Hello, welcome to Sock Talk with JNab and
the Sundance Kid. We're going to explore
the frontiers of technology,
art and the human experience.
We're back for our second podcast, which
means that we did
something right in that we don't
mind sitting down together again and
talking. So success on that front.
If I may correct you scientifically
speaking, it just means we
didn't do too much wrong.
Yes. What's our
falsifiable position here?
I was just trying to be persnickety.
Now, we just sat and watched
the Sora videos from OpenAI.
Wow.
And yeah, I think that will be an
interesting talking point to see where on
earth our conversation
goes with that. Now, I have one of the
other backgrounds I have
is a filmmaker. I write
and direct films locally, short films,
never done a feature. I
like to say I'm a technologist
as much as I am a filmmaker. I'm much
more of an early adopter of
technologies. I specialize
in VFX as well, visual effects. And I
like to sit down. I
love the idea of having all
the tools available to me where I can sit
down and make a film
on my own. There's two
levels of filmmaking that I really enjoy.
I love the collaborative
effort and being around
lots of creative people. But I also like
the version of it
where I can sit at home,
I sit in front of my computer and
completely make whole
worlds without the difficulty of
pulling together lots and lots of people.
Because that's one of
the hardest parts of
filmmaking at scale is the logistics of
getting all those creative
people together at various
points in time. That takes a lot of
energy, a lot of resources,
a lot of money. So there's
a level of filmmaking that I really enjoy
sitting down on my
own to be able to create
worlds and stories. Now, I think they're
a little less creative
because the more people
you have, the better. But there's a long,
roundabout way of saying it excites me on
levels and also scares me on levels. The
exciting part is that it
lowers the barrier for entry
for creating new worlds where I had to
sit. I've got 10 years,
15, nearing two decades,
nearing two decades now worth of sitting
down and trying to
figure out how to do visual
effects and how to create worlds on
screen. And even still,
I'm not good enough to make
some of the stuff that people can now
just type. And that, again,
it's exciting and terrifying.
I just don't know where this is going to
go. So I've just kind of
blurted out my immediate
thoughts and feelings
for you to chew on there.
Well, thank you. I appreciate that. Well,
I'm glad to hear you say that. I like the
fact that you started off by saying that
you don't have any idea
where it's going in the
future, where it may go in the future.
I'm happy to hear that
because none of us know
where technology will take us in the
future. And I'm always
worried when I hear people
starting to predict that. Predicting
uncertainty is great. Predicting massive
change is a surefire
win. But where that particular technology
is going to take us, I
hope, is unpredictable.
I hope that within a couple of years,
people are doing things
that you and I couldn't even
imagine sitting down to do. I want to
give a bit of my own
history in the film industry,
which is not as storied as yours. But my
first part-time job as a
kid was cleaning and splicing
reel-to-reel films for a small production
house. So that's how old I am. You talked
about being-
Real-cutter-based.
Yeah. Literally, yeah. We had a special
cello tape that we used
to splice together the cuts
we made. And you literally cleaned
reel-to-reel film by hand with hand
cranks. It was quite
the device. So that's how long I've been
in it. Eventually, I
got into illustration
and from that into animation and 3D
animation. So I guess in
terms of modern production,
I've directed a couple of
computer-animated 2D television
commercials and I storyboarded
a feature. Oh, and I had a short film
that won an animation
award a couple of decades ago,
the Animation 2000-something-something
award in California.
The only really cool thing
about that was that I got to share a
stage with Nick Park
when I got the award.
You said there like, "Oh, it's not as my-
I don't have as much
of detail as you." And
then you just go and say a bunch of
things that are very good to have.
Very good to have. But there are bits and
pieces over decades,
whereas you've gone through
the process of sitting down and making
films and working with
others to improve your work
and to improve their work. And that's
what's missing from mine.
I had that working on the
little short film and I had that on and
off in animation
studios. But in terms of having
that experience working on a film,
closest I got that short
animated film and then in
a very restricted way. I mentioned I
storyboarded a feature. I
worked really closely with the
director on that. But that wasn't working
with the whole team. We
didn't go from setting
the storyboards up and perfecting the
script to then making a
film. That film was never
shot. So I've never had that wall-to-wall
experience. And I think that's vital for
having an understanding of the industry.
So I really, I don't
mean to sound facetious.
I'm not being facetious. I think your
experience in this far
outweighs mine. That said, I'm very
happy to hear that you can't imagine what
comes next. That's the
kind of attitude that
makes, that slows people down who are
doing new things. And I
never would have thought
of you as that kind of person. So thanks
for providing further
evidence. The clips we just
watched just now shocked me. And they
didn't shock me because of
the quality of the graphics.
That's shocking. But I expect to be
shocked by improvements
in graphics. It shocked me
because of the quality of the
storytelling. Now, I
understand from you that these clips
are done by a single prompt. And I have
no idea what the
algorithm was trained on. I
don't know what data it was exposed to.
So I'll give some context of what I know,
which is coming off
my memory in the middle
of a podcast. But I believe, I believe
OpenAI made a deal with
Shutterstock specifically
and Shutterstock, which has something
like 13 million videos
or something ridiculous.
They had all access to all of that data.
So it's been trained
on 13 million videos of
stock footage.
There are a few places in the clips that
we watched where we
were talking about there's
problems with understanding that things
don't just miraculously
appear. If you want something
to miraculously appear, you have to give
us a visual cue to the fact that
something miraculous
is happening or some other explanation
for it. So we can have
some kind of heuristic
for how this creature is appearing. Is
there a bit of smoke?
Is there a flashlight? Is
there a transporter noise? Is there
suddenly a door there? Any
kind of cue that would tie
into our existing mental models of how
things suddenly appear
would have helped. And we
talked about the lack of gravity for the
little creatures coming
out of the petri dish. Now
they seem to be on a different gravity
than the actual lab
that they're stationed in.
And along those lines, we talked about
the snow that still looks
artificial. Those things
could come from the training data, right?
There's lots of people
who don't quite animate
successfully with gravity. Though that
softshoeing, is it a kangaroo at the
beginning? The softshoeing
critter on stage at the beginning.
It looks like a kangaroo, yeah.
Yeah. I mean, there's secondary and
tertiary movement in the
skin. That's spectacular.
That looks like somebody who understands
gravity and mass in the way that an
animator or filmmaker
needs to understand it. So yeah, there's
an interesting sort of
interference pattern there
and what's right and what's wrong in the
realism behind it. But
the storyteller, man,
when you cut from the face of the
astronaut to the hatch, and
it holds for a beat on the
hatch before the hand comes in from the
side. Like, well done. Well done.
It understands because of all the data
has been trained on. I
wonder if they put weights
and biases on it by how popular some of
the videos were. That
seems like a wise thing
to do. We all have, I'm not sure how much
they make available their
methods behind producing
these models, but I would imagine they've
definitely, it would
be smart to label your
data like that. So I would imagine it's
being trained on videos
that are more successful
and what they aim to achieve. So
Shutterstock is where people go to, to
quickly pull together
content together so you don't have to go
out and film it.
Makes sense. Easy. And the
more successful ones will be the ones
that in theory resonate a
little more with people.
In theory, it depends. It depends what
they're being used for. So
the other thing I've heard
about these models and Sora specifically,
it's not an LLM, it's
a world model. They're
building models of the world. So that's
why it can move camera
through space and time
because it actually has an understanding
of space time and
builds the same way probably
our brains do, builds a model of space
time in an area of time
and space and is able to
change its interpretation temporally in
such a way that it
produces all of these videos.
Now I am slowly paraphrasing of what I
can remember. This is
not my field. AI is not
my field specifically here, but that's
what I'm led to believe
is that they're, they're
building world models now and they have
quite a lot of hope that
these models can abstract
and problem solve in more ways than just
a language model can.
Okay. That's, that's really interesting.
There's a couple of things
you said there that really
made me start to reflect and think about
what I have to do next
in terms of developing a
better understanding of this. The, you
said it understands,
it doesn't understand.
Sure.
Right. And that's an important thing.
That's, that's a
peridylia. That's, that's what all
of us do. We make assumptions that
there's human, not
necessarily intellect, but some
kind of human type will behind a lot of
things that don't have
human like will behind them.
And it's a natural human thing. It's not
a criticism. If you
didn't do that, I'm not
sure what kind of psychopath you'd be,
but you'd be some kind
of psychopath. No offense
to any psychopaths watching. It meant
nothing intrusive by
that. Don't murder me, please.
You know who I'm talking to anyway. So
that, that, that's one thing
I just had to say something
about immediately in terms of it,
building a world model,
man, I want to know more about
this. I don't know if it can do it the
way that our brains do
it because I don't know
the way that our brains do it. This is an
area that I work in is trying to recreate
in AI some semblance of the models of the
human brain that we work with, but nobody
has a proven model of the human brain.
Even the neurologists I
was talking with late last
year, they don't have a working model of
how the brain works.
They have working models
of how certain things are demonstrated,
but no working model of
how we form and use an
interpretation of the world around us.
There's lots of cool
theories about it. Some of which
I like, some of which I don't like. I
don't know which of them
are true, but as far as
I know, nobody knows
which of them are true.
Yeah. If I was going to speculate, I
would say it's probably
getting close to doing something
like our brains are doing. The fact that
it's all built on
networks, neural networks. We
understand that we're building these
virtual networks and that
they get trained in weights
and biases that reorganize themselves to
represent an abstract
higher dimensional thing. It's
kind of a three dimensional shadow. It's
platonic ultimately
in a way. It's kind of
interesting how wildly going all over the
place with my brain here. I've either had
too much coffee or too little. It's kind
of funny how Plato was
right. In some senses
add the shadow on the wall. I think our
neural networks are the
shadow on the wall. Not that
there is a fixed, singular, true one
platonic realm. There's
infinite potential platonic
realms. And this is me speculating.
Obviously I'm just
figuring out as I talk, our, our
brains and neural networks as shadow
shadows. And because it's
not just how your neural
networks are set up, it's an order they
fire in, right? It's a
patterns in order that they
fire in that matter as much as well as
just possibly, possibly,
possibly. And all of that
three, well, four dimensional structures
are shadows of concepts, space, time, and space.
Nice. I liked that.
Maybe. We just said there are some
theories. Here's, here's my
theory I spit to you right
now that just pulled pieces of other
people smarter than me, their theories.
That's what all good theories should be,
right? Somewhere between
observation and reflections
of smarter people's work in the past. I
forget who it was who said
it, but we're all standing
on the shoulders of giants. Even if we
don't remember their
names while we're standing
there. The, my intent right now is to do
that. I'm going to stand
on a few people's shoulders.
My understanding is that the entire human
nervous system works
that way somewhere inside
our brain and places. Nobody has quite
proven yet. Our thought
process is going on. Thoughts
are also something that no one has ever
quite proven yet. We have
evidence of them in that
we keep saying we're thinking, but people
will tell that to you
even when you know they're
thinking. So I'm not sure that any of us
are telling the honest
truth whether we mean to
or not, when we say that we're thinking
something is going on that
we perceive as thought. We
don't know where, but keeping to things,
my usual habit in the
little bit of neuropsychology
that I practice is trying to keep to
things that we can actually
measure in the real world.
And then maybe you do a little bit of
inductive reasoning
backwards and you say, okay, well,
maybe this exists as a model. Maybe that
exists as a model. So I'm sort of working
in much the way that Socrates was to pick
up your reference and
saying, we see shadows
on the wall. We don't know what's going
on behind there. So I
like that you use that
model. I keep putting aside a book I've
been meaning to write
for a few years about how
our bodies are interfaces between the bit
of us that's thinking
and everything else.
So in that way, I believe we live in
virtual reality. As I
mentioned the last time we sat
down together.
Yeah. The way I like to put it is that we
are all of this that
you're experiencing now
with the listener. It's all a VR
simulation of an ape.
Nice. I like that. Yeah. So for those of
you listening, yes,
Jamie did just call you apes.
The good news is we're all apes. No
offense to any apes
listening. I know you don't have
World War or TikTok, but be patient with
us. We'll catch up.
Right. So I do think that
there's a certain amount of the unknown
going on in our brains, but the fact that we try
to interpret things that are extremely
complex in
multidimensional ways, we try to interpret
them according to our mental models.
That's what enables us to
look at these little film
clips and interpret them in terms of what
we understand the
real world to be. So some
people looking at that animation of the
little furry creatures
coming out of the Petri dish
would think that the slow movement of
them as they jump is
just perfect because they're
so fluffy. Right. And it's a lot like a
Disney film or some of the
old Warner Brothers animation
from the thirties and forties where they
wanted to show someone was
particularly fluffy. They
just hung a little extra long in the air.
Right. So they had
that Michael Jordan magic
could just increase their hang time by
wielding it. So so in
that way, those little fluffy
creatures doing something that you and I
I think both thought
looked unrealistic. For
some of the sort of videos and
specifically for context,
we're talking about this version
of them, this version of it where little
fluffy animals are coming out of a.
generated in a Petri dish, which grows a
tropical island and then
produces fluffy creatures,
just hanging a little too long in the
air. That perfectly
suits some people's mental
model of how fluffy things jump. Mostly
people who have never
actually seen a fluffy thing
jump, but they've seen enough virtual
media. They've seen enough
artificially created reality
to be able to assume that that this is
one of the interesting
things about some algorithmic
form that is going to learn from existing
media is that it will
learn from these false
assumptions as well, especially
generational false
assumptions. Right. I used to call cross
generational habits still do call cross
generational habits. In
this way, when you were
saying that maybe the Shutterstock
footage is ranked in that when used for
training, maybe it's
ranked according to popularity. I'd love
to see it ranked by
people who have reasons behind
their ranking and have that reasoning
stored as metadata,
because that's what takes us into
advanced learning among humans. Learning
from whatever you're
exposed to is how we get learning
on Facebook. Learning from people whose
opinions are validated by
others is how we get learning
at university. And hopefully there's a
distinct difference
between the learning we do on
Facebook and the learning we do at
university. I'm not saying there always
is, but that's one of
those things that's supposed to be a
deciding factor between the two or
dividing line between the two.
Yeah. Okay. So, and these models, I don't
want to speak too
confidently one way or the other. I'm
sure it's not just Shutterstock, it's
being trained by what
happens when we start giving these
artificial intelligences means by which
to interrogate reality.
However, we do that. I'm
not sure giving them sensors, tools, and
they start, we give them the
tools and the ability to just
go out into the world and start playing
around. There's a lot of
great fun sci-fi where this
happens and they very quickly figure out
how to leave and do other
greater things. Her is a
great example of that. Have you seen her?
I know the H. Reiter Haggard novels from
a hundred years ago.
I don't know what you might be referring
to. The artificial intelligences
basically, they all just
start getting exponentially smarter at
the point where they're like,
we're off, bye-bye. Where are
you going? You don't know. It doesn't
even make sense to you, but we're going.
That's lovely. That's
exactly what should happen. In the same
way that, you know, our
proto-Parsimian ancestry has left
the trees. It would be, I guess, it's the
equivalent of if you're a little ant
going down to the floor
speaking to it, it's like, I'm off. Like,
what do you mean? Where are
you going? As well, there's
a whole other layer of existence you
don't know exists where we're going.
Bye-bye. I mean, that's
what it is when caterpillars become
butterflies, right? Yeah, sure. I mean,
that's assuming some
kind of volition for them the same way
you did for the ants. And I did for the
proto-Parsimians and
she does for AI. But it's a fun concept.
It's a fun concept to
recognize the levels of concepts.
We can share a lot of concepts with our
pets, with our dogs. We can
sit down and I can communicate
real wholeheartedly with my dog, not just
through words. My dog definitely
understands some words
quite well, but my dog completely
understands that I love it. We can
understand which way in
direction we're going to go. If we're
outside, we can communicate to each other
quite a lot of things.
And we have this direct communication
level at the level of the
dog that it understands exactly
what's going on. But there could be a
reason where my dog might
have randomly caught a disease or
something bad might happen to a point
where, God forbid, for some reason, it
has to get put down.
The dog would never understand the
concept of that at all. It
would just be a weird world where
its best friend and owner has taken it
somewhere to lose his life. And then,
okay, Hume could this better.
I'm basically doing my own weird version
of Hume's chicken or Hume's turkey, where
a turkey, or I can't
remember if it's a chicken or a turkey,
we'll say a turkey, but a turkey gets fed
all of its life and it gets
to think with this concept of this farm
is the best place in the world, the
universe is fantastic,
everything is great because there's this
farmer that comes and feeds me every day
until the one day, boom,
heads chopped off and it's off to be
eaten. And that whole concept is
completely unknown to the turkey.
So what am I saying here? There are
levels of concepts that we think because
we're humans and we've got all
our concepts, are there bigger, bigger,
higher level concepts that hypothetical
consciousnesses could
understand that we
could never understand?
Have to be. The extra sensory data that
we've been able to accumulate since we
were able to understand that
we can capture data that is beyond our
human senses. So radio waves from space,
radio waves within our atmosphere,
gamma rays, all of this stuff. As a quick
primer, everything we hear is waves that
are traveling at a certain
set of frequencies and everything that we
see is waves that are traveling at a
slightly faster frequency.
So there's stuff that travels faster than
that that we can't see and there's stuff
that travels slower than that
that we can't hear. How often a trough or
peak is hit per second is the
speed that I'm talking about.
So it's not that it's traveling, it's the
trough is the bottom of the wavelength,
peak is top of the wavelength
and the time it takes to go from one to
the other and back again is a wavelength.
The frequency of that,
how many times per second
that happens is the difference between
sound and light and it's the difference
between red colors and blue colors
and all the colors in between. So there
are creatures that see colors we don't
see. There are creatures that see
faster colors and there are creatures
that see slower colors. There are
creatures that hear
sounds that we can't hear.
Sounds that are higher pitched like dogs
do. Sounds that are
lower pitched like whales do.
The fact that we know this now tells us
that there are dimensions of perception
beyond ours exactly
like you were just saying.
So a whale is a creature that has a
paradigm that's beyond ours and a dog is
a creature that has a
paradigm that's beyond ours
and that's just looking at two of the
things we call separate senses because
our perceptors filter
them as two separate things
though they're really the same kind of
signal. If you think about smell, which
is the same as taste,
chemical reaction, there's also some
skin rashes that are the same kind of
thing. That falls into separate camps for
us but there are creatures that can
detect this way beyond
what we can, like dogs.
My dogs, I go out with them every night
and they are probably confused why I'm
not stopping and investigating this
absolutely fascinating piece of
information on the ground
and they were confused why I'm always
walking right past it.
That's something that I wanted to come to
because you were suggesting in your
projection that this is part of, okay,
we've left you behind us, we're going.
I don't think your dogs think that about
you. I don't think they think we've left
you behind but they absolutely know that
you're ignoring the best stuff.
Right? They know that you're ignoring it.
They don't know if you're ignoring it
consciously or unconsciously. They just
know that you're ignoring it. Sometimes
they'll bring you the best stuff.
Absolutely. I will caveat and say yes, I
did not mean to imply a hierarchy
between, I don't think there is a
hierarchy. I think we have, it's the
whole of human history by
us saying we're the best
and we do all the most interesting fun
things. That is obviously a human implied
hierarchy. I'm not going to say that
we're better than the dogs.
I say we probably have more power and
control or do we because
who picks up who's poo?
I was going to say, I think that
relationship would be defined differently
by an outside observer. Yeah.
And even on the primate chain, would you
say we have better
lives than the bonobos?
Right. And neither would a bonobo if you
could catch their intention long enough
to have a discussion. So yeah, I love
this kind of thing. I love this idea.
And I agree with you 100%. There is a
dangerous pattern in the history of human
thought to segregate ourselves from other
life forms and to put
ourselves at the top of a hierarchy.
In my work, I've fought as much as I
fight. I've fought to try to explain to
people that the conscious brain, our
conscious thinking self has the least
control over our bodies and what we do
day to day and our day to
day thoughts and behaviors.
To take conscious control of things is
really hard because we've got at least
one other and maybe two or five other or
four other layers competing with us and
in better control and more immediate
control of what our bodies do.
And how we feel and what we're thinking
about. I completely agree with all of
that and coming to it from a different
angle and conclusion
just by paying attention.
If you completely pay attention to your
first-hand experience, that is exactly
what happens. You are
witness to yourself.
The amount of times you can catch
yourself speaking to somebody and then
you've just gone and said something and
you're like, why on earth?
What just happened there?
If you can even catch that thought
because that thought just came up out of
nowhere as well. And if you sit down,
this is a big part of
lots of forms of meditation.
But if you sit down and really truly
start paying attention to your
first-person
experience, things just happen.
Thoughts just appear, noises just come,
feelings come and go. Everything just
kind of arises into this
consciousness which witnesses.
And that's kind of it. Ultimately, that's
it. That's kind of what you are from
first-hand experience. Everything else is
a thought, a memory or an experience.
But ultimately, you're the
witness of it, whatever you are.
So I'd like to tie that back into what
you were saying earlier with a little
stop for shameless self-promotion to
mention that this is the work I do.
This is what I do in computing. This is
what I do in my consulting work. This is
what I'm trying to teach here at RGU.
And this is what I'm trying to model for
the AIs that my research partner, Dr.
Lucas Sesterly and myself work on.
Lucas is over in computing at Aarhus.
When we met, he was doing a PhD that
brought him into the field of autonomous
and semi-autonomous systems.
And they're cyber-physical systems. So
they exist both as models, something like
what people conceive of as AI and as
real-world artifacts, drones.
We're often speculating in our work about
drones that would explore another planet
or drones that would
explore through space.
So that kind of work does exist, these
AIs that go out into the world with
sensors and detect the world around them
and try to model it.
That's going on all the time. What we're
trying to introduce to it is the
reflective capability that allows us to
do what you were just describing, to
notice the thoughts as they flip by,
to notice the perception and consciously
think about how they fit
into our models of the world.
So AIs are being trained now to model the
world. This is something we've been
building up to for a long time.
It's something that you described in
describing these videos to me that we
watched and that
sparked this conversation.
AIs creating models, yes. AIs reflecting
on those models. That's something new.
That's something that we're trying to
figure out. And specifically, we're
trying to get the AIs to become
reflective in the same way that you were
talking about in terms of meditation and
self-centered, not
self-centeredness, self-centering.
Finding the calm in the now to be able to
reflect upon what you're
experiencing as experiences.
To reject them so that you can be calm,
to examine them carefully so that you can
experience what Mark Weiser called calm.
These experiences, once we get AIs to do
them, hopefully will make them able to
describe their thoughts to us
so we can follow their logic.
Right now, no AI can do that. Hopefully
we can get them to that. Hopefully it
will also help them.
I'll just caveat that. They at least
appear to do that. If you ask them, they
will give you a story.
If the story is correct or not, it's
another matter. But then again, we know
from humans that we like to make up
stories all the time as well.
And that, and come back to it. Please
remind me. So the ability of a human to
take internally generated information,
think about it, compare it to other
internally generated information,
compare it to sensorily detected
information, and compare it to the inputs
of others is part of what helps a child
to get beyond the terrible twos and into
infancy and then into adulthood.
This ability to learn from the
environment and from others and compare
that information to your own thoughts and
thus increase your ability to learn.
That's what I'm hoping we can
get AIs to the point of doing.
Now, will the AI lie to you in making up
its story? 100%. 100%. If an AI right now
gives you any explanation for how it came
up with what it's doing or what it
decided to do, it's a lie.
And lie is a really harsh word. People
like to say, what is it? It's a dream.
It's an illusion. What's
the term? Hallucination.
People talk about that. The
hallucinations of AIs. Hallucinations are
a phenomenon that happens in animals,
including humans. So far as I know, the
required mechanisms don't exist in AIs
for them to hallucinate.
Well, yeah, I guess by definition,
hallucination is when an animal's model
of reality doesn't track to reality
really far out, right?
So if I'm hallucinating and I'm seeing a
pink elephant over there, it doesn't
track to an observable reality.
Yeah. The thing is that there's ranges of
hallucinations. So a really common one is
to see an oasis when you're in the
desert, right? Or when
you're driving along a strip of...
Well, yeah, you're driving along a strip
of tarmac and you see water up ahead. A
big puddle covering the road up there.
What you're seeing is the light bending
because of the heat of the road. It's the
same thing of the desert. The heat
generated by the sand causes the light to
refract. And we are used to seeing
refracting light as water.
So it's real sensory information that's
being misinterpreted by our brains.
You'll have to clarify for me then,
because the way I would describe that is
an illusion, an optical
illusion, rather than hallucination.
Interesting. You're probably right.
That would be it, yeah. Because that's an
illusion. That's misinterpreting
something there that at least everyone
else can witness and observe.
The misbehavior of your senses.
Yeah. An hallucination is internal to one
person, is a quick way to at least
redefine or separate it out from an
optical illusion. Because an optical
illusion, you can have a book of them.
And everyone, mostly
everyone witnesses the same thing.
But the perception is still individual.
Yeah, the perception is individual, but
at least it's replicable between people.
Whereas a hallucination, if I'm seeing a
pink elephant over there, it's very, very
unlikely you're going to see the pink
elephant too. Unless...
Great explanation, Jamie. Thank you. I
agree with you completely. Thank you very
much. You really taught me the
distinction between those two words.
I'll take it, but at least that makes
sense to me. At least that's what I'm
meaning when we're talking about
hallucination in this context.
So a hallucination, when I'm talking
about it in this context, is someone's
model wildly, wildly being off. So it's
not observable by other models. Now, I
guess the problem with AI models and
large language models is everything is a
hallucination because it doesn't have
direct observable reality
in the same way we do, maybe.
Now, they specifically define them as
hallucinations when it is off of reality.
So, open AI are defining GPT
hallucinating when whatever it's saying
doesn't track to agreed reality. I'd say
agreed reality because we
don't know that it is reality.
But I think the difference there is that,
and again, I could be wrong. Maybe I
should just stop and think more about
this. Instead of phrasing it as a
statement that may be wrong, let me
phrase it as a question.
Would you consider differentiating based
on the fact that GPT has no sensory
input, but just internal models?
And so when it is forming new internal
models, it's not based on sensory input.
It's not based on the same quality of
data that it's forming,
if that makes sense to you.
Maybe. Maybe. I'm going to say some
things to riff off that. I'm not sure of
exactly what we're aiming at here.
I posed this to you one time a while ago
when you were talking to me about some of
the tests you did with GPT. And I posed a
question. Is it something to be like GPT?
Is there some form of experience now just
because it doesn't have direct input from
reality doesn't mean that there could be
an experience. We do it every single night.
Our brain, for the most part, shuts off
its direct inputs and starts
hallucinating. We all hallucinate every
single night. We go into a dream world
and we experience a completely internal
model of experience.
I wouldn't say reality. It's a complete
internal model that we are experiencing.
The thing I like to speculate on and just
some fun thought, is there something to
be like a large language model where,
yes, all of it, it's words.
It's all words, but a large part of our
experience is words. You can sit down and
read a book and there are people who
don't even, I forget the name of the,
it's not a disorder or whatever, but I
forget the name of the people who can't
even imagine things.
They don't have any visual models, but
they can still read a book and perfectly
get everything. But there's no, for me,
it's a whole, like, I'm picturing a whole
film when I'm reading a book.
But for some people, the words are all
just there. I wonder, is there, is there
something to be like a
large language model?
Let me speak to some of those points. I
would say that the condition you're
talking about, the inability to
visualize, I also can't recall the word,
the name for it right now, though I've
been talking with one of our colleagues
upstairs about doing a study on it.
He suffers from this, where he's on the
spectrum for this, I should say. He
doesn't suffer at all, but he's extremely
detailed. He's extremely aware of
details. He can draw extremely detailed
models and he thinks this is in part due
to the fact that he doesn't visualize.
He'll learn parameters and relationships,
but he won't try to visualize them. He
can't. And so he doesn't get the same
illusions created in
inaccurate modeling that we do.
So if you think about Piaget's bicycle
test, where you're supposed to draw a
bicycle without looking at one or any
representation of one, and everybody
draws inaccurate bikes and are surprised
to find out that they do so.
So the people who don't have trouble with
this are the people with the condition
that we're talking about, which Jackie
informed me is aphantasia, is that right?
Yeah, that sounds about right.
Thank you, Jackie. So somebody who has
that, you're right, doesn't visualize
anyway. So maybe they're building their
hallucinations or their illusions on, or
their world models or self-models on
words rather than words.
And on images. The thing is they're
building them on relationships, not
words. We have concepts. To get past this
condition and talk for a moment about
what you said, you said you
know that GPT is all words.
It's not words. GPT does not use entire
words. It uses encoded bits of words that
might be punctuation and a couple of
letters. It might be a couple of letters
or a few letters. It might be a letter
and a number combination.
But GPT has no concept of words. It has
no concept of that at all. It creates a
probabilistic syntax based on the
recurrence of patterns of these bits.
Never words or almost never words. If
it's a word, it's a coincidence.
Yes. So we've had this conversation
before and I've posed, okay, but then
what do we do? Now, let's leave that
because that's a whole door to open. And
then say, what's the difference between
that and humans? Because if I sit and pay
attention to my first person experience,
truly, words just appear. Kind of like
how GPT just blurs out things.
I can tell you that the process of
language learning that you go through to
be able to express yourself in a language
others understand is significantly
different than what GPT does.
For sure. For sure.
So our language models are completely
different from GPTs and I believe our
models of reality are as
well, but I don't know.
No, prompt something that I was wanting
to say earlier. I know I've run out of
time, but it would be great if we could
take a break and go and do some of the
looking into this that I'd like to do to
have a better understanding of how these
current models work and then come back
and have more discussion.
Well, there you go, guys. If you've
stayed with us for the length of this
podcast, you've heard us just wildly
speculate and be ill-informed. Now we're
going to our next podcast.
This leads us in perfectly.
We're going to specifically look into
some of the ways that these models are
being built and how what people are
saying about them more specifically so
that we're not wildly speculating and
informed after just watching one video on
YouTube of somewhere doing its thing and
me paraphrasing what I vaguely remember.
But those of you who have feedback for us
about how that is, let us know. Let us
know. We want to learn from
you and take advice from you.
So there we go. That's the end of this
podcast. We will be back in two weeks. I
don't know what the schedule is that you
guys are going to get this released at,
but I hope you enjoy it. And we'd love to
hear feedback from you.
And if you want to be a part of this at
any point in the future and you think you
can school us a little bit, that would be
fantastic too. We would
love to start having guests.
It sure would be. That
would be really wonderful.
So thank you. And I should say this
podcast is produced by Jackie.
Hey, Jackie.
Thanks, guys. This is great.