Ben Goertzel: Artificial General Intelligence
Guests: Ben Goertzel
Full transcript · Read the full conversation in order, with its original context.
Contents28 chapters
§01Transcript · 00:00–09:21§02Transcript · 09:23–18:51§03Transcript · 18:55–27:50§04Transcript · 27:54–37:34§05Transcript · 37:37–47:52§06Transcript · 47:54–56:27§07Transcript · 56:29–1:05:44§08Transcript · 1:05:49–1:15:29§09Transcript · 1:15:34–1:24:29§10Transcript · 1:24:32–1:33:06§11Transcript · 1:33:11–1:42:27§12Transcript · 1:42:29–1:51:09§13Transcript · 1:51:11–2:00:15§14Transcript · 2:00:16–2:09:05§15Transcript · 2:09:08–2:17:46§16Transcript · 2:17:51–2:27:10§17Transcript · 2:27:12–2:35:51§18Transcript · 2:35:54–2:44:53§19Transcript · 2:44:55–2:53:37§20Transcript · 2:53:40–3:03:13§21Transcript · 3:03:14–3:13:00§22Transcript · 3:13:06–3:22:16§23Transcript · 3:22:18–3:31:38§24Transcript · 3:31:38–3:40:49§25Transcript · 3:40:51–3:49:44§26Transcript · 3:49:48–3:59:12§27Transcript · 3:59:14–4:08:48§28Transcript · 4:08:50§01Transcript · 00:00–09:21
The following is a conversation with Ben Goertzel,
one of the most interesting minds
in the artificial intelligence community.
He's the founder of SingularityNet,
designer of OpenCog AI framework,
formerly a director of research
at the Machine Intelligence Research Institute,
and chief scientist of Hanson Robotics,
the company that created the Sophia robot.
He has been a central figure in the AGI community
for many years, including in his organizing
and contributing to the conference
and artificial general intelligence,
the 2020 version of which is actually happening this week,
Wednesday, Thursday, and Friday.
It's virtual and free.
I encourage you to check out the talks,
including by Yosha Bach from episode 101 of this podcast.
Quick summary of the ads.
Two sponsors, the Jordan Harbinger Show and Masterclass.
Please consider supporting this podcast
by going to jordanharbinger.com/lex
and signing up at masterclass.com/lex.
Click the links, buy all the stuff.
It's the best way to support this podcast
and the journey I'm on in my research and startup.
This is the Artificial Intelligence Podcast.
If you enjoy it, subscribe on YouTube,
review it with five stars on Apple Podcast,
support it on Patreon, or connect with me on Twitter,
@lexfriedman, spelled without the E, just F-R-I-D-M-A-N.
As usual, I'll do a few minutes of ads now
and never any ads in the middle
that can break the flow of the conversation.
This episode is supported by the Jordan Harbinger Show.
Go to jordanharbinger.com/lex.
It's how he knows I sent you.
On that page, there's links to subscribe to it
on Apple Podcast, Spotify, and everywhere else.
I've been binging on his podcast.
Jordan is great.
He gets the best out of his guests,
dives deep, calls them out when it's needed,
and makes the whole thing fun to listen to.
He's interviewed Kobe Bryant, Mark Cuban,
Neil deGrasse Tyson, Garry Kasparov, and many more.
His conversation with Kobe is a reminder
how much focus and hard work is acquired for greatness
in sport, business, and life.
I highly recommend the episode if you want to be inspired.
Again, go to jordanharbinger.com/lex.
It's how Jordan knows I sent you.
This show is sponsored by Masterclass.
Sign up at masterclass.com/lex to get a discount
and to support this podcast.
When I first heard about Masterclass,
I thought it was too good to be true.
For 180 bucks a year, you get an all-access pass
to watch courses from, to list some of my favorites.
Chris Hadfield on space exploration,
Neil deGrasse Tyson on scientific thinking
and communication, Will Wright, creator of
the greatest city-building game ever,
SimCity and Sims on game design,
Carlos Santana on guitar, Garry Kasparov,
the greatest chess player ever on chess,
Daniel Negreanu on poker, and many more.
Chris Hadfield explaining how rockets work
and the experience of being launched into space alone
is worth the money.
Once again, sign up at masterclass.com/lex
to get a discount and to support this podcast.
And now, here's my conversation with Ben Goertzel.
What books, authors, ideas had a lot of impact on you
in your life in the early days?
- You know, what got me into AI and science fiction
and such in the first place wasn't a book,
but the original Star Trek TV show,
which my dad watched with me, like, in its first run.
It would have been 1968, '69 or something.
And that was incredible 'cause every show
that visited a different alien civilization
with different culture and weird mechanisms.
But that got me into science fiction,
and there wasn't that much science fiction
to watch on TV at that stage.
So that got me into reading the whole literature
of science fiction, you know,
from the beginning of the previous century until that time.
And I mean, there was so many science fiction writers
who were inspirational to me.
I'd say if I had to pick two,
it would have been Stanislaw Lem, the Polish writer.
- Yeah. - Yeah.
Solaris, and then he had a bunch of more obscure writings
on superhuman AIs that were engineered.
Solaris was sort of a superhuman,
naturally occurring intelligence.
Then Philip K. Dick, who, you know,
ultimately my fandom for Philip K. Dick
is one of the things that brought me together
with David Hanson, my collaborator on robotics projects.
So, you know, Stanislaw Lem was very much an intellectual,
right, so he had a very broad view of intelligence
going beyond the human and into what I would call,
you know, open-ended superintelligence.
The Solaris superintelligent ocean was intelligent
in some ways more generally intelligent than people,
but in a complex and confusing way
so that human beings could never quite connect to it,
but it was still palpably very, very smart.
And then the Golem 4 supercomputer,
in one of Lem's books, this was engineered by people,
but eventually it became very intelligent
in a different direction than humans
and decided that humans were kind of trivial
and not that interesting.
So it put some impenetrable shield around itself,
shut itself off from humanity,
and then issued some philosophical screed
about the pathetic and hopeless nature of humanity
and all human thought, and then disappeared.
Now, Philip K. Dick, he was a bit different.
He was human-focused, right?
His main thing was, you know, human compassion
and the human heart and soul are going to be the constant
that will keep us going through whatever aliens we discover
or telepathy machines or super AIs or whatever it might be.
So he didn't believe in reality,
like the reality that we see may be a simulation
or a dream or something else we can't even comprehend,
but he believed in love and compassion
as something persistent
through the various simulated realities.
So those two science fiction writers
had a huge impact on me.
Then a little older than that,
I got into Dostoevsky and Friedrich Nietzsche
and Rimbaud and a bunch of more literary type writing.
- Can we talk about some of those things?
So on the Solaris side, Stanislav Lem,
this kind of idea of there being intelligences out there
that are different than our own,
do you think there are intelligences maybe all around us
that we're not able to even detect?
So this kind of idea of,
maybe you can comment also on Stephen Wolfram,
thinking that there's computations all around us
and we're just not smart enough
to kind of detect their intelligence
or appreciate their intelligence.
- Yeah, so my friend Hugo de Garis,
who I've been talking to about these things
for many decades, since the early '90s,
he had an idea he called SIPI,
the Search for Intra Particulate Intelligence.
So the concept there was,
as AIs get smarter and smarter and smarter,
assuming the laws of physics as we know them now
are still what these super intelligences
perceived to hold and are bound by,
as they get smarter and smarter,
they're gonna shrink themselves littler and littler
because special relativity makes it
so they can communicate between two spatially distant points.
So they're gonna get smaller and smaller,
but then ultimately, what does that mean?
The minds of the super, super, super intelligences,
they're gonna be packed into the interaction
of elementary particles or quarks
or the partons inside quarks or whatever it is.
So what we perceive as random fluctuations
on the quantum or subquantum level
may actually be the thoughts
of the micro, micro, micro miniaturized super intelligences
'cause there's no way we can tell random from structured
but with an algorithmic information
more complex than our brains, right?
We can't tell the difference.
So what we think is random could be the thought processes
of some really tiny super minds,
and if so, there's not a damn thing we can do about it
except try to upgrade our intelligences
and expand our minds so that we can perceive
more of what's around us.
- But if those random fluctuations,
even if we go to quantum mechanics,
if that's actually super intelligent systems,
aren't we then part of the soup of super intelligence?
Aren't we just like a finger of the entirety of the body
of the super intelligent system?
- We could be.
I mean, a finger is a strange metaphor.
I mean, we--
- Well, a finger is dumb is what I mean.
- But a finger is also useful
and is controlled with intent by the brain,
whereas we may be much less than that, right?
I mean, yeah, we may be just some random epiphenomenon
that they don't care about too much.
§02Transcript · 09:23–18:51
Like think about the shape of the crowd
emanating from a sports stadium or something, right?
There's some emergent shape to the crowd.
It's there.
You could take a picture of it.
It's kind of cool.
It's irrelevant to the main point of the sports event
or where the people are going
or what's on the minds of the people
making that shape in the crowd, right?
So we may just be some semi-arbitrary higher level pattern
popping out of a lower level
hyper intelligent self-organization.
And I mean, so be it, right?
I mean, that's one thing that--
- Still a fun ride.
- Yeah, I mean, the older I've gotten,
the more respect I've achieved
for our fundamental ignorance.
I mean, mine and everybody else's.
I mean, I look at my two dogs,
two beautiful little toy poodles,
and they watch me sitting at the computer typing.
They just think I'm sitting there
wiggling my fingers to exercise and maybe,
or guarding the monitor on the desk
that they have no idea that I'm communicating
with other people halfway around the world,
let alone creating complex algorithms
running in RAM on some computer server
in St. Petersburg or something, right?
Although they're right there in the room with me.
So what things are there right around us
that we're just too stupid or close-minded to comprehend?
Probably quite a lot.
- Your very poodle could also be communicating
across multiple dimensions with other beings,
and you're too unintelligent to understand
the kind of communication mechanism they're going through.
- There have been various TV shows
and science fiction novels, Puzzling Cats, Dolphins,
Mice and Whatnot are actually super intelligences.
Here to observe that.
I would guess, as one or the other quantum physics founders
said, those theories are not crazy enough to be true.
The reality's probably crazier than that.
- Beautifully put.
So on the human side, with Philip K. Dick and in general,
where do you fall on this idea that love
and just the basic spirit of human nature
persists throughout these multiple realities?
Are you on the side, like the thing that inspires you
about artificial intelligence,
is it the human side of somehow persisting
through all of the different systems we engineer,
or is AI inspire you to create something
that's greater than human, that's beyond human,
that's almost non-human?
- I would say my motivation to create AGI
comes from both of those directions, actually.
So when I first became passionate about AGI,
when I was, it would have been two or three years old
after watching robots on Star Trek,
I mean, then it was really a combination
of intellectual curiosity, like can a machine really think,
how would you do that?
And yeah, just ambition to create something much better
than all the clearly limited
and fundamentally defective humans I saw around me.
Then as I got older and got more enmeshed
in the human world and got married, had children,
so my parents begin to age, I started to realize,
well, not only will AGI let you go far beyond
the limitations of the human,
but it could also stop us from dying and suffering
and feeling pain and tormenting ourselves mentally.
So you can see AGI has amazing capability to do good
for humans, as humans, alongside with its capability
to go far, far beyond the human level.
So I mean, both aspects are there,
which makes it even more exciting and important.
- So you mentioned the Sievskii Nietzsche.
Where did you pick up from those guys?
I mean-- (laughs)
- That would probably go beyond the scope
of a brief interview, certainly.
But both of those are amazing thinkers
who one will necessarily have a complex relationship with.
So, I mean, Dostoevsky, on the minus side,
he's kind of a religious fanatic,
and he sort of helped squash the Russian nihilist movement,
which was very interesting,
'cause what nihilism meant originally
in that period of the mid-late 1800s in Russia
was not taking anything fully 100% for granted.
It was really more like what we'd call Bayesianism now,
where you don't want to adopt anything
as a dogmatic certitude and always leave your mind open.
And how Dostoevsky parodied nihilism was a bit different.
He parodied it as people who believe absolutely nothing,
so they must assign an equal probability weight
to every proposition, which doesn't really work.
So on the one hand, I didn't really agree with Dostoevsky
on his sort of religious point of view.
On the other hand, if you look at his understanding
of human nature and sort of the human mind
and heart and soul, it's really unparalleled.
He had an amazing view of how human beings
construct a world for themselves
based on their own understanding
and their own mental predisposition.
And I think if you look in "The Brothers Karamazov"
in particular, the Russian literary theorist
Mikhail Bakhtin wrote about this
as a polyphonic mode of fiction,
which means it's not third person,
but it's not first person from any one person really.
There are many different characters in the novel
and each of them is sort of telling part of the story
from their own point of view.
So the reality of the whole story is an intersection
like synergetically of the many different
characters' world views.
And that really, it's a beautiful metaphor
and even a reflection, I think,
of how all of us socially create our reality.
Like each of us sees the world in a certain way.
Each of us, in a sense, is making the world as we see it
based on our own minds and understanding,
but it's polyphony, like in music
where multiple instruments are coming together
to create the sound.
The ultimate reality that's created
comes out of each of our subjective understandings
intersecting with each other.
And that was one of the many beautiful things in Dostoyevsky.
- So maybe a little bit to mention,
you have a connection to Russia and the Soviet culture.
I mean, I'm not sure exactly what the nature
of the connection is, but at least the spirit
of your thinking is in there.
- Well, my ancestry is three quarters
Eastern European Jewish.
So, I mean, three of my great-grandparents
emigrated to New York from Lithuania
and sort of border regions of Poland,
which were in and out of Poland
around the time of World War I.
And they were socialists and communists
as well as Jews, mostly Menshevik, not Bolshevik.
And they sort of, they fled at just the right time
to the US for their own personal reasons.
And then almost all, or maybe all of my extended family
that remained in Eastern Europe was killed
either by Hitler's or Stalin's minions at some point.
So the branch of the family that emigrated to the US
was pretty much the only one that survived.
- So how much of the spirit of the people
is in your blood still?
Like, when you look in the mirror, do you see,
what do you see?
- Meat.
I see a bag of meat that I want to transcend
by uploading into some sort of superior reality.
(laughing)
I mean, yeah, very clearly.
- Well put.
- I mean, I'm not religious in a traditional sense,
but clearly the Eastern European Jewish tradition
was what I was raised in.
I mean, there was, my grandfather, Leo Zwell,
was a physical chemist who worked with Landis Pauling
and a bunch of the other early greats in quantum mechanics.
I mean, he was into X-ray diffraction.
He was on the material science side,
experimentalist rather than a theorist.
His sister was also a physicist.
And my father's father, Victor Goertzel,
was a PhD in psychology who had the unenviable job
of giving psychotherapy to the Japanese
in internment camps in the US in World War II,
like to counsel them why they shouldn't kill themselves,
even though they'd had all their stuff taken away
and been imprisoned for no good reason.
So I mean, yeah, there's a lot of Eastern European
Jewish tradition in my background.
One of my great uncles was, I guess,
conductor of San Francisco Orchestra.
So there's a lot of Mickey Salkind,
bunch of music in there also.
And clearly this culture was all about learning
and understanding the world
and also not quite taking yourself
too seriously while you do it, right?
There's a lot of Yiddish humor in there.
So I do appreciate that culture,
although the whole idea that the Jews
are the chosen people of God
never resonated with me too much.
- The graph of the Goertzel family,
§03Transcript · 18:55–27:50
I mean, just the people I've encountered
just doing some research and just knowing your work
through the decades, it's kind of fascinating.
Just the number of PhDs.
- Yeah, yeah.
I mean, my dad is a sociology professor
who recently retired from Rutgers University.
But clearly that gave me a head start in life.
I mean, my grandfather gave me all those
quantum mechanics books when I was like
seven or eight years old.
I remember going through them
and it was all the old quantum mechanics,
like Rutherford atoms and stuff.
So I got to the part of wave functions,
which I didn't understand,
although I was a very bright kid.
And I realized he didn't quite understand it either,
but at least, like he pointed me to some professor
he knew at UPenn nearby who understood these things.
So that's an unusual opportunity for a kid to have.
My dad, he was programming Fortran
when I was 10 or 11 years old
on like HP 3000 mainframes at Rutgers University.
So I got to do linear regression in Fortran
on punch cards when I was in middle school,
'cause he was doing, I guess,
analysis of demographic and sociology data.
So yes, certainly that gave me a head start
and a push towards science
beyond what would have been the case
with many, many different situations.
- When did you first fall in love with AI?
Is it the programming side of Fortran?
Is it maybe the sociology, psychology
that you picked up from your dad?
- I fell in love with AI
when I was probably three years old
when I saw a robot on Star Trek.
It was turning around in a circle going,
error, error, error, error,
because Spock and Kirk had tricked it
into a mechanical breakdown
by presenting it with a logical paradox.
And I was just like, well, this makes no sense.
This AI is very, very smart.
It's been traveling all around the universe,
but these people could trick it
with a simple logical paradox.
Like, if the human brain can get beyond that paradox,
why can't this AI?
So I felt the screenwriters of Star Trek
had misunderstood the nature of intelligence.
And I complained to my dad about it,
and he wasn't gonna say anything one way or the other.
But before I was born,
when my dad was at Antioch College
in the middle of the US,
he led a protest movement called SLAM,
Student League Against Mortality.
They were protesting against death
wandering across the campus.
So he was into some futuristic things, even back then,
but whether AI could confront logical paradoxes or not,
he didn't know.
But 10 years after that or something,
I discovered Douglas Hofstadter's book, "Gordal as Shabak."
And that was sort of to the same point of AI
and paradox and logic, right?
'Cause he was over and over
with Gordal's incompleteness theorem.
And can an AI really fully model itself reflexively,
or does that lead you into some paradox?
Can the human mind truly model itself reflexively,
or does that lead you into some paradox?
So I think that book, "Gordal as Shabak,"
which I think I read when it first came out.
I would've been 12 years old or something.
I remember it was like 16-hour day.
I read it cover to cover, and then re-read it.
I re-read it after that,
'cause there was a lot of weird things
with little formal systems in there
that were hard for me at the time.
But that was the first book I read
that gave me a feeling for AI
as like a practical academic or engineering discipline
that people were working in.
'Cause before I read "Gordal as Shabak,"
I was into AI from the point of view
of a science fiction fan.
And I had the idea, well, it may be a long time
before we can achieve immortality in superhuman AGI.
So I should figure out how to build a spacecraft
traveling close to the speed of light,
go far away, then come back to the Earth
in a million years when technology is more advanced
and we can build these things.
Reading "Gordal as Shabak,"
while it didn't all ring true to me, a lot of it did,
but I could see there are smart people right now
at various universities around me
who are actually trying to work on building
what I would now call AGI,
although Hofstadter didn't call it that.
So really, it was when I read that book,
which would have been probably middle school,
that then I started to think,
well, this is something that I could practically work on.
- Yeah, as opposed to flying away and waiting it out,
you can actually be one of the people
that actually builds the system.
- Yeah, exactly, and if you think about,
I mean, I was interested in what we'd now call
nanotechnology and in the human immortality and time travel,
all the same cool things as every other
science fiction-loving kid,
but AI seemed like if Hofstadter was right,
you just figure out the right program,
sit there and type it.
Like, you don't need to spin stars into weird configurations
or get government approval to cut people up
and fill it with their DNA or something, right?
It's just programming, and then, of course,
that can achieve anything else.
There's another book from back then,
which was by Feinbaum, Gerald Feinbaum,
who was a physicist at Princeton,
and that was the Prometheus Project,
and this book was written in the late 1960s,
so I encountered it in the mid '70s,
but what this book said is in the next few decades,
humanity is gonna create superhuman-thinking machines,
molecular nanotechnology, and human immortality,
and then the challenge we'll have is what to do with it.
Do we use it to expand human consciousness
in a positive direction,
or do we use it just to further vapid consumerism?
And what he proposed was that the UN
should do a survey on this,
and the UN should send people out to every little village
in remotest Africa or South America
and explain to everyone what technology
was gonna bring the next few decades
and the choice that we had about how to use it,
and let everyone on the whole planet vote
about whether we should develop
super AI nanotechnology and immortality
for expanded consciousness or for rampant consumerism.
And needless to say, that didn't quite happen,
and I think this guy died in the mid '80s,
so he didn't even see his ideas
start to become more mainstream,
but it's interesting, many of the themes
I'm engaged with now, from AGI and immortality,
even to trying to democratize technology,
as I've been pushing forward with Singularity
and my work in the blockchain world,
many of these themes were there in Feinbaum's book
in the late '60s even.
And of course, Valentin Turchin, a Russian writer,
and a great Russian physicist who I got to know
when we both lived in New York in the late '90s
and early aughts, he had a book in the late '60s in Russia
which was The Phenomenon of Science,
which laid out all these same things as well.
And Val died in, I don't remember,
2004 or five or something of Parkinson's-ism.
So yeah, it's easy for people to lose track now
of the fact that the futurist
and the singularitarian advanced technology ideas
that are now almost mainstream
and are on TV all the time, I mean,
these are not that new, right?
They're sort of new in the history of the human species,
but I mean, these were all around in fairly mature form
in the middle of the last century,
were written about quite articulately
by fairly mainstream people
who were professors at top universities.
It's just until the enabling technologies
got to a certain point, then you couldn't make it real.
So even in the '70s, I was sort of seeing that
and living through it, right?
From Star Trek to Douglas Hofstetter,
things were getting very, very practical
from the late '60s to the late '70s.
And the first computer I bought,
you could only program with hexadecimal machine code
and you had to solder it together.
And then a few years later, there's punch cards,
and a few years later, you could get Atari 400
and Commodore VIC-20, and you could type on a keyboard
and program in higher level languages
alongside the assembly language.
So these ideas have been building up a while,
and I guess my generation got to feel them build up,
which is different than people coming into the field now,
for whom these things have just been part of the ambiance
of culture for their whole career, or even their whole life.
§04Transcript · 27:54–37:34
- Well, it's fascinating to think about
there being all of these ideas kind of swimming
almost with a noise all around the world,
all the different generations,
and then some kind of nonlinear thing happens
where they percolate up
and capture the imagination of the mainstream.
And that seems to be what's happening with AI now.
- I mean, Nietzsche, who you mentioned,
had the idea of the Superman, right?
But he didn't understand enough about technology
to think you could physically engineer a Superman
by piecing together molecules in a certain way.
He was a bit vague about how the Superman would appear,
but he was quite deep in thinking about
what the state of consciousness
and the mode of cognition of a Superman would be.
He was a very astute analyst of how the human mind
constructs the illusion of a self,
how it constructs the illusion of free will,
how it constructs values like good and evil
out of its own desire to maintain
and advance its own organism.
He understood a lot about how human minds work.
Then he understood a lot
about how post-human minds would work.
I mean, the Superman was supposed to be a mind
that would basically have complete root access
to its own brain and consciousness
and be able to architect its own value system
and inspect and fine-tune all of its own biases.
So that's a lot of powerful thinking there,
which then fed in and sort of seeded
all of postmodern continental philosophy
and all sorts of things that have been very valuable
in development of culture and indirectly even of technology.
But of course, without the technology there,
it was all some quite abstract thinking.
So now we're at a time in history
when a lot of these ideas can be made real,
which is amazing and scary, right?
- It's kind of interesting to think,
what do you think Nietzsche would,
if he was born a century later,
or transported through time,
what do you think he would say about AI?
I mean-- - Well, those are
quite different.
If he's born a century later or transported through time--
- Well, he'd be on TikTok and Instagram
and he would never write the great works he's written,
so let's transport him through time.
- Maybe also Sprach Zarathustra would be a music video,
right, I mean, who knows?
- Yeah, but if he was transported through time,
do you think, that'd be interesting actually to go back,
you just made me realize that it's possible to go back
and read Nietzsche with an eye of,
is there some thinking about artificial beings?
I'm sure he had inklings, I mean, with Frankenstein,
before him, I'm sure he had inklings of artificial beings
somewhere in the text.
It'd be interesting to see, to try to read his work,
to see if he had an,
if Superman was actually an AGI system,
like if he had inklings of that kind of thinking.
- He didn't. - He didn't.
- No, I would say not.
I mean, he had a lot of inklings of modern cognitive science,
which are very interesting.
If you look in the third part of the collection
that's been titled The Will to Power,
I mean, in book three there,
there's very deep analysis of thinking processes,
but he wasn't so much of a physical tinkerer type guy,
right, it was very abstract.
- Do you think, what do you think about the will to power?
Do you think human, what do you think drives humans?
Is it--
- Oh, an unholy mix of things.
I don't think there's one pure, simple,
and elegant objective function driving humans by any means.
- What do you think, if we look at,
I know it's hard to look at humans in an aggregate,
but do you think overall humans are good?
Or do we have both good and evil within us
that depending on the circumstances,
depending on the whatever can percolate to the top?
- Good and evil are very ambiguous,
complicated, and in some ways silly concepts,
but if we could dig into your question
from a couple directions.
So I think if you look in evolution,
humanity is shaped both by individual selection
and what biologists would call group selection,
like tribe level selection, right?
So individual selection has driven us
in a selfish DNA sort of way,
so that each of us does to a certain approximation
what will help us propagate our DNA to future generations.
I mean, that's why I've got to have four kids so far,
and probably that's not the last one.
On the other hand--
- I like the ambition.
- Tribal, like group selection,
means humans in a way will do what will advocate
for the persistence of the DNA of their whole tribe,
or their social group.
And in biology, you have both of these, right?
And you can see, say, an ant colony or a beehive,
there's a lot of group selection
in the evolution of those social animals.
On the other hand, say, a big cat
or some very solitary animal,
it's a lot more biased toward individual selection.
Humans are an interesting balance,
and I think this reflects itself
in what we would view as selfishness versus altruism,
to some extent.
So we just have both of those objective functions
contributing to the makeup of our brains.
And then as Nietzsche analyzed in his own way,
and others have analyzed in different ways,
I mean, we abstract this as,
well, we have both good and evil within us, right?
'Cause a lot of what we view as evil
is really just selfishness.
A lot of what we view as good is altruism,
which means doing what's good for the tribe.
And on that level, we have both of those
just baked into us, and that's how it is.
Of course, there are psychopaths and sociopaths
and people who get gratified by the suffering of others,
and that's a different thing.
- Yeah, those are exceptions, but on the whole.
- Yeah, but I think at core, we're not purely selfish,
we're not purely altruistic.
We are a mix, and that's the nature of it.
And we also have a complex constellation of values
that are just very specific to our evolutionary history.
Like we love waterways and mountains,
and the ideal place to put a house
is in a mountain overlooking the water, right?
And we care a lot about our kids,
and we care a little less about our cousins,
and even less about our fifth cousins.
I mean, there are many particularities to human values,
which whether they're good or evil
depends on your perspective.
Say, I spent a lot of time in Ethiopia, in Addis Ababa,
where we have one of our AI development offices
for my Singularity Net project.
And when I walk through the streets in Addis,
there's people lying by the side of the road,
like just living there by the side of the road,
dying probably of curable diseases
without enough food or medicine.
And when I walk by them, I feel terrible, I give them money.
When I come back home to the developed world,
they're not on my mind that much.
I do donate some, but I mean,
I also spend some of the limited money I have
enjoying myself in frivolous ways
rather than donating it to those people
who are right now starving, dying,
and suffering on the roadside.
So does that make me evil?
I mean, it makes me somewhat selfish and somewhat altruistic
and we each balance that in our own way, right?
So whether that will be true of all possible AGIs
is a subtler question.
So that's how humans are.
- So you have a sense, you kind of mentioned
that there's a selfish,
I'm not gonna bring up the whole Ayn Rand idea
of selfishness being the core virtue.
That's a whole interesting kind of tangent
that I think we'll just distract ourselves on.
- I have to make one amusing comment.
- Sure.
- Comment that has amused me anyway.
So the, yeah, I have extraordinary negative respect
for Ayn Rand.
- Negative, what's a negative respect?
- But when I worked with a company called Genescient,
which was evolving flies to have extraordinary long lives
in Southern California.
So we had flies that were evolved by artificial selection
to have five times the lifespan of normal fruit flies.
But the population of super long-lived flies
was physically sitting in a spare room
at an Ayn Rand elementary school in Southern California.
So that was just like, well, if I saw this in a movie,
I wouldn't believe it.
- Well, yeah, the universe has a sense of humor
in that kind of way.
That fits in, humor fits in somehow
into this whole absurd existence.
But you mentioned the balance between selfishness
and altruism as kind of being innate.
§05Transcript · 37:37–47:52
Do you think it's possible that's kind of an emergent
phenomena, those peculiarities of our value system,
how much of it is innate?
How much of it is something we collectively,
kind of like a Dostoevsky novel,
bring to life together as a civilization?
- I mean, the answer to nature versus nurture
is usually both.
And of course, it's nature versus nurture
versus self-organization, as you mentioned.
So clearly, there are evolutionary roots
to individual and group selection
leading to a mix of selfishness and altruism.
On the other hand, different cultures manifest that
in different ways.
Well, we all have basically the same biology.
And if you look at sort of pre-civilized cultures,
you have tribes like the Yanomamo in Venezuela,
which their culture is focused on killing other tribes.
And you have other Stone Age tribes
that are mostly peaceable
and have big taboos against violence.
So you can certainly have a big difference
in how culture manifests these innate biological
characteristics, but still, there's probably limits
that are given by our biology.
I used to argue this with my great-grandparents
who were Marxists, actually,
because they believed in the withering away of the state.
Like, they believed that as you move from capitalism
to socialism to communism,
people would just become more social-minded
so that a state would be unnecessary
and people would just give,
everyone would give everyone else what they needed.
Now, setting aside that that's not what the various
Marxist experiments on the planet
seem to be heading toward in practice,
just as a theoretical point,
I was very dubious that human nature could go there.
Like, at that time, when my great-grandparents were alive,
I was just like, you know, I'm a cynical teenager.
I think humans are just jerks.
The state is not gonna wither away.
If you don't have some structure keeping people
from screwing each other over, they're gonna do it.
So now I actually don't quite see things that way.
I mean, I think my feeling now subjectively
is the culture aspect is more significant
than I thought it was when I was a teenager.
And I think you could have a human society
that was dialed dramatically further toward,
you know, self-awareness, other awareness, compassion,
and sharing than our current society.
And of course, greater material abundance helps,
but to some extent, material abundance
is a subjective perception also,
because many Stone Age cultures perceived themselves
as living in great material abundance,
that they had all the food and water they wanted,
they lived in a beautiful place,
that they had sex lives, that they had children.
I mean, they had abundance without any factories, right?
So I think humanity probably would be capable
of fundamentally more positive and joy-filled mode
of social existence than what we have now.
Clearly, Marx didn't quite have the right idea
about how to get there.
I mean, he missed a number of key aspects
of human society and its evolution.
And if we look at where we are in society now,
how to get there is a quite different question,
because there are very powerful forces
pushing people in different directions
than a positive, joyous, compassionate existence, right?
- So if we were to try to, you know,
Elon Musk is dreams of colonizing Mars at the moment,
so maybe he'll have a chance to start a new civilization
with a new governmental system.
And certainly, there's quite a bit of chaos.
We're sitting now, I don't know what the date is,
but this is June.
There's quite a bit of chaos in all different forms
going on in the United States and all over the world.
So there's a hunger for new types of governments,
new types of leadership, new types of systems.
And so what are the forces at play?
And how do we move forward?
- Yeah, I mean, colonizing Mars, first of all,
it's a super cool thing to do.
We should be doing it.
- So you love the idea?
- Yeah, I mean, it's more important than making
chocolatier chocolates and sexier lingerie
and many of the things that we spend
a lot more resources on as a species, right?
So I mean, we certainly should do it.
I think the possible futures in which a Mars colony
makes a critical difference for humanity are very few.
I mean, I think, I mean, assuming we make a Mars colony
and people go live there in a couple of decades,
I mean, their supplies are gonna come from Earth.
The money to make the colony came from Earth
and whatever powers are supplying the goods there
from Earth are gonna in effect be in control
of that Mars colony.
Of course, there are outlier situations where Earth
gets nuked into oblivion and somehow Mars
has been made self-sustaining by that point
and then Mars is what allows humanity to persist.
But I think that those are very, very, very unlikely.
- Do you don't think it could be a first step
on a long journey?
- Of course, it's a first step on a long journey,
which is awesome.
I'm guessing the colonization of the rest
of the physical universe will probably be done
by AGIs that are better designed to live in space
than by the meat machines that we are.
But I mean, who knows, we may cryopreserve ourselves
in some superior way to what we know now
and like shoot ourselves out to Alpha Centauri and beyond.
I mean, that's all cool, it's very interesting
and it's much more valuable than most things
that humanity is spending its resources on.
On the other hand, with AGI, we can get to a singularity
before the Mars colony becomes sustaining for sure,
possibly before it's even operational.
- So your intuition is that that's the problem
if we really invest resources and we can get to faster
than a legitimate full self-sustaining colonization of Mars.
- Yeah, and it's very clear that we will to me
because there's so much economic value
in getting from narrow AI toward AGI,
whereas the Mars colony, there's less economic value
until you get quite far out into the future.
So I think that's very interesting.
I just think it's somewhat off to the side.
I mean, just as I think say, art and music
are very, very interesting and I wanna see resources
go into amazing art and music being created
and I'd rather see that than a lot of the garbage
that society spends their money on.
On the other hand, I don't think Mars colonization
or inventing amazing new genres of music
is not one of the things that is most likely
to make a critical difference in the evolution
of human or non-human life in this part of the universe
over the next decade.
- Do you think AGI is really--
- AGI is by far the most important thing
that's on the horizon and then technologies
that have direct ability to enable AGI
or to accelerate AGI are also very important.
For example, say quantum computing.
I don't think that's critical to achieve AGI
but certainly you could see how the right quantum
computing architecture could massively accelerate AGI.
Similar other types of nanotechnology.
Now, the quest to cure aging and end disease
while not in the big picture as important as AGI,
of course it's important to all of us as individual humans
and if someone made a super longevity pill
and distributed it tomorrow, I mean, that would be huge
and a much larger impact than a Mars colony
is gonna have for quite some time.
- But perhaps not as much as an AGI system.
- No, because if you can make a benevolent AGI,
then all the other problems are solved.
I mean, then the AGI can be,
once it's as generally intelligent as humans,
it can rapidly become massively more generally intelligent
than humans and then that AGI should be able
to solve science and engineering problems
much better than human beings
as long as it is in fact motivated to do so.
That's why I said a benevolent AGI.
There could be other kinds.
- Maybe it's good to step back a little bit.
I mean, we've been using the term AGI.
People often cite you as the creator
or at least the popularizer of the term AGI,
artificial general intelligence.
Can you tell the origin story of the term?
- Sure, sure.
So yeah, I would say I launched the term AGI
upon the world for what it's worth
without ever fully being in love with the term.
What happened is I was editing a book
and this process started around 2001 or two.
I think the book came out 2005, funny.
I was editing a book which I provisionally
was titling Real AI.
And I mean, the goal was to gather together
fairly serious academicish papers
on the topic of making thinking machines
that could really think in the sense like people can
or even more broadly than people can.
So then I was reaching out to other folks
that I had encountered here or there
§06Transcript · 47:54–56:27
who were interested in that,
which included some other folks
who I knew from the transhumanist and singularitarian world
like Peter Vos, who has a company,
AGI Incorporated still in California
and included Shane Legg, who had worked for me
at my company WebMind in New York in the late '90s,
who by now has become rich and famous.
He was one of the co-founders of Google DeepMind.
But at that time, Shane was,
I think he may have just started doing his PhD
with Markus Hutter, who at that time
hadn't yet published his book, Universal AI,
which sort of gives a mathematical foundation
for artificial general intelligence.
So I reached out to Shane and Markus and Peter Vos
and Pei Wang, who was another former employee of mine
who had been Douglas Hofstadter's PhD student
who had his own approach to AGI
and a bunch of some Russian folks
reached out to these guys
and they contributed papers for the book.
But that was my provisional title,
but I never loved it because in the end,
I was doing some, what we would now call narrow AI,
as well like applying machine learning to genomics data
or chat data for sentiment analysis.
I mean, that work is real.
And in a sense, it's really AI.
It's just a different kind of AI.
Ray Kurzweil wrote about narrow AI versus strong AI.
But that seemed weird to me because,
first of all, narrow and strong are not antennas.
(laughing)
- That's right.
- But secondly, strong AI was used
in the cognitive science literature
to mean the hypothesis that digital computer AIs
could have true consciousness like human beings.
So there was already a meaning to strong AI,
which was complexly different but related, right?
So we were tossing around on an email list
whether what title it should be.
And so we talked about narrow AI, broad AI, wide AI,
narrow AI, general AI.
And I think it was either Shane Legg or Peter Vos
on the private email discussion we had.
He said, "Well, why don't we go with AGI,
"artificial general intelligence?"
And Pei Wang wanted to do GAI,
general artificial intelligence,
'cause in Chinese, it goes in that order.
But we figured gay wouldn't work
in US culture at that time, right?
So we went with the AGI.
We used it for the title of that book.
And part of Peter and Shane's reasoning
was you have the G factor in psychology,
which is IQ, general intelligence, right?
So you have a meaning of GI,
general intelligence in psychology.
So then you're looking like artificial GI.
So then--
- That makes a lot of sense.
- Yeah, we used that for the title of the book.
And so I think, maybe both Shane and Peter
think they invented the term.
But then later, after the book was published,
this guy Mark Gubrid came up to me,
and he's like, "Well, I published an essay
"with the term AGI in like 1997 or something.
"And so I'm just waiting for some Russian
"to come out and say they published that in 1953."
(laughing)
I mean, that term-- - For sure.
- That term is not dramatically innovative or anything.
It's one of these obvious in hindsight things,
which is also annoying in a way,
because Josh Habach, who you interviewed,
is a close friend of mine.
He likes the term synthetic intelligence,
which I like much better,
but it hasn't actually caught on, right?
Because artificial is a bit off to me,
'cause artifice is like a tool or something,
but not all AGIs are gonna be tools.
I mean, they may be now,
but we're aiming toward making them agents
rather than tools.
And in a way, I don't like the distinction
between artificial and natural,
because I mean, we're part of nature also,
and machines are part of nature.
I mean, you can look at evolved versus engineered,
but that's a different distinction.
Then it should be engineered general intelligence, right?
And then general, well, if you look at
Marcus Hutter's book, "Universal AI,"
what he argues there is,
within the domain of computation theory,
which is limited but interesting,
so if you assume computable environments
and computable reward functions,
then he articulates what would be
a truly general intelligence,
a system called AIXI, which is quite beautiful.
- AIXI.
- AIXI, and that's the middle name
of my latest child, actually.
- What's the first name?
- First name is QORXI, Q-O-R-X-I,
which my wife came up with,
but that's an acronym for
quantum organized rational expanding intelligence.
- AIXI's the middle name.
- His middle name is Aixiphanes, actually,
which means the former principal underlying AIXI.
But in any case--
- You're giving Elon Musk's new child a run for his money.
- Well, I did it first.
He copied me with this new freakish name,
but now if I have another baby,
I'm gonna have to outdo him.
- Outdo him.
- An arms race of weird, geeky baby names.
We'll see what the babies think about it.
But my oldest son, Zarathustra, loves his name,
and my daughter, Sherazade, loves her name.
So, so far, basically, if you give your kids weird names--
- They live up to it.
- Well, you're obliged to make the kids weird enough
that they like the names.
So it directs their upbringing in a certain way.
But anyway, what Marcus showed in that book
is that a truly general intelligence,
theoretically, is possible,
but would take infinite computing power.
So then the artificial is a little off.
The general is not really achievable within physics,
as we know it.
I mean, physics, as we know it, may be limited,
but that's what we have to work with now.
Intelligence--
- Infinitely general, you mean,
like from an information processing perspective, yeah.
- Yeah, intelligence is not very well-defined either.
I mean, what does it mean?
I mean, in AI now, it's fashionable to look at it
as maximizing an expected reward over the future.
But that sort of definition is pathological in various ways.
And my friend David Weinbaum, aka Weaver,
he had a beautiful PhD thesis on open-ended intelligence,
trying to conceive intelligence in a--
- Without a reward.
- Yeah, he's just looking at it differently.
He's looking at complex self-organizing systems
and looking at an intelligence system as being one
that revises and grows and improves itself
in conjunction with its environment
without necessarily there being one objective function
it's trying to maximize.
Although, over certain intervals of time,
it may act as if it's optimizing
a certain objective function.
Very much Solaris from Stanislav Lem's novels, right?
So yeah, the point is artificial general and intelligence--
- Don't work.
- They're all bad.
On the other hand, everyone knows what AI is,
and AGI seems immediately comprehensible
to people with a technical background.
So I think that the term has served
as sociological function.
Now it's out there everywhere, which baffles me.
- It's like KFC, I mean, that's it.
We're stuck with AGI probably for a very long time
until AGI systems take over and rename themselves.
- Yeah, and then we'll be--
- We're stuck with GPUs, too,
which mostly have nothing to do with graphics anymore.
- I wonder what the AGI system will call us humans.
That was maybe--
- Grandpa.
(laughing)
- GPs.
(laughing)
- Grandpa processing unit, yeah.
- Biological grandpa processing units.
Okay, so maybe also just a comment on AGI representing,
before even the term existed,
representing a kind of community.
You've talked about this in the past,
sort of AI has come in waves,
but there's always been this community of people
who dream about creating general human-level
superintelligent systems.
Can you maybe give your sense of the history
of this community as it exists today,
as it existed before this deep learning revolution,
all throughout the winters and the summers of AI?
§07Transcript · 56:29–1:05:44
- Sure, first I would say as a side point,
the winters and summers of AI
are greatly exaggerated by Americans.
And if you look at the publication record
of the artificial intelligence community
since say the 1950s,
you would find a pretty steady growth
in advance of ideas and papers.
And what's thought of as an AI winter or summer
was sort of how much money is the US military
pumping into AI, which was meaningful.
On the other hand, there was AI going on in Germany, UK,
and in Japan, and in Russia, all over the place,
while US military got more and less enthused about AI.
- That happened to be, just for people who don't know,
the US military happened to be the main source
of funding for AI research.
So another way to phrase that is it's up and down
of funding for artificial intelligence research.
- And I would say the correlation between funding
and intellectual advance was not 100%, right?
Because in Russia, as an example,
or in Germany, there was less dollar funding
than in the US, but many foundational ideas were laid out,
but it was more theory than implementation, right?
And US really excelled at sort of breaking through
from theoretical papers to working implementations,
which did go up and down somewhat
with US military funding.
But still, I mean, you can look in the 1980s,
Dietrich Doerner in Germany had self-driving cars
on the Autobahn, right?
And I mean, this, it was a little early
with regard to the car industry,
so it didn't catch on such as has happened now.
But I mean, that whole advancement
of self-driving car technology in Germany
was pretty much independent of AI military summers
and winters in the US.
So there's been more going on in AI globally
than not only most people on the planet realize,
but then most new AI PhDs realize,
because they've come up within a certain subfield of AI
and haven't had to look so much beyond that.
But I would say when I got my PhD in 1989 in mathematics,
I was interested in AI already.
- In Philadelphia, by the way.
- Yeah, I started at NYU, then I transferred to Philadelphia,
to Temple University, good old North Philly.
- North Philly, yeah.
- Yeah, yeah, yeah, the pearl of the US.
(laughing)
You never stopped at a red light then,
'cause you were afraid if you stopped at a red light,
someone will carjack you,
so you just drive through every red light.
- Yeah.
- Is it, every day driving or bicycling to Temple
from my house was like a new adventure, right?
But yeah, the reason I didn't do a PhD in AI
was what people were doing in the academic AI field then
was just astoundingly boring and seemed wrong-headed to me.
It was really like rule-based expert systems
and production systems.
And actually, I loved mathematical logic.
I had nothing against logic as the cognitive engine
for an AI, but the idea that you could type in the knowledge
that AI would need to think seemed just completely stupid
and wrong-headed to me.
I mean, you can use logic if you want,
but somehow the system has gotta be--
- Automated.
- Learning, right?
It should be learning from experience.
And the AI field then was not interested
in learning from experience.
I mean, some researchers certainly were.
I mean, I remember in mid '80s,
I discovered a book by John Andreas,
which was, it was about a reinforcement learning system
called PURR, P-U-R-R-P-U-S-S, which was an acronym
that I can't even remember what it was for,
but PURRpus, anyway.
But he, I mean, that was a system
that was supposed to be an AGI.
And basically, by some sort of fancy
like Markov decision process learning,
it was supposed to learn everything
just from the bits coming into it
and learn to maximize its reward
and become intelligent, right?
So that was there in academia back then,
but it was like isolated, scattered, weird people.
But all these isolated, scattered, weird people
in that period, I mean, they laid the intellectual grounds
for what happened later.
So you look at John Andreas at University of Canterbury
with his PURRpus reinforcement learning Markov system.
He was the PC supervisor for John Cleary in New Zealand.
Now, John Cleary worked with me
when I was at Waikato University in 1993 in New Zealand,
and he worked with Ian Witton there,
and they launched Weka,
which was the first open source machine learning toolkit,
which was launched in, I guess, '93 or '94
when I was at Waikato University.
- Written in Java, unfortunately.
- Written in Java,
which was a cool language back then, though, right?
- I guess it's still, well, it's not cool anymore,
but it's powerful.
- I find, like most programmers now,
I find Java unnecessarily bloated,
but back then it was like Java or C++, basically.
- Object-oriented, so it's nice.
- Java was easier for students.
Amusingly, a lot of the work on Weka
when we were in New Zealand
was funded by a US, sorry, a New Zealand government grant
to use machine learning
to predict the menstrual cycles of cows.
So in the US, all the grant funding for AI
was about how to kill people or spy on people.
In New Zealand, it's all about cows or kiwi fruits, right?
- Yeah.
- So yeah, anyway, I mean, John Andreas
had his probability theory-based reinforcement learning,
proto-AGI.
John Cleary was trying to do much more ambitious,
probabilistic AGI systems.
Now, John Cleary helped do Weka,
which was the first open-source machine learning tool.
It gets to the predecessor for TensorFlow and Torch
and all these things.
Also, Shane Legg was at Wikado
working with John Cleary and Ian Winton
and this whole group,
and then working with my own company's,
my company WebMind, an AI company I had in the late '90s
with a team there at Wikado University,
which is how Shane got his head full of AGI,
which led him to go on,
and with Demis Hassabis found DeepMind.
So what you can see through that lineage is,
you know, in the '80s and '70s,
John Andreas was trying to build
probabilistic reinforcement learning AGI systems.
The technology, the computers just weren't there to support.
His ideas were very similar to what people are doing now.
But, you know, although he's long since passed away
and didn't become that famous outside of Canterbury,
I mean, the lineage of ideas passed on
from him to his students to their students,
you can go trace directly from there to me
and to DeepMind, right?
So there was a lot going on in AGI
that did ultimately lay the groundwork for what we have today
but there wasn't a community, right?
And so when I started trying to pull together
an AGI community, it was in the, I guess,
the early aughts when I was living in Washington, DC
and making a living doing AI consulting
for various US government agencies.
And I organized the first AGI workshop in 2006.
And I mean, it wasn't like it was literally
in my basement or something.
I mean, it was in the conference room
at Marriott in Bethesda.
It's not that edgier underground, unfortunately, but still--
- How many people attended?
- About 60 or something.
- That's not bad.
- I mean, DC has a lot of AI going on,
probably until the last five or 10 years,
much more than Silicon Valley,
although it's just quiet because of the nature
of what happens in DC.
Their business isn't driven by PR.
Mostly when something starts to work really well,
it's taken black and becomes even more quiet.
But yeah, the thing is that really had the feeling
of a group of starry-eyed mavericks
huddled in a basement,
plotting how to overthrow the narrow AI establishment.
And for the first time, in some cases,
coming together with others who shared their passion
for AGI and a technical seriousness
about working on it.
I mean, that's very, very different
than what we have today.
I mean, now it's a little bit different.
We have AGI conference every year,
and there's several hundred people rather than 50.
Now it's more like this is the main gathering
of people who want to achieve AGI
and think that large-scale nonlinear regression
is not the golden path to AGI.
- So I mean it's-- - AKA neural network.
- Yeah, yeah, yeah.
Well, certain architectures for learning
§08Transcript · 1:05:49–1:15:29
using neural networks.
So yeah, the AGI conferences are sort of now
the main concentration of people not obsessed
with deep neural nets-- - The rebels.
- And deep reinforcement learning,
but still interested in AGI.
Not the only ones.
I mean, there's other little conferences and groupings
interested in human-level AI
and cognitive architectures and so forth.
But yeah, it's been a big shift.
Like back then, you couldn't really,
it would be very, very edgy then
to give a university department seminar
that mentioned AGI or human-level AI.
It was more like you had to talk about
something more short-term and immediately practical
than in the bar after the seminar,
you could bullshit about AGI in the same breath
as time travel or the simulation hypothesis or something.
Whereas now, AGI is not only in the academic seminar room.
Like you have Vladimir Putin knows what AGI is.
And he's like, Russia needs to become the leader in AGI.
So national leaders and CEOs of large corporations.
I mean, the CTO of Intel, Justin Ratner,
this was years ago, Singularity Summit Conference,
2008 or something.
He's like, we believe Ray Kurzweil,
the Singularity will happen in 2045
and it will have Intel inside.
(laughing)
I mean, so it's gone from being something
which is the pursuit of like crazed mavericks,
crackpots and science fiction fanatics
to being a marketing term for large corporations
and the national leaders, which is a astounding transition.
But yeah, in the course of this transition,
I think a bunch of sub-communities have formed.
And the community around the AGI conference series
is certainly one of them.
It hasn't grown as big as I might've liked it to.
On the other hand, sometimes a modest size community
can be better for making intellectual progress also.
You go to a Society for Neuroscience Conference,
you have 35 or 40,000 neuroscientists.
On the one hand, it's amazing.
On the other hand, you're not gonna talk to the leaders
of the field there if you're an outsider.
- Yeah, in the same sense, the AAAI,
the artificial intelligence,
the main kind of generic artificial intelligence conference
is too big, it's too amorphous.
Like it doesn't make sense.
- Well, yeah, but NIPS has become
a company advertising outlet now.
(laughing)
- So, I mean, to comment on the role of AGI
in the research community, I'd still,
if you look at NeurIPS, if you look at CVPR,
if you look at these iClear,
AGI is still seen as the outcast.
I would say in these main machine learning,
in these main artificial intelligence conferences
amongst the researchers,
I don't know if it's an accepted term yet.
What I've seen bravely, you mentioned Shane Legg,
is DeepMind and then OpenAI are the two places
that are, I would say unapologetically so far,
I think it's actually changing, unfortunately,
but so far they've been pushing the idea
that the goal is to create an AGI.
- Well, they have billions of dollars behind them,
so I mean, in the public mind,
that certainly carries some oomph, right?
I mean, Monial is-- - But they also
have really strong researchers, right?
- They do, they're great teams.
I mean, DeepMind in particular, yeah.
- And they have, I mean, DeepMind has Marcus Hodder
walking around, I mean, there's all these folks
who basically their full-time position involves
dreaming about creating AGI.
- I mean, Google Brain has a lot of amazing
AGI-oriented people also.
And I mean, so I'd say from a public marketing view,
DeepMind and OpenAI are the two large,
well-funded organizations that have put
the term and concept AGI out there
sort of as part of their public image.
But I mean, they're certainly not,
there are other groups that are doing research
that seems just as AGI-ish to me,
I mean, including a bunch of groups in Google's
main Mountain View office.
So yeah, it's true, AGI is somewhat
away from the mainstream now,
but if you compare it to where it was 15 years ago,
there's been an amazing mainstreaming.
You could say the same thing about super longevity research,
which is one of my application areas that I'm excited about.
I mean, I've been talking about this since the '90s,
but working on this since 2001.
And back then, really to say you're trying
to create therapies to allow people to live
hundreds or thousands of years,
you were way, way, way, way out of the industry
academic mainstream.
But now, Google had Project Calico,
Craig Venter had Human Longevity Incorporated,
and then once the suits come marching in, right?
I mean, once there's big money in it,
then people are forced to take it seriously,
'cause that's the way modern society works.
So it's still not as mainstream as cancer research,
just as AGI is not as mainstream
as automated driving or something.
But the degree of mainstreaming that's happened
in the last 10 to 15 years is astounding
to those of us who've been at it for a while.
- Yeah, but there's a marketing aspect to the term,
but in terms of actual full force research
that's going on under the header of AGI,
it's currently, I would say, dominated,
maybe you can disagree,
dominated by neural networks research,
that the nonlinear regression, as you mentioned.
Like, what's your sense with OpenCog, with your work,
but in general, I was logic-based systems
and expert systems.
For me, always seemed to capture a deep element
of intelligence that needs to be there.
Like you said, it needs to learn,
it needs to be automated somehow,
but that seems to be missing from a lot of research currently.
So what's your sense?
I guess one way to ask this question,
what's your sense of what kind of things
will an AGI system need to have?
- Yeah, that's a very interesting topic
that I've thought about for a long time.
And I think there are many, many different approaches
that can work for getting to human level AI.
So I don't think there's like one golden algorithm,
one golden design that can work.
And I mean, flying machines is the much-worn analogy here,
right, like I mean, you have airplanes,
you have helicopters, you have balloons,
you have stealth bombers
that don't look like regular airplanes.
You've got all blimps.
- Birds too.
- Birds, yeah, and bugs, right?
- Yeah.
- And I mean, there are certainly many kinds
of flying machines that--
- And there's a catapult that you can just launch.
(laughing)
- There's bicycle-powered flying machines, right?
- Nice, yeah.
- Yeah, so now these are all analyzable
by a basic theory of aerodynamics, right?
Now, so one issue with AGI is we don't yet have the analog
of the theory of aerodynamics.
And that's what Marcus Hutter was trying to make
with the AXE and his general theory of general intelligence.
But that theory, in its most clearly articulated parts,
really only works for either infinitely powerful machines
or almost, or insanely impractically powerful machines.
So I mean, if you were gonna take a theory-based approach
to AGI, what you would do is say,
well, let's take what's called, say, AXE-TL,
which is Hutter's AXE machine that can work
on merely insanely much processing power
rather than infinitely much--
- What does TL stand for?
- Time and length.
- Okay.
- So you're, basically, how--
- Like constraints and all.
- Yeah, yeah, yeah.
So how AXE works, basically, is each action
that it wants to take, before taking that action,
it looks at all its history, and then it looks
at all possible programs that it could use
to make a decision, and it decides which decision program
would have let it make the best decisions
according to its reward function over its history,
and it uses that decision program
to make the next decision, right?
- It's not afraid of infinite resources.
- It's searching through the space
of all possible computer programs
in between each action and each next action.
Now, AXE-TL searches through all possible
computer programs that have runtime less than T
and length less than L, which is still
an impracticably humongous space, right?
So what you would like to do to make an AGI,
and what will probably be done 50 years from now
to make an AGI, is say, okay, well, we have some constraints.
§09Transcript · 1:15:34–1:24:29
We have these processing power constraints,
and we have space and time constraints on the program.
We have energy utilization constraints,
and we have this particular class of environments
that we care about, which may be, say,
manipulating physical objects on the surface of the Earth,
communicating in human language, I mean,
whatever our particular, not annihilating humanity,
whatever our particular requirements happen to be,
if you formalize those requirements
in some formal specification language,
you should then be able to run
a automated program specializer on AXE-TL,
specialize it to the computing resource constraints
and the particular environment and goal,
and then it will spit out the specialized version
of AXE-TL to your resource restrictions
in your environment, which will be your AGI, right?
And that, I think, is how our super AGI
will create new AGI systems, right?
But that's a very--
- That just seems really inefficient.
- That's a very Russian approach, by the way.
Like, the whole field of program specialization
came out of Russia.
- Can you backtrack?
So what is program specialization?
So it's basically--
- Well, take sorting, for example.
You can have a generic program for sorting lists,
but what if all your lists you care about
are length 10,000 or less?
- Got it.
- You can run an automated program specializer
on your sorting algorithm,
and it will come up with the algorithm
that's optimal for sorting lists of length 10,000 or less.
- That's kind of like,
isn't that kind of the process of evolution,
is a program specializer to the environment?
So you're kind of evolving human beings or living creatures.
- Exactly, I mean, your Russian heritage is showing there.
So with Alexander Vityaev and Peter Anokhin and so on,
I mean, there's a long history
of thinking about evolution that way also, right?
Well, my point is that what we're thinking of
as a human-level general intelligence,
you know, if you start from narrow AIs,
like are being used in the commercial AI field now,
then you're thinking,
okay, how do we make it more and more general?
On the other hand, if you start from Aixi
or Schmidhuber's girdle machine,
or these infinitely powerful,
but practically infeasible AIs,
then getting to a human-level AGI
is a matter of specialization.
It's like, how do you take
these maximally general learning processes,
and how do you specialize them
so that they can operate
within the resource constraints that you have,
but will achieve the particular things that you care about?
'Cause we humans are not
maximally general intelligences, right?
If I ask you to run a maze in 750 dimensions,
you'll probably be very slow,
whereas at two dimensions,
you're probably way better, right?
So I mean, because our hippocampus
has a two-dimensional map in it, right?
And it does not have a 750-dimensional map in it.
So I mean, we're a peculiar mix
of generality and specialization, right?
- We probably start quite general at birth.
Not obviously still narrow,
but more general than we are at age 20,
and 30, and 40, and 50, and 60.
- I don't think that,
I think it's more complex than that,
because in some sense, a young child
is less biased,
and their brain has yet to sort of crystallize
into appropriate structures for processing aspects
of the physical and social world.
On the other hand,
the young child is very tied to their sensorium,
whereas we can deal with abstract mathematics,
like 750 dimensions,
and the young child cannot,
because they haven't grown
what Piaget called the formal capabilities.
They haven't learned to abstract yet, right?
And the ability to abstract
gives you a different kind of generality
than what a baby has.
So there's both more specialization
and more generalization
that comes with the development process, actually.
- I mean, I guess just the trajectories
of the specialization are most controllable
at the young age, I guess is one way to put it.
- Do you have kids?
- No.
They're not as controllable as you think.
- So you think it's interesting.
- I think, honestly, I think--
- Interesting.
- A human adult is much more generally intelligent
than a human baby.
Babies are very stupid.
I mean, they're cute,
which is why we put up with their repetitiveness
and stupidity.
And they have what the Zen guys would call
a beginner's mind, which is a beautiful thing,
but that doesn't necessarily correlate
with a high level of intelligence.
- So on the plot of cuteness and stupidity,
there's a process that allows us
to put up with their stupidity
as they become more intelligent.
- So by the time you're an ugly old man like me,
you gotta get really, really smart to compensate.
- To compensate, okay, cool.
- But yeah, going back to your original question,
so the way I look at human-level AGI
is how do you specialize
unrealistically inefficient superhuman
brute force learning processes
to the specific goals that humans need to achieve
and the specific resources that we have?
And both of these, the goals and the resources
and the environments, I mean, all this is important.
And on the resources side,
it's important that the hardware resources
we're bringing to bear are very different
than the human brain.
So the way I would want to implement AGI
on a bunch of neurons in a vat
that I could rewire arbitrarily
is quite different than the way I would want to create AGI
on say a modern server farm of CPUs and GPUs,
which in turn may be quite different
than the way I would wanna implement AGI
on whatever quantum computer we'll have in 10 years,
supposing someone makes a robust quantum Turing machine
or something, right?
So I think there's been co-evolution
of the patterns of organization in the human brain
and the physiological particulars
of the human brain over time.
And when you look at neural networks,
that is one powerful class of learning algorithms,
but it's also a class of learning algorithms
that evolved to exploit the particulars of the human brain
as a computational substrate.
If you're looking at the computational substrate
of a modern server farm,
you won't necessarily want the same algorithms
that you want on the human brain.
Then from the right level of abstraction,
you could look at maybe the best algorithms on the brain
and the best algorithms on a modern computer network
as implementing the same abstract learning
and representation processes,
but finding that level of abstraction
is its own AGI research project then, right?
So that's about the hardware side and the software side,
which follows from that.
Then regarding what are the requirements,
I wrote the paper years ago
on what I called the embodied communication prior,
which was quite similar in intent to Yoshua Bengio's
recent paper on the consciousness prior,
except I didn't wanna wrap up consciousness in it
because to me, the qualia problem and subjective experience
is a very interesting issue also,
which we can chat about,
but I would rather keep that philosophical debate
distinct from the debate of what kind of biases
do you wanna put in a general intelligence
to give it human-like general intelligence.
- And I'm not sure Yoshua Bengio
is really addressing that kind of consciousness.
He's just using the term.
- I love Yoshua to pieces.
He's by far my favorite of the lions of deep learning.
He's such a good-hearted guy.
- He's a good human being, yeah, for sure.
- I am not sure he has plumbed to the depths
of the philosophy of consciousness.
- No, he's using it as a sexy term.
- Yeah, yeah, yeah.
So what I called it was the embodied communication prior.
- Can you maybe explain it a little bit?
- Yeah, yeah, what I meant was,
what are we humans evolved for?
You can say being human, but that's very abstract, right?
I mean, our minds control individual bodies,
§10Transcript · 1:24:32–1:33:06
which are autonomous agents,
moving around in a world that's composed
largely of solid objects, right?
And we've also evolved to communicate via language
with other solid object agents
that are going around doing things collectively with us
in a world of solid objects.
And these things are very obvious,
but if you compare them to the scope
of all possible intelligences,
or even all possible intelligences
that are physically realizable,
that actually constrains things a lot.
So if you start to look at,
how would you realize some specialized
or constrained version of universal general intelligence
in a system that has limited memory
and limited speed of processing,
but whose general intelligence will be biased
toward controlling a solid object agent,
which is mobile in a solid object world
for manipulating solid objects
and communicating via language
with other similar agents in that same world, right?
Then starting from that,
you're starting to get a requirements analysis
for human level general intelligence.
And then that leads you into cognitive science.
And you can look at, say,
what are the different types of memory
that the human mind and brain has?
And this has matured over the last decades.
And I got into this a lot.
So after getting my PhD in math,
I was an academic for eight years.
I was in departments of mathematics,
computer science, and psychology.
When I was in the psychology department
at University of Western Australia,
I was focused on cognitive science
of memory and perception.
Actually, I was teaching neural nets
and deep neural nets,
and it was multilayer perceptrons, right?
- Psychology?
(laughs)
- Cognitive science, it was cross-disciplinary
among engineering, math, psychology,
philosophy, linguistics, computer science.
But yeah, we were teaching psychology students
to try to model the data
from human cognition experiments
using multilayer perceptrons,
which was the early version of a deep neural network.
Very, very, yeah, recurrent backprop
was very, very slow to train back then, right?
- So this is the study of these constraints.
- Yeah.
- Systems that are supposed to deal with physical objects.
- So if you look at cognitive psychology,
you can see there's multiple types of memory,
which are to some extent represented
by different subsystems in the human brain.
So we have episodic memory,
which takes into account our life history
and everything that's happened to us.
We have declarative or semantic memory,
which is like facts and beliefs
abstracted from the particular situations
that they occurred in.
There's sensory memory,
which to some extent is sense modality specific,
and then to some extent is unified across sense modalities.
There's procedural memory, memory of how to do stuff,
like how to swing the tennis racket, right?
Which is, there's motor memory,
but it's also a little more abstract than motor memory.
It involves cerebellum and cortex working together.
Then there's memory linkage with emotion,
which has to do with linkages of cortex and limbic system.
There's specifics of spatial and temporal modeling
connected with memory,
which has to do with the hippocampus and thalamus
connecting to cortex.
And the basal ganglia, which influences goals.
So we have specific memory of what goals,
sub-goals and sub-sub-goals we wanted to perceive
in which context in the past.
Human brain has substantially different subsystems
for these different types of memory,
and substantially differently tuned learning,
like differently tuned modes of long-term potentiation
to do with the types of neurons and neurotransmitters
in the different parts of the brain
corresponding to these different types of knowledge.
And these different types of memory and learning
in the human brain, I mean, you can back these all
into embodied communication for controlling agents
in worlds of solid objects.
Now, so if you look at building an AGI system,
one way to do it, which starts more from cognitive science
and neuroscience is to say,
okay, what are the types of memory that--
- That are necessary for this kind of world?
- Yeah, yeah, necessary for this sort of intelligence.
What types of learning work well
with these different types of memory?
And then how do you connect all these things together, right?
And of course, the human brain did it incrementally
through evolution because each of the sub-networks
of the brain, I mean, it's not really the lobes
of the brain, it's the sub-networks,
each of which is widely distributed.
Which of the, each of the sub-networks of the brain
co-evolved with the other sub-networks of the brain,
both in terms of its patterns of organization
and the particulars of the neurophysiology.
So they all grew up communicating and adapting to each other.
It's not like they were separate black boxes
that were then glommed together, right?
Whereas as engineers, we would tend to say,
let's make the declarative memory box here
and the procedural memory box here
and the perception box here and wire them together.
And when you can do that, it's interesting.
I mean, that's how a car is built, right?
But on the other hand, that's clearly not
how biological systems are made.
The parts co-evolve so as to adapt and work together.
So this-- - That's by the way
how every human engineered system that flies,
that was what we're using that analogy before,
is built as well. - Yes.
- So do you find this at all appealing?
Like there's been a lot of really exciting,
which I find strange that it's ignored,
work in cognitive architectures, for example,
throughout the last few decades.
Do you find that-- - Yeah, I mean,
I had a lot to do with that community.
And Paul Rosenblum, who was one of the,
and John Laird, who built the SOAR architecture,
are friends of mine.
And I learned SOAR quite well and ACT-R
and these different cognitive architectures.
How I was looking at the AI world about 10 years ago,
before this whole commercial deep learning explosion was,
on the one hand, you had these cognitive architecture guys
who were working closely with psychologists
and cognitive scientists who had thought a lot about
how the different parts of a human-like mind
should work together.
On the other hand, you had these learning theory guys
who didn't care at all about the architecture,
but were just thinking about like,
how do you recognize patterns in large amounts of data?
And in some sense, what you needed to do
was to get the learning that the learning theory guys
were doing and put it together with the architecture
that the cognitive architecture guys were doing.
And then you would have what you needed.
Now, you can't, unfortunately, when you look at the details,
you can't just do that without totally rebuilding
what is happening on both the cognitive architecture
and the learning side.
So, I mean, they tried to do that in SOAR,
but what they ultimately did is like take a deep neural net
or something for perception,
and you include it as one of the black boxes.
- Yeah, it becomes one of the boxes.
The learning mechanism becomes one of the boxes
as opposed to fundamental part of the system.
- That doesn't quite work.
You could look at some of the stuff DeepMind has done,
like the differential neural computer or something.
That sort of has a neural net for deep learning perception.
It has another neural net, which is like a memory matrix
that stores, say, the map of the London subway or something.
So, probably Demis Asabas was thinking about this
like part of cortex and part of hippocampus,
'cause hippocampus has a spatial map.
And when he was a neuroscientist,
he was doing a bunch on cortex-hippocampus interconnection.
So there, the DNC would be an example of folks
from the deep neural net world trying to take a step
in the cognitive architecture direction
by having two neural modules that correspond roughly
to two different parts of the human brain
that deal with different kinds of memory and learning.
But on the other hand, it's super, super, super crude
from the cognitive architecture view, right?
Just as what John Laird and Sohr did with neural nets
was super, super crude from a learning point of view,
'cause the learning was like off to the side,
not affecting the core representations, right?
And when you weren't learning the representation,
you were learning the data that feeds into the,
you were learning abstractions of perceptual data
to feed into the representation that was not learned, right?
So yeah, this was clear to me a while ago.
§11Transcript · 1:33:11–1:42:27
And one of my hopes with the AGI community
was to sort of bring people
from those two directions together.
That didn't happen much in terms of--
- Not yet.
- And what I was gonna say is it didn't happen
in terms of bringing the lines of cognitive architecture
together with the lines of deep learning.
It did work in the sense that a bunch of younger researchers
have had their heads filled with both of those ideas.
This comes back to a saying my dad,
who was a university professor, often quoted to me,
which was science advances one funeral at a time,
which I'm trying to avoid.
Like I'm 53 years old, and I'm trying to invent
amazing, weird-ass new things
that nobody ever thought about,
which we'll talk about in a few minutes.
But there is that aspect, right?
Like the people who've been at AI a long time
and have made their career at developing one aspect,
like a cognitive architecture or a deep learning approach,
it can be hard once you're old
and have made your career doing one thing.
It can be hard to mentally shift gears.
I mean, I try quite hard to remain flexible-minded.
- Have you been successful somewhat in changing,
maybe, have you changed your mind on some aspects
of what it takes to build an AGI, like technical things?
- The hard part is that the world doesn't want you to.
- The world or your own brain?
- The world, well, that one point
is that your brain doesn't want to.
The other part is that the world doesn't want you to.
Like the people who have followed your ideas
get mad at you if you change your mind.
And the media wants to pigeonhole you
as an avatar of a certain idea.
But yeah, I've changed my mind on a bunch of things.
I mean, when I started my career,
I really thought quantum computing
would be necessary for AGI.
And I doubt it's necessary now,
although I think it will be a super major enhancement.
But I mean, I'm also, I'm now in the middle
of embarking on a complete rethink and rewrite
from scratch of our OpenCog AGI system,
together with Alexey Podopov and his team in St. Petersburg,
who's working with me in SingularityNet.
So now we're trying to like go back to basics,
take everything we learned from working
with the current OpenCog system,
take everything everybody else has learned
from working with their proto-AGI systems
and design the best framework for the next stage.
And I do think there's a lot to be learned
from the recent successes with deep neural nets
and deep reinforcement systems.
I mean, people made these essentially trivial systems
work much better than I thought they would.
And there's a lot to be learned from that.
And I want to incorporate that knowledge appropriately
in our OpenCog 2.0 system.
On the other hand, I also think current deep neural net
architectures as such will never get you anywhere near AGI.
So I think you want to avoid the pathology
of throwing the baby out with the bathwater
and like saying, well, these things are garbage
because foolish journalists overblow them
as being the path to AGI,
and a few researchers overblow them as well.
There's a lot of interesting stuff to be learned there,
even though those are not the golden path.
- So maybe this is a good chance to step back.
You mentioned OpenCog 2.0, but--
- Go back to OpenCog 0.0, which exists now.
- Alpha, yeah.
Yeah, maybe talk through the history of OpenCog
and your thinking about these ideas.
- I would say OpenCog 2.0 is a term
worth throwing around sort of tongue in cheek
because the existing OpenCog system
that we're working on now is not remotely close
to what we'd consider a 1.0, right?
I mean, it's an early, it's been around,
what, 13 years or something,
but it's still an early stage research system, right?
And actually, we are going back to the beginning
in terms of theory and implementation
'cause we feel like that's the right thing to do,
but I'm sure what we end up with is gonna have
a huge amount in common with the current system.
I mean, we all still like the general approach.
So there's-- - So first of all,
what is OpenCog?
- Sure, OpenCog is an open-source software project
that I launched together with several others in 2008,
and probably the first code written toward that
was written in 2001 or two or something
that was developed as a proprietary code base
within my AI company, Novamente LLC,
and we decided to open-source it in 2008,
cleaned up the code, threw out some things,
added some new things, and--
- What language is it written in?
- It's C++.
Primarily, there's a bunch of scheme as well,
but most of it's C++.
- And it's separate from,
something we'll also talk about is SingularityNet.
So it was born as a non-networked thing.
- Correct, correct.
Well, there are many levels of networks,
and involved here, right?
- No connectivity to the internet.
Or no, at birth.
- Yeah, I mean, SingularityNet is a separate project
and a separate body of code,
and you can use SingularityNet as part of the infrastructure
for a distributed OpenCog system,
but there are different layers.
- Yeah, got it.
- So OpenCog, on the one hand, as a software framework,
could be used to implement a variety
of different AI architectures and algorithms.
But in practice, there's been a group of developers,
which I've been leading together with Linus Vepsdas,
Neil Geisweiler, and a few others,
which have been using the OpenCog platform
and infrastructure to implement certain ideas
about how to make an AGI.
So there's been a little bit of ambiguity
about OpenCog, the software platform,
versus OpenCog, the AGI design,
'cause in theory, you could use that software,
you could use it to make a neural net,
you could use it to make a lot of different AGI.
- What kind of stuff does the software platform provide?
Like in terms of utilities, tools, like what?
- Yeah, let me first tell about OpenCog
as a software platform,
and then I'll tell you the specific AGI R&D
we've been building on top of it.
So the core component of OpenCog as a software platform
is what we call the atom space,
which is a weighted labeled hypergraph.
- A-T-O-M, atom space.
- Atom space, yeah, yeah, not atom, like Adam and Eve,
although that would be cool too.
Yeah, so you have a hypergraph, which is like a,
so a graph in this sense is a bunch of nodes
with links between them.
A hypergraph is like a graph,
but links can go between more than two nodes.
You have a link between three nodes.
And in fact, OpenCog's atom space
would properly be called a metagraph
because you can have links pointing to links,
or you could have links pointing to whole subgraphs, right?
So it's an extended hypergraph or a metagraph.
- Is metagraph a technical term?
- It is now a technical term.
- Interesting.
- But I don't think it was yet a technical term
when we started calling this a generalized hypergraph.
But in any case, it's a weighted-labeled
generalized hypergraph or weighted-labeled metagraph.
The weights and labels mean that the nodes and links
can have numbers and symbols attached to them.
So they can have types on them.
They can have numbers on them that represent,
say, a truth value or an importance value
for a certain purpose.
- And of course, like with all things,
you can reduce that to a hypergraph,
and then the hypergraph can be reduced to a graph.
- You can reduce a hypergraph to a graph,
and you can reduce a graph to an adjacency matrix.
So, I mean, there's always multiple representations.
- But there's a layer of representation
that seems to work well here.
Got it.
- Right, right, right.
And so, similarly, you could have a link to a whole graph
because a whole graph could represent,
say, a body of information.
And I could say, I reject this body of information.
Then one way to do that is make that link
go to that whole subgraph
representing the body of information, right?
I mean, there are many alternate representations,
but that's, anyway, what we have in OpenCog,
we have an atom space,
which is this weighted-labeled generalized hypergraph.
Knowledge Store, it lives in RAM.
There's also a way to back it up to disk.
There are ways to spread it
among multiple different machines.
Then there are various utilities for dealing with that.
So there's a pattern matcher,
§12Transcript · 1:42:29–1:51:09
which lets you specify a sort of abstract pattern
and then search through a whole atom space,
weighted-labeled hypergraph,
to see what sub-hypergraphs may match that pattern,
for an example.
So that's, then there's something called
the COGS server in OpenCog,
which lets you run a bunch of different agents or processes
in a scheduler, and each of these agents,
basically, it reads stuff from the atom space
and it writes stuff to the atom space.
So this is sort of the basic operational model.
- That's the software framework.
- Right, and of course, there's a lot there,
just from a scalable software engineering standpoint.
- So you could use this, I don't know if you've,
have you looked into the Steven Wolfram's
physics project recently with the hypergraphs and stuff?
Could you theoretically use the software framework to--
- You certainly could, although Wolfram would rather die
than use anything but Mathematica for his work.
- Well, that's, yeah, but there's a big community of people
who are, you know, would love integration.
And like you said, the young minds love the idea
of integrating, of connecting things.
- Yeah, yeah, that's right, and I would add on that note,
the idea of using hypergraph-type models in physics
is not very new.
Like if you look at--
- The Russians did it first.
- Well, I'm sure they did, and a guy named Ben Dribis,
who's a mathematician, a professor in Louisiana or somewhere,
had a beautiful book on quantum sets and hypergraphs
and algebraic topology for discrete models of physics
and carried it much farther than Wolfram has,
but he's not rich and famous,
so it didn't get in the headlines.
But yeah, Wolfram aside, yes, certainly,
that's a good way to put it.
The whole OpenCog framework,
you could use it to model biological networks
and simulate biology processes.
You could use it to model physics
on discrete graph models of physics.
So you could use it to do, say,
biologically realistic neural networks, for example.
So that's a framework.
- What do agents and processes do?
Do they grow the graph?
What kind of computations, just to get a sense,
are they supposed to do?
- So in theory, they could do anything they want to do.
They're just C++ processes.
On the other hand, the computation framework
is sort of designed for agents
where most of their processing time
is taken up with reads and writes to the Adam space.
And so that's a very different processing model
than, say, the matrix multiplication-based model
as underlies most deep learning systems.
So you could create an agent
that just factored numbers for a billion years.
It would run within the OpenCog platform,
but it would be pointless.
I mean, the point of doing OpenCog
is because you want to make agents
that are cooperating via reading and writing
into this weighted labeled hypergraph, right?
And that has both cognitive architecture importance,
because then this hypergraph is being used
as a sort of shared memory
among different cognitive processes,
but it also has software
and hardware implementation implications,
'cause current GPU architectures
are not so useful for OpenCog,
whereas a graph chip would be incredibly useful, right?
And I think GraphCore has those now,
but they're not ideally suited for this.
But I think in the next, let's say, three to five years,
we're gonna see new chips
where a graph is put on the chip,
and the back and forth between multiple processes
acting SIMD and MIMD on that graph is gonna be fast.
And then that may do for OpenCog-type architectures
what GPUs did for deep neural architectures.
- As a small tangent,
can you comment on thoughts about neuromorphic computing?
So like hardware implementations
of all these different kind of,
are you excited by that possibility?
- I'm excited by graph processors,
because I think they can massively speed up--
- OpenCog. - OpenCog,
which is a class of architectures that I'm working on.
I think if, in principle,
neuromorphic computing should be amazing.
I haven't yet been fully sold
on any of the systems that are out there.
Like memristors should be amazing too, right?
So a lot of these things have obvious potential,
but I haven't yet put my hands on a system
that seemed to manifest that potential.
- Mark system should be amazing,
but the current systems have not been great.
(laughs)
For example, if you wanted to make
a biologically realistic hardware neural network,
like making a circuit in hardware
that emulated like the Hodgkin-Huxley equation
or the Izhekevich equation,
like differential equations
for a biologically realistic neuron,
and putting that in hardware on the chip,
that would seem that it would make more feasible
to make a large scale, truly,
biologically realistic neural network.
No, what's been done so far is not like that.
So I guess, personally, as a researcher,
I mean, I've done a bunch of work in cognitive,
sorry, in computational neuroscience,
where I did some work with IARPA
in DC Intelligence Advanced Research Project Agency.
We were looking at how do you make
a biologically realistic simulation
of seven different parts of the brain
cooperating with each other,
using realistic non-linear dynamical models of neurons,
and how do you get that to simulate
what's going on in the mind of a GEOINT intelligence analyst
while they're trying to find terrorists on a map, right?
So if you wanna do something like that,
having neuromorphic hardware that really let you simulate
like a realistic model of the neuron would be amazing,
but that's sort of with my
computational neuroscience hat on, right?
With an AGI hat on,
I'm just more interested in these
hypergraph knowledge representation-based architectures,
which would benefit more
from various types of graph processors,
'cause the main processing bottleneck
is reading and writing to RAM.
It's reading and writing to the graph in RAM.
The main processing bottleneck
for this kind of proto-AGI architecture
is not multiplying matrices,
and for that reason,
GPUs, which are really good at multiplying matrices,
don't apply as well.
There are frameworks like Gunrock and others
that try to boil down graph processing to matrix operations,
and they're cool,
but you're still putting a square peg
into a round hole in a certain way.
The same is true,
I mean, current quantum machine learning,
which is very cool,
it's also all about how to get matrix and vector operations
in quantum mechanics,
and I see why that's natural to do.
I mean, quantum mechanics is all unitary matrices,
and vectors, right?
On the other hand,
you could also try to make graph-centric quantum computers,
which I think is where things will go,
and then we can take the OpenCog implementation layer,
implement it in an uncollapsed state
inside a quantum computer,
but that may be the singularity squared, right?
(Luke laughs)
I'm not sure we need that to get to human level.
- Singularity squared. - Human level.
- That's already beyond the first singularity,
but can we just--
- Yeah, let's go back to OpenCog.
- No, no, yeah, and the hypergraph and OpenCog.
- Yeah, yeah, that's the software framework, right?
So then the next thing is,
our cognitive architecture
tells us particular algorithms to put there.
- Got it.
Can we backtrack on the kind of,
is this graph designed,
is it, in general, supposed to be sparse,
and the operations constantly grow and change the graph?
- Yeah, the graph is sparse.
- And, but is it constantly adding links and so on?
- Yeah, it is a self-modifying hypergraph.
- So it's not, so the write and read operations
you're referring to,
this isn't just a fixed graph to which you change the way,
it's just a constantly growing graph.
- Yeah, that's true.
So it is different model than, say,
current deep neural nets
and have a fixed neural architecture,
and you're updating the weights,
although there have been, like,
cascade correlational neural net architectures
§13Transcript · 1:51:11–2:00:15
that grow new nodes and links,
but the most common neural architectures now
have a fixed neural architecture,
you're updating the weights,
and in OpenCog, you can update the weights,
and that certainly happens a lot,
but adding new nodes,
adding new links, removing nodes and links
is an equally critical part of the system's operations.
- Got it.
So now, when you start to add these cognitive algorithms
on top of this OpenCog architecture,
what does that look like?
- Yeah, so within this framework, then,
creating a cognitive architecture is basically two things.
It's choosing what type system
you wanna put on the nodes and links in the hypergraph,
what types of nodes and links you want,
and then it's choosing what collection of agents,
what collection of AI algorithms or processes
are gonna run to operate on this hypergraph,
and of course, those two decisions
are closely connected to each other.
So in terms of the type system,
there are some links that are more neural net-like,
they just have weights to get updated by Hebbian learning,
and activation spreads along them.
There are other links that are more logic-like,
and nodes that are more logic-like,
so you could have a variable node,
and you can have a node representing
a universal or existential quantifier,
as in predicate logic or term logic.
So you can have logic-like nodes and links,
or you can have neural-like nodes and links.
You can also have procedure-like nodes and links,
as in, say, a combinatory logic or lambda calculus
representing programs.
So you can have nodes and links
representing many different types of semantics,
which means you could make a horrible, ugly mess,
or you could make a system
where these different types of knowledge
all interpenetrate and synergize
with each other beautifully, right?
- So the hypergraph can contain programs.
- Yeah, it can contain programs,
although in the current version,
it is a very inefficient way
to guide the execution of programs,
which is one thing that we are aiming to resolve
with our rewrite of the system now.
- So what to you is the most beautiful aspect
of OpenCog?
Just to you personally,
some aspect that captivates your imagination
from beauty or power?
- What fascinates me is finding a common representation
that underlies abstract declarative knowledge
and sensory knowledge and movement knowledge
and procedural knowledge and episodic knowledge,
finding the right level of representation
where all these types of knowledge are stored
in a sort of universal and interconvertible
yet practically manipulable way, right?
So to me, that's the core,
'cause once you've done that,
then the different learning algorithms
can help each other out.
Like what you want is, if you have a logic engine
that helps with declarative knowledge,
and you have a deep neural net
that gathers perceptual knowledge,
and you have say an evolutionary learning system
that learns procedures,
you want these to not only interact
on the level of sharing results
and passing inputs and outputs to each other,
you want the logic engine when it gets stuck
to be able to share its intermediate state
with the neural net
and with the evolutionary learning algorithm
so that they can help each other out of bottlenecks
and help each other solve combinatorial explosions
by intervening inside each other's cognitive processes.
But that can only be done
if the intermediate state of a logic engine,
evolutionary learning engine,
and a deep neural net are represented in the same form.
And that's what we figured out how to do
by putting the right type system
on top of this weighted labeled hypergraph.
- So is there, can you maybe elaborate
on what are the different characteristics
of a type system that can coexist
amongst all these different kinds of knowledge
that needs to be represented?
And is, I mean, like is it hierarchical?
Just any kind of insights you can give
on that kind of type system?
- Yeah, yeah, so this gets very nitty gritty
and mathematical, of course.
But one key part is switching
from predicate logic to term logic.
- What is predicate logic?
What is term logic?
- So term logic was invented by Aristotle,
or at least that's the oldest recollection we have of it.
But term logic breaks down basic logic
into basically simple links between nodes,
like an inheritance link between node A and node B.
So in term logic, the basic deduction operation
is A implies B, B implies C, therefore A implies C.
Whereas in predicate logic,
the basic operation is modus ponens,
like A, A implies B, therefore B.
So it's a slightly different way of breaking down logic.
But by breaking down logic into term logic,
you get a nice way of breaking logic
down into nodes and links.
So your concepts can become nodes,
the logical relations become links.
And so then inference is like,
so if this link is A implies B,
this link is B implies C,
then deduction builds a link A implies C.
And your probabilistic algorithm
can assign a certain weight there.
Now, you may also have like a Hebbian neural link
from A to C, which is the degree to which thinking,
the degree to which A being the focus of attention
should make B the focus of attention, right?
So you could have then a neural link,
and you could have a symbolic,
like logical inheritance link in your term logic.
And they have separate meaning,
but they could be used to guide each other as well.
Like if there's a large amount of neural weight
on the link between A and B,
that may direct your logic engine to think about,
well, what is the relation?
Are they similar?
Is there an inheritance relation?
Are they similar in some context?
On the other hand, if there's a logical relation
between A and B, that may direct your neural component
to think, well, when I'm thinking about A,
or should I be directing some attention to B also,
because there's a logical relation.
So in terms of logic, there's a lot of thought
that went into how do you break down logic relations,
including basic sort of propositional logic relations
as Aristotelian term logic deals with,
and then quantifier logic relations also.
How do you break those down elegantly into a hypergraph?
'Cause you, I mean, you can boil logic expression
into a graph in many different ways.
Many of them are very ugly, right?
We tried to find elegant ways of sort of hierarchically
breaking down complex logic expression into nodes and links,
so that if you have, say, different nodes representing,
you know, Ben, AI, Lex, interview, or whatever,
the logic relations between those things are compact
in the node and link representation,
so that when you have a neural net acting
on those same nodes and links,
the neural net and the logic engine
can sort of interoperate with each other.
- And also interpretable by humans?
Is that an important-- - That's tough.
Yeah, in simple cases, it's interpretable by humans,
but honestly, you know,
I would say logic systems give more potential
for transparency and comprehensibility
than neural net systems, but you still have to work at it,
because, I mean, if I show you a predicate logic proposition
with, like, 500 nested universal
and existential quantifiers and 217 variables,
that's no more comprehensible than the weight matrix
of a neural network, right?
So I'd say the logic expressions that an AI learns
from its experience are mostly totally opaque
to human beings, and maybe even harder to understand
than a neural net, 'cause, I mean,
when you have multiple nested quantifier bindings,
it's a very high level of abstraction.
There is a difference, though, in that within logic,
it's a little more straightforward to pose the problem
of, like, normalize this and boil this down
to a certain form.
I mean, you can do that in neural nets, too.
Like, you can distill a neural net to a simpler form,
but that's more often done to make a neural net
that'll run on an embedded device or something.
It's harder to distill a net to a comprehensible form
than it is to simplify a logic expression
to a comprehensible form, but it doesn't come for free.
Like, what's in the AI's mind is incomprehensible
to a human unless you do some special work
to make it comprehensible.
§14Transcript · 2:00:16–2:09:05
So on the procedural side, there's some different
and sort of interesting voodoo there.
I mean, if you're familiar, in computer science,
there's something called the Curry-Howard correspondence,
which is a one-to-one mapping between proofs and programs.
So every program can be mapped into a proof.
Every proof can be mapped into a program.
You can model this using category theory
and a bunch of nice math,
but we wanna make that practical, right?
So that if you have an executable program
that moves a robot's arm or figures out
in what order to say things in a dialogue,
that's a procedure represented in OpenCog's hypergraph.
But if you wanna reason on how to improve that procedure,
you need to map that procedure into logic
using Curry-Howard isomorphism,
so then the logic engine can reason
about how to improve that procedure
and then map that back into the procedural representation
that is efficient for execution.
So again, that comes down to not just
can you make your procedure into a bunch of nodes and links,
'cause I mean, that can be done trivially.
A C++ compiler has nodes and links inside it.
Can you boil down your procedure
into a bunch of nodes and links
in a way that's hierarchically decomposed and simple enough?
- That you can reason about.
- Yeah, yeah, that given the resource constraints at hand,
you can map it back and forth to your term logic,
like fast enough and without having
a bloated logic expression, right?
So there's just a lot of nitty gritty particulars there.
But by the same token, if you ask a chip designer,
like how do you make the Intel i7 chip so good?
There's a long list of technical answers there,
which will take a while to go through, right?
And this has been decades of work.
I mean, the first AI system of this nature I tried to build
was called WebMind in the mid 1990s.
And we had a big graph, a big graph operating in RAM
implemented with Java 1.1,
which was a terrible, terrible implementation idea.
And then each node had its own processing.
So like there, the core loop looped through all nodes
in the network and let each node enact
what its little thing was doing.
And we had logic and neural nets in there,
but an evolutionary learning,
but we hadn't done enough of the math
to get them to operate together very cleanly.
So it was really, it was quite a horrible mess.
So as well as shifting an implementation
where the graph is its own object
and the agents are separately scheduled,
we've also done a lot of work
on how do you represent programs?
How do you represent procedures?
You know, how do you represent genotypes for evolution
in a way that the interoperability
between the different types of learning
associated with these different types of knowledge
actually works?
And that's been quite difficult.
It's taken decades and it's totally off to the side
of what the commercial mainstream of the AI field is doing,
which isn't thinking about representation at all, really.
Although you could see like in the DNC,
they had to think a little bit
about how do you make representation of a map
in this memory matrix work together
with the representation needed
for say visual pattern recognition
in a hierarchical neural network.
But I would say we have taken that direction
of taking the types of knowledge you need
for different types of learning,
like declarative, procedural, attentional,
and how do you make these types of knowledge
represent in a way that allows cross learning
across these different types of memory.
We've been prototyping and experimenting with this
within OpenCog and before that WebMind
since the mid 1990s.
Now, disappointingly to all of us,
this has not yet been cashed out in an AGI system, right?
I mean, we've used this system
within our consulting business.
So we've built natural language processing
and robot control and financial analysis.
We've built a bunch of sort of vertical market specific
proprietary AI projects.
They use OpenCog on the backend,
but we haven't, that's not the AGI goal, right?
It's interesting, but it's not the AGI goal.
So now what we're looking at with our rebuild of the system--
- 2.0.
- Yeah, we're also calling it true AGI.
So we're not quite sure what the name is yet.
We made a website for 2agi.io,
but we haven't put anything on there yet.
We may come up with an even better name.
- It's kind of like the real AI starting point
for your AGI book.
- Yeah, but I like true better
because true has like, you can be true hearted, right?
You can be true to your girlfriend.
So true has a number and it also has logic in it, right?
Because logic is a key part of the system.
- I like it, yeah.
- So yeah, with the true AGI system,
we're sticking with the same basic architecture,
but we're trying to build on what we've learned.
And one thing we've learned is that,
we need type checking among dependent types
to be much faster and among probabilistic dependent types
to be much faster.
So as it is now,
you can have complex types on the nodes and links,
but if you want to put,
like if you want types to be first class citizens,
so that you can have the types can be variables
and then you do type checking
among complex higher order types,
you can do that in the system now, but it's very slow.
This is stuff like is done
in cutting edge program languages like Agda or something,
these obscure research languages.
On the other hand, we've been doing a lot
tying together deep neural nets with symbolic learning.
So we did a project for Cisco, for example,
which was on, this was street scene analysis,
but they had deep neural models
for a bunch of cameras watching street scenes,
but they trained a different model for each camera
because they couldn't get the transfer learning
to work between camera A and camera B.
So we took what came out of all the deep neural models
for the different cameras,
we fed it into an open code symbolic representation,
then we did some pattern mining and some reasoning
on what came out of all the different cameras
within the symbolic graph.
And that worked well for that application.
I mean, Hugo Latapie from Cisco gave a talk
touching on that at last year's AGI conference,
it was in Shenzhen.
On the other hand, we learned from there,
it was kind of clunky to get the deep neural models
to work well with the symbolic system
because we were using Torch
and Torch keeps a sort of state computation graph,
but you needed like real time access
to that computation graph within our hypergraph.
And we certainly did it.
Alexey Polypov, who leads our St. Petersburg team
wrote a great paper on cognitive modules in OpenCog,
explaining sort of how do you deal
with the Torch compute graph inside OpenCog.
But in the end, we realized like,
that just hadn't been one of our design thoughts
when we built OpenCog, right?
So between wanting really fast dependent type checking
and wanting much more efficient interoperation
between the computation graphs of deep neural net frameworks
and OpenCog's hypergraph,
and adding on top of that,
wanting to more effectively run an OpenCog hypergraph
distributed across RAM and 10,000 machines,
which is, we're doing dozens of machines now,
but it's just not, we didn't architect it
with that sort of modern scalability in mind.
So these performance requirements
are what have driven us to want to re-architect the base,
but the core AGI paradigm doesn't really change.
Like the mathematics is the same.
It's just, we can't scale to the level that we want
in terms of distributed processing
or speed of various kinds of processing
with the current infrastructure
that was built in the phase 2001 to 2008,
which is hardly shocking, right?
- Well, I mean, the three things you mentioned
are really interesting.
So what do you think about, in terms of interoperability,
communicating with computational graph of neural networks,
what do you think about the representations
that neural networks form?
- They're bad, but there's many ways
that you could deal with that.
So I've been wrestling with this a lot
in some work on supervised grammar induction,
and I have a simple paper on that
that I'll give at the next AGI conference,
the online portion of which is next week, actually.
- What is grammar induction?
- So this isn't AGI either,
but it's sort of on the verge
between Neural AI and AGI or something.
§15Transcript · 2:09:08–2:17:46
On supervised grammar induction is the problem,
throw your AI system a huge body of text
and have it learn the grammar of the language
that produced that text.
So you're not giving it labeled examples.
So you're not giving it like a thousand sentences
where the parses were marked up by graduate students.
So it's just got to infer the grammar from the text.
It's like the Rosetta Stone, but worse, right?
'Cause you only have the one language,
and you have to figure out what is the grammar.
So that's not really AGI because,
I mean, the way a human learns language is not that, right?
I mean, we learn from language that's used in context.
So it's a social embodied thing.
We see how a given sentence is grounded in observation.
- As an interactive element, I guess.
- Yeah, yeah, yeah.
On the other hand, so I'm more interested in that.
I'm more interested in making an AGI system learn language
from its social and embodied experience.
On the other hand, that's also more of a pain to do,
and that would lead us into Hanson Robotics
and their robotics work I've known,
which we'll talk about in a few minutes.
But just as an intellectual exercise,
as a learning exercise, trying to learn grammar
from a corpus is very, very interesting, right?
And that's been a field in AI for a long time.
No one can do it very well.
So we've been looking at transformer neural networks
and tree transformers, which are amazing.
These came out of Google Brain, actually.
And actually, on that team was Lucas Kaiser,
who used to work for me in the period 2005 through '08
or something.
So it's been fun to see my former AGI employees disperse
and do all these amazing things.
Way too many sucked into Google, actually.
(laughing)
But yeah, anyway--
- We'll talk about that, too.
- Lucas Kaiser and a bunch of these guys,
they create transformer networks,
that classic paper like attention is all you need
and all these things following on from that.
So we're looking at transformer networks.
And these are able to, I mean, this is what underlies
GPT-2 and GPT-3 and so on, which are very, very cool
and have absolutely no cognitive understanding
of any of the texts we're looking at.
Like, they're very intelligent idiots, right?
- So, sorry to take, but to bring us back,
but do you think GPT-3 understands language?
- No, not at all, it understands nothing.
It's a complete idiot.
But it's a brilliant idiot.
- You don't think GPT-20 will understand language?
- No, no, no, no.
- So size is not gonna buy you understanding?
- Any more than a faster car is gonna get you to Mars.
It's a completely different kind of thing.
- I mean, these networks are very cool.
And as an entrepreneur, I can see many highly valuable uses
for them and as an artist, I love them, right?
So I mean, we're using our own neural model,
which is along those lines to control
the Philip K. Dick robot now.
And it's amazing to like train a neural model
on the robot Philip K. Dick and see it come up
with like crazed stoned philosopher pronouncements,
very much like what Philip K. Dick might have said, right?
Like these models are super cool.
And I'm working with Hanson Robotics now
on using a similar but more sophisticated one for SOFIA,
which we haven't launched yet.
But so I think it's cool.
But no, these are-- - But it's not understanding.
- These are recognizing a large number of shallow patterns.
They're not forming an abstract representation.
And that's the point I was coming to
when we're looking at grammar induction.
We tried to mine patterns out of the structure
of the transformer network.
And you can, but the patterns aren't what you want.
They're nasty.
So I mean, if you do supervised learning,
if you look at sentences where you know
the correct parts of a sentence,
you can learn a matrix that maps
between the internal representation of the transformer
and the parts of the sentence.
And so then you can actually train something
that will output the sentence parts
from the transformer network's internal state.
And we did this, I think, Christopher Manning,
some others have now done this also.
But I mean, what you get is that the representation
is horribly ugly and is scattered all over the network
and doesn't look like the rules of grammar
that you know are the right rules of grammar, right?
It's kind of ugly.
So what we're actually doing
is we're using a symbolic grammar learning algorithm,
but we're using the transformer neural network
as a sentence probability oracle.
So like, if you have a rule of grammar
and you aren't sure if it's a correct rule of grammar
or not, you can generate a bunch of sentences
using that rule of grammar and a bunch of sentences
violating that rule of grammar.
And you can see the transformer model
doesn't think the sentences obeying the rule of grammar
are more probable than the sentences
disobeying the rule of grammar.
So in that way, you can use the neural model
as a sense probability oracle to guide
a symbolic grammar learning process.
And that seems to work better than trying to milk
the grammar out of the neural network
that doesn't have it in there.
So I think the thing is these neural nets
are not getting a semantically meaningful representation
internally by and large.
So one line of research is to try to get them to do that.
And Infogam was trying to do that.
So like, if you look back like two years ago,
there was all these papers on like Edward,
this probabilistic programming neural net framework
that Google had, which came out of Infogam.
So the idea there was like, you could train
an Infogam neural net model, which is a generative
associative network to recognize and generate faces.
And the model would automatically learn a variable
for how long the nose is and automatically learn a variable
for how wide the eyes are or how big the lips are
or something, right?
So it automatically learned these variables,
which have a semantic meaning.
So that was a rare case where a neural net
trained with a fairly standard GAN method
was able to actually learn the semantic representation.
So for many years, many of us tried to take that
the next step and get a GAN type neural network
that would have not just a list of semantic latent variables,
but would have say a base net of semantic latent variables
with dependencies between them.
The whole programming framework, Edward, was made for that.
I mean, no one got it to work, right?
And it could be--
- Do you think it's possible?
Yeah, do you think--
- I don't know.
It might be that back propagation just won't work for it
because the gradients are too screwed up.
Maybe you could get it to work using CMAES
or some like floating point evolutionary algorithm.
We tried, we didn't get it to work.
Eventually, we just paused that rather than gave it up.
We paused that and said, well, okay,
let's try more innovative ways to learn implicit,
to learn what are the representations implicit
in that network without trying to make it grow
inside that network.
And I described how we're doing that in language.
You can do similar things in vision, right?
So what--
- Use it as an oracle.
- Yeah, yeah, yeah.
So you can, that's one way is that you use
a structure learning algorithm, which is symbolic.
And then you use the deep neural net
as an oracle to guide the structure learning algorithm.
The other way to do it is like Infogel was trying to do
and try to tweak the neural network
to have the symbolic representation inside it.
I tend to think what the brain is doing
is more like using the deep neural net type thing
as an oracle.
Like I think the visual cortex or the cerebellum
are probably learning a non-semantically meaningful
opaque tangled representation.
And then when they interface with the more cognitive parts
of the cortex, the cortex is sort of using those
as an oracle and learning the abstract representation.
So if you do sports, take for example,
serving in tennis, right?
I mean, my tennis serve is okay, not great,
but I learned it by trial and error, right?
And I mean, I learned music by trial and error too.
I just sit down and play.
But then if you're an athlete, which I'm not a good athlete,
I mean, and then you'll watch videos of yourself serving
and your coach will help you think about what you're doing.
And you'll then form a declarative representation,
but your cerebellum maybe didn't have
a declarative representation.
Same way with music, like I will hear something in my head.
I'll sit down and play the thing like I heard it.
And then I will try to study what my fingers did
§16Transcript · 2:17:51–2:27:10
to see like, what did you just play?
Like, how did you do that, right?
Because if you're composing,
you may wanna see how you did it
and then declaratively morph that in some way
that your fingers wouldn't think of, right?
But the physiological movement may come out
of some opaque like cerebellar reinforcement learn thing.
And so that's, I think, trying to milk the structure
of a neural net by treating it as an oracle
may be more like how your declarative mind post-processes
what your visual or motor cortex.
So I mean, in vision, it's the same way.
Like you can recognize beautiful art
much better than you can say why
you think that piece of art is beautiful.
But if you're trained as an art critic,
you do learn to say why.
And some of it's bullshit, but some of it isn't, right?
Some of it is learning to map sensory knowledge
into declarative and linguistic knowledge,
yet without necessarily making the sensory system itself
use a transparent and easily communicable representation.
- Yeah, that's fascinating.
To think of neural networks as like dumb question answers
that you can just milk to build up a knowledge base.
And that could be multiple networks, I suppose,
from different--
- Yeah, yeah.
So I think if a group like DeepMind or OpenAI
were to build AGI,
and I think DeepMind is like 1,000 times more likely
from what I could tell,
'cause they've hired a lot of people with broad minds
and many different approaches and angles on AGI,
whereas OpenAI is also awesome,
but I see them as more of like a pure
deep reinforcement learning shop.
- Yeah, this time I got you.
- So far. - Yeah, there's a lot of,
you're right, there's, I mean,
there's so much interdisciplinary work at DeepMind,
like neuroscience--
- And you put that together with Google Brain,
which, granted, they're not working that closely together
now, but my oldest son, Zarathustra,
is doing his PhD in machine learning applied
to automated theorem proving in Prague under Joseph Urban.
So the first paper, DeepMath,
which applied deep neural nets to guide theorem proving
was out of Google Brain.
I mean, by now, the automated theorem proving community
has gone way, way, way beyond anything Google was doing.
But still, yeah, but anyway,
if that community was gonna make an AGI,
probably one way they would do it was,
you know, take 25 different neural modules
architected in different ways,
maybe resembling different parts of the brain,
like a basal ganglia model, cerebellum model,
a thalamus module, a few hippocampus models,
number of different models
representing parts of the cortex, right?
Take all of these and then wire them together
to co-train and learn them together.
Like that would be an approach to creating an AGI.
One could implement something like that efficiently
on top of our true AGI, like OpenCog 2.0 system,
once it exists, although obviously Google
has their own highly efficient implementation architecture.
So I think that's a decent way to build AGI.
I was very interested in that in the mid '90s.
But I mean, the knowledge about how the brain works
sort of pissed me off, like it wasn't there yet.
Like, you know, in the hippocampus,
you have these concept neurons,
like the so-called grandmother neuron,
which everyone laughed at, it's actually there.
Like I have some Lex Friedman neurons
that fire differentially when I see you
and not when I see any other person, right?
So how do these Lex Friedman neurons,
how do they coordinate with the distributed representation
of Lex Friedman I have in my cortex, right?
There's some back and forth between cortex and hippocampus
that lets these discrete symbolic representations
in hippocampus correlate and cooperate
with the distributed representations in cortex.
This probably has to do with how the brain
does its version of abstraction and quantifier logic, right?
Like you can have a single neuron in the hippocampus
that activates a whole distributed activation pattern
in cortex, well, this may be how the brain
does like symbolization and abstraction,
as in functional programming or something.
But we can't measure it, like we don't have enough electrodes
stuck between the cortex and the hippocampus
in any known experiment to measure it.
So I got frustrated with that direction,
not 'cause it's impossible--
- 'Cause we just don't understand enough yet.
- Of course, it's a valid research direction,
you can try to understand more and more.
And we are measuring more and more
about what happens in the brain now than ever before.
So it's quite interesting.
On the other hand, I sort of got more of an engineering
mindset about AGI, I'm like, well, okay,
we don't know how the brain works that well,
we don't know how birds fly that well yet either.
We have no idea how a hummingbird flies
in terms of the aerodynamics of it.
On the other hand, we know basic principles
of like flapping and pushing the air down,
and we know the basic principles
of how the different parts of the brain work.
So let's take those basic principles
and engineer something that embodies
those basic principles, but is well designed
for the hardware that we have on hand right now.
- Yeah, so do you think we can create AGI
before we understand how the brain works?
- I think that's probably what will happen.
And maybe the AGI will help us do better brain imaging
that will then let us build artificial humans,
which is very, very interesting to us
because we are humans, right?
I mean, building artificial humans is super worthwhile.
I just think it's probably not the shortest path to AGI.
- So it's a fascinating idea that we would build AGI
to help us understand ourselves.
A lot of people ask me if young people
interested in doing artificial intelligence,
they look at sort of doing graduate level,
even undergrads, but graduate level research,
and they see where the artificial intelligence
community stands now.
It's not really AGI type research for the most part.
So the natural question they ask is,
what advice would you give?
I mean, maybe I could ask if people were interested
in working on OpenCog or in some kind of direct
or indirect connection to OpenCog or AGI research,
what would you recommend?
- OpenCog, first of all, is open source project.
There's a Google group discussion list.
There's a GitHub repository.
So if anyone's interested in lending a hand
with that aspect of AGI,
introduce yourself on the OpenCog email list.
And there's a Slack as well.
I mean, we're certainly interested to have inputs
into our redesign process for a new version of OpenCog,
but also we're doing a lot of very interesting research.
I mean, we're working on data analysis
for COVID clinical trials.
We're working with Hanson Robotics.
We're doing a lot of cool things
with the current version of OpenCog now.
So there's certainly opportunity to jump into OpenCog
or various other open source AGI oriented projects.
- So would you say there's like masters
and PhD theses in there?
- Plenty, yeah, plenty, of course.
I mean, the challenge is to find a supervisor
who wants to foster that sort of research,
but it's way easier than it was when I got my PhD, right?
- So, okay, great.
So you talked about OpenCog,
which is kind of one, the software framework,
but also the actual attempt to build an AGI system.
And then there is this exciting idea of SingularityNet.
So maybe, can you say first, what is SingularityNet?
- Sure, sure.
SingularityNet is a platform
for realizing a decentralized network
of artificial intelligences.
So Marvin Minsky, the AI pioneer who I knew a little bit,
he had the idea of a society of minds.
Like you should achieve an AI
not by writing one algorithm or one program,
but you should put a bunch of different AIs out there
and the different AIs will interact with each other,
each playing their own role.
And then the totality of the society of AIs
would be the thing that displayed
the human level intelligence.
And when he was alive, I had many debates with Marvin
about this idea.
And I think he really thought the mind
was more like a society than I do.
Like, I think you could have a mind
that was as disorganized as a human society,
but I think a human-like mind
has a bit more central control than that, actually.
I mean, we have this thalamus
and the medulla and limbic system.
We have a sort of top-down control system
that guides much of what we do,
more so than a society does.
§17Transcript · 2:27:12–2:35:51
So I think he stretched that metaphor a little too far,
but I also think there's something interesting there.
And so in the '90s, when I started my first
sort of non-academic AI project, WebMind,
which was an AI startup in New York
in the Silicon Alley area in the late '90s,
what I was aiming to do there
was make a distributed society of AIs,
the different parts of which would live
on different computers all around the world,
and each one would do its own thinking
about the data local to it.
But they would all share information with each other
and outsource work with each other and cooperate,
and the intelligence would be in the whole collective.
And I organized a conference together
with Francis Heiligin at Free University of Brussels in 2001,
which was the Global Brain Zero Conference.
And we're planning the next version,
the Global Brain One Conference,
at the Free University of Brussels for next year, 2021,
so 20 years after.
And then maybe we can have the next one 10 years after that,
like exponentially faster until the singularity comes, right?
- The timing is right, yeah.
- Yeah, yeah, exactly.
So yeah, the idea with the global brain was,
maybe the AI won't just be in a program
on one guy's computer,
but the AI will be in the internet as a whole
with the cooperation of different AI modules
living in different places.
So one of the issues you face
when architecting a system like that
is how is the whole thing controlled?
Do you have a centralized control unit
that pulls the puppet strings
of all the different modules there?
Or do you have a fundamentally decentralized network
where the society of AIs is controlled
in some democratic and self-organized way
by all the AIs in that society, right?
And Francis and I had different view of many things,
but we both wanted to make a global society of AI minds
with a decentralized organizational mode.
Now, the main difference was he wanted the individual AIs
to be all incredibly simple
and all the intelligence to be on the collective level.
Whereas I thought that was cool,
but I thought a more practical way to do it might be
if some of the agents in the society of minds
were fairly generally intelligent on their own.
So like you could have a bunch of open cogs out there
and a bunch of simpler learning systems,
and then these are all cooperating, coordinating together,
sort of like in the brain.
Okay, the brain as a whole is the general intelligence,
but some parts of the cortex,
you could say have a fair bit of general intelligence
on their own,
whereas say parts of the cerebellum or limbic system
have very little general intelligence on their own,
and they're contributing to general intelligence
by way of their connectivity to other modules.
- Do you see instantiations of the same kind of,
maybe different versions of open cog,
but also just the same version of open cog
and maybe many instantiations
of it as being all parts of it?
- That's what David and Hans and I wanna do
with many Sophia and other robots.
Each one has its own individual mind living on a server,
but there's also a collective intelligence infusing them
and a part of the mind living on the edge in each robot.
So the thing is, at that time,
as well as WebMind being implemented in Java 1.1
as like a massive distributed system,
blockchain wasn't there yet.
So how did it have them do this decentralized control?
You know, we sort of knew it.
We knew about distributed systems.
We knew about encryption.
So I mean, we had the key principles
of what underlies blockchain now,
but I mean, we didn't put it together
in the way that it's been done now.
So when Vitalik Buterin and colleagues
came out with Ethereum blockchain,
many, many years later, like 2013 or something,
then I was like, well, this is interesting.
Like this is Solidity scripting language.
It's kind of dorky in a way,
and I don't see why you need a turn complete language
for this purpose.
But on the other hand,
this is like the first time I could sit down
and start to like script infrastructure
for decentralized control of the AIs
in a society of minds in a tractable way.
Like you can hack the Bitcoin code base,
but it's really annoying.
Whereas Solidity is Ethereum scripting language
is just nicer and easier to use.
I'm very annoyed with it by this point.
But like Java, I mean, these languages are amazing
when they first come out.
So then I came up with the idea
that turned into Singularity Net.
Okay, let's make a decentralized agent system
where a bunch of different AIs wrapped up
and say different Docker containers or LXC containers,
different AIs can each of them have their own identity
on the blockchain.
And the coordination of this community of AIs
has no central controller, no dictator, right?
And there's no central repository of information.
The coordination of the society of minds
is done entirely by the decentralized network
in a decentralized way by the algorithms, right?
'Cause the motto of Bitcoin is in math we trust, right?
And so that's what you need.
You need the society of minds to trust only in math,
not trust only in one centralized server.
- So the AI systems themselves are outside of the blockchain
but then the communication between--
- At the moment, yeah, yeah.
I would have loved to put the AI's operations
on chain in some sense,
but in Ethereum it's just too slow.
You can't deal with it.
- Somehow it's the basic communication
between AI systems that's--
- Yeah, yeah, so basically an AI is just some software
in Singularity, an AI is just some software process
living in a container.
And-- - There's input and output.
- There's a proxy that lives in that container
along with the AI that handles the interaction
with the rest of Singularity Net.
And then when one AI wants to contribute
with another one in the network,
they set up a number of channels.
And the setup of those channels
uses the Ethereum blockchain.
But once the channels are set up,
then data flows along those channels
without having to be on the blockchain.
All that goes on the blockchain is the fact
that some data went along that channel.
So you can do--
- So there's not a shared knowledge, it's--
- Well, the identity of each agent is on the blockchain.
- Right. - On the Ethereum blockchain.
If one agent rates the reputation of another agent,
that goes on the blockchain.
And agents can publish what APIs they will fulfill
on the blockchain.
But the actual data for AI and the results--
- It's not on the blockchain. - AI is not on the blockchain.
- Do you think it could be, do you think it should be?
- In some cases it should be.
In some cases maybe it shouldn't be.
But I mean, I think that--
- So I'll give you an example.
- Using Ethereum, you can't do it.
Using, now there's more modern and faster blockchains
where you could start to do that in some cases.
Two years ago that was less so.
It's a very rapidly evolving ecosystem.
- So like one example maybe you can comment on.
Something I worked a lot on is autonomous vehicles.
You can see each individual vehicle as a AI system.
And you can see vehicles from Tesla, for example,
and then Ford and GM and all these as also like larger.
I mean, they all are running the same kind of system
on each sets of vehicles.
So it's individual AI systems on individual vehicles,
but it's all different, the statiation is the same AI system
within the same company.
So you can envision a situation
where all of those AI systems are put on SingularityNet.
Right? - Yeah.
- And how do you see that happening
and what would be the benefit?
And could they share data?
I guess one of the biggest things is the power there
is in the decentralized control,
but the benefit would have been,
is really nice if they can somehow share the knowledge
in an open way if they choose to.
- Yeah, yeah, yeah.
Those are all quite good points.
So I think the benefit from being on the data side
on the decentralized network as we envision it
is that we want the AIs in the network
to be outsourcing work to each other
and making API calls to each other frequently.
- I got you.
- So the real benefit would be if that AI
wanted to outsource some cognitive processing
§18Transcript · 2:35:54–2:44:53
or data processing or data pre-processing, whatever,
to some other AIs in the network
which specialize in something different.
And this really requires a different way of thinking
about AI software development, right?
So just like object-oriented programming
was different than imperative programming,
and now object-oriented programmers
all use these frameworks to do things
rather than just libraries even.
You know, shifting to agent-based programming
where AI agent is asking other like live,
real-time evolving agents for feedback
in what they're doing, that's a different way of thinking.
I mean, it's not a new one.
There was loads of papers on agent-based programming
in the '80s and onward.
But if you're willing to shift
to an agent-based model of development,
then you can put less and less in your AI
and rely more and more on interactive calls
to other AIs running in the network.
And of course, that's not fully manifested yet
because although we've rolled out
a nice working version of Singularity Net Platform,
there's only 50 to 100 AIs running in there now.
There's not tens of thousands of AIs.
So we don't have the critical mass
for the whole society of mind to be doing what we want.
- Yeah, the magic really happens
when there's just a huge number of agents.
- Yeah, yeah, exactly.
In terms of data, we're partnering closely
with another blockchain project called Ocean Protocol.
And Ocean Protocol, that's the project of Trent McConaghy
who developed BigchainDB,
which is a blockchain-based database.
So Ocean Protocol is basically blockchain-based big data
and aims at making it efficient
for different AI processes or statistical processes
or whatever to share large data sets.
Or one process can send a clone of itself
to work on the other guy's data set
and send results back and so forth.
So by getting Ocean and you have data lakes,
so this is the data ocean, right?
That's the game.
By getting Ocean and SingularityNet to interoperate,
we're aiming to take into account of the big data aspect.
Also, but it's quite challenging
because to build this whole decentralized
blockchain-based infrastructure,
I mean, your competitors are like Google, Microsoft,
Alibaba, and Amazon, which have so much money
to put behind their centralized infrastructures.
Plus they're solving simpler algorithmic problems
'cause making it centralized in some ways is easier, right?
So they're very major computer science challenges.
And I think what you saw with the whole ICO boom
in the blockchain and cryptocurrency world
is a lot of young hackers who are hacking Bitcoin
or Ethereum and they see,
well, why don't we make this decentralized on blockchain?
Then after they raise some money through an ICO,
they realize how hard it is.
Like actually we're wrestling with incredibly hard
computer science and software engineering
and distributed systems problems, which can be solved,
but they're just very difficult to solve.
And in some cases, the individuals
who started those projects were not well-equipped
to actually solve the problems that they wanted to.
- So you think, would you say that's the main bottleneck?
If you look at the future of currency,
the question is--
- Currency, the main bottleneck is politics.
Like it's governments and the bands of armed thugs
that will shoot you if you bypass
their currency restrictions.
- That's right, so like your sense is that
versus the technical challenges,
'cause you kind of just suggested
the technical challenges are quite high as well.
- I mean, for making a distributed money,
you could do that on Algorand right now.
I mean, so that while Ethereum is too slow,
there's Algorand and there's a few other more modern,
more scalable blockchains that would work fine
for a decentralized global currency.
So I think there were technical bottlenecks
to that two years ago, and maybe Ethereum 2.0
will be as fast as Algorand.
I don't know, that's not fully written yet, right?
So I think the obstacle to currency
being put on the blockchain is that--
- Is the other stuff you mentioned.
- I mean, currency will be on the blockchain.
It'll just be on the blockchain in a way
that enforces centralized control
and government hegemony rather than otherwise.
Like the ERMB will probably be the first global,
the first currency on the blockchain.
The E-Ruble may be next.
There already-- - E-Ruble?
- Yeah, yeah, yeah.
I mean, the point is-- - Oh, that's hilarious.
- Digital currency makes total sense,
but they would rather do it in the way
that Putin and Xi Jinping have access
to the global keys for everything, right?
- So, and then the analogy to that
in terms of singularity net, I mean, there's echoes.
I think you've mentioned before that Linux gives you hope.
- AI is not as heavily regulated as money, right?
- Not yet, right?
Not yet.
- Oh, that's a lot slipperier than money too, right?
I mean, money is easier to regulate
'cause it's kind of easier to define,
whereas AI is, it's almost everywhere inside everything.
Where's the boundary between AI and software, right?
I mean, if you're gonna regulate AI,
there's no IQ test for every hardware device
that has a learning algorithm.
You're gonna be putting hegemonic regulation
on all software, and I don't rule out that that could happen.
- Any adaptive software.
- Yeah, but how do you tell if a software's adaptive?
Every software's gonna be adaptive, I mean.
- Or maybe we're living in the golden age of open source
that will not always be open.
Maybe it'll become centralized control
of software by governments.
- It is entirely possible, and part of what I think
we're doing with things like singularity net protocol
is creating a tool set that can be used
to counteract that sort of thing.
Say a similar thing about mesh networking, right?
Plays a minor role now, the ability to access internet
like directly phone to phone.
On the other hand, if your government starts trying
to control your use of the internet,
suddenly having mesh networking there
can be very convenient, right?
So right now, something like a decentralized
blockchain-based AGI framework, or narrow AI framework,
it's cool, it's nice to have.
On the other hand, if governments start trying
to tamp down on my AI interoperating with someone's AI
in Russia or somewhere, then suddenly having
a decentralized protocol that nobody owns or controls
becomes an extremely valuable part of the tool set.
And we've put that out there now.
It's not perfect, but it operates.
And it's pretty blockchain agnostic.
So we're talking to Algorand about making part
of singularity run on Algorand.
My good friend Tufi Saliba has a cool blockchain project
called Toda, which is a blockchain
without a distributed ledger.
It's like a whole other architecture.
- Interesting.
- So there's a lot of more advanced things
you can do in the blockchain world.
Singularity net could be ported to a whole bunch of,
it could be made multi-chain and ported
to a whole bunch of different blockchains.
And there's a lot of potential and a lot of importance
to putting this kind of tool set out there.
If you compare to OpenCog, what you could see
is OpenCog allows tight integration of a few AI algorithms
that share the same knowledge store in real time,
in RAM, right?
Singularity net allows loose integration
of multiple different AIs.
They can share knowledge, but they're mostly
not gonna be sharing knowledge in RAM on the same machine.
And I think what we're gonna have is a network
of network of networks, right?
Like, I mean, you have the knowledge graph
inside the OpenCog system.
And then you have a network of machines
inside distributed OpenCog mind.
But then that OpenCog will interface with other AIs
doing deep neural nets or custom biology data analysis
or whatever they're doing in Singularity net,
which is a looser integration of different AIs,
some of which may be their own networks, right?
And I think at a very loose analogy,
you could see that in the human body.
Like the brain has regions like cortex or hippocampus,
which tightly interconnect like conical columns
within the cortex, for example.
Then there's looser connection
within the different lobes of the brain.
And then the brain interconnects with the endocrine system
and different parts of the body even more loosely.
Then your body interacts even more loosely
with the other people that you talk to.
So you often have networks within networks
§19Transcript · 2:44:55–2:53:37
within networks with progressively looser coupling
as you get higher up in that hierarchy.
I mean, you have that in biology,
you have it in the internet as a just networking medium.
And I think that's what we're gonna have
in the network of software processes leading to AGI.
- That's a beautiful way to see the world.
Again, the same similar question is with OpenCog.
If somebody wanted to build an AI system
and plug into the Singularity net,
what would you recommend?
Like how would you go about it?
- So that's much easier.
I mean, OpenCog is still a research system.
So it takes some expertise and some time.
We have tutorials, but it's somewhat
cognitively labor intensive to get up to speed on OpenCog.
And I mean, what's one of the things we hope to change
with the true AGI OpenCog 2.0 version
is just make the learning curve more similar
to TensorFlow or Torch or something.
'Cause right now, OpenCog is amazingly powerful,
but not simple to do.
On the other hand, Singularity net,
as a open platform was developed
a little more with usability in mind,
although the blockchain is still kind of a pain.
So I mean, if you're a command line guy,
there's a command line interface.
It's quite easy to take any AI that has an API
and lives in a Docker container and put it online anywhere.
And then it joins the global Singularity net.
And anyone who puts a request for services
out into the Singularity net,
the peer to peer discovery mechanism will find your AI.
And if it does what was asked,
it can then start a conversation with your AI
about whether it wants to ask your AI to do something for it,
how much it would cost and so on.
So that's fairly simple.
If you wrote an AI and want it listed
on like official Singularity net marketplace,
which is on our website,
then we have a publisher portal
and then there's a KYC process to go through
'cause then we have some legal liability
for what goes on that website.
So that in a way that's been an education too.
There's sort of two layers.
Like there's the open decentralized protocol.
- And there's the market.
- Yeah, anyone can use the open decentralized protocol.
So say some developers from Iran
and there's brilliant AI guys
in University of Isfahan in Tehran,
they can put their stuff on Singularity net protocol
and just like they can put something on the internet.
I don't control it.
But if we're gonna list something
on the Singularity net marketplace
and put a little picture and a link to it,
then if I put some Iranian AI genius's code on there,
then Donald Trump can send a bunch of jackbooted thugs
to my house to arrest me for doing business with Iran.
So I mean, we already see in some ways
the value of having a decentralized protocol
'cause what I hope is that someone in Iran
will put online an Iranian Singularity net marketplace,
which you can pay in a cryptographic token,
which is not owned by any country.
And then if you're in like Congo or somewhere
that doesn't have any problem with Iran,
you can subcontract AI services
that you find on that marketplace,
even though US citizens can't by US law.
So right now, that's kind of a minor point.
As you alluded, if regulations go in the wrong direction,
it could become more of a major point.
But I think it also is the case
that having these workarounds to regulations in place
is a defense mechanism against those regulations
being put into place.
And you can see that in the music industry, right?
I mean, Napster just happened and BitTorrent just happened.
And now most people in my kids' generation,
they're baffled by the idea of paying for music, right?
I mean, my dad pays for music.
(laughing)
- Yeah, that's true.
- Because these decentralized mechanisms happened,
and then the regulations followed, right?
And the regulations would be very different
if they'd been put into place
before there was Napster and BitTorrent and so forth.
So in the same way, we got to put AI out there
in a decentralized vein,
and big data out there in a decentralized vein now,
so that the most advanced AI in the world
is fundamentally decentralized.
And if that's the case,
that's just the reality the regulators have to deal with.
And then as in the music case,
they're gonna come up with regulations
that sort of work with the decentralized reality.
- Beautiful.
You were the chief scientist of Hanson Robotics.
You're still involved with Hanson Robotics,
doing a lot of really interesting stuff there.
This is, for people who don't know,
the company that created Sophia the Robot.
Can you tell me who Sophia is?
- I'd rather start by telling you who David Hanson is.
(Luke laughs)
David is the brilliant mind behind the Sophia Robot,
and he remains, so far he remains more interesting
than his creation,
although she may be improving faster than he is, actually.
I mean, he's-- - Yeah.
- So yeah, I met-- - That's a good point.
- I met David maybe 2007 or something
at some futurist conference.
We were both speaking at.
And I could see we had a great deal in common.
I mean, we were both kind of crazy,
but we also, we both had a passion for AGI
and the singularity,
and we were both huge fans of the work of Philip K. Dick,
the science fiction writer.
And I wanted to create benevolent AGI
that would create massively better life
for all humans and all sentient beings,
including animals, plants, and superhuman beings.
And David, he wanted exactly the same thing,
but he had a different idea of how to do it.
He wanted to get computational compassion.
Like, he wanted to get machines
that would love people and empathize with people.
And he thought the way to do that
was to make a machine that could look people eye to eye,
face to face, look at people,
and make people love the machine,
and the machine loves the people back.
So I thought that was a very different way of looking at it,
'cause I'm very math-oriented,
and I'm just thinking, like,
what is the abstract cognitive algorithm
that will let the system internalize
the complex patterns of human values, blah, blah, blah,
whereas he's like, look you in the face and the eye
and love you, right?
So we hit it off quite well.
And we talked to each other off and on.
Then I moved to Hong Kong in 2011.
So I'd been, I mean, I've been living all over the place.
I've been in Australia and New Zealand
in my academic career, then in Las Vegas for a while,
was in New York in the late '90s,
starting my entrepreneurial career,
was in DC for nine years
doing a bunch of US government consulting stuff,
then moved to Hong Kong in 2011,
mostly 'cause I met a Chinese girl
who I fell in love with, and we got married.
She's actually not from Hong Kong,
she's from mainland China,
but we converged together in Hong Kong,
still married now, have a two-year-old baby.
- So went to Hong Kong to see about a girl, I guess.
- Yeah, pretty much, yeah.
And on the other hand,
I started doing some cool research there
with Jin You at Hong Kong Polytechnic University.
I got involved with a project called IDEA
using machine learning for stock and futures prediction,
which was quite interesting.
And I also got to know something
about the consumer electronics
and hardware manufacturer ecosystem
in Shenzhen across the border,
which is like the only place in the world
that makes sense to make complex consumer electronics
at large scale and low cost.
It's just, it's astounding,
the hardware ecosystem that you have in South China.
Like US, people here cannot imagine
what it's like.
So David was starting to explore that also.
I invited him to Hong Kong to give a talk
at Hong Kong PolyU,
and I introduced him in Hong Kong
to some investors who were interested in his robots.
And he didn't have Sophia then,
he had a robot of Philip K. Dick,
our favorite science fiction writer.
He had a robot Einstein,
he had some little toy robots
that looked like his son, Zeno.
So through the investors I connected him to,
he managed to get some funding
to basically port Hanson Robotics to Hong Kong.
§20Transcript · 2:53:40–3:03:13
And when he first moved to Hong Kong,
I was working on AGI research
and also on this machine learning trading project.
So I didn't get that tightly involved
with Hanson Robotics.
But as I hung out with David more and more,
as we were both there in the same place,
I started to get,
I started to think about what you could do
to make his robots smarter than they were.
And so we started working together.
And for a few years,
I was chief scientist and head of software
at Hanson Robotics.
Then when I got deeply into the blockchain side of things,
I stepped back from that and co-founded SingularityNet.
David Hanson was also one of the co-founders
of SingularityNet.
So part of our goal there had been
to make the blockchain based like cloud mind platform
for Sophia and the other--
- Sophia would be just one of the robots
in SingularityNet.
- Yeah, yeah, yeah, exactly.
Sophia, many copies of the Sophia robot
would be among the user interfaces
to the globally distributed SingularityNet cloud mind.
And I mean, David and I talked about that
for quite a while before co-founding SingularityNet.
- By the way, in his vision and your vision,
was Sophia tightly coupled to a particular AI system
or was the idea that you can plug,
you could just keep plugging in different AI systems
within the head of it?
- I think David's view was always that Sophia
would be a platform,
much like say the Pepper robot is a platform from SoftBank.
Should be a platform with a set of nicely designed APIs
that anyone can use to experiment
with their different AI algorithms on that platform.
And SingularityNet of course fits right into that, right?
'Cause SingularityNet, it's an API marketplace.
So anyone can put their AI on there.
OpenCog is a little bit different.
I mean, David likes it, but I'd say it's my thing.
It's not his.
David has a little more passion
for biologically based approaches to AI than I do,
which makes sense.
I mean, he's really into human physiology and biology.
He's a character sculptor, right?
So yeah, he's interested in,
but he also worked a lot with rule-based
and logic-based AI systems too.
So yeah, he's interested in not just Sophia,
but all the Hanson robots as a powerful social
and emotional robotics platform.
And what I saw in Sophia was a way to get AI algorithms
out there in front of a whole lot of different people
in an emotionally compelling way.
And part of my thought was really kind of abstract,
connected to AGI ethics.
And many people are concerned AGI
is gonna enslave everybody or turn everybody
into computronium to make extra hard drives
for their cognitive engine or whatever.
And emotionally, I'm not driven to that sort of paranoia.
I'm really just an optimist by nature,
but intellectually, I have to assign a non-zero probability
to those sorts of nasty outcomes.
'Cause if you're making something 10 times as smart as you,
how can you know what it's gonna do?
There's an irreducible uncertainty there,
just as my dog can't predict what I'm gonna do tomorrow.
So it seemed to me that based on our current state
of knowledge, the best way to bias the AGI's we create
toward benevolence would be to infuse them with love
and compassion the way that we do our own children.
So you want to interact with AIs in the context
of doing compassionate, loving, and beneficial things.
And in that way, as your children will learn,
by doing compassionate, beneficial, loving things
alongside you, and that way the AI will learn in practice
what it means to be compassionate, beneficial, and loving.
It will get a sort of ingrained intuitive sense of this,
which it can then abstract in its own way
as it gets more and more intelligent.
Now David saw this the same way.
That's why he came up with the name Sophia,
which means wisdom.
So it seemed to me making these beautiful, loving robots
to be rolled out for beneficial applications
would be the perfect way to roll out early stage AGI systems
so they can learn from people,
and not just learn factual knowledge,
but learn human values and ethics from people
while being their home service robots,
their education assistants, their nursing robots.
So that was the grand vision.
Now if you've ever worked with robots,
the reality is quite different, right?
Like the first principle is the robot is always broken.
(laughing)
I mean, I worked with robots in the '90s a bunch
when you had to solder them together yourself,
and I'd put neural nets doing reinforcement learning
on overturned solid bowl type robots in the '90s
when I was a professor.
Things, of course, advanced a lot, but--
- But the principle still holds.
- The principle of the robot's always broken still holds.
Yeah, so faced with the reality of making Sophia do stuff,
many of my robo-AGI aspirations were temporarily cast aside.
And I mean, there's just a practical problem
of making this robot interact in a meaningful way,
'cause you put nice computer vision on there,
but there's always glare.
And then, or you have a dialogue system,
but at the time I was there,
like no speech-to-text algorithm could deal
with Hong Kongese people's English accents.
So the speech-to-text was always bad,
so the robot always sounded stupid
because it wasn't getting the right text, right?
So I started to view that really as what,
what in software engineering you call a walking skeleton,
which is maybe the wrong metaphor to use for Sophia,
or maybe the right one.
I mean, where the walking skeleton is
in software development is,
if you're building a complex system, how do you get started?
Well, one way is to first build part one well,
then build part two well,
then build part three well, and so on.
And the other way is you make like a simple version
of the whole system and put something in the place
of every part the whole system will need,
so that you have a whole system that does something,
and then you work on improving each part
in the context of that whole integrated system.
So that's what we did on a software level in Sophia.
We made like a walking skeleton software system,
where so there's something that sees,
there's something that hears,
there's something that moves,
there's something that remembers,
there's something that learns.
You put a simple version of each thing in there,
and you connect them all together,
so that the system will do its thing.
So there's a lot of AI in there.
There's not any AGI in there.
I mean, there's computer vision to recognize people's faces,
recognize when someone comes in the room and leaves,
try to recognize whether two people are together or not.
I mean, the dialogue system,
it's a mix of like hand-coded rules with deep neural nets
that come up with their own responses.
And there's some attempt to have a narrative structure
and sort of try to pull the conversation
into something with a beginning, middle, and end
in this sort of story arc.
So it's.
- I mean, like if you look at the Lubner Prize
and the systems that beat the Turing test currently,
they're heavily rule-based,
because like you had said, narrative structure,
to create compelling conversations,
you currently, neural networks cannot do that well,
even with Google MENA.
When you actually look at full-scale conversations,
it's just not. - Yeah, this is the thing.
So we've been, I've actually been running an experiment
the last couple of weeks, taking Sophia's chatbot,
and then Facebook's transformer chatbot,
which they opened the model.
We've had them chatting to each other
for a number of weeks on the server, just.
- That's funny.
- We're generating training data of what Sophia says
in a wide variety of conversations.
But we can see, compared to Sophia's current chatbot,
the Facebook deep neural chatbot
comes up with a wider variety of fluent-sounding sentences.
On the other hand, it rambles like mad.
The Sophia chatbot, it's a little more repetitive
in the sentence structures it uses.
On the other hand, it's able to keep a conversation arc
over a much longer period, right?
So now, you can probably surmount that using Reformer
and using various other deep neural architectures
to improve the way these transformer models are trained.
But in the end, neither one of them
really understands what's going on.
And I mean, that's the challenge I had with Sophia,
is if I were doing a robotics project aimed at AGI,
I would wanna make like a robo-toddler
that was just learning about what it was seeing,
'cause then the language is grounded
in the experience of the robot.
§21Transcript · 3:03:14–3:13:00
But what Sophia needs to do to be Sophia
is talk about sports or the weather or robotics
or the conference she's talking at.
She needs to be fluent talking about
any damn thing in the world,
and she doesn't have grounding for all those things.
So there's this, just like, I mean,
Google Mina and Facebook's chat,
but don't have grounding
for what they're talking about either.
So in a way, the need to speak fluently
about things where there's no non-linguistic grounding
pushes what you can do for Sophia in the short term
a bit away from AGI.
- I mean, it pushes you towards IBM Watson situation
where you basically have to do heuristic
and hard-code stuff and rule-based stuff.
I have to ask you about this.
Okay, so because, in part, Sophia is an art creation
because it's beautiful.
She's beautiful because she inspires
through our human nature of anthropomorphize things.
We immediately see an intelligent being there.
- Because David is a great sculptor.
- Is a great sculptor, that's right.
So in fact, if Sophia just had nothing inside her head,
said nothing, if she just sat there,
we already prescribed some intelligence to her.
- There's a long selfie line in front of her
after every talk.
- That's right.
So it captivated the imagination of many people.
I wasn't gonna say the world,
but yeah, I mean, a lot of people.
- Billions of people, which is amazing.
- It's amazing, right.
Now, of course, many people have prescribed
essentially AGI type of capabilities to Sophia
when they see her.
And of course, friendly French folk like Yann LeCun
immediately see that, the people from the AI community,
and get really frustrated because--
- It's understandable.
- And so what, and then they criticize people like you
who sit back and don't say anything about,
basically allow the imagination of the world,
allow the world to continue being captivated.
So what's your sense of that kind of annoyance
that the AI community has?
- I think there's several parts to my reaction there.
First of all, if I weren't involved with Hanson Robots
and didn't know David Hanson personally,
I probably would have been very annoyed initially
at Sophia as well.
I mean, I can understand the reaction.
I would have been like, wait, all these stupid people
out there think this is an AGI, but it's not an AGI,
but they're tricking people that this very cool robot
is an AGI.
And now those of us trying to raise funding to build AGI,
people will think it's already there and already works.
So on the other hand, I think,
even if I weren't directly involved with it,
once I dug a little deeper into David and the robot
and the intentions behind it,
I think I would have stopped being pissed off,
whereas folks like Jan LeCun have remained pissed off
after their initial reaction.
- That's his thing, that's his thing.
- I think that in particular struck me as somewhat ironic
because Jan LeCun is working for Facebook,
which is using machine learning to program the brains
of the people in the world toward vapid consumerism
and political extremism.
So if your ethics allows you to use machine learning
in such a blatantly destructive way,
why would your ethics not allow you to use machine learning
to make a lovable theatrical robot
that draws some foolish people into its theatrical illusion?
Like if the pushback had come from Yoshua Bengio,
I would have felt much more humbled by it
because he's not using AI for blatant evil, right?
On the other hand, he also is a super nice guy
and doesn't bother to go out there
trashing other people's work for no good reason.
- Shots fired, but I get you.
- I mean, if you're gonna ask, I'm gonna answer.
- No, for sure.
I think we'll go back and forth.
I'll talk to Jan again.
- I would add on this, though.
David Hanson is an artist and he often speaks off the cuff,
and I have not agreed with everything
that David has said or done regarding Sophia,
and David also does not agree with everything
David has said or done about Sophia.
I mean, David is an artistic wild man,
and that's part of his charm.
That's part of his genius.
So certainly there have been conversations
within Hanson Robotics and between me and David
where I was like, let's be more open
about how this thing is working,
and I did have some influence in nudging Hanson Robotics
to be more open about how Sophia was working,
and David wasn't especially opposed to this,
and he was actually quite right about it.
What he said was, you can tell people
exactly how it's working, and they won't care.
They want to be drawn into the illusion,
and he was 100% correct.
I'll tell you what, this wasn't Sophia.
This was Philip K. Dick, but we did some interactions
between humans and Philip K. Dick robot
in Austin, Texas a few years back,
and in this case, the Philip K. Dick was just teleoperated
by another human in the other room.
So during the conversations, we didn't tell people
the robot was teleoperated.
We just said, here, have a conversation with Phil Dick.
We're gonna film you, right?
And they had a great conversation with Philip K. Dick,
teleoperated by my friend, Stephan Bugay.
After the conversation, we brought the people
in the back room to see Stephan,
who was controlling the Philip K. Dick robot,
but they didn't believe it.
These people were like, well, yeah,
but I know I was talking to Phil.
Maybe Stephan was typing, but the spirit of Phil
was animating his mind while he was typing.
So even though they knew it was a human in the loop,
even seeing the guy there, they still believed
that was Phil they were talking to.
A small part of me believes that they were right, actually.
Because our understanding--
- Well, we don't understand the universe.
- That's the thing. - I mean, there is
a cosmic mind field that we're all embedded in
that yields many strange synchronicities in the world,
which is a topic we don't have time to go into too much.
- Yeah, I mean, there's something to this
where our imagination about Sophia
and people like Jan LeCun being frustrated about it
is all part of this beautiful dance
of creating artificial intelligence that's almost essential.
You see with Boston Dynamics, I'm a huge fan of as well.
I mean, these robots are very far from intelligent.
- I played with their last one, actually.
- With the spot mini? - Yeah, very cool.
It reacts quite in a fluid and flexible way.
- But we immediately ascribe the kind of intelligence,
we immediately ascribe AGI to them.
- Yeah, yeah, if you kick it and it falls down
and goes, "Ow," you feel bad, right?
You can't help it. - You feel bad.
And I mean, that's part of,
that's gonna be part of our journey
in creating intelligent systems,
more and more and more and more.
Like, as Sophia starts out with a walking skeleton,
as you add more and more intelligence,
I mean, we're gonna have to deal with this kind of idea.
- Absolutely, and about Sophia, I would say,
I mean, first of all, I have nothing against Jan LeCun.
- No, no, this is fun, this is all for fun.
- He's a nice guy.
If he wants to play the media banter game,
I'm happy to play him.
He's a good researcher and a good human being,
and I'd happily work with the guy.
But the other thing I was gonna say is,
I have been explicit about how Sophia works,
and I've posted online, and what, H+ Magazine,
an online webzine, I mean, I posted
a moderately detailed article explaining,
like, there are three software systems
we've used inside Sophia.
There's a timeline editor,
which is like a rule-based authoring system,
where she's really just being an outlet
for what a human scripted.
There's a chatbot, which has some rule-based
and some neural aspects.
And then sometimes we've used OpenCog behind Sophia,
where there's more learning and reasoning.
And, you know, the funny thing is,
I can't always tell which system is operating here, right?
I mean, so whether she's really learning,
or thinking, or just appears to be,
over a half hour I could tell,
but over, like, three or four minutes of interaction,
I couldn't perhaps tell.
- So even having three systems
that's already sufficiently complex,
where you can't really tell right away.
- Yeah, the thing is, even if you get up on stage
and tell people how Sophia's working,
and then they talk to her,
they still attribute more agency and consciousness to her
§22Transcript · 3:13:06–3:22:16
than is really there.
So I think there's a couple levels of ethical issue there.
One issue is, should you be transparent
about how Sophia is working?
And I think you should.
And I think we have been.
I mean, there's articles online,
there's some TV special that goes through me
explaining the three subsystems behind Sophia.
So the way Sophia works is out there much more clearly
than how Facebook's AI works or something, right?
I mean, we've been fairly explicit about it.
The other is, given that telling people how it works
doesn't cause them to not attribute
too much intelligence agency to it anyway,
then should you keep fooling them
when they want to be fooled?
And I mean, the whole media industry
is based on fooling people the way they want to be fooled.
And we are fooling people 100% toward a good end.
I mean, we are playing on people's sense of empathy
and compassion so that we can give them
a good user experience with helpful robots,
and so that we can fill the AI's mind
with love and compassion.
So I've been talking a lot with Hanson Robotics lately
about collaborations in the area of medical robotics.
And we haven't quite pulled the trigger
on a project in that domain yet,
but we may well do so quite soon.
So we've been talking a lot about,
robots can help with elder care,
robots can help with kids.
David's done a lot of things with autism therapy
and robots before.
In the COVID era, having a robot
that can be a nursing assistant in various senses
can be quite valuable.
The robots don't spread infection,
and they can also deliver more attention
than human nurses can give, right?
So if you have a robot that's helping a patient with COVID,
if that patient attributes more understanding
and compassion and agency to that robot,
then it really has because it looks like a human.
I mean, is that really bad?
I mean, we can tell them it doesn't fully understand you,
and they don't care,
'cause they're lying there with a fever and they're sick,
but they'll react better to that robot
with its loving, warm facial expression
than they would to a pepper robot
or a metallic-looking robot.
So it's really, it's about how you use it, right?
If you made a human-looking, like, door-to-door sales robot
that used its human-looking appearance
to scam people out of their money,
then you're using that connection in a bad way,
but you could also use it in a good way.
But then that's the same problem with every technology, right?
- Beautifully put.
So like you said, we're living in the era of the COVID.
This is 2020, one of the craziest years in recent history.
So if we zoom out and look at this panel,
this pandemic, the coronavirus pandemic,
maybe let me ask you this kind of thing,
in viruses in general,
when you look at viruses,
do you see them as a kind of intelligence system?
- I think the concept of intelligence
is not that natural of a concept in the end.
I think human minds and bodies are a kind of complex,
self-organizing, adaptive system,
and viruses certainly are that, right?
They're a very complex, self-organizing, adaptive system.
If you wanna look at intelligence as Marcus Hutter defines it
as sort of optimizing computable reward functions
over computable environments,
for sure viruses are doing that, right?
And I mean, in doing so, they're causing some harm to us.
So the human immune system
is a very complex, self-organizing, adaptive system,
which has a lot of intelligence to it.
And viruses are also adapting
and dividing into new mutant strains and so forth.
And ultimately, the solution is gonna be nanotechnology.
I mean, the solution is gonna be making little nanobots that--
- Fight the viruses?
- Well, people will use them to make nastier viruses,
but hopefully we can also use them
to just detect, combat, and kill the viruses.
But I think now we're stuck with the biological mechanisms
to combat these viruses.
And yeah, we've been,
AGI is not yet mature enough to use against COVID,
but we've been using machine learning
and also some machine reasoning in OpenCog
to help some doctors
to do personalized medicine against COVID.
So the problem there is given the person's genomics
and given their clinical medical indicators,
how do you figure out which combination of antivirals
is gonna be most effective against COVID for that person?
And so that's something
where machine learning is interesting,
but also we're finding the abstraction.
We get an OpenCog with machine reasoning is interesting
'cause it can help with transfer learning
when you have not that many different cases to study
and qualitative differences
between different strains of a virus
or people of different ages who may have COVID.
- So there's a lot of different disparate data to work with
and it's small data sets and somehow integrating them.
- You know, this is one of the shameful things
that's very hard to get that data.
So, I mean, we're working with a couple groups
doing clinical trials
and they're sharing data with us under non-disclosure.
But what should be the case is like
every COVID clinical trial
should be putting data online somewhere,
like suitably encrypted to protect patient privacy
so that anyone with the right AI algorithms
should be able to help analyze it.
And any biologist should be able to analyze it by hand
to understand what they can, right?
Instead, that data is like siloed
inside whatever hospital is running the clinical trial,
which is completely asinine and ridiculous.
So why the world works that way?
I mean, we could all analyze why,
but it's insane that it does.
You look at this hydrochloroquine, right?
All these clinical trials were done
were reported by Surgisphere,
some little company no one ever heard of.
And everyone paid attention to this.
So they were doing more clinical trials based on that.
Then they stopped doing clinical trials based on that.
Then they started again.
And why isn't that data just out there
so everyone can analyze it and see what's going on, right?
- Do you have hope that we'll move,
that data will be out there eventually
for future pandemics?
I mean, do you have hope that our society
will move in the direction of--
- Not in the immediate future
because the US and China frictions are getting very high.
So it's hard to see US and China
as moving in the direction
of openly sharing data with each other, right?
It's not, there's some sharing of data,
but different groups are keeping their data private
till they've milked the best results from it
and then they share it, right?
So it's, so yeah, we're working with some data
that we've managed to get our hands on,
something we're doing to do good for the world.
And it's a very cool playground
for like putting deep neural nets and open code together.
So we have like a bioatom space
full of all sorts of knowledge
from many different biology experiments
about human longevity
and from biology knowledge bases online.
And we can do like graph to vector type embeddings
where we take nodes from the hypergraph,
embed them into vectors.
We can then feed into neural nets
for different types of analysis.
And we were doing this in the context of a project
called the Rejuve that we spun off from SingularityNet
to do longevity analytics,
like understand why people live to 105 years or over
and other people don't.
And then we had this spin-off Singularity Studio
where we're working with some healthcare companies
on data analytics.
But so there's bioatom space
that we built for these more commercial
and longevity data analysis purposes.
We're repurposing and feeding COVID data
into the same bioatom space
and playing around with like graph embeddings
from that graph into neural nets for bioinformatics.
So it's both being a cool testing ground
for some of our bio AI learning and reasoning.
And it seems we're able to discover things
that people weren't seeing otherwise.
'Cause the thing in this case is
for each combination of antivirals,
you may have only a few patients
who've tried that combination.
And those few patients may have
their particular characteristics.
Like this combination of three
was tried only on people age 80 or over.
This other combination of three,
§23Transcript · 3:22:18–3:31:38
which has an overlap with the first combination,
was tried more on young people.
So how do you combine those different pieces of data?
It's a very dodgy transfer learning problem,
which is the kind of thing
that the probabilistic reasoning algorithms
we have inside OpenCog are better at
than deep neural networks.
On the other hand, you have gene expression data
where you have 25,000 genes
and the expression level of each gene
in the peripheral blood of each person.
So that sort of data, either deep neural nets
or tools like XGBoost or CatBoost,
these decision forestries,
are better at dealing with than OpenCog.
'Cause it's just these huge,
huge messy floating point vectors
that are annoying for a logic engine to deal with,
but are perfect for a decision forest or neural net.
So it's a great playground for hybrid AI methodology.
And we can have a singularity net,
have OpenCog in one agent,
and XGBoost in a different agent,
and they talk to each other.
But at the same time, it's highly practical, right?
'Cause we're working with, for example,
some physicians on this project,
physicians in the group called Enthopinion,
based out of Vancouver and Seattle,
who are, these guys are working every day
like in the hospital with patients dying of COVID.
So it's quite cool to see like neural symbolic AI,
like where the rubber hits the road,
trying to save people's lives.
I've been doing bio AI since 2001,
but mostly human longevity research
and fly longevity research,
trying to understand why some organisms
really live a long time.
This is the first time like race against the clock
and try to use the AI to figure out stuff that,
like if we take two months longer to solve the AI problem,
some more people will die
because we don't know what combination
of antivirals to give them.
- Yeah.
At the societal level, at the biological level,
at any level, are you hopeful about us
as a human species getting out of this pandemic?
What are your thoughts on it in general?
- The pandemic will be gone in a year or two months.
There's a vaccine for it.
So I mean, that's--
- A lot of pain and suffering can happen in that time.
So that could be irreversible.
- I think if you spend much time in sub-Saharan Africa,
you can see there's a lot of pain
and suffering happening all the time.
Like you walk through the streets
of any large city in sub-Saharan Africa,
and there are loads,
I mean, tens of thousands,
probably hundreds of thousands of people
lying by the side of the road,
dying mainly of curable diseases without food or water,
and either ostracized by their families
or they left their family house
'cause they didn't want to infect their family, right?
I mean, there's tremendous human suffering on the planet
all the time, which most folks in the developed world
pay no attention to, and COVID is not remotely the worst.
How many people are dying of malaria all the time?
I mean, so COVID is bad.
It is by no mean the worst thing happening,
and setting aside diseases,
I mean, there are many places in the world
where you're at risk of having your teenage son
kidnapped by armed militias
and forced to get killed in someone else's war
fighting tribe against tribe.
I mean, so humanity has a lot of problems,
which we don't need to have
given the state of advancement of our technology right now.
And I think COVID is one of the easier problems to solve
in the sense that there are many brilliant people
working on vaccines.
We have the technology to create vaccines,
and we're gonna create new vaccines.
We should be more worried
that we haven't managed to defeat malaria after so long
and after the Gates Foundation and others
putting so much money into it.
I mean, I think clearly the whole global medical system,
global health system,
and the global political and socioeconomic system
are incredibly unethical and unequal and badly designed.
And I mean, I don't know how to solve that directly.
I think what we can do indirectly to solve it
is to make systems that operate in parallel
and off to the side of the governments
that are nominally controlling the world
with their armies and militias.
And to the extent that you can make compassionate
peer-to-peer decentralized frameworks for doing things,
these are things that can start out unregulated.
And then if they get traction before the regulators come in,
then they've influenced the way the world works, right?
SingularityNet aims to do this with AI.
Rajuv, which is a spinoff from SingularityNet,
you can see it, rajuv.io.
- How do you spell that?
- R-E-J-U-V-E, rajuv.io.
That aims to do the same thing for medicine.
So it's like peer-to-peer sharing of medical data.
So you can share medical data into a secure data wallet.
You can get advice about your health and longevity
through apps that Rajuv will launch
within the next couple months.
And then SingularityNet AI can analyze all this data,
but then the benefits from that analysis
are spread among all the members of the network.
But I mean, of course, I'm gonna hawk my particular projects
but I mean, whether or not SingularityNet and Rajuv
are the answer, I think it's key to create
decentralized mechanisms for everything.
I mean, for AI, for human health, for politics,
for jobs and employment, for sharing social information.
And to the extent decentralized peer-to-peer methods
designed with universal compassion at the core
can gain traction, then these will just decrease the role
that government has.
And I think that's much more likely to do good
than trying to explicitly reform
the global government system.
I mean, I'm happy other people are trying to explicitly
reform the global government system.
On the other hand, you look at how much good the internet
or Google did or mobile phones did.
Even you're making something that's decentralized
and throwing it out everywhere and it takes hold,
then government has to adapt.
And I mean, that's what we need to do with AI and with health.
And in that light, I mean, the centralization of healthcare
and of AI is certainly not ideal, right?
Like most AI PhDs are being sucked in by a half dozen
to a dozen big companies.
Most AI processing power is being bought
by a few big companies for their own proprietary good.
And most medical research is within a few
pharmaceutical companies and clinical trials run
by pharmaceutical companies will stay solid
within those pharmaceutical companies.
Now, these large centralized entities,
which are intelligences in themselves, these corporations,
but they're mostly malevolent, psychopathic
and sociopathic intelligences.
Not saying the people involved are,
but the corporations as self-organizing entities
on their own, which are concerned with maximizing
shareholder value as a sole objective function.
I mean, AI and medicine are being sucked
into these pathological corporate organizations
with government cooperation and Google cooperating
with British and US government on this
as one among many, many different examples.
23andMe providing you the nice service
of sequencing your genome and then licensing the genome
to GlaxoSmithKline on an exclusive basis, right?
- Right.
- Now you can take your own DNA
and do whatever you want with it,
but the pooled collection of 23andMe sequenced DNA
is just to GlaxoSmithKline.
Someone else could reach out to everyone
who had worked with 23andMe to sequence their DNA
and say, give us your DNA for our open
and decentralized repository that we'll make available
to everyone, but nobody's doing that
'cause it's a pain to get organized.
And the customer list is proprietary to 23andMe, right?
- Interesting.
- So, yeah, I mean, this I think is a greater risk
to humanity from AI than rogue AGIs turning the universe
into paperclips or computronium.
'Cause what you have here is mostly good-hearted
and nice people who are sucked into a mode of organization
of large corporations, which has evolved
just for no individual's fault,
just because that's the way society has evolved.
- It's not all choice to get self-interested
and become psychopathic, like you said.
- The corporation is psychopathic,
even if the people are not.
And that's really the disturbing thing about it
because the corporations can do things
that are quite bad for society,
even if nobody has a bad intention.
- Right, no individual member of that corporation
has a bad intention.
§24Transcript · 3:31:38–3:40:49
- No, some probably do, but it's not necessary
that they do for the corporation.
Like, I mean, Google, I know a lot of people in Google,
and there are, with very few exceptions,
they're all very nice people who genuinely want
what's good for the world.
And Facebook, I know fewer people,
but it's probably mostly true.
It's probably like fun young geeks
who wanna build cool technology.
- I actually tend to believe that even the leaders,
even Mark Zuckerberg, one of the most disliked people
in tech, also wants to do good for the world.
- What do you think about Jamie Dimon?
- Who's Jamie Dimon?
- The heads of the great banks.
May have a different psychology.
- Oh boy, yeah, well.
I tend to be naive about these things
and see the best.
I tend to agree with you that I think the individuals
wanna do good by the world,
but the mechanism of the company can sometimes
be its own intelligence system.
- I mean, there's a, my cousin Mario Goetzel
has worked for Microsoft since 1985 or something,
and I can see for him, I mean,
as well as just working on cool projects,
you're coding stuff that gets used by,
like, billions and billions of people.
And do you think if I improve this feature,
that's making billions of people's lives easier, right?
So of course that's cool,
and the engineers are not in charge
of running the company anyway.
And of course, even if you're Mark Zuckerberg or Larry Page,
I mean, you still have a fiduciary responsibility.
And I mean, you're responsible to the shareholders,
your employees, who you want to keep paying them
and so forth.
So yeah, you're enmeshed in this system.
And when I worked in DC,
I worked a bunch with INSCOM, US Army Intelligence,
and I was heavily politically opposed
to what the US Army was doing in Iraq at that time,
like torturing people in Abu Ghraib.
But everyone I knew in US Army and INSCOM,
when I hung out with them, was a very nice person.
They were friendly to me,
they were nice to my kids and my dogs, right?
And they really believed that the US
was fighting the forces of evil.
And if you ask them about Abu Ghraib,
they're like, "Well, but these Arabs
"will chop us into pieces,
"so how can you say we're wrong to waterboard them a bit?"
Right?
Like, that's much less than what they would do to us.
It's just in their worldview,
what they were doing was really genuinely
for the good of humanity.
Like, none of them woke up in the morning and said,
"I want to do harm to good people
"because I'm just a nasty guy," right?
So yeah, most people on the planet,
setting aside a few genuine psychopaths and sociopaths,
I mean, most people on the planet
have a heavy dose of benevolence and wanting to do good,
and also a heavy capability to convince themselves
whatever they feel like doing,
or whatever is best for them,
is for the good of humankind, right?
- So the more we can decentralize control of--
- Decentralization, you know, democracy is horrible,
but this is like Winston Churchill said,
"It's the worst possible system of government
"except for all the others," right?
I mean, I think the whole mess of humanity
has many, many very bad aspects to it,
but so far the track record of elite groups
who know what's better for all of humanity
is much worse than the track record
of the whole teeming democratic participatory
mess of humanity, right?
I mean, none of them is perfect by any means.
The issue with a small elite group that knows what's best
is even if it starts out as truly benevolent
and doing good things in accordance
with its initial good intentions,
you find out you need more resources,
you need a bigger organization,
you pull in more people, internal politics arises,
difference of opinions arise, and bribery happens.
Like some opponent organization
takes a second in command now to make the first in command
of some other organization,
and I mean, there's a lot of history
of what happens with elite groups
thinking they know what's best for the human race.
So if I have to choose,
I'm gonna reluctantly put my faith
in the vast democratic decentralized mass,
and I think corporations have a track record
of being ethically worse
than their constituent human parts,
and democratic governments have a more mixed track record,
but there are at least--
- But it's the best we got.
- Yeah, I mean, you can, there's Iceland,
very nice country, right?
I mean, democratic for 800 plus years,
very benevolent, beneficial government,
and I think, yeah, there are track records
of democratic modes of organization.
Linux, for example, some of the people in charge of Linux
are overtly complete assholes, right?
And trying to reform themselves, in many cases,
in other cases not, but the organization as a whole,
I think it's done a good job overall.
It's been very welcoming in the third world, for example,
and it's allowed advanced technology
to roll out on all sorts of different embedded devices
and platforms in places where people couldn't afford
to pay for proprietary software.
So I'd say the internet, Linux,
and many democratic nations are examples
of how certain open decentralized democratic methodology
can be ethically better than the sum of the parts
rather than worse, and corporations,
that has happened only for a brief period,
and then it goes sour, right?
I mean, I'd say a similar thing about universities.
Like, university is a horrible way to organize research
and get things done, yet it's better than anything else
we've come up with, right?
A company can be much better, but for a brief period of time
and then it stops being so good, right?
So then I think if you believe that AGI
is gonna emerge sort of incrementally
out of AIs doing practical stuff in the world,
like controlling humanoid robots, or driving cars,
or diagnosing diseases, or operating killer drones,
or spying on people and reporting under the government,
then what kind of organization
creates more and more advanced narrow AI
verging toward AGI may be quite important
because it will guide what's in the mind
of the early stage AGI as it first gains the ability
to rewrite its own code base
and project itself toward superintelligence.
And if you believe that AI may move toward AGI
out of this sort of synergetic activity
of many agents cooperating together,
rather than just have one person's project,
then who owns and controls that platform
for AI cooperation becomes also very, very important.
And is that platform AWS?
Is it Google Cloud?
Is it Alibaba?
Or is it something more like the internet
or SingularityNet, which is open and decentralized?
So if all of my weird machinations come to pass,
I mean, we have the Hanson robots
being a beautiful user interface,
gathering information on human values
and being loving and compassionate to people
in medical home service robot office applications.
You have SingularityNet in the backend
networking together many different AIs
toward cooperative intelligence,
fueling the robots among many other things.
You have OpenCog 2.0 and TrueAGI
as one of the sources of AI
inside this decentralized network,
powering the robot and medical AIs
helping us live a long time
and cure diseases among other things.
And this whole thing is operating
in a democratic and decentralized way, right?
I think if anyone can pull something like this off,
you know, whether using the specific technologies
I've mentioned or something else,
I mean, then I think we have a higher odds
of moving toward a beneficial technological singularity
rather than one in which the first super AGI
is indifferent to humans
and just considers us an inefficient use of molecules.
- That was a beautifully articulated vision for the world.
So thank you for that.
Well, let's talk a little bit about life and death.
- I'm pro-life and anti-death.
For most people, there's few exceptions
that I won't mention here.
- I'm glad just like your dad,
you're taking a stand against death.
You have, by the way, you have a bunch of awesome music
where you play piano online.
One of the songs that I believe you've written,
the lyrics go, by the way, I like the way it sounds.
People should listen to it, it's awesome.
§25Transcript · 3:40:51–3:49:44
I considered, I probably will cover it.
It's a good song.
Tell me why do you think it is a good thing
that we all get old and die?
It's one of the songs.
I love the way it sounds.
But let me ask you about death first.
Do you think there's an element to death
that's essential to give our life meaning?
Like the fact that this thing ends?
- Let me say, I'm pleased and a little embarrassed
you've been listening to that music I put online.
- Oh, it's awesome.
- One of my regrets in life recently
is I would love to get time to really produce music well.
Like I haven't touched my sequencer software
in like five years.
I would love to rehearse and produce and edit.
But with a two-year-old baby
and trying to create the singularity, there's no time.
So I just made the decision to,
when I'm playing random shit in an off moment--
- Just record it.
- Just record it.
- Oh, you're still thinking.
- Put it out there like whatever.
Maybe if I'm unfortunate enough to die,
maybe that can be input to the AGI
when it tries to make an accurate mind upload of me, right?
Death is bad.
(laughing)
I mean, that's very simple.
It's baffling we should have to say that.
I mean, of course, people can make meaning out of death.
And if someone is tortured,
maybe they can make beautiful meaning out of that torture
and write a beautiful poem
about what it was like to be tortured, right?
I mean, we're very creative.
We can milk beauty and positivity
out of even the most horrible and shitty things.
But just because if I was tortured,
I could write a good song
about what it was like to be tortured,
doesn't make torture good.
And just because people are able to derive meaning
and value from death,
doesn't mean they wouldn't derive even better meaning
and value from ongoing life without death,
which I very--
- Indefinite.
- Yeah, yeah.
- So if you could live forever, would you live forever?
Forever.
- My goal with longevity research
is to abolish the plague of involuntary death.
I don't think people should die unless they choose to die.
If I had to choose forced immortality versus dying,
I would choose forced immortality.
On the other hand, if I chose,
if I had the choice of immortality
with the choice of suicide whenever I felt like it,
of course, I would take that instead.
And that's the more realistic choice.
I mean, there's no reason you should have forced immortality.
You should be able to live
until you get sick of living, right?
I mean, that's, and that will seem insanely obvious
to everyone 50 years from now.
And they will be so, I mean,
people who thought death gives meaning to life,
so we should all die,
they will look at that 50 years from now,
the way we now look at the Anabaptists in the year 1000,
who gave away all their positions,
went on top of the mountain for Jesus,
for Jesus to come and bring them to the ascension.
I mean, it's ridiculous that people think death is good,
because you gain more wisdom as you approach dying.
I mean, of course it's true.
I mean, I'm 53,
and the fact that I might have only a few more decades left,
it does make me reflect on things differently.
It does give me a deeper understanding of many things.
But I mean, so what?
You could get a deep understanding
in a lot of different ways.
Pain is the same way.
Like, we're gonna abolish pain.
And that's even more amazing than abolishing death.
I mean, once we get a little better at neuroscience,
we'll be able to go in and adjust the brain
so that pain doesn't hurt anymore.
And people will say that's bad,
because there's so much beauty
in overcoming pain and suffering.
Well, sure, and there's beauty in overcoming torture too.
And some people like to cut themselves,
but not many, right?
- That's an interesting, so, but to push,
I mean, to push back, again, this is the Russian side of me,
I do romanticize suffering.
It's not obvious, I mean, the way you put it,
it seems very logical, it's almost absurd
to romanticize suffering or pain or death.
But to me, a world without suffering,
without pain, without death, it's not obvious
what that world would say. - Well, then you can stay
in the people's zoo.
- People's zoo with suffering. - With the people
torturing each other, right? (laughs)
- No, but what I'm saying is, I don't,
well, that's, I guess what I'm trying to say,
I don't know if I was presented with that choice
of what I would choose, because to me--
- No, this is a subtler, it's a subtler matter,
and I've posed it in this conversation
in an unnecessarily extreme way.
So I think the way you should think about it
is what if there's a little dial on the side of your head,
and you could turn how much pain hurt,
turn it down to zero, turn it up to 11,
like in spinal tap, if it wants,
maybe through an actual spinal tap, right?
So, I mean, would you opt to have that dial there or not?
That's the question.
The question isn't whether you would turn the pain
down to zero all the time.
Would you opt to have the dial or not?
My guess is that in some dark moment of your life,
you would choose to have the dial implanted,
and then it would be there.
- Just to confess a small thing, don't ask me why,
but I'm doing this physical challenge currently,
where I'm doing 680 pushups and pull-ups a day,
and my shoulder is currently, as we sit here,
in a lot of pain, and I don't know.
I would certainly right now, if you gave me a dial,
I would turn that sucker to zero as quickly as possible.
- Good.
- But I think the whole point of this journey is,
I don't know.
- Well, because you're a twisted human being.
- I'm a twisted, so the question is,
am I somehow twisted because I created some kind
of narrative for myself so that I can deal
with the injustice and the suffering in the world,
or is this actually going to be a source of happiness?
- Well, this is, to an extent, is a research question
that humanity will undertake, right?
So, I mean, human beings do have a particular
biological makeup, which sort of implies
a certain probability distribution
over motivational systems, right?
So, I mean, we, and that is there.
- I'll put.
- That is there.
Now, the question is, how flexibly can that morph
as society and technology change, right?
So, if we're given that dial, and we're given a society
in which, say, we don't have to work for a living,
and in which there's an ambient, decentralized,
benevolent AI network that will warn us
when we're about to hurt ourself,
you know, if we're in a different context,
can we consistently, with being genuinely and fully human,
can we consistently get into a state of consciousness
where we just want to keep the pain dial turned
all the way down, and yet we're leading
very rewarding and fulfilling lives, right?
Now, I suspect the answer is yes, we can do that,
but I don't know that.
- It's a research question, like you said.
- I don't know that for certain.
Yeah, now, I'm more confident that we could create
a non-human AGI system, which just didn't need
an analog of feeling pain, and I think that AGI system
will be fundamentally healthier and more benevolent
than human beings, so I think it might or might not be true
that humans need a certain element of suffering
to be satisfied humans, consistent with the human physiology.
If it is true, that's one of the things
that makes us fucked and disqualified
to be the super AGI, right?
I mean, the nature of the human motivational system
is that we seem to gravitate towards situations
where the best thing in the large scale
is not the best thing in the small scale,
according to our subjective value system.
So we gravitate towards subjective value judgments,
where to gratify ourselves in the large,
we have to ungratify ourselves in the small,
and we do that in, you see that in music,
there's a theory of music which says
the key to musical aesthetics is the surprising
fulfillment of expectations.
Like you want something that will fulfill
the expectations enlisted in the prior part of the music,
but in a way with a bit of a twist that surprises you.
And I mean, that's true not only in outdoor music
§26Transcript · 3:49:48–3:59:12
like my own or that of Zappa or Steve Vai or Buckethead
or Christoph Penderecki or something,
it's even there in Mozart or something.
It's not there in elevator music too much,
but that's why it's boring, right?
But wrapped up in there is, you know,
we want to hurt a little bit so that we can feel the pain
go away, like we wanna be a little confused
by what's coming next, so then when the thing
that comes next actually makes sense,
it's so satisfying, right?
- It's the surprising fulfillment of expectations,
is that what you said?
- Yeah, yeah, yeah. - So beautifully put.
Is there, we've been skirting around a little bit,
but if I were to ask you the most ridiculous big question
of what is the meaning of life, what would your answer be?
- Three values, joy, growth, and choice.
I think you need joy, I mean, that's the basis
of everything if you want the number one value.
On the other hand, I'm unsatisfied with a static joy
that doesn't progress, perhaps because of some
elemental element of human perversity,
but the idea of something that grows
and becomes more and more and better
and better in some sense appeals to me.
But I also sort of like the idea of individuality,
that as a distinct system, I have some agency,
so there's some nexus of causality within this system,
rather than the causality being wholly evenly distributed
over the joyous growing mass.
So you start with joy, growth, and choice
as three basic values.
- And those three things could continue indefinitely.
That's something that could last forever.
Is there some aspect of something you called,
which I like, super longevity that you find exciting?
Research-wise, is there ideas in that space?
- I mean, I think, yeah, in terms of the meaning of life,
this really ties into that, because for us as humans,
probably the way to get the most joy, growth,
and choice is transhumanism,
and to go beyond the human form that we have right now.
I mean, I think human body is great,
and by no means do any of us maximize the potential
for joy, growth, and choice imminent in our human bodies.
On the other hand, it's clear that other configurations
of matter could manifest even greater amounts
of joy, growth, and choice than humans do,
maybe even finding ways to go beyond the realm of matter
as we understand it right now.
So I think in a practical sense,
much of the meaning I see in human life
is to create something better than humans
and go beyond human life.
But certainly that's not all of it for me
in a practical sense, right?
Like I have four kids and a granddaughter
and many friends and parents and family,
and just enjoying everyday human life.
Human social existence.
- But we can do even better.
- Yeah, yeah, and I mean, I love,
I've always, when I could live near nature,
I spend a bunch of time out in nature in the forest
and on the water every day and so forth.
So I mean, enjoying the pleasant moment is part of it,
but the growth and choice aspect are severely limited
by our human biology.
In particular, dying seems to inhibit your potential
for personal growth considerably as far as we know.
I mean, there's some element of life after death perhaps,
but even if there is, why not also continue going
in this biological realm, right?
In super longevity, I mean,
you know, we haven't yet cured aging.
We haven't yet cured death.
Certainly there's very interesting progress all around.
I mean, CRISPR and gene editing
can be an incredible tool.
And I mean, right now, stem cells
could potentially prolong life a lot.
Like if you got stem cell injections
of just stem cells for every tissue of your body
injected into every tissue,
and you can just have replacement of your old cells
with new cells produced by those stem cells,
I mean, that could be highly impactful at prolonging life.
Now we just need slightly better technology
for having them grow, right?
So using machine learning to guide procedures
for stem cell differentiation and trans differentiation,
it's kind of nitty gritty,
but I mean, that's quite interesting.
So I think there's a lot of different things being done
to help with prolongation of human life,
but we could do a lot better.
So for example, the extracellular matrix,
which is the bunch of proteins
in between the cells in your body,
they get stiffer and stiffer as you get older.
And the extracellular matrix transmits information
both electrically, mechanically,
and to some extent, biophotonically.
So there's all this transmission
through the parts of the body,
but the stiffer the extracellular matrix gets,
the less the transmission happens,
which makes your body get worse coordinated
between the different organs as you get older.
So my friend, Christian Schaffmeister,
my alumnus organization, my alma mater,
the great Temple University,
Christian Schaffmeister has a potential solution to this
where he has these novel molecules called spiral ligamers,
which are like polymers that are not organic.
They're specially designed polymers
so that you can algorithmically predict
exactly how they'll fold very simply.
So he designed the molecular scissors
that have spiral ligamers that you could eat
and would then cut through all the glucosamine
and other cross-linked proteins
in your extracellular matrix, right?
But to make that technology really work
and be mature is several years of work.
As far as I know, no one's funding it at the moment.
So there's so many different ways
that technology could be used to prolong longevity.
What we really need, we need an integrated database
of all biological knowledge about human beings
and model organisms, like hopefully a massively distributed
open-cog bioatom space,
but it can exist in other forms too.
We need that data to be opened up
in a suitably privacy-protecting way.
We need massive funding into machine learning,
AGI, proto-AGI statistical research
aimed at solving biology,
both molecular biology and human biology
based on this massive, massive dataset, right?
And then we need regulators not to stop people
from trying radical therapies on themselves
if they so wish to,
as well as better cloud-based platforms
for automated experimentation on microorganisms,
flies and mice and so forth.
And we could do all this.
You look, after the last financial crisis,
Obama, who I generally like pretty well,
but he gave $4 trillion to large banks
and insurance companies.
Now in this COVID crisis, trillions are being spent
to help everyday people and small businesses.
In the end, we'll probably will find many more trillions
being given to large banks and insurance companies anyway.
Like, could the world put $10 trillion
into making a massive, holistic bio-AI and bio-simulation
and experimental biology infrastructure?
We could, we could put $10 trillion into that
without even screwing us up too badly,
just as in the end, COVID and the last financial crisis
won't screw up the world economy so badly.
We're not putting $10 trillion into that.
Instead, all this research is siloed
inside a few big companies and government agencies.
And most of the data that comes from our individual bodies,
personally, that could feed this AI to solve aging and death,
most of that data is sitting in some hospital's database
doing nothing, right?
- I got two more quick questions for you.
One, I know a lot of people are gonna ask me,
you're on the Joe Rogan podcast
wearing that same amazing hat.
Do you have a origin story for the hat?
Does the hat have its own story that you're able to share?
- The hat story has not been told yet,
so we're gonna have to come back
and you can interview the hat.
- The hat. (laughs)
- We'll leave that for the hat, so an interview.
It's too much to pack into a few seconds.
- Is there a book, is the hat gonna write a book?
Okay, we'll-- (laughs)
- It may transmit the information
through direct neural transmission.
- Okay, so it's actually,
there might be some neural link competition there.
Beautiful, we'll leave it as a mystery.
Maybe one last question.
If you build an AGI system,
you're successful at building the AGI system
that could lead us to the singularity
and you get to talk to her and ask her one question,
what would that question be?
- We're not allowed to ask,
what is the question I should be asking?
(both laugh)
- Yeah, that would be cheating,
but I guess that's a good question.
§27Transcript · 3:59:14–4:08:48
- I'm thinking of a,
I wrote a story with Stefan Bugaj once
where these AI developers,
they created a super smart AI aimed at answering
all the philosophical questions that had been worrying them,
like what is the meaning of life, is there free will,
what is consciousness and so forth.
So they got the super AGI built
and it churned a while, it said,
those are really stupid questions,
and then it puts off on a spaceship and left the Earth.
(Luke laughs)
- So you'd be afraid of scaring it off.
(Luke laughs)
- That's it, yeah.
I mean, honestly, there is no one question
that rises among all the others, really.
I mean, what interests me more is upgrading
my own intelligence so that I can absorb
the whole worldview of the super AGI.
But I mean, of course, if the answer could be
like what is the chemical formula for the immortality pill,
then I would do that or emit a bit string,
which will be the code for a super AGI
on the Intel i7 processor.
So those would be good questions.
- So if your own mind was expanded
to become super intelligent, like you're describing,
I mean, there's kind of a notion
that intelligence is a burden,
that it's possible that with greater and greater intelligence
that other metric of joy that you mentioned
becomes more and more difficult.
What's your--
- That's a pretty stupid idea.
(both laugh)
- So you think if you're super intelligent,
you can also be super joyful?
- I think getting root access to your own brain
will enable new forms of joy that we don't have now.
And I think, as I've said before,
what I aim at is really make multiple versions of myself.
So I would like to keep one version,
which is basically human like I am now,
but keep the dial to turn pain up and down
and get rid of death, right?
And make another version, which fuses its mind
with superhuman AGI,
and then will become massively transhuman.
And whether it will send some messages back
to the human me or not will be interesting to find out.
The thing is, once you're super AGI,
like one subjective second to a human
might be like a million subjective years
to that super AGI, right?
So it would be on a whole different basis.
I mean, at very least those two copies will be good to have,
but it could be interesting to put your mind
into a dolphin or a space amoeba
or all sorts of other things.
You can imagine one version
that doubled its intelligence every year,
and another version that just became a super AGI
as fast as possible, right?
So, I mean, now we're sort of constrained
to think one mind, one self, one body, right?
But I think we actually,
we don't need to be that constrained
in thinking about future intelligence
after we've mastered AGI and nanotechnology
and longevity biology.
I mean, then each of our minds
is a certain pattern of organization, right?
And I know we haven't talked about consciousness,
but I sort of, I'm panpsychist.
I sort of view the universe as conscious.
And so, a light bulb or a quark or an ant
or a worm or a monkey
have their own manifestations of consciousness.
And the human manifestation of consciousness,
it's partly tied to the particular meat
that we're manifested by,
but it's largely tied to the pattern of organization
in the brain, right?
So, if you upload yourself into a computer
or a robot or whatever else it is,
some element of your human consciousness may not be there
because it's just tied to the biological embodiment.
But I think most of it will be there,
and these will be incarnations of your consciousness
in a slightly different flavor.
And creating these different versions will be amazing,
and each of them will discover meanings of life
that have some overlap,
but probably not total overlap
with the human being's meaning of life.
The thing is, to get to that future
where we can explore different varieties of joy,
different variations of human experience and values
and transhuman experiences and values,
to get to that future,
we need to navigate through a whole lot of human,
bullshit of companies and governments
and killer drones and making and losing money
and so forth, right?
And that's the challenge we're facing now,
is if we do things right,
we can get to a benevolent singularity,
which is levels of joy, growth, and choice
that are literally unimaginable to human beings.
If we do things wrong,
we could either annihilate all life on the planet,
or we could lead to a scenario where, say,
all humans are annihilated,
and there's some super AGI that goes on
and does its own thing unrelated to us,
except via our role in originating it.
And we may well be at a bifurcation point now, right?
Where what we do now has significant causal impact
on what comes about,
and yet, most people on the planet
aren't thinking that way whatsoever.
They're thinking only about their own narrow aims
and aims and goals, right?
Now, of course, I'm thinking about my own narrow aims
and goals to some extent also,
but I'm trying to use as much of my energy and mind
as I can to push toward this more benevolent alternative,
which will be better for me,
but also for everybody else.
And it's weird that so few people understand
what's going on.
I know you interviewed Elon Musk,
and he understands a lot of what's going on,
but he's much more paranoid than I am, right?
'Cause Elon gets that AGI is gonna be
way, way smarter than people,
and he gets that an AGI does not necessarily
have to give a shit about people,
because we're a very elementary mode
of organization of matter,
compared to many AGIs.
But I don't think he has a clear vision
of how infusing early stage AGIs
with compassion and human warmth
can lead to an AGI that loves and helps people,
rather than viewing us as a historical artifact
and a waste of mass energy.
But on the other hand,
while I have some disagreements with him,
like he understands way, way more of the story
than almost anyone else in such a large-scale
corporate leadership position, right?
It's terrible how little understanding
of these fundamental issues exists out there now.
That may be different five or 10 years from now, though,
'cause I can see understanding of AGI and longevity
and other such issues is certainly much stronger
and more prevalent now than 10 or 15 years ago, right?
So I mean, humanity is, as a whole,
can be slow learners relative to what I would like,
but on a historical sense, on the other hand,
you could say the progress is astoundingly fast.
- But Elon also said, I think on the Joe Rogan podcast,
that love is the answer.
So maybe in that way, you and him are both on the same page
of how we should proceed with AGI.
I think there's no better place to end it.
I hope we get to talk again about the hat
and about consciousness and about a million topics
we didn't cover.
Ben, it's a huge honor to talk to you.
Thank you for making it out.
Thank you for talking today.
I really loved it. - No, thanks for having me.
This was really, really good fun,
and we dug deep into some very important things.
So thanks for doing this.
- Thanks very much.
Awesome.
Thanks for listening to this conversation with Ben Goertzel,
and thank you to our sponsors,
the Jordan Harbinger Show and Masterclass.
Please consider supporting the podcast
by going to jordanharbinger.com/lex
and signing up to Masterclass at masterclass.com/lex.
Click the links, buy the stuff.
It's the best way to support this podcast
and the journey I'm on in my research and startup.
If you enjoy this thing, subscribe on YouTube,
review it with five stars on Apple Podcast,
support it on Patreon, or connect with me on Twitter,
@lexfriedman, spelled without the E,
just F-R-I-D-M-A-N.
I'm sure eventually you will figure it out.
And now let me leave you with some words from Ben Goertzel.
"Our language for describing emotions is very crude.
"That's what music is for."
Thank you for listening, and hope to see you next time.
(upbeat music)
§28Transcript · 4:08:50
(upbeat music)