny
1:02:59
it to another great ape, or to an octopus, or to a computer.
1:03:02
I like it. Well, this has been so much fun, and
1:03:06
I feel like we should We should do this again sometime and continue having this
1:03:09
discussion about sort of how a consciousness and sendience applies to computers. I think
1:03:14
we didn't even really get to some of the more of the specific pitfalls and
1:03:16
benefits that the AI may have. I splattered a whole long outline all over
1:03:22
the page, the most of it. You know what I will, I
1:03:28
will gladly return anytime you'll have me. I'll gladly come and stand on my
1:03:31
soapbox. And I would love to listen to more of your thoughts. Well,
1:03:36
great, this has been so much fun. Yeah, I do like
1:03:38
to I've pontificated about in the past about the relative standing of part of what
1:03:46
creates the human experience that I think is worthy of further discussion and in line
1:03:51
with you know, the things that you've talked about that Shane's brought up as
1:03:53
well, and how we even think about what it would mean for something to
1:03:58
be conscious when it doesn't even have a central point in space, which is
1:04:02
something that is uniquely situated to all other biological organisms and even non biological organisms
1:04:08
as they are situated at a point in space, and AI is not really
1:04:13
bound by that in a similar way it is in the sense that is on
1:04:16
the same planet, but that could change and not that long so that I
1:04:21
think would be something worth talking about. Anyway, as we wrap this up,
1:04:25
we like to give some recommendations. They are often unrelated to our topic,
1:04:29
but it can certainly be related to our topic. And so what I'll
1:04:32
do is I'll do a quick read through our credits here and then if you
1:04:36
have something you'd like to recommend, if you'd like to take a second to
1:04:40
think about it, I'll be able to just go through these the necessary information
1:04:44
that everybody must know, which includes the fact that if you would like to
1:04:47
support us, as I mentioned at the top of this episode, you can
1:04:49
pick up some merch you can join us on Patreon. If you do join
1:04:53
us on Patreon, you get some extra benefits like ad free episodes, early
1:04:56
access to episodes, behind the scenes content, etc. And we just you
1:04:59
also get our adoration and love of you for so long as long as you
1:05:02
keep being a generous, kind person who joins us. And honestly, you
1:05:05
know, we appreciate all the people who have ever joined us. There The
1:05:09
current list of people includes Mike, m Megan, Leyla, Mike t,
1:05:13
Justin, Kim, Joshua, Brad, Stephanie, Olivia, Brian and Ashley.
1:05:16
Thank you all so much for your continued support. Thank you everybody for
1:05:19
listening. Thank you, of course for our guest Joseph Daily for joining us
1:05:23
on this episode today. And then if you would like to reach out to
1:05:26
us and talk to us about your thoughts about AI, we certainly would enjoy
1:05:29
hearing from you. You can email us directly at info at wwdwwdpodcast dot com.
1:05:34
You can also reach out to us on the social media platforms as I
1:05:38
said, like in subscribe, tell a friends, do all the fun stuff.
1:05:40
Is there anything that I'm missing, Shane, or that you'd like to
1:05:43
add? Oh? Well, hold on, Before I do that, I
1:05:45
do need to thank some more people because I've got a team of people who
1:05:47
helped makes this that makes this podcast happen. One of them is Shane so
1:05:51
writing in fact checking from Shane myself, Thanks for recording with me today Shane,
1:05:56
Music and editing and production from Justin. And then our social media coordinator
1:06:00
is Emma Wilson. Now, Shane, is there anything that I missed or
1:06:02
that you'd like to add before we get to our recommendations. I think I'd
1:06:05
just like to put a pin in talking about Asmov's laws of robotics. It's
1:06:10
a short time Yeah, we missed that. I think that's for a part
1:06:13
two. I want to make sure that we hit on that part. Yeah,
1:06:15
if a robot cuts down a tree, that seems harmless enough in and
1:06:19
of itself, but if it falls across a freeway, does it know that
1:06:21
that's a problem. Let's talk about our recommendations. Then recommendations. I'll go
1:06:35
first because mine is kind of a I usually recommend a movie or a TV
1:06:39
show, or a book, or more often than not, a game.
1:06:42
But today I'm going to recommend that people go on hike, find a nice
1:06:45
park or a nice mountain or to climb or something, and just to go
1:06:48
out in nature. Experience the cold, wonderful world around you if you're in
1:06:51
the Northern Hemisphere, or the beautiful balmy summer that has added to zenith if
1:06:56
you're in the Southern hemisphere. But you know, just get on the nature.
1:07:00
That's my recommendation. There's definitely a difference between asking somebody to go on
1:07:02
a hike versus telling somebody to take a hike, which I feel like is
1:07:06
a really unique like language thing to be like, go on a hike versus
1:07:11
take a hike, Like it's it's kind of mean to tell somebody to go
1:07:14
do extra work if they don't want to. Let's appropriate that in order to
1:07:15
make it make it mean nice things. Again. Well, I always like,
1:07:19
I always like those insults that are like old insults, like get lost,
1:07:23
which is like just like inherently mean, and just tell somebody to go
1:07:26
somewhere where they don't know where they are and and nobody else knows where they
1:07:30
are. Like that in itself is just kind of mean, cruel. Anyway,
1:07:32
we're gonna take it back. So my recommendation this week is scheduling time
1:07:36
to write. So I a few years ago, I'd read a book by
1:07:40
Questlove. It was called a Creative Quest, and it talked about like different
1:07:43
ways to stay creative. And one of the things he just said, wash
1:07:46
do it even if it doesn't like make sense or it's you're not motivated,
1:07:50
just so you can kind of continue to grease the wheels a little bit.
1:07:55
This year for twenty twenty four, I'm gonna take some more time to schedule
1:07:58
time to write, because right now it looks like I've got one, two,
1:08:01
three, four or five six six kind of book projects in the works
1:08:05
and about seven articles that I'm in the process of writing. So based on
1:08:11
necessity, I've got a schedule time to write, but I really enjoy writing,
1:08:15
and I think if you like that type of thing, schedule some time
1:08:17
to do it every day and it'll be nice and easy. There's a book
1:08:21
by Paul Silvia called How to Write a Lot, and that is essentially his
1:08:25
thesis as well. Yeah, a very funny book. Also, Joey,
1:08:29
do you have a recommendation you'd like to share with folks? I guess in
1:08:31
the spirit of our you know, enjoy life, self improvement New Year's the
1:08:36
recommendations. I would like to encourage everybody to take a second and listen to
1:08:42
your body and take a good stretch, because I have done far too little
1:08:45
of that as I've been glued to my chair trying to write a thesis this
1:08:51
fast year, and it's a really wonderful thing. And so yeah, just
1:08:56
take a second, be kind to yourself. Whatever that means to you,
1:09:00
I promise it's worth it. Well, I love that. All right,
1:09:03
Well, this has been a really fun conversation. As we said, we
1:09:05
must return at some point. We have so much more to talk about,
1:09:09
but this I think that I'll do it for today. Thank you. Again
1:09:11
for joining us. Anything you'd like to add before WED I hear Shane not
1:09:15
on my end? Cool? Well, thanks again, Joey anytime. All
1:09:16
right, This is Apraham, this is Shane. We're out, Seya.
1:09:20
You've been listening to Why We Do What We Do. You can learn more
1:09:24
about this and other episodes by going to WWDWWD podcast dot com. Thanks for
1:09:30
listening, and we hope you have an awesome day