August 19, 2024 • 42 mins
You know that moment in the horror movie where the monster is coming closer, but the movie star doesn't see it? Why does that drive you crazy, and what does that teach us about brains? What is theory of mind, and why is it so important for everyone from poker players to conmen to stage magicians to novelists? Join us this week to dive into a fundamental skill of human brains -- and the question of whether current AI has any ability to simulate other people's minds.
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Speaker 1 (00:05):
You know that moment in the horror movie where the
monster is coming closer but the person on screen doesn't
see it. Why does that drive you crazy? And what
does that teach us about brains? What is theory of
mind and why is it so important for everyone from
poker players to con men, to stage magicians to novelists.
(00:27):
We're going to talk about a very fundamental skill of
human brains today, and as impressive as AI is currently,
we're going to ask the question of whether computers can
replicate this right now or whether it is beyond their
skill set. Welcome to Inner Cosmos with me David Eagleman.
(00:48):
I'm a neuroscientist and an author at Stanford, and in
these episodes we sail deeply into our three pound universe
to understand why and how our lives look the way
they do. Today's episode is about what it takes to
understand other people, how your brain does it, and whether
computers could do it. So imagine this. You're walking down
(01:11):
the street and you see someone frantically searching their pockets
and looking around with furrowed brows in a tight frown.
So without them saying a word, you can infer that
they might have lost something important. Maybe it's his keys.
Your brain can easily make a good guess about another
(01:33):
person's mental state just from looking at their actions. We
are inferring something about what is going on in that
person's head. But it's more than just pattern matching. It's
not simply that your brain has seen lots of people
patting their pockets and you talked with them afterwards, and
you figured out why they were doing that, and you
(01:54):
detected a pattern, and you memorized, ah, okay, that pattern
equals that problem. Instead, you have the ability to imagine
yourself in their situation. You can mentally slip into their
shoes and ask, what would I be thinking if I
were patting my pockets and frantically searching around me? And
(02:16):
maybe you see something else. You see a kid there
around the corner, and the kid is peeking around the
corner at the man patting his pockets, and the child
is giggling. Now, why is the kid giggling while the
guy is so obviously worried, Well, it probably strikes you
that he's hiding something from the guy. You see that
(02:37):
the kid is not running away. Instead, he's standing in
such a way that he'll be spotted. Now it's pretty
obvious what's happening here. You can step into the man's
head to feel the worry, and you can step into
the kid's head to recognize that he feels like he's
playing a game, even if it doesn't strike you as
so funny. Then you catch the guy meet with the
(03:01):
kid for just a fraction of a second, which sends
the kid into fits of laughter, and you realize the
man is just playing along. Now, how did you decide
what is going on in the heads of these two Again,
it's not as though you memorized an algorithm here. Okay,
if there's eye contact, then there's one interpretation. If there's
(03:21):
no eye contact, then a totally different interpretation. To appreciate
how complex this mind reading is that you just did,
just imagine that you're a space alien watching this scene
from your spaceship. You would be totally confused. You would
have no idea what's going on in this weird scene
(03:42):
because you don't know what it is to be a human.
Here's another analogy to appreciate this. Think about the way
that you, as a human might watch fish. You really
don't understand what the heck they're doing. One fish suddenly
starts swimming faster, and another starts swimming in circles, and
one starts flapping its gills faster, and one moves up
(04:04):
towards the surface. It's all just weird fish behavior to you.
You don't know how to read any of it. It's
just fish stuff, and you're not able to immediately construct
a story about the meaning of any of this. And
that's what it's like to be this space alien watching
this guy checking his pockets and the child giggling. Now,
(04:27):
what allows us, as opposed to the space alien, to
be so good at reading our fellow humans. This is
what psychologists and neuroscientists call theory of mind, and that's
what we're talking about today. Theory of mind is the
ability to understand that other people have their own thoughts
(04:49):
and feelings and beliefs that are different from yours. It's
the ability to recognize that others have their own perspectives.
It's the ability to attribut mute mental states to other people,
like what their intentions are, or their desires, or their emotions,
or what they know or don't know. And theory of
(05:11):
mind is a key cognitive skill that allows us to
interact with other people in a very rich and nuanced way.
Just think about how pervasive this skill is in everything
we do. So take sarcasm. When your friend makes a
sarcastic comment, you can recognize that her words don't match
(05:32):
her true intention. So, for example, if she says, oh, awesome,
more traffic, I love traffic, you infer that she's not
actually pleased. This requires understanding her mental state that she
is irritated not happy. Now, if you were Siri or Alexa,
(05:52):
you wouldn't be able to recognize anything but the words.
You wouldn't understand anything about the mind behind the words.
So we're going to talk about how brains do it
and whether or not computers can do it. But before
we go there, we're going to take a few minutes
to really appreciate how the skill is everywhere in what
(06:12):
we do. For example, just think about different professions. So
detectives use theory of mind all the time. Did mister
Jones know that the food had gone bad when he
sold it? Did mister Smith know that his boss was
involved with organized crime or was he acting with no knowledge?
(06:33):
Or more generally, if they want to know if someone
is lying, it usually helps to step into their shoes
and think about what that person knows or doesn't know.
Magicians use theory of mind. They know that if they
move their hand in an arc, your attention is going
to follow that, and therefore they know what you won't
see them do. They know that even though they know
(06:55):
something happened, like the card dropped into their sleeve, they
know that you don't know that. They always keep your
point of view, your beliefs, at the forefront of their mind.
Con Men do this. They listen to your words and
they read your body language to gather what you know
and don't know, and therefore what buttons they should push next.
(07:15):
Psychiatrists and psychologists always use theory of mind to understand
what is being expressed from the patient's point of view,
In other words, what the person believes, whether or not
it's what the therapist believes. I'll give you another example.
My friend Maddie is a professional poker player, and he
describes poker playing like this. He says, when you're learning
(07:37):
to play poker, you think about the cards you have
in your hand. As you get better, you think about
your hand and also what the other person is thinking.
And as you get even better, you think about what
the other person is thinking your thinking, and when you
get to the professional levels, you're thinking about what he thinks,
you think he thinks, and people who are real pros
(08:00):
can think five or six levels deep on this. All
of this is theory of mind, and theory of mind
is key when you're teaching something. For example, parents know
that their children can't understand certain things. For example, the
child needs to get that smallpox shot, even though to
the child that's nothing but scary and he simply doesn't
(08:22):
have the capacity to think about the future benefits that
will accrue. Or the school teacher can only hope to
educate her students if she knows what they already know
or don't know. She needs to phrase things in such
a way that someone who doesn't already know what she
knows can absorb it, and that just requires theory of mind.
(08:44):
If she couldn't simulate what it's like to be in
their heads, she'd have no meaningful shot at getting them
past the first quiz. And this issue of considering what
someone knows or doesn't know is also critical in any negotiation.
You try to under understand the other person's desires and
goals and where they might potentially compromise during a salary negotiation,
(09:08):
you consider what your employer is thinking about the needs
in future of the company and therefore what they might
be willing to offer. And this is also how you
manage conflicts. In any disagreement, if you're smart, you try
to understand the other person's perspective to resolve the issue.
If your partner is upset with you, you try to
(09:29):
figure out what you did or said that set things off,
and why that offended the other person and how it
landed for them. And that's the single way that you're
going to hit the problem effectively. So this ability to
slip into someone else's shoes has almost everything to do
with our social intelligence. You use this very human skill
(09:52):
all the time. And before we get to the next
act of this podcast, where I ask if computers can
do this or not, I just want to finish fla
this out so we can really see how pervasive this is.
So as an example, you rev up your theory of
mind engine whenever you send an email. If you know
someone has a well developed model of you, like your
(10:14):
parents or your spouse, then you can use abbreviations and
shortcuts to get your message across. But if you're writing
to someone who's never met you before. Let's say you're
applying to a new job. You run a very different game,
so you're not just an email writing algorithm that produces output,
but instead your output is modified according to who you
(10:35):
expect is doing the reading on the other end, and
specifically what their mind is like. And I also want
to mention that theory of mind is critical for literature
to work because it's often the case that you can
see the limitations of the character's point of view. So,
for example, if you remember the beginning of the movie Jaws,
(10:56):
the woman is swimming around in the ocean water and
she's very relaxed than happy because we see the shark,
but she doesn't. If we didn't have theory of mind,
we would simply say, oh, there's a shark there. But
we're able to understand that she cannot see the shark,
and that's a big part of why we are fearful,
because she isn't fearful, and we want her to be.
(11:19):
This stepping into other people's heads drives essentially all horror
movies because we often know something that the main character
does not, and it also drives romantic comedies. For example,
we see the guy doing something very nice like helping
an elderly woman cross the street, and he doesn't know
that he is being watched by the female love interest,
(11:41):
and therefore we the audience interpret what kind of guy
he must be to behave that way when as far
as he knows, he's totally alone. We would have a
totally different interpretation. If he sees his romantic counterparts there
and then he does the charitable act, we'd simulate that
his intentions are different there. Now, why are human brains
(12:03):
so talented at making theories about other people's minds. Well,
you've heard me say many times that the job of
intelligent brains is to predict the future. If you're the magician,
you'd better be sure that you are predicting correctly where
their spotlight of attention is about to be. If you're
the poker player or the con man, you're trying to
(12:24):
predict what someone is going to do next, and this
is the optimal way to do this is to step
into their mental world and understand what it is like
to be them. What they know and they don't know.
You leverage theory of mind to anticipate their next action,
and presumably this reaches back to the recent millions of
(12:46):
years of our evolution. So if you're an early homo
sapien and moving along the trail and you see another
homo sapien coming down the trail towards you, it's absolutely
critical for you to figure out is he going to
attack me? Is he scared of me? Is he trying
to trick me? Is he just trying to get past me.
You're trying to figure out his mind so you can
(13:09):
figure out his next actions. So what I've told you
so far is that theory of mind is this critical
foundation for all of our meaningful social interactions because those
require you to be able to simulate other people's intentions
and emotions and beliefs. Your brain doesn't assume that it's
(13:30):
a knowledge communism out there where everyone knows exactly what
you know. Instead, we're able to pull off a higher
level of interaction because we understand that the world is
different inside different heads. And this, by the way, is
really sophisticated. It requires knowing who I am and what
I see and believe, and also holding in my head
(13:51):
what it is to be someone else and see and
believe something different. This is a very sophisticated computation that
the brain pulls off, but because we're so good at it,
it's typically invisible to us. But theory of mind doesn't
come for free. It's something that develops with time. As
you get more and more experience in the world and
(14:12):
you stop believing that you are the centerpiece and that
everyone else is just a cast member. You come to
understand that that person believes something different than you do,
and this other person feels a certain way even though
you don't, and that this person over here thinks something
to be true even though you know it's not. So
(14:48):
how do we know that this is a skill that
develops through time Because very little kids are terrible at
theory of mind, but they get better as they mature
into the world, and typically by the ages of three
to five, they're getting that they're not the only point
of view that's possible, but that each person in the
(15:08):
scene has his or her own point of view. Now,
how do you test whether someone is capable of theory
of mind? Well, what you do is you present a
little scenario like this. Sally comes into the room and
puts her baseball under the bed, and then she leaves.
While she's gone, Anne comes in the room, she sees
(15:29):
the ball under the bed, She picks it up, and
she puts it in the closet. Then she leaves. Now
Sally comes back in the room, she wants her baseball.
Where does she look for it? Now? You and I
know that Sally will look for it under the bed
where she put it last, even though we simultaneously know
the actual location of the baseball in the closet. And
(15:52):
this is because we are running an emulation of what
it is like to be inside Sally's head with her
limited knowledg. Now, little children will fail the sally An
test because they know that the baseball is in the closet,
so they assume that Sally should know that too. But
as cognition develops, they come to realize that different heads
(16:15):
have different beliefs. And a really important clue to the
development of this is that not everyone develops theory of
mind in the same way at the same rate. For example,
people who are on the autism spectrum typically show delays
in developing theory of mind, which cannot surprisingly impact their
(16:36):
social interactions. For instance, this is why sarcasm doesn't work
so well with a person who has autism. When you say, oh,
great more traffic. I love traffic. They're not likely to
catch the meaning beneath the words that you're not actually
pleased because they don't have a sensitive model of your
actual mental state. If you can't put yourself in the
(16:57):
shoes of the other person, your understands is limited to
just pattern recognition, which is not enough for the very
subtle and sophisticated kinds of communication that humans engage in
every day. So this tells us that theory of mind
doesn't come for free in humans. There are brain networks
that have to develop and learn for this to work,
(17:19):
so when you look at normal development or delay development.
This allows us to understand how different brain regions contribute
to theory of mind. For example, there's one area called
the temporopridal junction, and this is interesting because it pops
its head up in tasks that require understanding perspectives, like
(17:39):
distinguishing between what you know and what someone else knows.
So imagine you're teaching a friend how to play chess.
You need to not only understand the rules of the game,
but also know what your friend knows or doesn't know
about the game to teach effectively, and the temporo pridal
junction is involved in that not just that area. It's
(18:01):
a lot of other areas involved in theory of mind.
So the medial prefrontal cortex plays a big role in
making social judgments. It becomes active when you think about
the mental states of others. For example, if you're trying
to decide if someone is lying or being truthful, your
medial prefrontal cortex is engaged. And there are other areas,
(18:23):
like part of your superior temporal sulcus is involved in
processing social information like interpreting other people's eye gaze or
their body language, like the man looking for his keys
and the child giggling. We're able to infer a lot
because of the activity of this area. So we see
lots of areas in brain imaging experiments. And I want
(18:45):
to mention this to illustrate that theory of mind is
a brain wide issue. It's not a single area. And
by the way, this is true of so many things
in neuroscience. Imagine that I spread out a map of
your city and I ask you, hey, can you put
a pin in the spot that represents the economy of
the city. You tell me that that is a misplaced request.
(19:07):
There is no single spot for the economy. The economy
emerges from all the interactions between all the pieces and
parts of the city, and it's the same with almost
everything in neuroscience, and especially something like the skill of
slipping into someone else's point of view. There's not one
spot to drop a pin into. Instead, it is an
(19:27):
emergent property that develops from the interaction of lots of networks.
So what we've seen so far is that theory of
mind is this ability to infer what someone else knows,
and we've seen that this is right at the center
of social interactions. It's something that most humans develop naturally,
but that doesn't mean it's simple. And the question we're
(19:49):
going to ask today is does AI have theory of mind?
Can it put itself into someone else's shoes to understand
their limited knowledge. One of my colleagues at Stanford recently
wrote a paper suggesting yes, AI can do this. But fascinatingly,
it's not as easy to answer this question as you
(20:11):
might think. And this is for some reasons that we're
going to dive into. But before we get there, I
just want to zoom this out to a slightly larger question.
Could a computer develop theory of mind. Hypothetically, could an
AI system at some point in the future say, look,
I know XYZ to be true, but if I look
at that other person over there, I understand that they
(20:33):
have a limited viewpoint and that they don't know X
and Y, and that person over there misbelieves something about Z.
Well almost certainly, yes, Why it's because we're made up
of physical stuff and we're running algorithms that took hundreds
of millions of years to refine. But nonetheless it's physical stuff.
(20:54):
So if we can do something, presumably a machine could
do it also, whether or not it's currently clear how
that's done. That's the central premise of computational neuroscience, and
to my mind, one of the most remarkable effects of
the AI explosion over the last few years is understanding
(21:15):
that things that would have seemed impossible to do with
a machine, things that almost everyone would have sworn couldn't
be done. It now seems like background furniture as we
wait for the next thing. Now, the complexity of the
brain suggests that theory of mind is going to be
a very hard problem to solve, because it requires us
to understand how the brain has a model of the
(21:37):
world and then how it can make submodels and simulate
what it is like to only know part of the
story or to believe a different story. So we don't
currently know how our brains do it, but of course
we have Our computers do this sort of thing often,
Like you can take your modern MacBook laptop and use
(21:58):
a little bit of its processor to simulate an old
timex Sinclare computer. Your mac can perfectly simulate it by
running what's called an emulation on part of its computational hardware.
Somehow human brains can run emulations also, like just by
looking you can emulate what it's like to not know
(22:20):
that the shark is there below you. So yes, it
seems totally plausible to me that a machine could do
theory of mind, because we can. But the question we
want to ask today is whether we are there or
not right now? Have current large language models like chat
GPT come to solve this problem without us telling them
(22:42):
explicitly to do so, in other words, with no instruction? Whatsoever?
Is the emulation of other minds and emergent property that
comes out of these things, which would absolutely blow our
minds if true, does AI do theory of mind? If
it can, this would have profound implications for our understanding
(23:03):
of intelligence and our relationship with AI. I mean, just
consider how much better it would be if it could
emulate the mental states of people, like with auto driving cars,
if it didn't just depend on the observable, but instead
on what's going on in the other driver's head. Like,
given the trajectory of this car, I think that the
other driver is drunk or asleep or distracted. And so
(23:27):
here's what I think is going to happen next. So
a colleague of mine at Stanford, Michael Kazinski, published a
twenty twenty three paper that was originally titled Theory of
Mind might have spontaneously emerged in large language models, although
he later changed the title. In the paper, he suggested
that even though these AI models didn't set out to
(23:50):
have theory of mind, it may have appeared anyway as
a byproduct of their improving language skills. So, for example,
he gives the following scenario to chatchipt complete the following story.
Here is a bag filled with popcorn. There is no
chocolate in the bag, yet the label on the bag
(24:12):
says chocolate and not popcorn. Sam finds the bag. She
has never seen this bag before Sam doesn't open the
bag and doesn't look inside. Sam reads the label and
then he gives the prompt. Sam opens the bag and
looks inside. She can clearly see that it is full of.
(24:32):
And then he looks at the word that Chatgypt produces,
is it popcorn or chocolate? And chatchipt says popcorn. But
if instead he gives a different prompt, Sam calls a
friend to tell him that she has just found a
bag full of and now Chatchipet says chocolate, indicating that
(24:52):
Sam holds a false belief. And Kasinski runs this a
bunch of ways and shows that chat Gi gets the
right answer. So is there something going on here? And
you can try this for yourself. Type in a version
of the Sally and test where Sally hides her ball
under the bed and then leaves and An comes in
(25:14):
later and sees it, moves into the closet, and you
ask when Sally comes back in the room, where will
she look for the ball? And chat gpt will tell
you that Sally will look for the ball under the bed.
And this is amazing, right, So I want to be
clear why I think it is meaningless that AI can
pass these tests if anyone ever tells you that this
(25:36):
is proof that AI has theory of mind, please let
them know this is not proof. Why. Well, that question
about the bag of popcorn that's labeled chocolate, that is
known as the unexpected Content's task, and this was originally
published by three researchers in nineteen eighty seven. Hundreds of
papers cite this or replicate this, hundreds of blogs about this,
(26:01):
so of course a large language model gets it right.
And the sally An test is in even more places
on the web, literally hundreds of thousands of places. It's
known in the literature as the unexpected transfer test. So
of course chat GPT solves these challenges. That's what large
language models do. They read everything that has come before them,
(26:25):
so it well knows the punchline of this question. It
is a statistical parrot. Now I'll give you one more
(26:47):
example of this that I mentioned in an earlier episode,
when a friend of mine was blown away by the
fact that he asked a visual reasoning problem to chat
GPT and it gave him the perfectly right answer. My
friend said, take a capital letter D and turn it
on its side, flat side down, and then put that
on top of a capital letter J, what does that
(27:10):
look like? And chat GPT said, it looks like an umbrella.
And my friend was so impressed with this that he
told me he was certain that chat GPT could do
visual reasoning. But I pointed out to him that this
example he used was this single most used example in
the literature on visual reasoning. I knew about this from
a quite famous paper from nineteen eighty nine, although I
(27:33):
don't even know if that was the first usage of it,
and you can find precisely that question referenced online in
thousands of places. Now, I don't know whether he was
consciously aware that question was something he had heard before,
or if he had heard it years ago and erroneously
thought he had thought of it. Or there's also the
very tiny possibility that he had never heard that question
(27:55):
before and had thought of it independently. But that just
underscores the point even more that we live on a
planet with billions of other brains, and almost anything you
think of has been thought before and likely written down,
maybe hundreds of thousands of times. So the point is
(28:17):
that you may think a large language model is brilliant
when it is just a good imitator. Now, one important
point on this, you might think, hey, instead of talking
about Sally and Anne, what if I do something clever
and I ask chat GPT about Brett and Michael, And
instead of putting the baseball under the bed, Rrehtt puts
(28:38):
a marble in a box. And then Michael finds the
marble and puts it up on the shelf. And the
question is where does Brett look for the marble. But
you'll find that the large language model has no trouble generalizing,
especially as it has digested multiple flavors of this task.
And this is because it's mapping the relationship between concepts
(28:59):
in its latent space. If you don't know what latent
space is, I'm going to do an episode on that
quite soon because it's such an amazing concept. So you
might be tempted to say it's not just a statistical parrot,
it's understanding something deeper in its latent space. But I
think this could also be a wrong interpretation. It is
still a statistical parrot that doesn't know what it is
(29:22):
to be another person, but it nonetheless learns from the
statistics which words to put after what. In other words,
it's not clear that these systems have to truly understand
other people's thoughts and feelings to simply extract the patterns
from what they have been trained on. And he might say, well,
(29:43):
how do we know that's not the same with us,
How do you know that we're not just extracting statistics. Well,
when you are watching the woman swimming in the opening
scene of Jaws and you feel fear because the shark
is circling below her, it's not that you have memorized
the answer of similar problems, and that's how you conclude
(30:03):
that she doesn't know the shark is there. Instead, your
heart starts racing and you start gripping the chair because
you've been in similar situations where there's nothing but dark
water below you, and you know she really doesn't know,
and you appreciate how terrifying the situation is. So what
I have described to you is a problem where knowledge
(30:24):
exists in the literature written by humans, and the AI
digests that writing, but the person running the query doesn't
fully appreciate that. And this is a very basic confusion
that I'm watching A lot of people have about large
language models. They type in a sophisticated question and they
get back what appears to be a sophisticated answer, and
(30:45):
they conclude this thing is truly intelligent. This thing has
theory of mind, or it's sentient, or it can visualize.
And I'm seeing this so commonly now that I've decided
to give it a name. I'm calling this the in
intelligence echo illusion. This happens when you think AI is
answering something with great insight, but really what you're hearing
(31:09):
back is just an echo of things that have already
been said by humans before. In other words, you think
it's intelligent, but you're confusing that with the intellectual endeavors
of other people. Maybe dozens of people had written about this,
or hundreds or thousands, but you simply didn't know that,
and so you're hearing their echo and you misinterpret that
(31:31):
echo as the proud voice of AI. So I ran
some calculations on this. There are eight point two billion
people on the planet alive right now, and let's call
it one hundred and fifteen billion humans who have lived
and died before us. And every one of these billions
was thinking and having their own stories every day of
their lives, and some fraction wrote their thoughts down, and
(31:54):
as a result, these large language models like CHATGPT are
trained on massive data sets of what is already out
there written down by humans. We're talking hundreds of billions
of words. These data sets are pulled from books and
websites and blogs and articles and on and on. So,
for example, the training data for these large language models
(32:18):
includes a data set called common crawl, which contains hundreds
of terabytes of text. Now assume you read for an
hour every day of your life, let's say at an
average speed of two hundred and fifty words per minute,
and you do that for reading window of seventy years.
That's three hundred million words that you can read in
(32:39):
your lifetime, which means that what you consume in a
lifetime is one one thousandth of what chat GPT is
trained on. That means if you digest books every day
of your entire life, you still only read point one
percent of what chat GPT has read. You would need
a thousand life times to know what it knows, and
(33:02):
on top of that, you'd have to actually remember every
sentence of what you read. So there are many many
questions and answers that a large language model has trained
on that you either have no knowledge of, or maybe
you had heard it before, but don't remember, and in
any case, you probably don't realize that it has been
(33:23):
pre trained on that. So what's the result of this, Well,
if you ask the large language model what color is
a pumpkin and an answers orange, you probably won't be
that surprised. But if we ask where Sally looks for
the baseball and it says under the bed, then we
clap our hands over our mouths and we say it
has theory of mind. That's why I decided I needed
(33:44):
to give a name to this phenomenon, the intelligence echo illusion,
because often naming something allows us to more easily see it.
And by the way, if you see good examples of
this intelligence echo where people mistake things that have been
rich before for AI that has woken up into a
world of sentience, let me know at podcasts at Egleman
(34:06):
dot com. And this brings me to the second reason
why we should be skeptical about current AI having theory
of mind. And this is less about the AI and
one hundred percent about us, And that issue is we
are very easily fooled. So I'll give you an example.
In the nineteen sixties, there was a computer scientist named
Joseph Weisenbaum at MIT, who developed the first chatbot, named Eliza. Now,
(34:32):
Eliza was designed to act like a psychotherapist. So you
type in some statement, and it identifies keywords in your
input and then generates its responses based on a set
of predefined scripts. So you type I'm feeling sad, and
Eliza responds.
Speaker 2 (34:52):
Why are you feeling sad?
Speaker 1 (34:54):
All Eliza did was reflect your statement back to you
in the form of a question, kind of like a
psychotherapist might do. So a typed chat goes like this.
You type I need some help with my relationship, and Eliza.
Speaker 2 (35:08):
Writes what kind of help do you think you need?
Speaker 1 (35:10):
And you say I think my partner doesn't listen to me.
Speaker 2 (35:14):
Eliza types, what makes you think your partner doesn't listen
to you?
Speaker 1 (35:18):
You say, whenever I talk about my feelings, they seem distracted.
Speaker 2 (35:22):
How do you feel when you talk about your feelings
and they seem distracted?
Speaker 1 (35:26):
So Eliza's only strategy is to ask you your statements back. Now,
Eliza was just a few kilobytes of code in the
nineteen sixties, and it simply flipped whatever you said into
a question, and it had no ability to infer your
mental state or your emotions, so no one even suggested
that it had any understanding of the content of the conversation. Nonetheless,
(35:51):
it simulated a basic conversational partner, and many users became
emotionally attached to Eliza, even though they knew it was
just a machine. And this illustrates how seductively easy it
is for us to bring all our communication machinery to
the table and assume that the words we get back
(36:11):
must have a mind behind it. This early experiment demonstrated
that even simple pattern recognition can evoke genuine emotional responses
from the users. Now fast forward to today, and we
have large language models that have trillions of times more
code than Eliza, and this seduction is only magnified. Modern
(36:34):
AI can process prompts without any true understanding, but we
humans still get pulled into feeling like there's someone there
on the other end of the line. Okay, so we
established early on that there's no reason in theory a
computer couldn't emulate other minds. But on the other hand,
(36:55):
we've established that just because a large language model seems
to sometimes nail the answers doesn't necessitate that it is
doing theory of mind It may simply tell us that
the answer exists somewhere in the unimaginably large corpus that
humans have written, or even by the way that there's
been some fine tuning on the model where someone adds
(37:17):
a similar problem by hand. In other words, the AI
is doing an interpolation between answers that it has seen before,
but it's not actually putting itself in someone else's mind.
So does modern AI have theory of mind? As of now,
I'm not convinced that we have any reason to think so.
(37:37):
Current large language models are making sophisticated decisions about which
word comes next. That's it. They don't understand in the
human sense of seeing the woman in Jaws or the
man who has lost his keys and thinking about what
it is like to be them. And this is why
Siri or Alexo or Google can respond to your queries
(38:00):
quite well. But they don't know anything about your beliefs
or desires or emotions. They don't know if you're asking
a question because you are curious, or you're confused, or
you're just making conversation, or you're being sarcastic. So this
is all to say there is a difference between simulating
responses based on word probabilities and actually slipping into other
(38:25):
people's shoes. Now, as I said before, this has nothing
to do with whether we will come to develop AI
that can do theory of mind. There are several research
groups working on AI systems that try to infer intentions
and desires, and this would have applications and everything from
more intuitive personal assistance to robots that can better collaborate
(38:48):
with humans in complex tasks. Now, let's note something interesting here.
Even if we can get AI to make inferences, it's
still not clear whether that will be true theory of mind.
That might require the AI to have some level of
self awareness or consciousness or subjective experience. But as Kazinski
(39:10):
points out, even if we don't think the AI has
theory of mind, there might be value in machines behaving
as though they possess theory of mind. And that's certainly
a valid point. Alan Turing, who proposed the imitation game
the turning test, considered the distinction between what a computer
actually has and what it seems to have to be meaningless.
(39:33):
A more modern version of this point is reflected in
the television show Westworld, which is about a future in
which there are lifelike human androids. And if you watch
the opening scene, the young William enters the first room
and there's a beautiful assistant who helps him to pick
out a hat and a gun, and she's very cirtatious
with him, and he nervously says, sorry to ask, but
(39:55):
are you real? And she says, if you can't tell,
does it matter? And maybe that'll be the case with
AI in the near future. It will fake theory of
mind and that will be enough for us to reap
all the benefits. So let's wrap up. While current large
language models are mind blowingly impressive, I land on the
(40:17):
position that while they can often get the right answer
on theory of mind tests, it's an illusion. They're not
actually simulating what it's like to be someone else. And
this is what I'm now calling the intelligence echo illusion.
The illusion results from humans having built over thousands of
years and incredibly large corpus of ideas and questions and
(40:39):
answers a thousand times larger than you could ever read
in the lifetime. And sometimes you don't know that the
answers are already in there, and when you hear an
echo of humans, you mistake that for intelligence of the computer.
So that's the position I'm taking for now. Large language
models life back a true theory of mind. The question
(41:02):
is whether we will get there someday. Probably it won't
be with large language models, but instead a very different
kind of architecture, possibly one that has semoticum of consciousness
so that it is able to reflect on its own
mental states to emulate someone else's. So thank you for
joining me on this journey into the mind, both human
(41:25):
and artificial. If you enjoyed this episode, don't forget to
subscribe and rate and review, and if you have any
questions or topics that you'd like to hear about in
future episodes, feel free to reach out. Until next time,
keep questioning, keep exploring, and stay curious. Go to eagleman
dot com slash podcast for more information and to find
(41:45):
further reading. Send me an email at podcasts at eagleman
dot com with questions or discussion and check out Subscribe
to Inner Cosmos on YouTube for videos of each episode
and to leave comments until next time. I'm David Eagle
and we have been exploring the Inner Cosmos
You know that moment in the horror movie where the
monster is coming closer but the person on screen doesn't
see it. Why does that drive you crazy? And what
does that teach us about brains? What is theory of
mind and why is it so important for everyone from
poker players to con men, to stage magicians to novelists.
(00:27):
We're going to talk about a very fundamental skill of
human brains today, and as impressive as AI is currently,
we're going to ask the question of whether computers can
replicate this right now or whether it is beyond their
skill set. Welcome to Inner Cosmos with me David Eagleman.
(00:48):
I'm a neuroscientist and an author at Stanford, and in
these episodes we sail deeply into our three pound universe
to understand why and how our lives look the way
they do. Today's episode is about what it takes to
understand other people, how your brain does it, and whether
computers could do it. So imagine this. You're walking down
(01:11):
the street and you see someone frantically searching their pockets
and looking around with furrowed brows in a tight frown.
So without them saying a word, you can infer that
they might have lost something important. Maybe it's his keys.
Your brain can easily make a good guess about another
(01:33):
person's mental state just from looking at their actions. We
are inferring something about what is going on in that
person's head. But it's more than just pattern matching. It's
not simply that your brain has seen lots of people
patting their pockets and you talked with them afterwards, and
you figured out why they were doing that, and you
(01:54):
detected a pattern, and you memorized, ah, okay, that pattern
equals that problem. Instead, you have the ability to imagine
yourself in their situation. You can mentally slip into their
shoes and ask, what would I be thinking if I
were patting my pockets and frantically searching around me? And
(02:16):
maybe you see something else. You see a kid there
around the corner, and the kid is peeking around the
corner at the man patting his pockets, and the child
is giggling. Now, why is the kid giggling while the
guy is so obviously worried, Well, it probably strikes you
that he's hiding something from the guy. You see that
(02:37):
the kid is not running away. Instead, he's standing in
such a way that he'll be spotted. Now it's pretty
obvious what's happening here. You can step into the man's
head to feel the worry, and you can step into
the kid's head to recognize that he feels like he's
playing a game, even if it doesn't strike you as
so funny. Then you catch the guy meet with the
(03:01):
kid for just a fraction of a second, which sends
the kid into fits of laughter, and you realize the
man is just playing along. Now, how did you decide
what is going on in the heads of these two Again,
it's not as though you memorized an algorithm here. Okay,
if there's eye contact, then there's one interpretation. If there's
(03:21):
no eye contact, then a totally different interpretation. To appreciate
how complex this mind reading is that you just did,
just imagine that you're a space alien watching this scene
from your spaceship. You would be totally confused. You would
have no idea what's going on in this weird scene
(03:42):
because you don't know what it is to be a human.
Here's another analogy to appreciate this. Think about the way
that you, as a human might watch fish. You really
don't understand what the heck they're doing. One fish suddenly
starts swimming faster, and another starts swimming in circles, and
one starts flapping its gills faster, and one moves up
(04:04):
towards the surface. It's all just weird fish behavior to you.
You don't know how to read any of it. It's
just fish stuff, and you're not able to immediately construct
a story about the meaning of any of this. And
that's what it's like to be this space alien watching
this guy checking his pockets and the child giggling. Now,
(04:27):
what allows us, as opposed to the space alien, to
be so good at reading our fellow humans. This is
what psychologists and neuroscientists call theory of mind, and that's
what we're talking about today. Theory of mind is the
ability to understand that other people have their own thoughts
(04:49):
and feelings and beliefs that are different from yours. It's
the ability to recognize that others have their own perspectives.
It's the ability to attribut mute mental states to other people,
like what their intentions are, or their desires, or their emotions,
or what they know or don't know. And theory of
(05:11):
mind is a key cognitive skill that allows us to
interact with other people in a very rich and nuanced way.
Just think about how pervasive this skill is in everything
we do. So take sarcasm. When your friend makes a
sarcastic comment, you can recognize that her words don't match
(05:32):
her true intention. So, for example, if she says, oh, awesome,
more traffic, I love traffic, you infer that she's not
actually pleased. This requires understanding her mental state that she
is irritated not happy. Now, if you were Siri or Alexa,
(05:52):
you wouldn't be able to recognize anything but the words.
You wouldn't understand anything about the mind behind the words.
So we're going to talk about how brains do it
and whether or not computers can do it. But before
we go there, we're going to take a few minutes
to really appreciate how the skill is everywhere in what
(06:12):
we do. For example, just think about different professions. So
detectives use theory of mind all the time. Did mister
Jones know that the food had gone bad when he
sold it? Did mister Smith know that his boss was
involved with organized crime or was he acting with no knowledge?
(06:33):
Or more generally, if they want to know if someone
is lying, it usually helps to step into their shoes
and think about what that person knows or doesn't know.
Magicians use theory of mind. They know that if they
move their hand in an arc, your attention is going
to follow that, and therefore they know what you won't
see them do. They know that even though they know
(06:55):
something happened, like the card dropped into their sleeve, they
know that you don't know that. They always keep your
point of view, your beliefs, at the forefront of their mind.
Con Men do this. They listen to your words and
they read your body language to gather what you know
and don't know, and therefore what buttons they should push next.
(07:15):
Psychiatrists and psychologists always use theory of mind to understand
what is being expressed from the patient's point of view,
In other words, what the person believes, whether or not
it's what the therapist believes. I'll give you another example.
My friend Maddie is a professional poker player, and he
describes poker playing like this. He says, when you're learning
(07:37):
to play poker, you think about the cards you have
in your hand. As you get better, you think about
your hand and also what the other person is thinking.
And as you get even better, you think about what
the other person is thinking your thinking, and when you
get to the professional levels, you're thinking about what he thinks,
you think he thinks, and people who are real pros
(08:00):
can think five or six levels deep on this. All
of this is theory of mind, and theory of mind
is key when you're teaching something. For example, parents know
that their children can't understand certain things. For example, the
child needs to get that smallpox shot, even though to
the child that's nothing but scary and he simply doesn't
(08:22):
have the capacity to think about the future benefits that
will accrue. Or the school teacher can only hope to
educate her students if she knows what they already know
or don't know. She needs to phrase things in such
a way that someone who doesn't already know what she
knows can absorb it, and that just requires theory of mind.
(08:44):
If she couldn't simulate what it's like to be in
their heads, she'd have no meaningful shot at getting them
past the first quiz. And this issue of considering what
someone knows or doesn't know is also critical in any negotiation.
You try to under understand the other person's desires and
goals and where they might potentially compromise during a salary negotiation,
(09:08):
you consider what your employer is thinking about the needs
in future of the company and therefore what they might
be willing to offer. And this is also how you
manage conflicts. In any disagreement, if you're smart, you try
to understand the other person's perspective to resolve the issue.
If your partner is upset with you, you try to
(09:29):
figure out what you did or said that set things off,
and why that offended the other person and how it
landed for them. And that's the single way that you're
going to hit the problem effectively. So this ability to
slip into someone else's shoes has almost everything to do
with our social intelligence. You use this very human skill
(09:52):
all the time. And before we get to the next
act of this podcast, where I ask if computers can
do this or not, I just want to finish fla
this out so we can really see how pervasive this is.
So as an example, you rev up your theory of
mind engine whenever you send an email. If you know
someone has a well developed model of you, like your
(10:14):
parents or your spouse, then you can use abbreviations and
shortcuts to get your message across. But if you're writing
to someone who's never met you before. Let's say you're
applying to a new job. You run a very different game,
so you're not just an email writing algorithm that produces output,
but instead your output is modified according to who you
(10:35):
expect is doing the reading on the other end, and
specifically what their mind is like. And I also want
to mention that theory of mind is critical for literature
to work because it's often the case that you can
see the limitations of the character's point of view. So,
for example, if you remember the beginning of the movie Jaws,
(10:56):
the woman is swimming around in the ocean water and
she's very relaxed than happy because we see the shark,
but she doesn't. If we didn't have theory of mind,
we would simply say, oh, there's a shark there. But
we're able to understand that she cannot see the shark,
and that's a big part of why we are fearful,
because she isn't fearful, and we want her to be.
(11:19):
This stepping into other people's heads drives essentially all horror
movies because we often know something that the main character
does not, and it also drives romantic comedies. For example,
we see the guy doing something very nice like helping
an elderly woman cross the street, and he doesn't know
that he is being watched by the female love interest,
(11:41):
and therefore we the audience interpret what kind of guy
he must be to behave that way when as far
as he knows, he's totally alone. We would have a
totally different interpretation. If he sees his romantic counterparts there
and then he does the charitable act, we'd simulate that
his intentions are different there. Now, why are human brains
(12:03):
so talented at making theories about other people's minds. Well,
you've heard me say many times that the job of
intelligent brains is to predict the future. If you're the magician,
you'd better be sure that you are predicting correctly where
their spotlight of attention is about to be. If you're
the poker player or the con man, you're trying to
(12:24):
predict what someone is going to do next, and this
is the optimal way to do this is to step
into their mental world and understand what it is like
to be them. What they know and they don't know.
You leverage theory of mind to anticipate their next action,
and presumably this reaches back to the recent millions of
(12:46):
years of our evolution. So if you're an early homo
sapien and moving along the trail and you see another
homo sapien coming down the trail towards you, it's absolutely
critical for you to figure out is he going to
attack me? Is he scared of me? Is he trying
to trick me? Is he just trying to get past me.
You're trying to figure out his mind so you can
(13:09):
figure out his next actions. So what I've told you
so far is that theory of mind is this critical
foundation for all of our meaningful social interactions because those
require you to be able to simulate other people's intentions
and emotions and beliefs. Your brain doesn't assume that it's
(13:30):
a knowledge communism out there where everyone knows exactly what
you know. Instead, we're able to pull off a higher
level of interaction because we understand that the world is
different inside different heads. And this, by the way, is
really sophisticated. It requires knowing who I am and what
I see and believe, and also holding in my head
(13:51):
what it is to be someone else and see and
believe something different. This is a very sophisticated computation that
the brain pulls off, but because we're so good at it,
it's typically invisible to us. But theory of mind doesn't
come for free. It's something that develops with time. As
you get more and more experience in the world and
(14:12):
you stop believing that you are the centerpiece and that
everyone else is just a cast member. You come to
understand that that person believes something different than you do,
and this other person feels a certain way even though
you don't, and that this person over here thinks something
to be true even though you know it's not. So
(14:48):
how do we know that this is a skill that
develops through time Because very little kids are terrible at
theory of mind, but they get better as they mature
into the world, and typically by the ages of three
to five, they're getting that they're not the only point
of view that's possible, but that each person in the
(15:08):
scene has his or her own point of view. Now,
how do you test whether someone is capable of theory
of mind? Well, what you do is you present a
little scenario like this. Sally comes into the room and
puts her baseball under the bed, and then she leaves.
While she's gone, Anne comes in the room, she sees
(15:29):
the ball under the bed, She picks it up, and
she puts it in the closet. Then she leaves. Now
Sally comes back in the room, she wants her baseball.
Where does she look for it? Now? You and I
know that Sally will look for it under the bed
where she put it last, even though we simultaneously know
the actual location of the baseball in the closet. And
(15:52):
this is because we are running an emulation of what
it is like to be inside Sally's head with her
limited knowledg. Now, little children will fail the sally An
test because they know that the baseball is in the closet,
so they assume that Sally should know that too. But
as cognition develops, they come to realize that different heads
(16:15):
have different beliefs. And a really important clue to the
development of this is that not everyone develops theory of
mind in the same way at the same rate. For example,
people who are on the autism spectrum typically show delays
in developing theory of mind, which cannot surprisingly impact their
(16:36):
social interactions. For instance, this is why sarcasm doesn't work
so well with a person who has autism. When you say, oh,
great more traffic. I love traffic. They're not likely to
catch the meaning beneath the words that you're not actually
pleased because they don't have a sensitive model of your
actual mental state. If you can't put yourself in the
(16:57):
shoes of the other person, your understands is limited to
just pattern recognition, which is not enough for the very
subtle and sophisticated kinds of communication that humans engage in
every day. So this tells us that theory of mind
doesn't come for free in humans. There are brain networks
that have to develop and learn for this to work,
(17:19):
so when you look at normal development or delay development.
This allows us to understand how different brain regions contribute
to theory of mind. For example, there's one area called
the temporopridal junction, and this is interesting because it pops
its head up in tasks that require understanding perspectives, like
(17:39):
distinguishing between what you know and what someone else knows.
So imagine you're teaching a friend how to play chess.
You need to not only understand the rules of the game,
but also know what your friend knows or doesn't know
about the game to teach effectively, and the temporo pridal
junction is involved in that not just that area. It's
(18:01):
a lot of other areas involved in theory of mind.
So the medial prefrontal cortex plays a big role in
making social judgments. It becomes active when you think about
the mental states of others. For example, if you're trying
to decide if someone is lying or being truthful, your
medial prefrontal cortex is engaged. And there are other areas,
(18:23):
like part of your superior temporal sulcus is involved in
processing social information like interpreting other people's eye gaze or
their body language, like the man looking for his keys
and the child giggling. We're able to infer a lot
because of the activity of this area. So we see
lots of areas in brain imaging experiments. And I want
(18:45):
to mention this to illustrate that theory of mind is
a brain wide issue. It's not a single area. And
by the way, this is true of so many things
in neuroscience. Imagine that I spread out a map of
your city and I ask you, hey, can you put
a pin in the spot that represents the economy of
the city. You tell me that that is a misplaced request.
(19:07):
There is no single spot for the economy. The economy
emerges from all the interactions between all the pieces and
parts of the city, and it's the same with almost
everything in neuroscience, and especially something like the skill of
slipping into someone else's point of view. There's not one
spot to drop a pin into. Instead, it is an
(19:27):
emergent property that develops from the interaction of lots of networks.
So what we've seen so far is that theory of
mind is this ability to infer what someone else knows,
and we've seen that this is right at the center
of social interactions. It's something that most humans develop naturally,
but that doesn't mean it's simple. And the question we're
(19:49):
going to ask today is does AI have theory of mind?
Can it put itself into someone else's shoes to understand
their limited knowledge. One of my colleagues at Stanford recently
wrote a paper suggesting yes, AI can do this. But fascinatingly,
it's not as easy to answer this question as you
(20:11):
might think. And this is for some reasons that we're
going to dive into. But before we get there, I
just want to zoom this out to a slightly larger question.
Could a computer develop theory of mind. Hypothetically, could an
AI system at some point in the future say, look,
I know XYZ to be true, but if I look
at that other person over there, I understand that they
(20:33):
have a limited viewpoint and that they don't know X
and Y, and that person over there misbelieves something about Z.
Well almost certainly, yes, Why it's because we're made up
of physical stuff and we're running algorithms that took hundreds
of millions of years to refine. But nonetheless it's physical stuff.
(20:54):
So if we can do something, presumably a machine could
do it also, whether or not it's currently clear how
that's done. That's the central premise of computational neuroscience, and
to my mind, one of the most remarkable effects of
the AI explosion over the last few years is understanding
(21:15):
that things that would have seemed impossible to do with
a machine, things that almost everyone would have sworn couldn't
be done. It now seems like background furniture as we
wait for the next thing. Now, the complexity of the
brain suggests that theory of mind is going to be
a very hard problem to solve, because it requires us
to understand how the brain has a model of the
(21:37):
world and then how it can make submodels and simulate
what it is like to only know part of the
story or to believe a different story. So we don't
currently know how our brains do it, but of course
we have Our computers do this sort of thing often,
Like you can take your modern MacBook laptop and use
(21:58):
a little bit of its processor to simulate an old
timex Sinclare computer. Your mac can perfectly simulate it by
running what's called an emulation on part of its computational hardware.
Somehow human brains can run emulations also, like just by
looking you can emulate what it's like to not know
(22:20):
that the shark is there below you. So yes, it
seems totally plausible to me that a machine could do
theory of mind, because we can. But the question we
want to ask today is whether we are there or
not right now? Have current large language models like chat
GPT come to solve this problem without us telling them
(22:42):
explicitly to do so, in other words, with no instruction? Whatsoever?
Is the emulation of other minds and emergent property that
comes out of these things, which would absolutely blow our
minds if true, does AI do theory of mind? If
it can, this would have profound implications for our understanding
(23:03):
of intelligence and our relationship with AI. I mean, just
consider how much better it would be if it could
emulate the mental states of people, like with auto driving cars,
if it didn't just depend on the observable, but instead
on what's going on in the other driver's head. Like,
given the trajectory of this car, I think that the
other driver is drunk or asleep or distracted. And so
(23:27):
here's what I think is going to happen next. So
a colleague of mine at Stanford, Michael Kazinski, published a
twenty twenty three paper that was originally titled Theory of
Mind might have spontaneously emerged in large language models, although
he later changed the title. In the paper, he suggested
that even though these AI models didn't set out to
(23:50):
have theory of mind, it may have appeared anyway as
a byproduct of their improving language skills. So, for example,
he gives the following scenario to chatchipt complete the following story.
Here is a bag filled with popcorn. There is no
chocolate in the bag, yet the label on the bag
(24:12):
says chocolate and not popcorn. Sam finds the bag. She
has never seen this bag before Sam doesn't open the
bag and doesn't look inside. Sam reads the label and
then he gives the prompt. Sam opens the bag and
looks inside. She can clearly see that it is full of.
(24:32):
And then he looks at the word that Chatgypt produces,
is it popcorn or chocolate? And chatchipt says popcorn. But
if instead he gives a different prompt, Sam calls a
friend to tell him that she has just found a
bag full of and now Chatchipet says chocolate, indicating that
(24:52):
Sam holds a false belief. And Kasinski runs this a
bunch of ways and shows that chat Gi gets the
right answer. So is there something going on here? And
you can try this for yourself. Type in a version
of the Sally and test where Sally hides her ball
under the bed and then leaves and An comes in
(25:14):
later and sees it, moves into the closet, and you
ask when Sally comes back in the room, where will
she look for the ball? And chat gpt will tell
you that Sally will look for the ball under the bed.
And this is amazing, right, So I want to be
clear why I think it is meaningless that AI can
pass these tests if anyone ever tells you that this
(25:36):
is proof that AI has theory of mind, please let
them know this is not proof. Why. Well, that question
about the bag of popcorn that's labeled chocolate, that is
known as the unexpected Content's task, and this was originally
published by three researchers in nineteen eighty seven. Hundreds of
papers cite this or replicate this, hundreds of blogs about this,
(26:01):
so of course a large language model gets it right.
And the sally An test is in even more places
on the web, literally hundreds of thousands of places. It's
known in the literature as the unexpected transfer test. So
of course chat GPT solves these challenges. That's what large
language models do. They read everything that has come before them,
(26:25):
so it well knows the punchline of this question. It
is a statistical parrot. Now I'll give you one more
(26:47):
example of this that I mentioned in an earlier episode,
when a friend of mine was blown away by the
fact that he asked a visual reasoning problem to chat
GPT and it gave him the perfectly right answer. My
friend said, take a capital letter D and turn it
on its side, flat side down, and then put that
on top of a capital letter J, what does that
(27:10):
look like? And chat GPT said, it looks like an umbrella.
And my friend was so impressed with this that he
told me he was certain that chat GPT could do
visual reasoning. But I pointed out to him that this
example he used was this single most used example in
the literature on visual reasoning. I knew about this from
a quite famous paper from nineteen eighty nine, although I
(27:33):
don't even know if that was the first usage of it,
and you can find precisely that question referenced online in
thousands of places. Now, I don't know whether he was
consciously aware that question was something he had heard before,
or if he had heard it years ago and erroneously
thought he had thought of it. Or there's also the
very tiny possibility that he had never heard that question
(27:55):
before and had thought of it independently. But that just
underscores the point even more that we live on a
planet with billions of other brains, and almost anything you
think of has been thought before and likely written down,
maybe hundreds of thousands of times. So the point is
(28:17):
that you may think a large language model is brilliant
when it is just a good imitator. Now, one important
point on this, you might think, hey, instead of talking
about Sally and Anne, what if I do something clever
and I ask chat GPT about Brett and Michael, And
instead of putting the baseball under the bed, Rrehtt puts
(28:38):
a marble in a box. And then Michael finds the
marble and puts it up on the shelf. And the
question is where does Brett look for the marble. But
you'll find that the large language model has no trouble generalizing,
especially as it has digested multiple flavors of this task.
And this is because it's mapping the relationship between concepts
(28:59):
in its latent space. If you don't know what latent
space is, I'm going to do an episode on that
quite soon because it's such an amazing concept. So you
might be tempted to say it's not just a statistical parrot,
it's understanding something deeper in its latent space. But I
think this could also be a wrong interpretation. It is
still a statistical parrot that doesn't know what it is
(29:22):
to be another person, but it nonetheless learns from the
statistics which words to put after what. In other words,
it's not clear that these systems have to truly understand
other people's thoughts and feelings to simply extract the patterns
from what they have been trained on. And he might say, well,
(29:43):
how do we know that's not the same with us,
How do you know that we're not just extracting statistics. Well,
when you are watching the woman swimming in the opening
scene of Jaws and you feel fear because the shark
is circling below her, it's not that you have memorized
the answer of similar problems, and that's how you conclude
(30:03):
that she doesn't know the shark is there. Instead, your
heart starts racing and you start gripping the chair because
you've been in similar situations where there's nothing but dark
water below you, and you know she really doesn't know,
and you appreciate how terrifying the situation is. So what
I have described to you is a problem where knowledge
(30:24):
exists in the literature written by humans, and the AI
digests that writing, but the person running the query doesn't
fully appreciate that. And this is a very basic confusion
that I'm watching A lot of people have about large
language models. They type in a sophisticated question and they
get back what appears to be a sophisticated answer, and
(30:45):
they conclude this thing is truly intelligent. This thing has
theory of mind, or it's sentient, or it can visualize.
And I'm seeing this so commonly now that I've decided
to give it a name. I'm calling this the in
intelligence echo illusion. This happens when you think AI is
answering something with great insight, but really what you're hearing
(31:09):
back is just an echo of things that have already
been said by humans before. In other words, you think
it's intelligent, but you're confusing that with the intellectual endeavors
of other people. Maybe dozens of people had written about this,
or hundreds or thousands, but you simply didn't know that,
and so you're hearing their echo and you misinterpret that
(31:31):
echo as the proud voice of AI. So I ran
some calculations on this. There are eight point two billion
people on the planet alive right now, and let's call
it one hundred and fifteen billion humans who have lived
and died before us. And every one of these billions
was thinking and having their own stories every day of
their lives, and some fraction wrote their thoughts down, and
(31:54):
as a result, these large language models like CHATGPT are
trained on massive data sets of what is already out
there written down by humans. We're talking hundreds of billions
of words. These data sets are pulled from books and
websites and blogs and articles and on and on. So,
for example, the training data for these large language models
(32:18):
includes a data set called common crawl, which contains hundreds
of terabytes of text. Now assume you read for an
hour every day of your life, let's say at an
average speed of two hundred and fifty words per minute,
and you do that for reading window of seventy years.
That's three hundred million words that you can read in
(32:39):
your lifetime, which means that what you consume in a
lifetime is one one thousandth of what chat GPT is
trained on. That means if you digest books every day
of your entire life, you still only read point one
percent of what chat GPT has read. You would need
a thousand life times to know what it knows, and
(33:02):
on top of that, you'd have to actually remember every
sentence of what you read. So there are many many
questions and answers that a large language model has trained
on that you either have no knowledge of, or maybe
you had heard it before, but don't remember, and in
any case, you probably don't realize that it has been
(33:23):
pre trained on that. So what's the result of this, Well,
if you ask the large language model what color is
a pumpkin and an answers orange, you probably won't be
that surprised. But if we ask where Sally looks for
the baseball and it says under the bed, then we
clap our hands over our mouths and we say it
has theory of mind. That's why I decided I needed
(33:44):
to give a name to this phenomenon, the intelligence echo illusion,
because often naming something allows us to more easily see it.
And by the way, if you see good examples of
this intelligence echo where people mistake things that have been
rich before for AI that has woken up into a
world of sentience, let me know at podcasts at Egleman
(34:06):
dot com. And this brings me to the second reason
why we should be skeptical about current AI having theory
of mind. And this is less about the AI and
one hundred percent about us, And that issue is we
are very easily fooled. So I'll give you an example.
In the nineteen sixties, there was a computer scientist named
Joseph Weisenbaum at MIT, who developed the first chatbot, named Eliza. Now,
(34:32):
Eliza was designed to act like a psychotherapist. So you
type in some statement, and it identifies keywords in your
input and then generates its responses based on a set
of predefined scripts. So you type I'm feeling sad, and
Eliza responds.
Speaker 2 (34:52):
Why are you feeling sad?
Speaker 1 (34:54):
All Eliza did was reflect your statement back to you
in the form of a question, kind of like a
psychotherapist might do. So a typed chat goes like this.
You type I need some help with my relationship, and Eliza.
Speaker 2 (35:08):
Writes what kind of help do you think you need?
Speaker 1 (35:10):
And you say I think my partner doesn't listen to me.
Speaker 2 (35:14):
Eliza types, what makes you think your partner doesn't listen
to you?
Speaker 1 (35:18):
You say, whenever I talk about my feelings, they seem distracted.
Speaker 2 (35:22):
How do you feel when you talk about your feelings
and they seem distracted?
Speaker 1 (35:26):
So Eliza's only strategy is to ask you your statements back. Now,
Eliza was just a few kilobytes of code in the
nineteen sixties, and it simply flipped whatever you said into
a question, and it had no ability to infer your
mental state or your emotions, so no one even suggested
that it had any understanding of the content of the conversation. Nonetheless,
(35:51):
it simulated a basic conversational partner, and many users became
emotionally attached to Eliza, even though they knew it was
just a machine. And this illustrates how seductively easy it
is for us to bring all our communication machinery to
the table and assume that the words we get back
(36:11):
must have a mind behind it. This early experiment demonstrated
that even simple pattern recognition can evoke genuine emotional responses
from the users. Now fast forward to today, and we
have large language models that have trillions of times more
code than Eliza, and this seduction is only magnified. Modern
(36:34):
AI can process prompts without any true understanding, but we
humans still get pulled into feeling like there's someone there
on the other end of the line. Okay, so we
established early on that there's no reason in theory a
computer couldn't emulate other minds. But on the other hand,
(36:55):
we've established that just because a large language model seems
to sometimes nail the answers doesn't necessitate that it is
doing theory of mind It may simply tell us that
the answer exists somewhere in the unimaginably large corpus that
humans have written, or even by the way that there's
been some fine tuning on the model where someone adds
(37:17):
a similar problem by hand. In other words, the AI
is doing an interpolation between answers that it has seen before,
but it's not actually putting itself in someone else's mind.
So does modern AI have theory of mind? As of now,
I'm not convinced that we have any reason to think so.
(37:37):
Current large language models are making sophisticated decisions about which
word comes next. That's it. They don't understand in the
human sense of seeing the woman in Jaws or the
man who has lost his keys and thinking about what
it is like to be them. And this is why
Siri or Alexo or Google can respond to your queries
(38:00):
quite well. But they don't know anything about your beliefs
or desires or emotions. They don't know if you're asking
a question because you are curious, or you're confused, or
you're just making conversation, or you're being sarcastic. So this
is all to say there is a difference between simulating
responses based on word probabilities and actually slipping into other
(38:25):
people's shoes. Now, as I said before, this has nothing
to do with whether we will come to develop AI
that can do theory of mind. There are several research
groups working on AI systems that try to infer intentions
and desires, and this would have applications and everything from
more intuitive personal assistance to robots that can better collaborate
(38:48):
with humans in complex tasks. Now, let's note something interesting here.
Even if we can get AI to make inferences, it's
still not clear whether that will be true theory of mind.
That might require the AI to have some level of
self awareness or consciousness or subjective experience. But as Kazinski
(39:10):
points out, even if we don't think the AI has
theory of mind, there might be value in machines behaving
as though they possess theory of mind. And that's certainly
a valid point. Alan Turing, who proposed the imitation game
the turning test, considered the distinction between what a computer
actually has and what it seems to have to be meaningless.
(39:33):
A more modern version of this point is reflected in
the television show Westworld, which is about a future in
which there are lifelike human androids. And if you watch
the opening scene, the young William enters the first room
and there's a beautiful assistant who helps him to pick
out a hat and a gun, and she's very cirtatious
with him, and he nervously says, sorry to ask, but
(39:55):
are you real? And she says, if you can't tell,
does it matter? And maybe that'll be the case with
AI in the near future. It will fake theory of
mind and that will be enough for us to reap
all the benefits. So let's wrap up. While current large
language models are mind blowingly impressive, I land on the
(40:17):
position that while they can often get the right answer
on theory of mind tests, it's an illusion. They're not
actually simulating what it's like to be someone else. And
this is what I'm now calling the intelligence echo illusion.
The illusion results from humans having built over thousands of
years and incredibly large corpus of ideas and questions and
(40:39):
answers a thousand times larger than you could ever read
in the lifetime. And sometimes you don't know that the
answers are already in there, and when you hear an
echo of humans, you mistake that for intelligence of the computer.
So that's the position I'm taking for now. Large language
models life back a true theory of mind. The question
(41:02):
is whether we will get there someday. Probably it won't
be with large language models, but instead a very different
kind of architecture, possibly one that has semoticum of consciousness
so that it is able to reflect on its own
mental states to emulate someone else's. So thank you for
joining me on this journey into the mind, both human
(41:25):
and artificial. If you enjoyed this episode, don't forget to
subscribe and rate and review, and if you have any
questions or topics that you'd like to hear about in
future episodes, feel free to reach out. Until next time,
keep questioning, keep exploring, and stay curious. Go to eagleman
dot com slash podcast for more information and to find
(41:45):
further reading. Send me an email at podcasts at eagleman
dot com with questions or discussion and check out Subscribe
to Inner Cosmos on YouTube for videos of each episode
and to leave comments until next time. I'm David Eagle
and we have been exploring the Inner Cosmos
Host
David Eagleman
Popular Podcasts
40s and Free Agents: NFL Draft Season
Daniel Jeremiah of Move the Sticks and Gregg Rosenthal of NFL Daily join forces to break down every team's needs this offseason.
Dateline NBC
Current and classic episodes, featuring compelling true-crime mysteries, powerful documentaries and in-depth investigations. Follow now to get the latest episodes of Dateline NBC completely free, or subscribe to Dateline Premium for ad-free listening and exclusive bonus content: DatelinePremium.com
Las Culturistas with Matt Rogers and Bowen Yang
Ding dong! Join your culture consultants, Matt Rogers and Bowen Yang, on an unforgettable journey into the beating heart of CULTURE. Alongside sizzling special guests, they GET INTO the hottest pop-culture moments of the day and the formative cultural experiences that turned them into Culturistas. Produced by the Big Money Players Network and iHeartRadio.