Joscha Bach: Nature of Reality, Dreams, and Consciousness
嘉宾：Joscha Bach
全文文稿 · 按对话顺序保留完整内容，深入理解上下文。
本篇目录22 章节
§01Transcript · 00:00–08:23§02Transcript · 08:25–17:55§03Transcript · 17:57–26:33§04Transcript · 26:35–34:11§05Transcript · 34:13–42:36§06Transcript · 42:39–51:08§07Transcript · 51:10–59:35§08Transcript · 59:39–1:07:52§09Transcript · 1:07:54–1:16:35§10Transcript · 1:16:37–1:25:29§11Transcript · 1:25:34–1:34:07§12Transcript · 1:34:10–1:42:46§13Transcript · 1:42:47–1:51:46§14Transcript · 1:51:49–2:00:49§15Transcript · 2:00:52–2:09:19§16Transcript · 2:09:21–2:18:59§17Transcript · 2:19:02–2:28:18§18Transcript · 2:28:20–2:37:31§19Transcript · 2:37:34–2:47:30§20Transcript · 2:47:33–2:56:20§21Transcript · 2:56:21–3:04:38§22Transcript · 3:04:40–3:12:15§01Transcript · 00:00–08:23
The following is a conversation with Yosha Bach,
his second time on the podcast.
Yosha is one of the most fascinating minds in the world,
exploring the nature of intelligence,
cognition, computation, and consciousness.
To support this podcast, please check out our sponsors,
Coinbase, Codecademy, Linode, NetSuite, and ExpressVPN.
Their links are in the description.
This is the Lex Friedman Podcast,
and here is my conversation with Yosha Bach.
Thank you for once again coming on
to this particular Russian program,
and sticking to the theme of a Russian program.
Let's start with the darkest of topics.
- Привет.
(both laughing)
- So this is inspired by one of your tweets.
You wrote that, quote, "When life feels unbearable,
"I remind myself that I'm not a person.
"I'm a piece of software running on the brain
"of a random ape for a few decades.
"It's not the worst brain to run on."
Have you experienced low points in your life?
Have you experienced depression?
- Of course, we all experience low points in our life,
and we get appalled by the things,
by the ugliness of stuff around us.
We might get desperate about our lack of self-regulation,
and sometimes life is hard,
and I suspect you don't get to your life,
nobody does to get to their life without low points,
and without moments where they're despairing.
And I thought that, let's capture this state,
and how to deal with that state.
And I found that very often,
you realize that when you stop taking things personally,
when you realize that this notion of a person is a fiction,
similar as it is in Westworld,
where the robots realize that their memories and desires
are the stuff that keeps them in the loop,
and they don't have to act on those memories and desires,
that our memories and expectations is what make us unhappy.
And the present rarely does.
The day in which we are, for the most part, it's okay, right?
When we are sitting here, right here, right now,
we can choose how we feel.
And the thing that affects us is the expectation
that something is going to be different
from what we want it to be,
or the memory that something was different
from what you wanted it to be.
And once we basically zoom out from all this,
what's left is not a person.
What's left is this state of being conscious,
which is a software state.
And software doesn't have an identity.
It's a physical law.
And it's a law that acts in all of us,
and it's embedded in a suitable substrate.
And we didn't pick that substrate, right?
We are mostly randomly instantiated on it.
And there are all these individuals,
and everybody has to be one of them.
And eventually you're stuck on one of them,
and have to deal with that.
- So you're like a leaf floating down the river.
You just have to accept that there's a river,
and you just float wherever it takes you.
- You don't have to do this.
The thing is that the illusion
that you are an agent is a construct.
What part of that is actually under your control?
And I think that our consciousness
is largely a control model for our own attention.
So we notice where we are looking,
and we can influence what we are looking,
how we are disambiguating things,
how we put things together in our mind.
And the whole system that runs us
is this big cybernetic motivational system.
So we're basically like a little monkey
sitting on top of an elephant.
And we can prod this elephant here and there
to go this way or that way.
And we might have the illusion that we are the elephant,
or that we are telling it what to do.
And sometimes we notice that it walks
into a completely different direction.
And we didn't set this thing up.
It just is the situation that we find ourselves in.
- How much prodding can we actually do of the elephant?
- A lot.
But I think that our consciousness
cannot create the motive force.
- Is the elephant consciousness in this metaphor?
- No, the monkey is the consciousness.
The monkey is the attentional system
that is observing things.
There is a large perceptual system
combined with a motivational system
that is actually providing the interface to everything
and our own consciousness,
I think is the tool that directs the attention
of that system, which means it singles out features
and performs conditional operations
for which it needs an index memory.
But this index memory is what we perceive
as our stream of consciousness.
But the consciousness is not in charge.
That's an illusion.
- So everything outside of that consciousness
is the elephant.
So it's the physics of the universe,
but it's also society that's outside of your...
- I would say the elephant is the agent.
So there is an environment to which the agent is stomping
and you are influencing a little part of that agent.
- So can you, is the agent a single human being?
What's, which object has agency?
- That's an interesting question.
I think a way to think about an agent
is that it's a controller with a set point generator.
The notion of a controller comes from cybernetics
and control theory.
Control system consists out of a system
that is regulating some value
and the deviation of that value from a set point.
And it has a sensor that measures the system's deviation
from that set point and an effector
that can be parametrized by the controller.
So the controller tells the effector to do a certain thing.
And the goal is to reduce the distance
between the set point and the current value of the system.
And there's environment which disturbs the regulated system,
which brings it away from that set point.
So simplest case is a thermostat.
The thermostat is really simple
because it doesn't have a model.
The thermostat is only trying to minimize
the set point deviation in the next moment.
And if you want to minimize the set point deviation
over a longer time span, you need to integrate it.
You need to model what is going to happen.
So for instance, when you think about
that your set point is to be comfortable in life,
maybe you need to make yourself uncomfortable first.
Right, so you need to make a model
of what's going to happen when.
This is task of the controller is to use its sensors
to measure the state of the environment
and the system that is being regulated
and figure out what to do.
And if the task is complex enough,
the set points are complicated enough.
And if the controller has enough capacity
and enough sensor feedback,
then the task of the controller is to make a model
of the entire universe that it's in,
the conditions under which it exists and of itself.
And this is a very complex agent.
And we are in that category.
And an agent is not necessarily a thing in the universe.
It's a class of models that we use
to interpret aspects of the universe.
And when we notice the environment around us,
a lot of things only make sense
at the level that we're entangled with them
if we interpret them as control systems
that make models of the world
and try to minimize their own set points.
- So but the models are the agents.
- The agent is a class of model.
And we notice that we are an agent ourself.
We are the agent that is using our own control model
to perform actions.
We notice we produce a change in the model
and things in the world change.
And this is how we discover the idea that we have a body,
that we are situated in an environment,
and that we have a first-person perspective.
- Still don't understand what's the best way to think
of which object has agency with respect to human beings.
Is it the body?
Is it the brain?
Is it the contents of the brain that has agency?
Like what's the actuators that you're referring to?
What is the controller and where does it reside?
Or is it these impossible things?
'Cause I keep trying to ground it to space-time,
the three-dimensional space and the one dimension of time.
What's the agent in that for humans?
- There is not just one.
It depends on the way in which you're looking at the thing
and which you're framing it.
Imagine that you are, say, Angela Merkel,
and you are acting on behalf of Germany.
Then you could say that Germany is the agent.
And in the mind of Angela Merkel,
she is Germany to some extent,
because in the way in which she acts,
§02Transcript · 08:25–17:55
the destiny of Germany changes.
There are things that she can change
that basically affect the behavior of that nation state.
- Okay, so it's hierarchies of...
To go to another one of your tweets
with, I think you were playfully mocking Jeff Hawkins
with saying his brain's all the way down.
So it's like, it's agents all the way down?
It's agents made up of agents made up of agents?
Like if Angela Merkel's Germany
and Germany's made up of a bunch of people
and the people are themselves agents in some kind of context
and then the people are made up of cells, each individual.
So is it agents all the way down?
- I suspect that has to be like this
in a world where things are self-organizing.
Most of the complexity that we are looking at,
everything in life is about self-organization.
So I think up from the level of life, you have agents.
And below life, you rarely have agents
because sometimes you have control systems
that emerge randomly in nature
and try to achieve a set point,
but they're not that interesting agents that make models.
And because to make an interesting model of the world,
you typically need a system that is Turing complete.
- Can I ask you a personal question?
What's the line between life and non-life?
It's personal because you're a life form.
So where do you think in this emerging complexity,
at which point does the thing start being living
and have agency?
- Personally, I think that the simplest answer
that is that life is cells because--
- Life is what?
- Cells.
- Cells.
- Biological cells.
So it's a particular kind of principle
that we have discovered to exist in nature.
It's modular stuff that consists out of basically
this DNA tape with a redried head on top of it
that is able to perform arbitrary computations
and state transitions within the cell.
And it's combined with a membrane
that insulates the cell from its environment.
And there are chemical reactions inside of the cell
that are in disequilibrium.
And the cell is running in such a way
that this disequilibrium doesn't disappear.
And if the cell goes into an equilibrium state, it dies.
And it requires something like a neck entropy extractor
to maintain this disequilibrium.
So it's able to harvest neck entropy from its environment
and keep itself running.
- Yeah, so there's information and there's a wall
to maintain this disequilibrium.
But isn't this very earth-centric?
Like what you're referring to as-
- I'm not making a normative notion.
You could say that there are probably other things
in the universe that are cell-like and life-like,
and you could also call them life,
but eventually it's just a willingness
to find an agreement of how to use the terms.
I like cells because it's completely coextensional
with the way that we use the word
even before we knew about cells.
So people were pointing at some stuff
and saying this is somehow animate,
and this is very different from the non-animate stuff,
and what's the difference between the living
and the dead stuff?
And it's mostly whether the cells are working or not.
And also this boundary of life where we say that,
for instance, a virus is basically an information packet
that is subverting the cell and not life by itself.
That makes sense to me.
And it's somewhat arbitrary.
You could, of course, say that systems
that permanently maintain a disequilibrium
and can self-replicate are always life.
And maybe that's a useful definition too,
but this is eventually just how you want to use the word.
- Is it so useful for conversation,
but is it somehow fundamental to the universe?
Do you think there's an actual line
to eventually be drawn between life and non-life,
or is it all a kind of continuum?
- I don't think it's a continuum,
but there's nothing magical that is happening.
Living systems are a certain type of machine.
- What about non-living systems?
Is it also a machine?
- There are non-living machines,
but the question is at which point is a system able
to perform arbitrary state transitions
to make representations?
And living things can do this.
And of course, we can also build non-living things
that can do this, but we don't know anything in nature
that is not a cell and is not created by cellular life
that is able to do that.
Not only do we not know,
I don't think we have the tools to see otherwise.
I always worry that we look at the world too narrowly.
Like there could be life of a very different kind
right under our noses that we're just not seeing
because we're not either limitations
of our cognitive capacity,
or we're just not open-minded enough,
either with the tools of science
or just the tools of our mind.
- Yeah, that's possible.
I find this thought very fascinating.
And I suspect that many of us ask ourselves since childhood,
what are the things that we are missing?
What kind of systems and interconnections exist
that are outside of our gaze?
But we are looking for it,
and physics doesn't have much room at the moment
for opening up something that would not violate
the conservation of information as we know it.
- Yeah, but I wonder about time scale and scale,
spatial scale, whether we just need to open up our idea
of how life presents itself.
It could be operating at a much slower time scale,
a much faster time scale.
And it's almost sad to think that there's all this life
around us that we're not seeing,
because we're just not thinking in terms of the right scale,
both time and space.
- What is your definition of life?
What do you understand as life?
- Entities of sufficiently high complexity
that are full of surprises.
(both laughing)
I don't know, I don't have a free will,
so that just came out of my mouth.
I'm not sure that even makes sense.
There's certain characteristics.
So complexity seems to be a necessary property of life.
And I almost want to say it has ability
to do something unexpected.
- It seems to me that life is the main source
of complexity on earth.
- Yes.
- And complexity is basically a bridgehead
that order builds into chaos by modeling,
by processing information in such a way
that you can perform reactions
that would not be possible for dump systems.
And this means that you can harvest like entropy
that dump systems cannot harvest.
And this is what complexity is mostly about.
In some sense, the purpose of life is to create complexity.
- Yeah, increasing.
I mean, there seems to be some kind of universal drive
towards increasing pockets of complexity.
I don't know what that is.
That seems to be like a fundamental,
I don't know if it's a property of the universe
or it's just a consequence of the way the universe works.
But there seems to be this small pockets
of emergent complexity that builds on top of each other
and starts having like greater and greater complexity
by having like a hierarchy of complexity.
Little organisms building up a little society
that then operates almost as an individual organism itself.
And all of a sudden you have Germany and Merkel.
- But that's not obvious to me.
Everything that goes up has to come down at some point.
Right, so if you see this big exponential curve somewhere,
it's usually the beginning of an S-curve.
Or something eventually reaches saturation
and the S-curve is the beginning of some kind of bump
that goes down again.
And there is just this thing that when you are
in sight of an evolution of life,
you are on top of a puddle of negentropy
that is being sucked dry by life.
And during that happening, you see an increase in complexity
because life forms are competing with each other
to get more and more and finer and finer corner
of that negentropy extraction.
- But I feel like that's a gradual, beautiful process
that almost follows a process akin to evolution.
And the way it comes down is not the same way it came up.
The way it comes down is usually harshly and quickly.
So usually there's some kind of catastrophic event.
- Well, the Roman Empire took a long time.
- But would you classify that
as a decrease in complexity though?
- Yes, I think that this size of the cities
that could be fed has decreased dramatically.
And you could see that the quality of the art decreased
and it did so gradually.
And maybe future generations,
when they look at the history of the United States
in the 21st century,
§03Transcript · 17:57–26:33
will also talk about the gradual decline,
not something that suddenly happens.
- Do you have a sense of where we are?
Are we on the exponential rise?
Are we at the peak?
Or are we at the downslope of the United States empire?
- It's very hard to say from a single human perspective,
but it seems to me that we are probably at the peak.
- I think that's probably the definition
of like optimism and cynicism.
So my nature of optimism is I think we're on the rise.
(both laughing)
But I think this is just all a matter of perspective.
Nobody knows, but I do think that erring
on the side of optimism,
like you need a sufficient number,
you need a minimum number of optimists
in order to make that up thing actually work.
And so I tend to be on the side of the optimists.
- I think that we are basically a species of grasshoppers
that have turned into locusts.
And when you are in that locust mode,
you see an amazing rise of population numbers
and of the complexity of the interactions
between the individuals.
But it's ultimately the question is, is it sustainable?
- See, I think we're a bunch of lions and tigers
that have become domesticated cats,
to use a different metaphor.
And so I'm not exactly sure we're so destructive,
we're just softer and nicer and lazier.
- I think we have monkeys and not the cats.
And if you look at the monkeys, they are very busy.
- The ones that have a lot of sex, those monkeys?
- Not just the bonobos.
I think that all the monkeys are basically
a discontent species that always needs to meddle.
- Well, the gorillas seem to have
a little bit more of a structure,
but it's a different part of the tree.
Okay, you mentioned the elephant
and the monkey riding the elephant.
And consciousness is the monkey.
And there's some prodding that the monkey gets to do.
And sometimes the elephant listens.
I heard you got into some contentious,
maybe you can correct me,
but I heard you got into some contentious
free will discussions.
Is this with Sam Harris or something like that?
- Not that I know of.
- Some people on Clubhouse told me
you made a bunch of big debate points about free will.
Well, let me just then ask you,
where in terms of the monkey and the elephant,
do you think we land in terms of the illusion of free will?
How much control does the monkey have?
- We have to think about what the free will is
in the first place.
We are not the machine.
We are not the thing that is making the decisions.
We are a model of that decision-making process.
And there is a difference between
making your own decisions
and predicting your own decisions.
And that difference is the first person perspective.
And what basically makes decision-making
under conditions of free will distinct
from just automatically doing the best thing
is that we often don't know what the best thing is.
We make decisions under uncertainty.
We make informed bets using a betting algorithm
that we don't yet understand
because we haven't reverse engineered
our own minds sufficiently.
We don't know the expected rewards.
We don't know the mechanism by which we estimate
the rewards and so on.
- But there is an algorithm.
- We observe ourselves performing
where we see that we weigh facts and factors
and the future, and then some kind of possibility,
some motive gets raised to an intention.
And that's informed bet that the system is making.
And that making of the informed bet,
the representation of that is what we call free will.
And it seems to be paradoxical
because we think that's the crucial thing
is that it's somehow indeterministic.
And yet if it was indeterministic, it would be random.
And it cannot be random because if it was random,
if just dice were being thrown in the universe
randomly forces you to do things, it would be meaningless.
So the important part of the decisions
is always the deterministic stuff.
But it appears to be indeterministic to you
because it's unpredictable.
Because if it was predictable,
you wouldn't experience it as a free will decision.
You would experience it as just doing
the necessary right thing.
And you see this continuum between the free will
and the execution of automatic behavior
when you're observing other people.
So for instance, when you are observing your own children,
if you don't understand them,
you will use this agent model
where you have an agent with a set point generator.
And the agent is doing the best it can
to minimize the difference to the set point.
And it might be confused and sometimes impulsive or whatever,
but it's acting on its own free will.
And when you understand what happens
in the mind of the child, you see that it's automatic.
And you can outmodel the child,
you can build things around the child
that will lead the child to making exactly the decision
that you are predicting.
And under these circumstances,
like when you are a stage musician
or somebody who is dealing with people
that you sell a car to,
and you completely understand the psychology
and the impulses and the space of thoughts
that this individual can have at that moment.
Under these circumstances,
it makes no sense to attribute free will
because it's no longer decision-making under uncertainty.
You are already certain.
For them, there's uncertainty,
but you already know what they're doing.
- But what about for you?
So is this akin to systems like cellular automata
where it's deterministic,
but when you squint your eyes a little bit,
it starts to look like there's agents making decisions
at the higher, sort of when you zoom out
and look at the entities that are composed
by the individual cells.
Even though there's underlying simple rules
that make the system evolve in deterministic ways,
it looks like there's organisms making decisions.
Is that where the illusion of free will emerges,
that jump in scale?
- It's a particular type of model,
but this jump in scale is crucial.
The jump in scale happens whenever
you have too many parts to count
and you cannot make a model at that level,
and you try to find some higher level regularity.
And the higher level regularity is a pattern
that you project into the world to make sense of it.
And agency is one of these patterns, right?
You have all these cells that interact with each other,
and the cells in our body are set up in such a way
that they benefit if their behavior is coherent,
which means that they act as if
they were serving a common goal.
And that means that they will evolve regulation mechanisms
that act as if they were serving a common goal.
And now you can make sense of all these cells
by projecting the common goal into them.
- Right, so for you then, free will is an illusion.
- No, it's a model, and it's a construct.
It's basically a model that the system
is making of its own behavior,
and it's the best model that it can come up with
under the circumstances, and it can get replaced
by a different model, which is automatic behavior,
when you fully understand the mechanism
under which you are acting.
- Yeah, but another word for model is what?
Story.
So it's the story you're telling.
I mean, do you actually have control?
Is there such a thing as a you,
and is there such a thing as you having control?
So like, are you manifesting your evolution as an entity?
- In some sense, the you is the model of the system
that is in control.
It's a story that the system tells itself
about somebody who is in control.
- Yeah.
- And the contents of that model are being used
to inform the behavior of the system.
- Okay.
- So the system is completely mechanical,
and the system creates that story like a loom,
and then it uses the contents of that story
to inform its actions and writes the results
of that action into the story.
- So how is that not an illusion?
The story is written then,
or rather, we're not the writers of the story.
- Yes, but we always knew that.
- No, we don't know that.
When did we know that?
- I think that's mostly a confusion about concepts.
The conceptual illusion in our culture comes from the idea
that we live in physical reality,
§04Transcript · 26:35–34:11
and that we experience physical reality,
and that we have ideas about it.
- Yep.
- And then you have this dualist interpretation
where you have two substances, res extensa,
the world that you can touch
and that is made of extended things,
and res cogitans, which is the world of ideas.
And in fact, both of them are mental representations.
One is the representations of the world as a game engine
that your mind generates to make sense
of the perceptual data.
And the other one's-- - That's the physical world?
- Yes, that's what we perceive as the physical world.
But we already know that the physical world
is nothing like that, right?
Quantum mechanics is very different
from what you and me perceive as the world.
The world that you and me perceive is a game engine.
- Yeah.
- And there are no colors and sounds in the physical world.
They only exist in the game engine generated by your brain.
And then you have ideas that cannot be mapped
onto extended regions, right?
So the objects that have a spatial extension
in the game engine are res extensa,
and the objects that don't have a physical extension
in the game engine are ideas.
And they both interact in our mind
to produce models of the world.
- Yep, but when you play video games,
I understand that what's actually happening
is zeros and ones inside of a computer,
inside of a CPU and a GPU,
but you're still seeing the rendering of that.
And you're still making decisions
whether to shoot to turn left or to turn right
if you're playing a shooter.
Every time I start thinking about Skyrim
and Elder Scrolls and walking around in beautiful nature
and swinging a sword.
But it feels like you're making decisions
inside that video game.
So even though you don't have direct access
in terms of perception to the bits, to the zeros and ones,
it still feels like you're making decisions
and your decisions actually feels like they're being applied
all the way down to the zeros and ones.
- Yes.
- So it feels like you have control
even though you don't have direct access to reality.
So there is basically a special character in the video game
that is being created by the video game engine.
- Yeah.
- And this character is serving the aesthetics
of the video game.
And that is you.
- Yes, but I feel like I have control inside the video game.
Like all those 12 year olds
that kick my ass on the internet.
- So when you play the video game,
it doesn't really matter that there's zeros and ones.
You don't care about the widths of the bus.
You don't care about the nature of the CPU that it runs on.
What you care about are the properties of the game
that you're playing.
And you hope that the CPU is good enough.
- Yes.
- And a similar thing happens when we interact with physics.
The world that you and me are in is not the physical world.
The world that you and me are in is a dream world.
- How close is it to the real world though?
- We know that it's not very close,
but we know that the dynamics of the dream world
match the dynamics of the physical world
to a certain degree of resolution.
- Right.
- But the causal structure of the dream world is different.
- So you see, for instance,
waves crashing on your feet, right?
But there are no waves in the ocean.
There's only water molecules that have tangents
between the molecules that are the result of electrons
in the molecules interacting with each other.
- Aren't they like very consistent?
We're just seeing a very crude approximation.
Isn't our dream world very consistent?
Like to the point of being mapped directly one-to-one
to the actual physical world,
as opposed to us being completely tricked.
This is like where you have like Donald Trump.
- It's not a trick.
That's my point.
It's not an illusion.
It's a form of data compression.
- Yeah, yeah.
- It's an attempt to deal with the dynamics
of too many parts to count
at the level at which we're entangled
with the best model that you can find.
- Yeah, so we can act in that dream world
and our actions have impact in the real world,
in the physical world.
- Yes.
- To which we don't have access.
- Yes, but it's basically like accepting the fact
that the software that we live in,
the dream that we live in,
is generated by something outside of this world
that you and me are in.
- So is the software deterministic
and do we not have any control?
Do we have...
So free will is having a conscious being.
Free will is the monkey being able to steer the elephant.
- No, it's slightly different.
Basically in the same way as you are modeling
the water molecules in the ocean
that engulf your feet when you are walking on the beach
as waves and there are no waves,
but only the atoms on more complicated stuff
underneath the atoms and so on.
And you know that, right?
You would accept, yes,
there is a certain abstraction that happens here.
It's a simplification of what happens.
And the simplification that is designed
in such a way that your brain can deal with it,
temporarily and spatially in terms of resources
and tuned for the predictive value.
So you can predict with some accuracy
whether your feet are going to get wet or not.
- But it's a really good interface and approximation.
- Yes.
- It's like E equals MC squared is a good...
Equations are a good approximation
for what they're much better approximation.
So to me, waves is a really nice approximation
of what's all the complexity that's happening underneath.
- Basically it's a machine learning model
that is constantly tuned to minimize surprises.
So it basically tries to predict as well as it can
what you're going to perceive next.
- Are we talking about...
Which is the machine learning,
our perception system or the dream world?
- The machine world is...
Dream world is the result of the machine learning process
of the perception system.
- That's doing the compression.
- Yes.
And the model of you as an agent
is not a different type of model or it's a different type,
but not different as in its model-like nature
from the model of the ocean, right?
Some things are oceans, some things are agents.
And one of these agents is using your own control model,
the output of your model,
the things that you perceive yourself as doing.
And that is you.
- What about the fact that like when you're standing
with the water on your feet
and you're looking out into the vast,
like open water of the ocean,
and then there's a beautiful sunset.
And it...
The fact that it's beautiful
and then maybe you have like friends or a loved one with you
and like you feel love.
What is that?
As the dream world or what is that?
- Yes, it's all happening inside of the dream.
- Okay.
But see, the word dream makes it seem like it's not real.
- No, of course it's not real.
The physical universe is real,
but the physical universe is incomprehensible
and it doesn't have any feeling of realness.
The feeling of realness that you experience
gets attached to certain representations
where your brain assesses,
this is the best model of reality that I have.
- So the only thing that's real to you
is the thing that's happening at the very base of reality,
like the...
- Yeah, for something to be real,
it needs to be implemented.
So the model that you have of reality
is real in as far as it is a model, right?
It's an appropriate description of the world
to say that there are models that are being experienced.
But the world that you experience
is not necessarily implemented.
There is a difference between a reality,
a simulation, and a simulacrum.
The reality that we're talking about
is something that fully emerges
over a causally closed lowest layer.
And the idea of physicalism is that we are in that layer,
that basically our world emerges over that.
§05Transcript · 34:13–42:36
Every alternative to physicalism is a simulation theory,
which basically says that we are
in some kind of simulation universe
and the real world needs to be an apparent universe of that,
where the actual causal structure is, right?
And when you look at the ocean in your own mind,
you are looking at a simulation
that explains what you're going to see next.
- So we are living in a simulation.
- Yes, but a simulation generated by our own brains.
- Yeah.
- And this simulation is different from the physical reality
because the causal structure that is being produced,
what you are seeing,
is different from the causal structure of physics.
- But consistent.
- Hopefully.
If not, then you are going to end up
in some kind of institution
where people will take care of you
because your behavior will be inconsistent, right?
Your behavior needs to work in such a way
that it's interacting
with an accurately predictive model of reality.
And if your brain is unable
to make your model of reality predictive,
you will need help.
- So what do you think about Donald Hoffman's argument
that it doesn't have to be consistent,
the dream world to what he calls like the interface
to the actual physical reality,
where there could be evolution.
I think he makes an evolutionary argument,
which is like it could be an evolutionary advantage
to have the dream world drift away from physical reality.
- I think that only works if you have tenure.
As long as you're still interacting with the ground truth,
your model needs to be somewhat predictive.
- Well, in some sense,
humans have achieved a kind of tenure in the animal kingdom.
- Yeah, and at some point we became too big to fail,
so we became post-modernist.
(laughing)
- It all makes sense now.
- It's a version of reality that we like.
- Oh man, okay.
- Yeah, but basically you can do magic.
You can change your assessment of reality,
but eventually reality is going to come bite you in the ass
if it's not predictive.
- Do you have a sense
of what is that base layer of physical reality?
You have these attempts at the theories of everything,
the very, very small of like string theory,
or what Stephen Wolfram talks about with a hypergrass.
These are these tiny, tiny, tiny, tiny objects.
And then there is more like quantum mechanics
that's talking about objects that are much larger,
but still very, very, very tiny.
Do you have a sense of where the tiniest thing is
that is like at the lowest level,
the turtle at the very bottom?
Do you have a sense of what that turtle is?
- I don't think that you can talk about where it is
because space is emergent over the activity of these things.
So space, the coordinates only exist
in relation to the other things.
And so you could, in some sense, abstract it into locations
that can hold information and trajectories
that the information can take
between the different locations.
And this is how we construct our notion of space.
And physicists usually have a notion of space
that is continuous.
And this is a point where I tend to agree
with people like Stephen Wolfram,
who are very skeptical of the geometric notions.
I think that geometry is the dynamics
of too many parts to count.
And when there are no infinities,
if there were two infinities,
you would be running into contradictions,
which is in some sense what Gödel and Turing discovered
in response to Hilbert's call.
- So there are no infinities.
- There are no infinities.
- Infinity is fake.
- There is unboundedness,
but if you have a language that talks about infinity,
at some point the language is going to contradict itself,
which means it's no longer valid.
In order to deal with infinities in mathematics,
you have to postulate the existence initially.
You cannot construct the infinities.
And that's an issue, right?
You cannot build up an infinity from zero.
But in practice, you never do this, right?
When you perform calculations,
you only look at the dynamics of too many parts to count.
And usually these numbers are not that large.
They're not googles or something.
The infinities that we are dealing with in our universe
are mathematically speaking, relatively small integers.
And still, what we're looking at is dynamics
where a trillion things behave similar
to a hundred trillion things,
or something that is very, very large,
because they're converging.
And these convergent dynamics, these operators,
this is what we deal with when we are doing the geometry.
Geometry is stuff where we can pretend that it's continuous,
because if we subdivide the space sufficiently fine-grained,
these things approach a certain dynamic.
And this approach dynamic, that is what we mean by it.
But I don't think that infinity would work.
So to speak that you would know the last digit of pi,
and that you have a physical process
that rests on knowing the last digit of pi.
- Yeah, that could be just a peculiar quirk
of human cognition that we like discreet.
Discreet makes sense to us.
Infinity doesn't.
So in terms of our intuitions.
- No, the issue is that everything that we think about
needs to be expressed in some kind of mental language,
not necessarily a natural language,
but some kind of mathematical language
that your neurons can speak,
that refers to something in the world.
And what we have discovered is that
we cannot construct a notion of infinity
without running into contradictions,
which means that such a language is no longer valid.
And I suspect this is what made Pythagoras so unhappy
when somebody came up with the notion of irrational numbers
before it was time, right?
There's this myth that he had this person killed
when he blepped out the secret,
that not everything can be expressed
as a ratio between two numbers,
but there are numbers between the ratios.
The world was not ready for this.
And I think he was right.
That has confused mathematicians very seriously
because these numbers are not values, they are functions.
And so you can calculate these functions
to a certain degree of approximation,
but you cannot pretend that pi has actually a value.
Pi is a function that would approach this value
to some degree, but nothing in the world
rests on knowing pi.
- How important is this distinction
between discrete and continuous for you to get to the bottom?
'Cause there's a, I mean,
in discussion of your favorite flavor
of the theory of everything, there's a few on the table.
So there's string theory, there's particular,
there's a loop quantum gravity,
which focused on one particular unification.
There's just a bunch of favorite flavors
of different people trying to propose
a theory of everything.
Eric Weinstein and a bunch of people throughout history.
And then of course, Stephen Wolfram,
who I think is one of the only people doing a discrete.
- No, no, there's a bunch of physicists
who do this right now.
And like Toffoli and Tomasello.
And digital physics is something that is, I think,
growing in popularity.
But the main reason why this is interesting is
because it's important sometimes to settle disagreements.
I don't think that you need infinities at all
and you never needed them.
You can always deal with very large numbers
and you can deal with limits, right?
We are fine with doing that.
You don't need any kind of infinity.
You can build your computer algebra systems just as well
without believing in infinity in the first place.
- So you're okay with limits?
- Yeah, so basically a limit means that something
is behaving pretty much the same
if you make the number larger.
Because it's converging to a certain value
and at some point the difference becomes negligible
and you can no longer measure it.
And in this sense, you have things that,
if you have an n-gon which has enough corners,
then it's going to behave like a circle at some point.
And it's only going to be in some kind of esoteric thing
that cannot exist in the physical universe
that you would be talking about this perfect circle.
And now it turns out that it also wouldn't work
in mathematics because you cannot construct mathematics
that has infinite resolution
without running into contradictions.
So that is itself not that important
because we never did that, right?
It's just a thing that some people thought we could.
§06Transcript · 42:39–51:08
And this leads to confusion.
So for instance, Roger Penrose uses this as an argument
to say that there are certain things
that mathematicians can do dealing with infinities.
And by extension, our mind can do
that computers cannot do.
- Yeah, he talks about that there's the human mind
can do certain mathematical things
that the computer as defined
by the universal Turing machine cannot.
- Yes.
- So that it has to do with infinity.
- Yes, it's one of the things.
So he is basically pointing at the fact
that there are things that are possible
in the mathematical mind and in pure mathematics
that are not possible in machines
that can be constructed in the physical universe.
And because he's an honest guy,
he thinks this means that present physics
cannot explain operations that happen in our mind.
- Do you think he's right on the...
So let's leave his discussion of consciousness aside
for the moment.
Do you think he's right about just
what he's basically referring to as intelligence?
So is the human mind fundamentally more capable
as a thinking machine than a universal Turing machine?
- No.
- So he's suggesting that, right?
- So our mind is actually less than a Turing machine.
There can be no Turing machine
because it's defined as having an infinite tape.
And we always only have a finite tape.
- But he's saying it's better.
- Our minds can only perform finitely many operations.
Yes, he thinks so.
- It can do the kind of computation
that a Turing machine cannot.
- And that's because he thinks that our minds
can do operations that have infinite resolution
in some sense.
And I don't think that's the case.
Our minds are just able to discover these limit operators
over too many parts to count.
- I see.
What about his idea that consciousness
is more than a computation?
So it's more than something that a Turing machine
can do.
So again, saying that there's something special
about our mind that cannot be replicated in the machine.
- The issue is that I don't even know
how to construct a language
to express this statement correctly.
- Well, the basic statement is,
there's a human experience that includes intelligence,
that includes self-awareness,
that includes the hard problem of consciousness.
And the question is, can that be fully simulated
in the computer, in the mathematical model of the computer
as we understand it today?
Roger Penrose says no.
So the universal Turing machine
cannot simulate the universe.
- So the interesting question is,
and you have to ask him this, is why not?
What is the specific thing that cannot be modeled?
And when I looked at his writings,
and I haven't read all of it,
but when I read, for instance,
the section that he writes in the introduction
to "A Road to Infinity,"
the thing that he specifically refers to
is the way in which human minds deal with infinities.
And that itself can, I think, easily be deconstructed.
A lot of people feel that our experience
cannot be explained in a mechanical way,
and therefore it needs to be different.
And I concur, our experience is not mechanical.
Our experience is simulated.
It exists only in a simulation.
Only a simulation can be conscious.
Physical systems cannot be conscious
because they're only mechanical.
Cells cannot be conscious.
Neurons cannot be conscious.
Brains cannot be conscious.
People cannot be conscious,
as far as if you understand them as physical systems.
What can be conscious is the story of the system
in the world where you write all these things
into the story.
You have experiences for the same reason
that a character in a novel has experiences,
because it's written into the story.
And now the system is acting on that story.
And it's not a story that is written in a natural language.
It's written in a perceptual language,
in this multimedia language of the game engine.
And in there, you write in what kind of experience you have
and what this means for the behavior of the system,
for your behavior tendencies, for your focus,
for your attention, for your experience of valence,
and so on.
And this is being used to inform the behavior of the system
in the next step.
And then the story updates with the reactions of the system
and the changes in the world and so on.
And you live inside of that model.
You don't live inside of the physical reality.
- And, I mean, just to linger on it,
like you see, okay, it's in the perceptual language,
the multimodal perceptual language.
That's the experience.
That's what consciousness is within that model,
within that story.
But do you have agency?
When you play a video game, you can turn left
and you can turn right in that story.
So in that dream world, how much control do you,
is there such a thing as you in that story?
Is it right to say the main character,
everybody's NPCs, and then there's the main character,
and you're controlling the main character?
Or is that an illusion?
Is there a main character that you're controlling?
I'm getting to the point of the free will point.
- Imagine that you are building a robot that plays soccer.
And you've been to MIT computer science,
you basically know how to do that.
And so you would say the robot is an agent
that solves a control problem.
How to get the ball into the goal.
And it needs to perceive the world,
and the world is disturbing him in trying to do this.
So he has to control many variables to make that happen
and to project itself and the ball into the future
and understand its position on the field
relative to the ball and so on,
and the position of its limbs
or in the space around it and so on.
So it needs to have an adequate model
that abstracting reality in a useful way.
And you could say that this robot does have agency
over what it's doing in some sense.
And the model is going to be a control model.
And inside of that control model,
you can possibly get to a point
where this thing is sufficiently abstract
to discover its own agency.
Our current robots don't do that.
They don't have a unified model of the universe.
But there's not a reason why we shouldn't be getting there
at some point in the not too distant future.
And once that happens,
you will notice that the robot tells a story
about a robot playing soccer.
So the robot will experience itself playing soccer
in a simulation of the world that it uses
to construct a model of the locations of its legs
and limbs in space on the field
with relationship to the ball.
And it's not going to be at the level of the molecules.
It will be an abstraction that is exactly at the level
that is most suitable for past planning
of the movements of the robot.
Right, it's going to be a high-level abstraction,
but a very useful one that is as predictive
as you can make it.
And in that side of that story,
there is a model of the agency of that system.
So this model can accurately predict
that the contents of the model are going to be driving
the behavior of the robot in the immediate future.
- But there's the hard problem of consciousness,
which I would also,
there's a subjective experience of free will as well,
that I'm not sure where the robot gets that,
where that little leap is.
Because for me right now,
everything I imagine with that robot,
as it gets more and more and more sophisticated,
the agency comes from the programmer of the robot still,
of what was programmed in.
- You could probably do an end-to-end learning system.
You maybe need to give it a few priors,
so you nudge the architecture in the right direction
that it converges more quickly.
But ultimately discovering the suitable hyper parameters
of the architecture is also only a search process, right?
And as the search process was evolution,
that has informed our brain architecture,
so we can converge in a single lifetime
on useful interaction with the world.
- See, the problem is,
if we define hyper parameters broadly,
so it's not just the parameters that control
this end-to-end learning system,
but the entirety of the design of the robot.
§07Transcript · 51:10–59:35
You have to remove the human completely from the picture.
And then in order to build the robot,
you have to create an entire universe.
'Cause you have to go, you can't just shortcut evolution,
you have to go from the very beginning.
In order for it to have,
'cause I feel like there's always a human
pulling the strings,
and that makes it seem like the robot is cheating,
it's getting a shortcut to consciousness.
- And you are looking at the current Boston Dynamics robots,
it doesn't look as if there is somebody pulling the strings,
it doesn't look like cheating anymore.
- Okay, so let's go there,
'cause I got to talk to you about this.
So obviously with the case of Boston Dynamics,
as you may or may not know,
it's always either hard-coded or remote-controlled.
There's no intelligence.
- I don't know how the current generation
of Boston Dynamics robots works,
but what I've been told about the previous ones
was that it's basically all cybernetic control,
which means you still have feedback mechanisms and so on,
but it's not deep learning for the most part
as it's currently done.
It's for the most part just identifying a control hierarchy
that is congruent to the limbs that exist
and the parameters that need to be optimized
for the movement of these limbs,
and then there is a convergence progress.
So it's basically just regression
that you would need to control this.
But again, I don't know whether that's true,
that's just what I've been told about how they work.
- We have to separate several levels of discussions here.
So the only thing they do is pretty sophisticated control
with no machine learning
in order to maintain balance or to right itself.
It's a control problem in terms of using the actuators
to when it's pushed or when it steps on a thing
that's uneven, how to always maintain balance.
And there's a tricky set of heuristics around that,
but that's the only goal.
Everything you see Boston Dynamics doing
in terms of that to us humans is compelling,
which is any kind of higher order movement,
like turning, wiggling its butt,
like jumping back on its two feet, dancing.
Dancing is even worse because dancing is hard coded in.
It's choreographed by humans, it's choreography software.
So there is no, of all that high level movement,
there's no anything that you can call,
certainly can't call AI,
there's no even like basic heuristics,
it's all hard coded in.
And yet we humans immediately project agency onto them,
which is fascinating.
- So the gap here is it doesn't necessarily have agency.
What it has is cybernetic control.
And the cybernetic control means you have a hierarchy
of feedback loops that keep the behavior
in certain boundaries so the robot doesn't fall over
and it's able to perform the movements.
And the choreography cannot really happen
with motion capture because the robot would fall over
because the physics of the robot,
the weight distribution and so on is different
from the weight distribution in the human body.
So if you were using the directly motion captured movements
of a human body to project it into this robot,
it wouldn't work.
You can do this with a computer animation,
it will look a little bit off, but who cares.
But if you want to correct for the physics,
you need to basically tell the robot
where it should move its limbs,
and then the control algorithm is going
to approximate a solution that makes it possible
within the physics of the robot.
And you have to find the basic solution
for making that happen,
and there's probably going to be some regression necessary
to get the control architecture to make these movements.
- But those two layers are separate.
- Yes.
- So the thing, the higher level instruction
of how you should move and where you should move
is a higher level.
- Yes, so I expect that the control level of these robots
at some level is dumb.
This is just the physical control movement,
the motor architecture.
But it's a relatively smart motor architecture.
It's just that there is no high level deliberation
about what decisions to make necessarily.
- But see, it doesn't feel like free will or consciousness.
- No, no, that was not where I was trying to get to.
I think that in our own body, we have that too.
So we have a certain thing that is basically
just a cybernetic control architecture
that is moving our limbs.
And deep learning can help in discovering
such an architecture if you don't have it
in the first place.
If you already know your hardware,
you can maybe handcraft it.
But if you don't know your hardware,
you can search for such an architecture.
And this work already existed in the '80s and '90s.
People were starting to search for control architectures
by motor babbling and so on,
and just use reinforcement learning architectures
to discover such a thing.
And now imagine that you have
the cybernetic control architecture already inside of you.
And you extend this a little bit.
So you are seeking out food, for instance,
or rest, and so on.
And you get to have a baby at some point.
And now you add more and more control layers to this.
And the system is reverse engineering
its own control architecture,
and builds a high level model to synchronize
the pursuit of very different conflicting goals.
And this is how I think you get to purposes.
Purposes are models of your goals.
Your goals may be intrinsic as the result
of the different set point violations that you have,
hunger and thirst for very different things,
and rest and pain avoidance and so on.
And you put all these things together,
and eventually you need to come up with a strategy
to synchronize them all.
And you don't need just to do this alone by yourself,
because we are state building organisms.
We cannot function as isolation
the way that Homo sapiens is set up.
So our own behavior only makes sense
when you zoom out very far into a society,
or even into ecosystemic intelligence on the planet,
and our place in it.
So the individual behavior only makes sense
in these larger contexts.
And we have a number of priors built into us.
So we are behaving as if we were acting
on these high level goals,
pretty much right from the start.
And eventually in the course of our life,
we can reverse engineer the goals that we're acting on,
what actually are our higher level purposes.
And the more we understand that,
the more our behavior makes sense.
But this is all at this point,
complex stories within stories
that are driving our behavior.
- Yeah, I just don't know how big of a leap it is
to start create a system that's able to tell stories
within stories.
Like how big of a leap that is
from where currently Boston Dynamics is,
or any robot that's operating in the physical space.
And that leap might be big
if it requires to solve the hard problem of consciousness,
which is telling a hell of a good story.
- I suspect that consciousness itself is relatively simple.
What's hard is perception,
and the interface between perception and reasoning.
That's for instance, the idea of the consciousness prior
that would be built into such a system by Joshua Bengio.
And what he describes, and I think that's accurate,
is that our own model of the world
can be described through something like an energy function.
The energy function is modeling the contradictions
that exist within the model at any given point.
And you try to minimize these contradictions,
the tangents in the model.
And to do this, you need to sometimes test things.
You need to conditionally disambiguate figure and ground.
You need to distinguish whether this is true
or that is true, and so on.
Eventually you get to an interpretation,
but you will need to manually depress a few points
in your model to let it snap into a state that makes sense.
And this function that tries to get the biggest dip
in the energy function in your model,
according to Joshua Bengio, is related to consciousness.
It's a low dimensional discrete function
that tries to maximize this dip in the energy function.
- Yeah, I think I would need to dig into details
because I think the way he uses the word consciousness
is more akin to self-awareness,
like modeling yourself within the world,
as opposed to the subjective experience, the hard problem.
- No, it's not even the self is in the world.
The self is the agent, and you don't need to be aware
of yourself in order to be conscious.
The self is just a particular content that you can have,
but you don't have to have.
But you can be conscious in, for instance, a dream at night
§08Transcript · 59:39–1:07:52
or during a meditation state, but you don't have a self.
- Right.
- You're just aware of the fact that you are aware.
And what we mean by consciousness in the colloquial sense
is largely this reflexive self-awareness,
that we become aware of the fact
that we are paying attention,
that we are the thing that pays attention.
- We are the thing that pays attention, right.
I don't see where the awareness that we're aware,
the hard problem doesn't feel like it's solved.
I mean, it's called a hard problem for a reason
because it seems like there needs to be a major leap.
- Yeah, I think the major leap is to understand
how it is possible that a machine can dream,
that a physical system is able to create a representation
that the physical system is acting on,
and that is spun force and so on.
But once you accept the fact that you are not in physics,
but that you exist inside of the story,
I think the mystery disappears.
Everything is possible in a story.
- You exist inside the story.
Okay, so the machine--
- Consciousness is being written into the story.
The fact that you experience things
is written to the story.
You ask yourself, is this real what I'm seeing?
And your brain writes into the story, yes, it's real.
- So what about the perception of consciousness?
So to me, you look conscious.
So the illusion of consciousness,
the demonstration of consciousness,
I ask for the legged robot,
how do we make this legged robot conscious?
So there's two things,
and maybe you can tell me if they're neighboring ideas.
One is actually make it conscious,
and the other is make it appear conscious to others.
Are those related?
- Let's ask it from the other direction.
What would it take to make you not conscious?
So when you are thinking about how you perceive the world,
can you decide to switch from looking at qualia
to looking at representational states?
- And it turns out you can.
There is a particular way in which you can look at the world
and recognize its machine nature, including your own.
And in that state, you don't have that conscious experience
in this way anymore.
It becomes apparent as a representation.
Everything becomes opaque.
And I think this thing that you recognize
everything as a representation,
this is typically what we mean with enlightenment states.
And it can happen at the motivational level,
but you can also do this on the experiential level,
on the perceptual level.
- See, but then I can come back to a conscious state.
Okay, I particularly,
I'm referring to the social aspect,
that the demonstration of consciousness
is a really nice thing at a party
when you're trying to meet a new person.
It's a nice thing to know that they're conscious,
and they can, I don't know how fundamental consciousness
is in human interaction,
but it seems like to be at least an important part.
And I ask that in the same kind of way for robots.
You know, in order to create a rich,
compelling human robot interaction,
it feels like there needs to be elements of consciousness
within that interaction.
- And my cat is obviously conscious.
And so my cat can do this party trick.
She also knows that I am conscious,
be able to have feedback about the fact
that we are both acting on models of our own awareness.
- The question is how hard is it for the robot,
artificially created robot,
to achieve cat-level and party tricks?
- Yes, so the issue for me is currently not so much
on how to build a system that creates a story
about a robot that lives in the world,
but to make an adequate representation of the world.
And the model that you and me have is a unified one.
It's one where you basically make sense of everything
that you can perceive.
Every feature in the world that enters your perception
can be relationally mapped
to a unified model of everything.
And we don't have an AI that is able to construct
such a unified model yet.
- So you need that unified model to do the party trick?
- Yes, I think that it doesn't make sense
if this thing is conscious,
but not in the same universe as you,
because you could not relate to each other.
- So what's the process, would you say,
of engineering consciousness in a machine?
Like, what are the ideas here?
- So you probably want to have
some kind of perceptual system.
This perceptual system is a processing agent
that is able to track sensory data
and predict the next frame and the sensory data
from the previous frames of the sensory data
and the current state of the system.
So the current state of the system is,
in perception, instrumental
to predicting what happens next.
And this means you build lots and lots of functions
that take all the blips that you feel on your skin
and that you see on your retina, or that you hear,
and puts them into a set of relationships
that allows you to predict what kind of sensory data,
what kind of sensor of blips,
vector of blips you're going to perceive
in the next frame, right?
This is tuned, and it's constantly tuned
until it gets as accurate as it can.
- You build a very accurate prediction mechanism
that is step one of the perception.
So first you predict, then you perceive
and see the error in your prediction.
- And you have to do two things to make that happen.
One is you have to build a network of relationships
that are constraints,
that take all the variants in the world,
put each of the variances into a variable
that is connected with relationships to other variables.
And these relationships are computable functions
that constrain each other.
So when you see a nose
that points at a certain direction in space,
you have a constraint that says
there should be a face nearby that has the same direction.
And if that is not the case,
you have some kind of contradiction
that you need to resolve
because it's probably not a nose what you're looking at.
It just looks like one.
So you have to reinterpret the data
until you get to a point where your model converges.
And this process of making the sensory data
fit into your model structure
is what Piaget calls the assimilation.
And accommodation is the change of the models
where you change your model in such a way
that you can assimilate everything.
- So you're talking about building
a hell of an awesome perception system
that's able to do prediction and perception correct
and keep improving.
- No, wait, just figure it out.
- Wait, there's more.
- Yes, there's more.
So the first thing that we wanted to do
is we want to minimize the contradictions in the model.
And of course, it's very easy to make a model
in which you minimize the contradictions
just by allowing that it can be
in many, many possible states, right?
So if you increase degrees of freedom,
you will have fewer contradictions.
But you also want to reduce the degrees of freedom
because degrees of freedom mean uncertainty.
You want your model to reduce uncertainty
as much as possible.
But reducing uncertainty is expensive.
So you have to have a trade-off
between minimizing contradictions and reducing uncertainty.
And you have only a finite amount of compute
and experimental time and effort available
to reduce uncertainty in the world.
So you need to assign value to what you observe.
So you need some kind of motivational system
that is estimating what you should be looking at
and what you should be thinking about it,
how you should be applying your resources
to model what that is, right?
So you need to have something like convergence links
that tell you how to get from the present state of the model
to the next one.
You need to have these compatibility links
that tell you which constraints exist
and which constraint violations exist.
And you need to have some kind of motivational system
that tells you what to pay attention to.
So now we have a second agent next to the perceptual agent.
We have a motivational agent.
This is a cybernetic system
that is modeling what the system needs,
what's important for the system,
and that interacts with the perceptual system
to maximize the expected reward.
- And you're saying the motivational system
is some kind of, like,
what is it, a higher level narrative over some lower level?
- No, it's just your brainstem stuff,
§09Transcript · 1:07:54–1:16:35
the limbic system stuff that tells you,
okay, now you should get something to eat
because I've just measured your blood sugar.
- So you mean like motivational system,
like the lower level stuff, like hungry?
- Yes, there's basically physiological needs
and some cognitive needs and some social needs,
and they all interact.
And they all implemented different parts
in your nervous system as the motivational system.
But they're basically cybernetic feedback loops.
It's not that complicated.
It's just a lot of code.
And so you now have a motivational agent
that makes your robot go for the ball,
or that makes your worm go to eat food and so on.
And you have the perceptual system
that lets it predict the environment,
so it's able to solve that control problem to some degree.
And now what we learned is that it's very hard
to build a machine learning system
that looks at all the data simultaneously
to see what kind of relationships could exist between them.
So you need to selectively model the world.
You need to figure out,
where can I make the biggest difference
if I would put the following things together?
Sometimes you find a gradient for that, right?
When you have a gradient,
you don't need to remember where you came from.
You just follow the gradient
until it doesn't get any better.
But if you have a world where the problems are discontinuous
and the search spaces are discontinuous,
you need to retain memory of what you explored.
You need to construct a plan of what to explore next.
And this thing means that you have,
next to this perceptual construction system
and the motivational cybernetics,
an agent that is paying attention
to what it should select at any given moment
to maximize reward.
And this scanning system,
this attention agent is required for consciousness
and consciousness is its control model.
So it's the index memories that this thing retains
when it manipulates the perceptual representations
to maximize the value and minimize the conflicts
and to increase coherence.
So the purpose of consciousness is to create coherence
in your perceptual representations,
remove conflicts, predict the future,
construct counterfactual representations
so you can coordinate your actions and so on.
And in order to do this, it needs to form memories.
These memories are partial binding states
of the working memory contents
that are being revisited later on to backtrack,
to undo certain states, to look for alternatives.
And these index memories that you can recall,
that is what you perceive as your stream of consciousness.
And being able to recall these memories,
this is what makes you conscious.
If you could not remember what you paid attention to,
you wouldn't be conscious.
- So consciousness is the index in the memory database.
Okay.
But let me sneak up to the questions of consciousness
a little further.
So we usually relate suffering to consciousness.
So the capacity to suffer.
I think to me, that's a really strong sign of consciousness,
is a thing that can suffer.
How is that useful?
Suffering.
And like in your model, what you just described,
which is indexing of memories,
and what is the coherence with the perception,
with this predictive thing that's going on in the perception,
how does suffering relate to any of that?
You know, the higher level suffering that humans do?
- Basically pain is a reinforcement signal.
Pain is a signal that one part of your brain
sends to another part of your brain,
or in an abstract sense, part of your mind
sends to another part of the mind to regulate its behavior,
to tell it the behavior that you're currently exhibiting
should be improved.
And this is the signal that I tell you
to move away from what you're currently doing
and push into a different direction.
So pain gives part of you an impulse
to do something differently.
But sometimes this doesn't work,
because the training part of your brain
is talking to the wrong region,
or because it has the wrong model
of the relationships in the world.
Maybe you're mismodeling yourself,
or you're mismodeling the relationship
of yourself to the world,
or you're mismodeling the dynamics of the world.
So you're trying to improve something
that cannot be improved by generating more pain.
But the system doesn't have any alternative.
So it doesn't get better.
What do you do if something doesn't get better,
and you want it to get better?
You increase the strength of the signal.
And when the signal becomes chronic,
when it becomes permanent,
without a change inside, this is what we call suffering.
And the purpose of consciousness
is to deal with contradictions,
with things that cannot be resolved.
The purpose of consciousness, I think,
is similar to a conductor in an orchestra.
When everything works well,
the orchestra doesn't need much of a conductor,
as long as it's coherent.
But when there is a lack of coherence,
or something is consistently producing disharmony
and mismatches, then the conductor becomes alert
and interacts with it.
So suffering attracts the activity of our consciousness.
And the purpose of that is ideally
that we bring new layers online,
new layers of modeling that are able
to create a model of the dysregulation,
so we can deal with it.
And this means that we typically get
higher level consciousness, so to speak.
We get some consciousness above our pay grade, maybe,
if we have some suffering early in our life.
Most of the interesting people
had trauma early on in their childhood.
And trauma means that you are suffering an injury
for which the system is not prepared,
which it cannot deal with,
which it cannot insulate itself from.
So something breaks.
And this means that the behavior of the system
is permanently disturbed
in a way that some mismatch exists now in the regulation,
that just by following your impulses,
by following the pain in the direction which it hurts,
the situation doesn't improve, but get worse.
And so what needs to happen is that you grow up.
And that part that has grown up is able to deal
with the part that is stuck in this earlier phase.
- Yeah, so it leads to growth,
you're adding extra layers to your cognition.
Let me ask you then, 'cause I gotta stick on suffering,
the ethics of the whole thing.
So not our consciousness, but the consciousness of others.
You've tweeted, "One of my biggest fears
is that insects could be conscious.
The amount of suffering on earth would be unthinkable."
So when we think of other conscious beings,
is suffering a property of consciousness
that we're most concerned about?
So I'm still thinking about robots,
how to make sense of other non-human things
that appear to have the depth of experience
that humans have.
And to me, that means consciousness
and the darkest side of that, which is suffering,
the capacity to suffer.
And so I started thinking, how much responsibility
do we have for those other conscious beings?
That's where the definition of consciousness
becomes most urgent.
Like having to come up with a definition of consciousness
becomes most urgent, is who should we,
and should we not be torturing?
- There's no general answer to this.
Was Genghis Khan doing anything wrong?
It depends on how you look at it.
- Well, he drew a line somewhere
where this is us and that's them.
It's the circle of empathy.
It's like these, you don't have to use the word consciousness
but these are the things that matter to me
if they suffer or not.
And these are the things that don't matter to me.
- Yeah, but when one of his commanders failed him,
he broke his spine and let him die in a horrible way.
And so in some sense, I think he was indifferent
to suffering or he was not indifferent in the sense
that he didn't see it as useful if he inflicted suffering,
but he did not see it as something that had to be avoided.
That was not the goal.
The question was, how can I use suffering
and the infliction of suffering to reach my goals
from his perspective?
- I see, so like different societies throughout history
put different value on the--
- Different individuals, different psyches.
- But also even the objective of avoiding suffering.
Like some societies probably,
§10Transcript · 1:16:37–1:25:29
I mean, this is where like religious belief really helps
that afterlife, that doesn't matter that you suffer or die,
what matters is you suffer honorably, right?
So that you enter the afterlife.
- It seems to be superstitious to me,
basically beliefs that assert things
for which no evidence exists,
are incompatible with sound epistemology.
And I don't think that religion has to be superstitious,
otherwise it should be condemned in all cases.
- You're somebody who's saying we live in a dream world,
we have zero evidence for anything.
- So it's not the case.
There are limits to what languages can be constructed.
Mathematics breaks solid evidence for its own structure.
And once we have some idea of what languages exist
and how a system can learn
and what learning itself is in the first place,
and so on, we can begin to realize that our intuitions
that we are able to learn about the regularities
of the world and minimize appraisal
and understand the nature of our own agency
to some degree of abstraction.
That's not an illusion.
It's a useful approximation.
- Just because we live in a dream world
doesn't mean mathematics can't give us a consistent glimpse
of physical, of objective reality.
- We can basically distinguish useful encodings
from useless encodings.
And when we apply our truth-seeking to the world,
we know we usually cannot find out
whether a certain thing is true.
What we typically do is we take the state vector
of the universe, separate it into separate objects
that interact with each other, so interfaces.
And this distinction that we are making
is not completely arbitrary.
It's done to optimize the compression
that we can apply to our models of the universe.
So we can predict what's happening
with our limited resources.
In this sense, it's not arbitrary.
But the separation of the world into objects
that are somehow discrete and interacting with each other
is not the true reality, right?
The boundaries between the objects
are projected into the world, not arbitrarily projected,
but still, it's only an approximation
of what's actually the case.
And we sometimes notice that we run into contradictions
when we try to understand high-level things
like economic aspects of the world and so on,
or political aspects or psychological aspects
where we make simplifications.
And the objects that we are using to separate the world
are just one of many possible projections
of what's going on.
And so it's not in this postmodernist sense
completely arbitrary and you're free to pick
what you want or dismiss what you don't like
because it's all stories.
No, that's not true.
You have to show for every model
of how well it predicts the world.
So the confidence that you should have
in the entities of your models
should correspond to the evidence that you have.
- Can I ask you on a small tangent
to talk about your favorite set of ideas and people,
which is postmodernism?
- What? (laughs)
- What is postmodernism?
How would you define it?
And why to you is it not a useful framework of thought?
- Postmodernism is something
that I'm really not an expert on.
And postmodernism is a set of philosophical ideas
that is difficult to lump together,
that is characterized by some useful thinkers,
some of them poststructuralist and so on.
And I'm mostly not interested in it
because I think that it's not leading me anywhere
that I find particularly useful.
It's mostly, I think, born out of the insight
that the ontologies that we impose on the world
are not literally true,
and that we can often get to a different interpretation
by the world by using a different ontology
that is different separation of the world
into interacting objects.
But the idea that this makes the world
and a set of stories that are arbitrary, I think is wrong.
And the people that are engaging in this type of philosophy
are working in an area
that I largely don't find productive.
There's nothing useful coming out of this.
So this idea that truth is relative
is not something that has in some sense,
informed physics or theory of relativity.
And there is no feedback between those.
There is no meaningful influence
of this type of philosophy on the sciences
or in engineering or in politics.
But there is a very strong information of this on ideology,
because it basically has become an ideology
that is justifying itself by the notion
that truth is a relative concept.
And it's not being used in such a way
that the philosophers or sociologists
that take up these ideas say,
oh, I should doubt my own ideas
because maybe my separation of the world into objects
is not completely valid,
and I should maybe use a different one
and be open to a pluralism of ideas.
But it mostly exists to dismiss the ideas of other people.
- It becomes, yeah, it becomes a political weapon of sorts.
- Yes. - To achieve power.
- Basically, there's nothing wrong, I think,
with developing a philosophy around this,
but to develop norms around the idea
that truth is something that is completely negotiable
is incompatible with the scientific project.
And I think if the academia has no defense
against the ideological parts
of the postmodernist movement, it's doomed.
- Right, you have to acknowledge
the ideological part of any movement, actually,
including postmodernism.
- Well, the question is what an ideology is.
And to me, an ideology is basically a viral memeplex
that is changing your mind in such a way
that reality gets warped.
It gets warped in such a way that you're being cut off
from the rest of human thought space,
and you cannot consider things outside
of the range of ideas of your own ideology
as possibly true.
- Right, so, I mean, there's certain properties
to an ideology that make it harmful.
One of them is that dogmatism of just certainty,
dogged certainty in that you're right,
you have the truth, and nobody else does.
- Yeah, but what is creating the certainty?
It's very interesting to look at the type of model
that is being produced.
Is it basically just a strong prior?
And you tell people, oh, this idea
that you consider to be very true,
the evidence for this is actually just much weaker
than you thought, and look here at some studies.
No, this is not how it works.
It's usually normative, which means some thoughts
are unthinkable because they would change your identity
into something that is no longer acceptable.
And this cuts you off from considering an alternative,
and many de facto religions use this trick
to lock people into a certain mode of thought,
and this removes agency over your own thoughts,
and it's very ugly to me.
It's basically not just a process of domestication,
but it's actually an intellectual castration that happens.
It's an inability to think creatively
and to bring forth new thoughts.
- Can I ask you about substances, chemical substances
that affect the video game, the dream world?
So psychedelics that increasingly have been getting
a lot of research done on them.
So in general, psychedelics, psilocybin, MDMA,
but also a really interesting one, the big one,
which is DMT.
What and where are the places that these substances take
the mind that is operating in the dream world?
Do you have an interesting sense how this throws a wrinkle
into the prediction model?
Is it just some weird little quirk,
or is there some fundamental expansion
of the mind going on?
- I suspect that a way to look at psychedelics
is that they induce particular types
of lucid dreaming states.
So it's a state in which certain connections
are being severed in your mind, are no longer active.
Your mind basically gets free to move in a certain direction
because some inhibition, some particular inhibition
doesn't work anymore.
And as a result, you might stop having a self,
or you might stop perceiving the world as three-dimensional.
And you can explore that state.
And I suppose that for every state that can be induced
with psychedelics, there are people that are naturally
in that state.
So sometimes psychedelics shift you through a range
of possible mental states, and they can also shift you
out of the range of permissible mental states,
that is where you can make predictive models of reality.
And what I observe in people that use psychedelics a lot
is that they tend to be overfitting.
Overfitting means that you are using more bits
§11Transcript · 1:25:34–1:34:07
for modeling the dynamics of a function than you should.
And so you can fit your curve to extremely detailed things
in the past, but this model is no longer predictive
for the future.
- What is it about psychedelics that forces that?
I thought it would be the opposite.
I thought that it's a good mechanism
for generalization, for regularization.
So it feels like psychedelics expansion of the mind,
like taking you outside of, like forcing your model
to be non-predictive is a good thing.
Meaning like, it's almost like, okay,
what I would say is psychedelics are akin to is traveling
to a totally different environment.
Like going, if you've never been to like India
or something like that from the United States,
very different set of people, different culture,
different food, different roads and values
and all those kinds of things.
- Yeah, so psychedelics can, for instance,
teleport people into a universe that is hyperbolic,
which means that if you imagine a room that you're in,
you can turn around 360 degrees
and you didn't go full circle.
You need to go 720 degrees to go full circle.
- Exactly.
- So the things that people learn in that state
cannot be easily transferred in this universe
that we are in.
It could be that if they're able to abstract
and understand what happened to them,
that they understand that some part
of their spatial cognition has been desynchronized
and has found a different synchronization.
And this different synchronization
happens to be a hyperbolic one, right?
So you learn something interesting about your brain.
It's difficult to understand what exactly happened,
but we get a pretty good idea once we understand
how the brain is representing geometry.
- Yeah, but doesn't give you a fresh perspective
on the physical reality?
(knocking)
Who's making that sound?
Is it inside my head or is it external?
- Well, there is no sound outside of your mind,
but it's making sense of phenomenon physics.
(laughing)
- Yeah, in the physical reality, there's sound waves
traveling through air.
Okay.
- That's our model of what's happened.
- That's our model of what happened, right.
So, don't psychedelics give you a fresh perspective
on this physical reality?
Not this physical reality, but this more,
what do you call the dream world?
That's mapped directly to--
- The purpose of dreaming at night, I think,
is data augmentation.
- Well, exactly.
So that's very different.
That's very similar to psychedelics.
- So you basically change parameters
about the things that you have learned.
And for instance, when you are young,
you have seen things from certain perspectives,
but not from others.
So your brain is generating new perspectives
of objects that you already know,
which means they can learn to recognize them later
from different perspectives.
And I suspect that's the reason that many of us
remember to have flying dreams as children,
because it's just different perspectives
of the world that you already know.
And that it starts to generate
these different perspective changes,
and then it fluidly turns this into a flying dream
to make sense of what's happening, right?
So you fill in the gaps,
and suddenly you see yourself flying.
And similar things can happen with semantic relationships.
So it's not just spatial relationships,
but it can also be the relationships
between ideas that are being changed.
And it seems that the mechanisms
that make that happen during dreaming
are interacting with these same receptors
that are being simulated by psychedelics.
So I suspect that there is a thing
that I haven't read really about,
the way in which dreams are induced in the brain.
It's not just that the activity of the brain
gets tuned down because your eyes are closed
and you no longer get enough data from your eyes,
but there is a particular type of neurotransmitter
that is saturating your brain during these phases,
during the RM phases,
and you produce controlled hallucinations.
And psychedelics are linking into these mechanisms,
I suspect.
- So isn't that another trickier form of data augmentation?
- Yes, but it's also data augmentation
that can happen outside of the specification
that your brain is tuned to.
So basically people are overclocking their brains,
and that produces states
that are subjectively extremely interesting.
- Yeah, I just-
- But from the outside, very suspicious.
- So I think I'm over applying the metaphor
of a neural network in my own mind,
which I just think that doesn't lead to overfitting, right?
But you were just sort of anecdotally saying
my experiences with people that have done psychedelics
are that kind of quality.
- I think it typically happens.
So if you look at people like Timothy Leary,
and he has written beautiful manifestos
about the effect of LSD on people.
He genuinely believed, he writes in his manifestos,
that in the future, science and art
will only be done on psychedelics
because it's so much more efficient and so much better.
And he gave LSD to children in this community
of a few thousand people that he had near San Francisco.
And basically he was losing touch with reality.
He did not understand the effects,
the things that he was doing would have
on the reception of psychedelics by society,
because he was unable to think critically
about what happened.
What happened was that he got in a euphoric state.
That euphoric state happened because he was overfitting.
He was taking this sense of euphoria
and translating it into a model
of actual success in the world, right?
He was feeling better.
Limitations had disappeared,
that he experienced to be existing,
but he didn't get superpowers.
- I understand what you mean by overfitting now.
There's a lot of interpretation
to the term overfitting in this case, but I got you.
So he was getting positive rewards
from a lot of actions that he shouldn't have been doing.
- But not just this.
So if you take, for instance, John Lilly,
who was studying dolphin languages and aliens and so on,
a lot of people that use psychedelics became very loopy.
And the typical thing that you notice
when people are on psychedelics is that they are in a state
where they feel that everything can be explained now.
Everything is clear.
Everything is obvious.
And sometimes they have indeed discovered
a useful connection, but not always.
Very often these connections are over-interpretations.
- I wonder, you know, there's a question
of correlation versus causation.
And also I wonder if it's the psychedelics
or if it's more the social, like being the outsider
and having a strong community of outside
and having a leadership position
in an outsider cult-like community,
that could have a much stronger effect of overfitting
than do psychedelics themselves, the actual substances,
because it's a counterculture thing.
So it could be that as opposed to the actual substance.
If you're a boring person who wears a suit and tie
and works at a bank and takes psychedelics,
that could be a very different effect
of psychedelics on your mind.
I'm just sort of raising the point
that the people you referenced are already weirdos.
I'm not sure exactly.
- No, not necessarily.
A lot of the people that tell me
that they use psychedelics in a useful way
started out as squares and were liberating themselves
because they were stuck.
They were basically stuck in local optimum
of their own self-model, of their relationship to the world.
And suddenly they had data augmentation.
They basically saw and experienced a space of possibilities.
They experienced what it would be like to be another person.
And they took important lessons
from that experience back home.
(inhales deeply)
- Yeah.
I mean, I love the metaphor of data augmentation
because that's been the primary driver
of self-supervised learning in the computer vision domain
is data augmentation.
So it's funny to think of data augment,
like chemically induced data augmentation in the human mind.
- There's also a very interesting effect that I noticed.
I know several people who are severe to me
that LSD has cured their migraines.
§12Transcript · 1:34:10–1:42:46
So severe cluster eight headaches or migraines
that didn't respond to standard medication
that disappeared after a single dose.
And I don't recommend anybody doing this,
especially not in the US where it's illegal.
And there are no studies on this for that reason.
But it seems that anecdotally
that it basically can reset the serotonergic system.
So it's basically pushing them
outside of their normal boundaries.
And as a result, it needs to find a new equilibrium.
And in some people, that equilibrium is better.
But it also follows that in other people, it might be worse.
So if you have a brain that is already teetering
on the boundary to psychosis,
it can be permanently pushed over that boundary.
- Well, that's why you have to do good science,
which they're starting to do
on all these different substances
of how well it actually works for the different conditions
like MDMA seems to help with PTSD, same with psilocybin.
That you need to do good science,
meaning large studies of large N.
- Yeah, so based on the existing studies with MDMA,
it seems that if you look at Rick Doblin's work
and what he has published about this and talks about,
MDMA seems to be a psychologically relatively safe drug,
but it's physiologically not very safe.
That is, there is neurotoxicity
if you would use too large dose.
And if you combine this with alcohol,
which a lot of kids do in party settings during raves
and so on, it's very hepatotoxic.
So basically you can kill your liver.
And this means that it's probably something that is best
and most productively used in a clinical setting
by people who really know what they're doing.
And I suspect that's also true for the other psychedelics.
That is, while the other psychedelics
are probably not as toxic as say alcohol,
the effects on Nisaki can be much more profound and lasting.
- Yeah, well, as far as I know, psilocybin,
so mushrooms, magic mushrooms,
as far as I know in terms of the studies they're running,
I think have no, like they're allowed to do
what they're calling heroic doses.
So that one does not have a toxicity.
So they could do like huge doses in a clinical setting
when they're doing study on psilocybin,
which is kind of fun.
- Yeah, it seems that most of the psychedelics
work in extremely small doses,
which means that the effect on the rest of the body
is relatively low.
And MDMA is probably the exception.
Maybe ketamine can be dangerous in larger doses
because it can depress breathing and so on.
But the LSD and psilocybin work in very, very small doses,
at least the active part of them,
of psilocybin and LSD is only the active part.
And the, but the effect that it can have
on your mental wiring can be very dangerous, I think.
- Let's talk about AI a little bit.
What are your thoughts about GPT-3 and language models
trained with self-supervised learning?
It came out quite a bit ago,
but I wanted to get your thoughts on it.
- Yeah.
In the '90s, I was in New Zealand
and I had an amazing professor, Ian Witton,
who realized I was bored in class and put me in his lab.
And he gave me the task to discover grammatical structure
in an unknown language.
And the unknown language that I picked was English
because it was the easiest one to find,
Corpus 4, construct one.
And he gave me the largest computer at the whole university.
It had two gigabytes of RAM, which was amazing.
And I wrote everything in C
with some in-memory compression to do statistics
over the language.
And I first would create a dictionary of all the words,
which basically tokenizes everything and compresses things
so that I don't need to store the whole word,
but just a code for every word.
And then I was taking this all apart in sentences
and I was trying to find all the relationships
between all the words in the sentences
and do statistics over them.
And that proved to be impossible
because the complexity is just too large.
So if you want to discover the relationship
between an article and a noun,
and there are three adjectives in between,
you cannot do N-gram statistics
and look at all the possibilities that can exist,
at least not with the resources that we had back then.
So I realized I need to make some statistics
over what I need to make statistics over.
So I wrote something that was pretty much a hack
that did this for at least first order relationships.
And I came up with some kind of mutual information graph
that was indeed discovering something
that looks exactly like the grammatical structure
of the sentence, just by trying to encode the sentence
in such a way that the words would be written
in the optimal order inside of the model.
And what I also found is that if we would be able
to increase the resolution of that
and not just use this model
to reproduce grammatically correct sentences,
we would also be able to correct
stylistically correct sentences
by just having more bits in these relationships.
And if we wanted to have meaning,
we would have to go much higher order.
And I didn't know how to make higher order models back then
without spending way more years in research
on how to make the statistics
over what we need to make statistics over.
And this thing that we cannot look at the relationships
between all the bits in your input
is being solved in different domains in different ways.
So in computer graphics, the computer vision,
standard methods for many years now
is convolutional neural networks.
Convolutional neural networks are hierarchies of filters
that exploit the fact that neighboring pixels in images
are usually semantically related
and distance pixels in images
are usually not semantically related.
So you can just by grouping the pixels
that are next to each other hierarchically together,
reconstruct the shape of objects.
And this is an important prior
that we built into these models
so they can converge quickly.
But this doesn't work in language
for the reason that adjacent words are often
but not always related
and distant words are sometimes related
while the words in between are not.
So how can you learn the topology of language?
And I think for this reason that this difficulty existed,
the transformer was invented
in natural language processing, not in vision.
And what the transformer is doing,
it's a hierarchy of layers
where every layer learns what to pay attention to
in the given context in the previous layer.
So what to make the statistics over.
- And the context is significantly larger
than the adjacent word.
- Yes.
So the context that GPT-3 has been using,
the transformer itself is from 2017
and it wasn't using that large of a context.
OpenAI has basically scaled up this idea
as far as they could at the time.
And the context is about 2048 symbols,
tokens in the language.
These symbols are not characters,
but they take the words and project them
into a vector space where words
that are statistically co-occurring a lot
are neighbors already.
So it's already a simplification
of the problem a little bit.
And so every word is basically a set of coordinates
in a high dimensional space.
And then they use some kind of trick
to also encode the order of the words in a sentence
or in the not just sentence,
but 2048 tokens is about a couple pages of text
or two and a half pages of text.
And so they managed to do pretty exhaustive statistics
over the potential relationships
between two pages of text, which is tremendous, right?
I was just using a single sentence back then
and I was only looking for first order relationships
and they were really looking for
much, much higher level relationships.
And what they discover after they fed this
with an enormous amount of training data,
pretty much the written internet
or a subset of it that had some quality,
but substantial portion of the common crawl
that they're not only able to reproduce style,
but they're also able to reproduce
some pretty detailed semantics,
like being able to add three digit numbers
and multiply two digit numbers
or to translate between pro-Eng languages
and things like that.
So the results that GPT-3 got, I think were amazing.
- By the way, I actually didn't check carefully.
It's funny you just mentioned
how you coupled semantics to the multiplication.
Is it able to do some basic math on two digit numbers?
- Yes.
§13Transcript · 1:42:47–1:51:46
- Okay, interesting.
I thought there's a lot of failure cases.
- Yeah, it basically fails if you take larger digit numbers.
So four digit numbers and so on
makes carrying mistakes and so on.
And if you take larger numbers,
you don't get useful results at all.
And this could be an issue of the training set,
where there are not many examples
of successful long form addition
and standard human written text.
- And humans aren't very good
at doing three digit numbers either.
- Yeah, and you're not writing a lot about it.
And the other thing is that the loss function
that is being used is only minimizing surprises.
So it's predicting what comes next in a typical text.
It's not trying to go for causal closure first as we do.
- Yeah.
- But the fact that that kind of prediction works
to generate text that's semantically rich
and consistent is interesting.
- Yeah.
- So yeah, so it's amazing that it's able
to generate semantically consistent text.
- It's not consistent.
So the problem is that it loses coherence at some point.
But it's also, I think, not correct to say
that GPT-3 is unable to deal with semantics at all,
because you ask it to perform certain transformations
in text and it performs these transformation in text.
And the kind of additions that it's able
to perform are transformations in text, right?
And there are proper semantics involved.
You can also do more.
There was a paper that was generating lots and lots
of mathematically correct text
and was feeding this into a transformer.
And as a result, it was able to learn how
to do differentiation integration in race
that according to the authors, Mathematica could not.
To which some of the people in Mathematica responded
that they were not using Mathematica in the right way
and so on.
I have not really followed the resolution of this conflict.
- This part, as a small tangent,
I really don't like in machine learning papers,
which they often do anecdotal evidence.
They'll find like one example in some kind
of specific use of Mathematica and demonstrate,
look, here's, they'll show successes and failures,
but they won't have a very clear representation
of how many cases this actually represents.
- Yes, but I think as a first paper,
this is a pretty good start.
And so the take home message, I think,
is that the authors could get better results from this
and their experiments than they could get from the vein,
which they were using computer algebra systems,
which means that was not nothing.
And it's able to perform substantially better
than GPT-3 can based on a much larger amount
of training data using the same underlying algorithm.
- Well, let me ask again.
So I'm using your tweets as if this is like Plato, right?
(both laughing)
As if this is well thought out novels that you've written.
You tweeted, "GPT-4 is listening to us now."
This is one way of asking,
what are the limitations of GPT-3 when it scales?
So what do you think will be the capabilities of GPT-4,
GPT-5, and so on?
What are the limits of this approach?
- So obviously when we are writing things right now,
everything that we are writing now
is going to be training data
for the next generation of machine learning models.
So yes, of course, GPT-4 is listening to us.
And I think the tweet is already a little bit older
and we now have WUDAO and we have a number of other systems
that basically are placeholders for GPT-4.
Don't know what OpenAI's plans are in this regard.
- I read that tweet in several ways.
So one is obviously everything you put on the internet
is used as training data.
But in a second way, I read it is in a,
we talked about agency.
I read it as almost like GPT-4 is intelligent enough
to be choosing to listen.
So not only did a programmer tell it to collect this data
and use it for training,
I almost saw the humorous angle,
which is like it has achieved AGI kind of thing.
- Well, the thing is,
could we be already be living in GPT-5?
(both laughing)
- So GPT-4 is listening and GPT-5 actually constructing
the entirety of the reality.
- Of course, in some sense,
what everybody is trying to do right now in AI
is to extend the transformer to be able to deal with video.
And there are very promising extensions, right?
There's a book by Google that is called Perceiver,
and that is overcoming some of the limitations
of the transformer by letting it learn the topology
of the different modalities separately,
and by training it to find better input features.
So the basically feature abstractions that are being used
by this successor to GPT-3 are chosen such a way
that it's able to deal with video input.
And there is more to be done.
So one of the limitations of GPT-3 is that it's amnesiac.
So it forgets everything beyond the two pages
that it currently reads, also during generation,
not just during learning.
- Do you think that's fixable
within the space of deep learning?
Can you just make a bigger, bigger, bigger input?
- No, I don't think that our own working memory
is infinitely large.
It's probably also just a few thousand bits.
But what you can do is you can structure
this working memory.
So instead of just force feeding this thing,
a certain thing that it has to focus on,
and it's not allowed to focus on anything else
as its network, you allow it to construct
its own working memory, as we do, right?
When we are reading a book,
it's not that we are focusing our attention
in such a way that we can only remember
the current page.
We will also try to remember other pages
and try to undo what we learned from them
or modify what we learned from them.
We might get up and take another book from the shelf.
We might go out and ask somebody,
and we can edit our working memory in any way
that is useful to put a context together
that allows us to draw the right inferences
and to learn the right things.
So this ability to perform experiments on the world
based on an attempt to become fully coherent
and to achieve causal closure,
to achieve a certain aesthetic of your modeling,
that is something that eventually needs to be done.
And at the moment, we are skirting this in some sense
by building systems that are larger and faster
so they can use dramatically larger resources
and human beings can do much more training data
to get to models that in some sense
are already very superhuman,
and in other ways are laughingly incoherent.
- So do you think sort of making the systems like,
what would you say, multi-resolutional?
So like some of the language models are focused on two pages,
some are focused on two books,
some are focused on two years of reading,
some are focused on a lifetime.
So it's like stacks, it's a GPT-3s all the way down.
- You want to have gaps in between them.
So it's not necessarily two years, there's no gaps.
It's things out of two years or out of 20 years
or 2,000 years or 2 billion years
where you are just selecting those bits
that are predicted to be the most useful ones
to understand what you're currently doing.
And this prediction itself requires a very complicated model
and that's the actual model that you need to be making.
It's not just that you are trying to understand
the relationships between things,
but what you need to make relationships,
discover relationships over.
- I wonder what that thing looks like,
what the architecture for the thing
that's able to have that kind of model.
- I think it needs more degrees of freedom
than the current models have.
So it starts out with the fact that you possibly
don't just want to have a feed-forward model,
but you want it to be fully recurrent.
And to make it fully recurrent,
you probably need to loop it back into itself
and allow it to skip connections.
Once you do this, when you are predicting the next frame
and your internal next frame in every moment,
and you are able to skip connection,
it means that signals can travel from the output
of the network into the middle of the network
faster than the inputs do.
- Do you think it could still be differentiable?
Do you think it still could be a neural network?
- Sometimes it can, and sometimes it cannot.
So it can still be a neural network,
but not a fully differentiable one.
And when you want to deal with non-differentiable ones,
you need to have an attention system
that is discrete and two-dimensional
and can perform grammatical operations.
You need to be able to perform program synthesis.
§14Transcript · 1:51:49–2:00:49
You need to be able to backtrack in these operations
that you perform on this thing.
And this thing needs a model of what it's currently doing.
And I think this is exactly the purpose
of our own consciousness.
- Yeah, the program things
that trick you on your own networks.
So let me ask you, it's not quite program synthesis,
but the application of these language models
to generation, to program synthesis,
but generation of programs.
So if you look at GitHub OpenPilot,
which is based on OpenAI's codecs,
I don't know if you got a chance to look at it,
but it's the system that's able to generate code
once you prompt it with, what is it?
Like the header of a function with some comments.
It seems to do an incredibly good job,
or not a perfect job, which is very important,
but an incredibly good job of generating functions.
What do you make of that?
Are you, is this exciting,
or is this just a party trick, a demo?
Or is this revolutionary?
- I haven't worked with this yet,
so it's difficult for me to judge it,
but I would not be surprised
if it turns out to be revolutionary.
And that's because the majority of programming tasks
that are being done in the industry right now
are not creative.
- Yeah.
- People are writing code that other people have written,
or they're putting things together
from code fragments that others have had.
And a lot of the work that programmers do in practice
is to figure out how to overcome the gaps
in their current knowledge,
in the things that people have already done.
- How to copy and paste from Stack Overflow, that's right.
- Yes, and so of course we can automate that.
- Yeah, to make it much faster
to copy and paste from Stack Overflow.
- Yes, but it's not just copying and pasting,
it's also basically learning which parts you need to modify
to make them fit together.
- Yeah, like literally sometimes as simple
as just changing the variable names
so it fits into the rest of your code.
- Yes, but this requires that you understand
the semantics of what you're doing to some degree.
- Yeah, and you can automate some of those things.
The thing that makes people nervous, of course,
is that a little bit wrong in a program
can have a dramatic effect
on the actual final operation of that program.
So that's one little error,
which in the space of language doesn't really matter,
but in the space of programs can matter a lot.
- Yes, but this is already what is happening
when humans program code.
- Yeah, this is--
- So we have a technology to deal with this.
- Somehow it becomes scarier
when you know that a program generated code
that's running a nuclear power plant.
It becomes scarier.
You know humans have errors too.
- Exactly.
- But it's scarier when a program is doing it because,
why, why?
- I mean, there's a fear that a program,
like a program may not be as good as humans
to know when stuff is important to not mess up.
Like there's a misalignment of priorities,
of values, that's potential.
Maybe that's the source of the worry.
I mean, okay, if I give you code generated by
GitHub OpenPilot and code generated by a human
and say, here, use one of these,
which, how do you select today and in the next 10 years,
which code to use?
Wouldn't you still be comfortable with the human?
- At the moment, when you go to Stanford to get an MRI,
they will write a bill to the insurance over $20,000.
And of this, maybe half of that gets paid by the insurance
and a quarter gets paid by you.
And the MRI cost them $600 to make, maybe, probably less.
And what are the values of the person
that writes the software and deploys this process?
It's very difficult for me to say whether I trust people.
I think that what happens there is a mixture
of proper Anglo-Saxon Protestant values
where somebody is trying to serve an abstract greater whole
and organized crime.
- Well, that's a very harsh,
I think that's a harsh view of humanity.
There's a lot of bad people, whether incompetent
or just malevolent in this world, yes.
But it feels like the more malevolent,
so the more damage you do to the world,
the more resistance you have in your own human heart.
- But don't explain with malevolence or stupidity
what can be explained by just people
acting on their incentives.
So what happens in Stanford is not that somebody is evil.
It's just that they do what they're being paid for.
- No, it's not evil.
I tend to, no, I see that as malevolence.
I see as I, even like being a good German,
as I told you offline, is some,
it's not absolute malevolence, but it's a small amount.
It's cowardice.
I mean, when you see there's something wrong with the world,
it's either incompetence and you're not able to see it,
or it's cowardice that you're not able to stand up,
not necessarily in a big way, but in a small way.
So I do think that is a bit of malevolence.
I'm not sure the example you're describing
is a good example of that. - So the question is,
what is it that you are aiming for?
And if you don't believe in the future,
if you, for instance, think that the dollar
is going to crash, why would you try to save dollars?
If you don't think that humanity will be around
in 100 years from now because global warming
will wipe out civilization,
why would you need to act as if it were?
So the question is, is there an overarching aesthetics
that is projecting you and the world into the future,
which I think is the basic idea of religion,
that you understand the interactions
that we have with each other
as some kind of civilization level agent
that is projecting itself into the future.
If you don't have that shared purpose,
what is there to be ethical for?
So I think when we talk about ethics and AI,
we need to go beyond the insane bias discussions and so on,
where people are just measuring the distance
between a statistic to their preferred current world model.
- The optimism, wait, wait, wait,
I was a little confused by the previous thing,
just to clarify.
There is a kind of underlying morality
to having an optimism that human civilization
will persist for longer than 100 years.
I think a lot of people believe
that it's a good thing for us to keep living.
- Yeah, of course. - And thriving.
- This morality itself is not an end to itself.
It's instrumental to people living in 100 years from now.
- Right. - Or 500 years from now.
So it's only justifiable if you actually think
that it will lead to people
or increase the probability of people being around
in that timeframe.
And a lot of people don't actually believe that,
at least not actively.
- But believe what exactly?
So I was-- - Most people don't believe
that they can afford to act on such a model.
Basically what happens in the US
is I think that the healthcare system
is for a lot of people no longer sustainable,
which means that if they need the help
of the healthcare system,
they're often not able to afford it.
And when they cannot help it,
they are often going bankrupt.
I think the leading cause of personal bankruptcy
in the US is the healthcare system.
- Yeah. - And that would not
be necessary.
It's not because people are consuming
more and more medical services
and are achieving a much, much longer life as a result.
That's not actually the story that is happening
because you can compare it to other countries.
And life expectancy in the US is currently not increasing
and it's not as high as in all the other
industrialized countries.
So some industrialized countries are doing better
with a much cheaper healthcare system.
And what you can see is, for instance, administrative load.
The healthcare system has maybe to some degree deliberately
set up its job placement program to allow people
to continue living a middle-class existence,
despite not having a useful use case in productivity.
So they are being paid to push paper around.
And the number of administrators in the healthcare system
has been increasing much faster
than the number of practitioners.
And this is something that you have to pay for, right?
And also the revenues that are being generated
in the healthcare system are relatively large
and somebody has to pay for them.
And the result why they are so large
is because market mechanisms are not working.
The FDA is largely not protecting people
§15Transcript · 2:00:52–2:09:19
from malpractice of healthcare providers.
The FDA is protecting healthcare providers
from competition.
- Right, okay, okay.
- So this is a thing that has to do with values.
And this is not because people are malicious on all levels.
It's because they are not incentivized to act
on a greater whole, on this idea that you treat somebody
who comes to you as a patient
like you would treat a family member.
- Yeah, yeah, but we're trying, I mean,
you're highlighting a lot of the flaws
of the different institutions
the systems we're operating under.
But I think there's a continued throughout history
mechanism design of trying to design incentives
in such a way that these systems behave better
and better and better.
I mean, it's a very difficult thing to operate
a society of hundreds of millions of people
effectively with--
- Yes, so do we live in a society
that is ever correcting?
Is this, do we observe that our models
of what we are doing are predictive of the future
and when they are not, we improve them.
Our laws adjudicated with clauses
that you put into every law,
what is meant to be achieved by that law
and the law will be automatically repealed
if it's not achieving that, right?
If you are optimizing your own laws,
if you're writing your own source code,
you probably make an estimate of what is the thing
that's currently wrong in my life?
What is it that I should change about my own policies?
What is the expected outcome?
And if that outcome doesn't manifest,
I will change the policy back, right?
Or I would change it into something different.
Are we doing this on a societal level?
- I think so.
I think it's easy to sort of highlight the,
I think we're doing it in the way that,
like I operate my current life.
I didn't sleep much last night.
You would say that, Lex, the way you need
to operate your life is you need to always get sleep.
The fact that you didn't sleep last night
is totally the wrong way to operate in your life.
Like you should have gotten all your shit done in time
and gotten to sleep because sleep is very important
for health and you're highlighting,
look, this person is not sleeping.
Look, the medical, the healthcare system is operating poor.
But the point is that we just,
it seems like this is the way,
especially in the capitalist society we operate,
we keep running into trouble and last minute,
we try to get our way out through innovation
and it seems to work.
You have a lot of people that ultimately are trying
to build a better world and get urgency about them
when the problem becomes more and more imminent.
And that's the way this operates.
But if you look at the history, the long arc of history,
it seems like that operating on deadlines produces progress
and builds better and better systems.
- You probably agree with me that the US
should have engaged in mask production in January, 2020.
And that we should have shut down the airports early on
and that we should have made it mandatory
that the people that work in nursing homes
on campus, rather than living at home
and then coming in and infecting people in the nursing homes
that had no immune response to COVID.
And that is something that was, I think, visible back then.
The correct decisions haven't been made.
We would have the same situation again.
How do we know that these wrong decisions
are not being made again?
Have the people that made the decisions
to not protect the nursing homes been punished?
Have the people that made the wrong decisions
with respect to testing that prevented the development
of testing by startup companies and the importing of tests
from countries that already had them,
have these people been held responsible?
- First of all, so what do you want to put
before the firing squad?
I think they are being held responsible.
- No, just make sure that this doesn't happen again.
- No, but it's not that.
Yes, they're being held responsible by many voices,
by people being frustrated.
There's new leaders being born now
that are going to see rise to the top in 10 years.
This moves slower than, there's obviously a lot
of older incompetence and bureaucracy
and these systems move slowly.
They move like science, one death at a time.
So yes, I think the pain that's been felt
in the previous year is reverberating throughout the world.
- Maybe I'm getting old.
I suspect that every generation in the US
after the war has lost the plot even more.
I don't see this development.
- The war, World War II?
- Yes, so basically there was a time
when we were modernist.
And in this modernist time, the US felt actively threatened
by the things that happened in the world.
The US was worried about possibility of failure.
And this imminence of possible failure led to decisions.
There was a time when the government would listen
to physicists about how to do things.
And the physicists were actually concerned
about what the government should be doing.
So they would be writing letters to the government.
And so for instance, the decision
for the Manhattan Project was something
that was driven in a conversation
between physicists and the government.
I don't think such a discussion would take place today.
- I disagree.
I think if the virus was much deadlier,
we would see a very different response.
I think the virus was not sufficiently deadly.
And instead, because it wasn't very deadly,
what happened is the current system
started to politicize it.
The mask, this is what I realized with masks early on.
They were not very quickly became not as a solution,
but they became a thing that politicians used
to divide the country.
So the same things happened with vaccines, same thing.
So like nobody's really,
people weren't talking about solutions to this problem
because I don't think the problem was bad enough.
When you talk about the war,
I think our lives are too comfortable.
I think in the developed world, things are too good
and we have not faced severe dangers.
When the danger, the severe dangers,
existential threats are faced, that's when we step up
on a small scale and a large scale.
Now, I don't, that's sort of my argument here,
but I did think the virus,
I was hoping that it was actually sufficiently dangerous
for us to step up because especially in the early days,
it was unclear.
It still is unclear because of mutations,
how bad it might be, right?
And so I thought we would step up and even,
so the masks point is a tricky one because to me,
the manufacture of masks isn't even the problem.
I'm still to this day and I was involved
with a bunch of this work,
have not seen good science done on whether masks work or not.
Like there still has not been a large scale study.
To me, that should be, there should be large scale studies
on every possible solution, like aggressive,
in the same way that the vaccine development was aggressive.
There should be masks, which tests,
what kind of tests work really well,
what kind of, like even the question
of how the virus spreads,
there should be aggressive studies on that to understand.
I'm still, as far as I know,
there's still a lot of uncertainty about that.
Nobody wants to see this as an engineering problem
that needs to be solved.
It's that I was surprised about,
but I would-- - I find that our views
are largely convergent, but not completely.
So I agree with the thing that,
because our society in some sense perceives itself
as too big to fail.
- Right. - And the virus
did not alert people to the fact
that we are facing possible failure.
That basically put us into the postmodernist mode.
And I don't mean in a philosophical sense,
but in a societal sense,
the difference between the postmodern society
and the modern society is that the modernist society
has to deal with the ground truths,
and the postmodernist society has to deal with appearances.
Politics becomes a performance,
and the performance is done for an audience,
and the organized audience is the media,
and the media evaluates itself via other media, right?
So you have an audience of critics
that evaluate themselves.
And I don't think it's so much the failure
of the politicians, because to get in power
and to stay in power, you need to be able
to deal with the published opinion.
- Well, I think it goes in cycles,
because what's going to happen
§16Transcript · 2:09:21–2:18:59
is all of the small business owners,
all the people who truly are suffering
and will suffer more because the effects
of the closure of the economy
and the lack of solutions to the virus,
they're going to apprise.
And hopefully, I mean, this is where charismatic leaders
can get the world in trouble.
But hopefully, we'll elect great leaders
that will break through this postmodernist idea
of the media and the perception
and the drama on Twitter and all that kind of stuff.
- But you know this can go either way.
- Yeah.
- When the Weimar Republic was unable
to deal with the economic crisis that Germany was facing,
there was an option to go back.
And there were people which thought,
let's get back to a constitutional monarchy
and let's get this to work,
because democracy doesn't work.
And eventually, there was no way back.
People decided there was no way back.
They needed to go forward.
And the only options for going forward
was to become a Stalinist communist,
basically an option to completely expropriate
the factories and so on and nationalize them
and to reorganize Germany in communist terms
and ally itself with Stalin and fascism.
And both options were obviously very bad.
And the one that the Germans picked led to a catastrophe
that devastated Europe.
And I'm not sure if the US has an immune response
against that.
I think that the far right is currently very weak in the US,
but this can easily change.
- Do you think from a historical perspective,
Hitler could have been stopped from within Germany
or from outside or this?
Well, depends on who you want to focus,
whether you want to focus on Stalin or Hitler,
but it feels like Hitler was the one
as a political movement that could have been stopped.
- I think that the point was that
a lot of people wanted Hitler.
So he got support from a lot of quarters.
It was a number of industrialists who supported him
because they thought that the democracy
is obviously not working and unstable
and you need a strong man.
And he was willing to play that part.
There were also people in the US
who thought that Hitler would stop Stalin
and would act as a bulwark against Bolshevism,
which he probably would have done, right?
But at which cost?
And then many of the things that he was going to do,
like the Holocaust,
was something where people thought this is rhetoric.
He's not actually going to do this.
Especially many of the Jews themselves,
which were humanists.
And for them, this was outside of the scope
that was thinkable.
- Right.
I mean, I wonder if Hitler is uniquely,
I want to carefully use this term, but uniquely evil.
So if Hitler was never born,
if somebody else would come in his place.
So like, just thinking about the progress of history,
how important are those singular figures
that lead to mass destruction and cruelty?
Because my sense is Hitler was unique.
It wasn't just about the environment
and the context that gave him.
Another person would not come in his place
to do as destructive of the things that he did.
There was a combination of charisma,
of madness, of psychopathy, of just ego,
all of those things,
which are very unlikely to come together
in one person in the right time.
- It also depends on the context of the country
that you're operating in.
If you tell the Germans that they have a historical destiny
in this romantic country,
the effect is probably different
than it is in other countries.
But Stalin has killed a few more people than Hitler did.
And if you look at the probability
that you survived under Stalin,
Hitler killed people if he thought
they were not worth living,
or if they were harmful to his racist project.
He basically felt that the Jews would be too cosmopolitan
and would not be willing to participate
in the racist redefinition of society
and the value of society and an ethnostate in this way,
as he wanted it to have it.
So he saw them as a harmful danger,
especially since they played such an important role
in the economy and culture of Germany.
And so he had basically had some radical,
but rational reason to murder them.
And Stalin just killed everyone.
He basically, the Stalinist purges were such a random thing
where he said that there's a certain possibility
that this particular part of the population
has a number of German collaborators or something,
and we just kill them all.
Or if you look at what Mao did,
the number of people that were killed in absolute numbers
were much higher under Mao than they were under Stalin.
So it's super hard to say.
The other thing is that you look at Genghis Khan and so on,
how many people he killed.
When you see there are a number of things
that happen in human history
that actually really put a substantial dent
in the existing population or Napoleon.
And it's very difficult to eventually measure it
because what's happening is basically evolution
on a human scale,
where one monkey figures out a way to become viral
and is using this viral technology
to change the patterns of society
at the very, very large scale.
And what we find so abhorrent about these changes
is the complexity that is being destroyed by this.
That it's basically like a big fire
that burns out a lot of the existing culture and structure
that existed before.
- Yeah, and it all just starts with one monkey,
one charismatic ape,
and there's a bunch of them throughout history.
- Yeah, but it's in a given environment.
It's basically similar to wildfires in California, right?
The temperature is rising, there is less rain falling,
and then suddenly a single spark can have an effect
that in other times would be contained.
- Okay, speaking of which,
I love how we went to Hitler and Stalin
from 20, 30 minutes ago,
GPT-3 generating, doing program synthesis.
The argument was about morality of AI versus human.
And specifically in the context of writing programs,
specifically in the context of programs
that can be destructive.
So running nuclear power plants
or autonomous weapons systems, for example.
And I think your inclination was to say
that it's not so obvious
that AI would be less moral than humans,
or less effective at making a world
that would make humans happy.
- So I'm not talking about self-directed systems
that are making their own goals at a global scale.
If you just talk about the deployment
of technological systems that are able
to see order and patterns and use this as control models
to act on the goals that we give them,
then if you have the correct incentives
to set the correct incentives for these systems,
I'm quite optimistic.
- But so humans versus AI, let me give you an example.
Autonomous weapons systems.
Let's say there's a city somewhere in the Middle East
that has a number of terrorists.
And the question is, what's currently done
with drone technology is you have information
about the location of a particular terrorist,
and you have a targeted attack,
you have a bombing of that particular building.
And that's all directed by humans
at the high level strategy,
and also at the deployment of individual bombs
and missiles like that the actual,
everything is done by human except the final targeting.
And the like the, it's like with spot similar thing,
like control the flight.
Okay, what if you give AI control and saying,
write a program that says,
here's the best information I have available
about the location of these five terrorists.
Here's the city, make sure it's,
all the bombing you do is constrained to the city.
Make sure it's precision based, but you take care of it.
So you do one level of abstraction out and saying,
take care of the terrorists in the city.
Which are you more comfortable with,
the humans or the JavaScript GPT-3 generated code
that's doing the deployment?
I mean, that's, this is the kind of question I'm asking,
is the kind of bugs that we see in human nature,
are they better or worse than the kind of bugs we see in AI?
- They're different bugs.
There is an issue that if people are creating
an imperfect automation of a process
that normally requires a mobile judgment.
§17Transcript · 2:19:02–2:28:18
And this mobile judgment is the reason why
it cannot be automated often.
It's not because the computation is too expensive,
but because the model that you give the AI
is not an adequate model of the dynamics
of the world because the AI does not understand
the context that it's operating in the right way.
And this is something that already happens with Excel.
Right, you don't need to have an AI system to do this.
If you have an automated process in place
where humans decide using automated criteria,
whom to kill when and whom to target when,
which already happens, right?
And you have no way to get off the kill list
once that happens.
Once you have been targeted according
to some automatic criterion by people, right?
In a bureaucracy.
That is the issue.
The issue is not the AI, it's the automation.
- So there's something about, right, it's automation.
But there's something about the,
there's a certain level of abstraction
where you give control to AI to do the automation.
There's a scale that could be achieved
that it feels like the scale of bug and scale mistake
and scale of destruction that could be achieved
of the kind that humans cannot achieve.
So AI is much more able to destroy
an entire country accidentally versus humans.
It feels like the more civilians die as a react
or suffer as the consequences of your decisions,
the more weight there is on the human mind
to make that decision.
And so like, it becomes more and more unlikely
to make that decision for humans.
For AI, it feels like it's harder
to encode that kind of weight.
- In a way, the AI that we're currently building
is automating statistics, right?
Intelligence is the ability to make models
so you can act on them.
And AI is the tool to make better models.
So in principle, if you're using AI wisely,
you're able to prevent more harm.
And I think that the main issue is not on the side
of the AI, it's on the side of the human command hierarchy
that is using technology irresponsibly.
- So the question is, how hard is it to encode,
to properly encode the right incentives into the AI?
- So for instance, there's this idea,
what happens if we let our airplanes being flown
with AI systems and then neural network is a black box
and so on.
And it turns out our neural networks
are actually not black boxes anymore.
There are functional approximators using linear algebra
and there are performing things that we can understand.
But we can also, instead of letting the neural network fly
the airplane, use the neural network to generate approval
of the correct program.
There's a degree of accuracy of the proof
that a human could not achieve.
And so we can use our AI by combining different technologies
to build systems that are much more reliable
than the systems that a human being could create.
And so in this sense, I would say that if you use
an early stage of technology to save labor
and don't employ competent people,
but just to hack something together, because you can,
that is very dangerous.
And if people are acting under these incentives
that they get away with delivering shoddy work
more cheaply using AI, there's less human oversight
than before, that's very dangerous.
- The thing is though, AI is still going to be unreliable,
perhaps less so than humans,
but it will be unreliable in novel ways.
And-
- Yeah, but this is an empirical question
and it's something that we can figure out and work with.
So the issue is, do we trust the systems,
the social systems that we have in place
and the social systems that we can build and maintain
that they're able to use AI responsibly?
If they can, then AI is good news.
If they cannot, then it's going to make
the existing problems worse.
- Well, and also who creates the AI, who controls it,
who makes money from it, because it's ultimately humans.
And then you start talking about
how much you trust the humans.
- So the question is, what does who mean?
I don't think that we have identity per se.
I think that the story of a human being is somewhat random.
What happens is more or less that everybody
is acting on their local incentives,
what they perceive to be their incentives.
And the question is, what are the incentives
that the one that is pressing the button
is operating under?
- Yeah.
It's nice for those incentives to be transparent.
So for example, I'll give you an example.
There seems to be a significant distrust of tech,
like entrepreneurs in the tech space,
or people that run, for example, social media companies,
like Mark Zuckerberg.
There's not a complete transparency of incentives
under which that particular human being operates.
We can listen to the words he says,
or what the marketing team says for a company,
but we don't know.
And that becomes a problem when the algorithms
and the systems created by him and other people
in that company start having more and more impact on society.
If the incentives were somehow,
the definition and the explainability of the incentives
was decentralized such that nobody can manipulate it,
no propaganda type manipulation
of how these systems actually operate could be done,
then yes, I think AI could achieve much fairer,
much more effective solutions
to difficult ethical problems.
But when there's humans in the loop
manipulating the dissemination,
the communication of how the system actually works,
that feels like you can run into a lot of trouble.
And that's why there's currently a lot of distrust
for people at the heads of companies
that have increasingly powerful AI systems.
- I suspect what happened traditionally in the US
was that since our decision-making is much more central,
or decentralized than in an authoritarian state, right?
People are making decisions autonomously
at many, many levels in a society.
What happened that was we created coherence
and cohesion in society by controlling what people thought
and what information they had.
The media synchronized public opinion
and social media have disrupted this.
It's not, I think, so much Russian influence or something.
It's everybody's influence.
It's that a random person can come up
with a conspiracy theory and disrupt what people think.
And if that conspiracy theory is more compelling
or more attractive than the standardized
public conspiracy theory that we give people as a default,
then it might get more traction, right?
You suddenly have the situation that a single individual
somewhere on a farm in Texas has more listeners than CNN.
- Which particular farmer are you referring to in Texas?
(both laughing)
- Probably no one.
- Yes, I had dinner with him a couple of times.
- Okay. - Right.
It's an interesting situation because you cannot get
to be an anchor in CNN if you don't go
through a complicated gatekeeping process.
And suddenly you have random people
without that gatekeeping process
just optimizing for attention.
Not necessarily with a lot of responsibility
for the long-term effects of projecting these theories
into the public.
And now there is a push of making social media
more like traditional media, which means that the opinion
that is being projected in social media
is more limited to an acceptable range.
With the goal of getting society into safe waters
and increase the stability and cohesion of society again,
which I think is a laudable goal.
But of course it also is an opportunity
to seize the means of indoctrination.
And the incentives that people are under when they do this
are in such a way that the AI ethics that we would need
becomes very often something like AI politics,
which is basically partisan and ideological.
And this means that whatever one side says,
another side is going to be disagreeing with.
In the same way as when you turn masks
or the vaccine into a political issue,
if you say that it is politically virtuous
to get vaccinated, it will mean that the people
that don't like you will not want to get vaccinated.
And as soon as you have this partisan discourse,
it's going to be very hard to make the right decisions
because the incentives get to be the wrong ones.
AI ethics needs to be super boring.
It needs to be done by people who do statistics all the time
and have extremely boring, long-winded discussions
that most people cannot follow
because they are too complicated, but that are dead serious.
These people need to be able to be better at statistics
than the leading machine learning researchers.
And at the moment, the AI ethics debate is the one
where you don't have any barrier to entry.
Everybody who has a strong opinion
and is able to signal that opinion in the right way-
- Strong words from Joshua Bach.
§18Transcript · 2:28:20–2:37:31
(Laura laughs)
- And to me, that's a very frustrating thing
because the field is so crucially important to our future.
- It's so crucially important,
but the only qualification you currently need
is to be outraged by the injustice in the world.
- It's more complicated, right?
Everybody seems to be outraged.
But let's just say that the incentives
are not always the right ones.
So basically, I suspect that a lot of people
that enter this debate don't have a vision
for what society should be looking like
in a way that is non-violent,
where we preserve liberal democracy,
where we make sure that we all get along
and we are around in a few hundred years from now,
preferably with a comfortable technological civilization
around us.
- I generally have a very foggy view of that world,
but I tend to try to follow,
and I think society should in some degree
follow the gradient of love,
increasing the amount of love in the world.
And whenever I see different policies or algorithms
or ideas that are not doing so,
obviously that's the ones that kind of resist.
- So the thing that terrifies me about this notion
is I think that German fascism was driven by love.
It was just a very selective love.
It was a love that basically-
- But now you're just manipulating.
I mean, that's, you have to be very careful.
You're talking to the wrong person in this way about love.
- So let's talk about what love is.
And I think that love is the discovery of shared purpose.
It's the recognition of the sacred and the other.
And this enables non-transactional interactions.
- But the size of the other that you include
needs to be maximized.
So it's basically appreciation,
like deep appreciation of the world around you fully,
like including the people that are very different than you,
the people that disagree with you completely,
including people, including living creatures
outside of just people, including ideas.
And it's like appreciation of the full mess of it.
And also it has to do with like empathy,
which is coupled with a lack of confidence,
uncertainty about, of your own rightness.
It's like an open, a radical open-mindedness
to the way forward.
- I agree with every part of what you said.
And now if you scale it up, what you recognize
is that love is in some sense the service
to a next level agency, to the highest level agency
that you can recognize.
It could be, for instance, life on earth
or beyond that, where you could say intelligent complexity
in the universe that you try to maximize in a certain way.
But when you think it's true,
it basically means a certain aesthetic.
And there is not one possible aesthetic.
There are many possible aesthetics.
And once you project an aesthetic into the future,
you can see that there are some which defect from it,
which are in conflict with it,
that are corrupt, that are evil.
You and me would probably agree that Hitler was evil
because the aesthetic of the world that he wanted
is in conflict with the aesthetic of the world
that you and me have in mind.
And so the thing that he destroyed,
we want to keep them in the world.
- There's kind of ways to deal.
I mean, Hitler is an easier case,
but perhaps he wasn't so easy in the '30s
to understand who is Hitler and who is not.
- No, it was just that there was no consensus
that the aesthetics that he had in mind were unacceptable.
- Yeah.
I mean, it's difficult.
Love is complicated because you can't just be so open-minded
that you let evil walk into the door,
but you can't be so self-assured
that you can always identify evil perfectly
because that's what leads to Nazi Germany,
having a certainty of what is and wasn't evil,
like always drawing lines of good versus evil.
There seems to be,
there has to be a dance between
like hard stances extending up against what is wrong,
and at the same time, empathy and open-mindedness
of towards not knowing what is right and wrong,
and like a dance between those.
- I found that when I watched the Miyazaki movies
that there is nobody who captures my spirituality
as well as he does.
It's very interesting and just vicious.
There is something going on in his movies
that is very interesting.
So for instance, Mononoke is discussing
not only an answer to Disney's simplistic notion of Mowgli,
the jungle boy who was raised by wolves,
and as soon as he sees people, realizes that he's one of them
and the way in which the moral life and nature
is simplified and romanticized and turned into kitsch.
It's disgusting in the Disney movie,
and he answers to this.
You see, he's replaced by Mononoke,
this wolf girl who was raised by wolves
and who was fierce and dangerous
and who cannot be socialized because she cannot be tamed,
cannot be part of human society.
And you see, human society,
it's something that is very, very complicated.
You see people extracting resources and destroying nature,
but the purpose is not to be evil,
but to be able to have a life that is free from,
for instance, oppression and violence
and to curb death and disease.
And you basically see this conflict,
which cannot be resolved in a certain way.
You see this moment when nature is turned into a garden
and it loses most of what it actually is
and humans no longer submitting to life
and death and nature.
And to these questions, there is no easy answer.
So he just turns it into something that is being observed
as a journey that happens.
And that happens with a certain degree of inevitability.
And the nice thing about all his movies
is there's a certain main character,
and it's the same in all movies.
It's this little girl that is basically Heidi.
And I suspect that happened because when he did field work
for working on the Heidi movies,
back then the Heidi animations,
before he did his own movies,
he traveled to Switzerland and Southeastern Europe
and the Adriatic and so on,
and got an idea about a certain aesthetic
and a certain way of life that informed his future thinking.
And Heidi has a very interesting relationship
to herself and to the world.
There's nothing that she takes for herself.
She is in a way fearless because she is committed
to a service, to a greater whole.
Basically, she is completely committed to serving God.
And it's not an institutionalized God.
It has nothing to do with the Roman Catholic Church
or something like this.
But in some sense, Heidi is an embodiment
of the spirit of European Protestantism.
It's this idea of a being
that is completely perfect and pure.
And it's not a feminist vision
because she is not a girl boss or something like this.
She is the justification for the men in the audience
to protect her, to build a civilization around her
that makes her possible.
So she is not just the sacrifice of Jesus,
who is innocent and therefore nailed to the cross.
She is not being sacrificed.
She is being protected by everybody around her
who recognizes that she is sacred.
And there are enough around her to see that.
So that's a very interesting perspective.
There's a certain notion of innocence.
And this notion of innocence is not universal.
It's not in all cultures.
Hitler wasn't innocent.
His idea of Germany was not that there is an innocence
that is being protected.
There was a predator that was going to triumph.
And it's also something that is not
at the core of every religion.
There are many religions which don't care about innocence.
They might care about increasing the status of something.
And that's a very interesting notion that is quite unique
and not claiming it's the optimal one.
It's just a particular kind of aesthetic,
which I think makes Miyazaki
into the most relevant Protestant philosopher today.
- And you're saying in terms of all the ways
that a society can operate,
perhaps the preservation of innocence
might be one of the best.
- No, it's just my aesthetic.
- Your aesthetic, gotcha.
- It's a particular way in which I feel
that I relate to the world
that is natural to my own specialization.
And maybe it's not an accident
that I have cultural roots in Europe,
in a particular world.
And so maybe it's a natural convergence point
and it's not something that you will find
in all other times in history.
- So I'd like to ask you about Solzhenitsyn
§19Transcript · 2:37:34–2:47:30
and our individual role as ants in this very large society.
So he says that some version of the line
between good and evil runs to the heart of every man.
Do you think all of us are capable of good and evil?
What's our role in this play,
in this game we're all playing?
Is all of us capable to play any role?
Is there an ultimate responsibility to,
you mentioned maintaining innocence
or whatever the highest ideal for a society you want,
are all of us capable of living up to that?
And that's our responsibility.
Or is there significant limitations
to what we're able to do in terms of good and evil?
- So there is a certain way, if you are not terrible,
if you are committed to some kind of civilizational agency,
the next level agent that you are serving,
some kind of transcendent principle.
In the eyes of that transcendental principle,
you are able to discern good from evil.
Otherwise you cannot,
otherwise you have just individual aesthetics.
The cat that is torturing a mouse
is not evil because the cat does not envision
or no part of the world,
of the cat is envisioning a world
where there is no violence and nobody is suffering.
If you have an aesthetic
where you want to protect innocence,
then torturing somebody needlessly is evil,
but only then.
- No, but within, I guess the question is within the aesthetic,
within your sense of what is good and evil,
are we still,
it seems like we're still able to commit evil.
- Yes, so basically if you are committing
to this next level agent,
you are not necessarily are this next level agent, right?
You are a part of it.
You have a relationship to it,
like a cell does to its organism, its hyperorganism.
And it only exists to the degree
that it's being implemented by you and others.
And that means that you're not completely fully serving it.
You have freedom in what you decide,
whether you are acting on your impulses and local incentives
and your feral impulses, so to speak,
or whether you're committing to it.
And what you perceive then is a tension
between what you would be doing
with respect to the thing that you recognize
as the sacred if you do,
and what you're actually doing.
And this is the line between good and evil, right?
Where you see, oh, I'm here acting
on my local incentives or impulses.
And here I'm acting on what I consider to be sacred.
And there's a tension between those.
And this is the line between good and evil
that might run through your heart.
And if you don't have that,
if you don't have this relationship
to a transcendental agent,
you could call this relationship
to the next level agent soul, right?
It's not a thing.
It's not an immortal thing that is intrinsically valuable.
It's a certain kind of relationship
that you project to understand what's happening.
Somebody is serving this transcendental sacredness
or they're not.
If you don't have this soul, you cannot be evil.
You're just a complex, natural phenomenon.
- So if you look at life, like starting today
or starting tomorrow, when we leave here today,
there's a bunch of trajectories
that you can take through life, maybe countless.
Do you think some of these trajectories
in your own conception of yourself,
some of those trajectories are the ideal life?
A life that if you were to be the hero of your life story,
you would want to be?
Like, is there some Josh or Bach that you're striving to be?
Like, this is the question I ask myself
as an individual trying to make a better world
in the best way that I could conceive of.
What is my responsibility there?
And how much am I responsible for the failure to do so?
'Cause I'm lazy and incompetent too often
in my own perception.
- In my own world view, I'm not very important.
So I don't have place for me as a hero in my own world.
I'm trying to do the best that I can,
which is often not very good.
And so it's not important for me to have status
or to be seen in a particular way.
It's helpful if others can see me,
if a few people can see me, that can be my friends.
- No, sorry, I want to clarify.
The hero, I didn't mean status or perception
or like some kind of marketing thing,
but more in private, in the quiet of your own mind.
Is there the kind of man you want to be
and would consider it a failure if you don't become that?
That's what I meant by hero.
- Yeah, not really.
I don't perceive myself as having such an identity.
And it's also sometimes frustrating,
but it's basically a lack of having this notion
of father that I need to be emulating.
- It's interesting.
I mean, it's the leaf floating down the river.
I worry that...
- Sometimes it's more like being the river.
- I'm just a fat frog sitting in a leaf
on a dirty, muddy lake.
I wish I was-- - Full of love.
- Waiting for a princess to kiss me.
Or the other way, I forgot which way it goes.
Somebody kisses somebody.
Can I ask you, I don't know if you know who Michael Malice is
but in terms of constructing systems of incentives,
it's interesting to ask.
I don't think I've talked to you about this before.
Malice espouses anarchism.
So he sees all government as fundamentally
getting in the way or even being destructive
to collaborations between human beings thriving.
What do you think?
What's the role of government in a society that thrives?
Is anarchism at all compelling to you as a system?
So like not just small government, but no government at all.
- Yeah, I don't see how this would work.
The government is an agent that imposes an offset
on your reward function, on your payout metrics.
So your behavior becomes compatible with the common good.
- So the argument there is that you can have collectives
like governing organizations, but not government.
Like where you're born on a particular set of land
and therefore you must follow this rule or else.
You're forced by what they call violence
because there's an implied violence here.
So with government, the key aspect of government
is it protects you from the rest of the world with an army
and with police, right?
So it has a monopoly on violence.
It's the only one that's able to do violence.
- So there are many forms of government,
not all governments do that, right?
But we find that in successful countries,
the government has a monopoly on violence.
And that means that you cannot get ahead
by starting your own army because the government
will come down on you and destroy you if you try to do that.
And in countries where you can build your own army
and get away with it, some people will do it, right?
In these countries is what we call failed countries
in a way.
And if you don't want to have violence,
the point is not to appeal to the moral intentions of people
because some people will use strategies
if they get ahead with them that feel a particular kind
of ecological niche.
So you need to destroy that ecological niche.
And if effective government has a monopoly on violence,
it can create a world where nobody is able to use violence
and get ahead, right?
So you want to use that monopoly on violence,
not to exert violence, but to make violence impossible,
to raise the cost of violence.
So people need to get ahead with nonviolent means.
- So the idea is that you might be able to achieve that
in an anarchist state with companies.
So with the forces of capitalism is create security companies
where the one that's most ethically sound
rises to the top, basically,
it would be a much better representative of the people
because there is a less sort of stickiness
to the big military force sticking around,
even though it's long overlived, outlived.
- So you have groups of militants
that are hopefully efficiently organized
because otherwise they're going to lose
against the other groups of militants.
And they are coordinating themselves
with the rest of society
until they are having a monopoly on violence.
How is that different from a government?
So it's basically converging to the same thing.
- So I think it always,
I was trying to argue with Malice,
I feel like it always converges towards government at scale.
But I think the idea is you can have a lot of collectives
that are, you basically never let anything scale too big.
So one of the problems with governments is it gets too big
in terms of the size of the group
over which it has control.
My sense is that would happen anyway.
So a successful company like Amazon or Facebook,
I mean, it starts forming a monopoly
§20Transcript · 2:47:33–2:56:20
over entire populations,
not over just the hundreds of millions,
but billions of people.
So I don't know.
But there is something about the abuses of power,
the government can have
when it has a monopoly on violence.
And so that's a tension there.
- So the question is how can you set the incentives
for government correctly?
And this mostly applies at the highest levels of government.
And because we haven't found a way to set them correctly,
we made the highest levels of government relatively weak.
And this is, I think, part of the reason
why we had difficulty to coordinate the pandemic response.
And China didn't have that much difficulty.
And there is, of course, a much higher risk
of the abuse of power that exists in China
because the power is largely unchecked.
And that's basically what happens in the next generation.
For instance, imagine that we would agree
that the current government of China
is largely correct and benevolent.
And maybe we don't agree on this, but if we did,
how can we make sure that this stays like this?
And if you don't have checks and balances
and division of power, it's hard to achieve.
We don't have a solution for that problem.
But the abolishment of government basically
would remove the control structure.
From a cybernetic perspective,
there is an optimal point in the system
that the regulation should be happening, right?
But you can measure the current incentives
and the regulator would be properly incentivized
to make the right decisions
and change the payout metrics of everything below it
in such a way that the local prisoners' dilemmas
get resolved, right?
You cannot resolve the prisoners' dilemma
without some kind of eternal control
that emulates an infinite game in a way.
- Yeah, I mean, there's a sense in which
it seems like the reason government,
the parts of government that don't work well currently
is because there's not good mechanisms
through which to interact for the citizenry
to interact with government.
It's basically, it hasn't caught up in terms of technology.
And I think once you integrate
some of the digital revolution
of being able to have a lot of access to data,
be able to vote on different ideas at a local level,
at all levels, at the optimal level, like you're saying,
that can resolve the prisoner dilemmas,
and to integrate AI to help you out,
automate things that are like,
that don't require the human ingenuity,
I feel like that's where government could operate that well
and can also break apart the inefficient bureaucracies
if needed.
There'll be a strong incentive
to be efficient and successful.
- So our human history, we see an evolution
and evolutionary competition of modes of government
and of individual governments is in these modes.
And every nation state in some sense
is some kind of organism that has found different solutions
for the problem of government.
And you could look at all these different models
and the different scales at which it exists
as empirical attempts to validate the idea
of how to build a better government.
And I suspect that the idea of anarchism,
similar to the idea of communism,
is the result of being disenchanted
with the ugliness of the real existing solutions
and the attempt to get to an utopia.
And I suspect that communism originally was not a utopia.
I think that in the same way as original Christianity,
it had a particular kind of vision.
And this vision is a society,
a mode of organization within the society
in which humans can coexist at scale without coercion.
The same way as we do in a healthy family, right?
In a good family,
you don't terrorize each other into compliance,
but you understand what everybody needs
and what everybody is able to contribute
and what the intended future of the whole thing is.
And everybody coordinates their behavior in the right way
and informs each other about how to do this.
And all the interactions that happen
are instrumental to making that happen, right?
Could this happen at scale?
And I think this is the idea of communism.
Communism is opposed to the idea
that we need economic terror
or other forms of terror to make that happen.
But in practice, what happened
is that the proto-communist countries,
the real existing socialism,
replaced a part of the economic terror with moral terror.
So we were told to do the right thing for moral reasons.
And of course it didn't really work
and the economy eventually collapsed.
And the moral terror had actual real cost, right?
People were in prison
because they were morally non-compliant.
And the other thing is that the idea of communism
became a utopia.
So it basically was projected into the afterlife.
We were told in my childhood
that communism was a hypothetical society
to which we were in a permanent revolution
that justified everything
that was presently wrong with society morally.
But it was something that our grandchildren
probably would not ever see
because it was too ideal and too far in the future
to make it happen right now.
And people were just not there yet morally.
And the same thing happened with Christianity, right?
This notion of heaven was mythologized
and projected into an afterlife.
And I think this was just the idea of God's kingdom,
of this world in which we instantiate
the next level transcendental agent in the perfect form.
So everything goes smoothly and without violence
and without conflict and without this human messiness
on this economic messiness and the terror and coercion
that existed in the present societies.
And the idea of whether humans can exist at scale
in a harmonious way and non-coercively is untested.
A lot of people tested it,
but didn't get it to work so far.
And the utopia is a world in where you get
all the good things without any of the bad things.
And you are, I think, very susceptible to believe
in utopias when you are very young
and don't understand that everything has to happen
in causal patterns, that there's always feedback loops
that ultimately are closed.
There's nothing that just happens because it's good or bad.
Good or bad don't exist in isolation.
They only exist with respect to larger systems.
- So can you intuit why utopias fail as systems?
So like having a utopia that's out there
beyond the horizon, is it because then,
so it's not only because it's impossible to achieve utopias,
but it's because what certain humans,
certain small number of humans start to
sort of greedily attain power and money
and control and influence as they become,
as they see the power in using this idea
of a utopia for propaganda.
- That's a bit like saying, why is my garden not perfect?
It's because some evil weeds are overgrowing it
and they always do.
- Yeah.
- But this is not how it works.
A good garden is a system that is in balance
and requires minimal interactions by the gardener.
And so you need to create a system
that is designed to self-stabilize.
And the design of social systems requires
not just the implementation of the desired functionality,
but the next level design, also in biological systems.
You need to create a system that wants to converge
to the intended function.
And so instead of just creating an institution
like the FDA that is performing a particular kind of role
in society, you need to make sure that the FDA
is actually driven by a system that wants
to do this optimally, that is incentivized
to do it optimally and then makes the performance
that is actually enacted in every generation
instrumental to that thing, that actual goal, right?
And that is much harder to design and to achieve.
- So you have to design a system where,
I mean, listen, communism also was quote unquote
incentivized to be a feedback loop system
that achieves that utopia.
It's just, it wasn't working given human nature.
The incentives were not correct given human nature.
- So how do you incentivize people
when they are getting coal off the ground
to work as hard as possible?
Because it's a terrible job
and it's very bad for your health.
And right, how do you do this?
And you can give them prices and metals
and status to some degree, right?
There's only so much status to give for that.
And most people will not fall for this, right?
Or you can pay them and you probably have to pay them
in an asymmetric way because if you pay everybody the same
and you nationalize the coal mines,
eventually people will figure out
that they can game the system.
§21Transcript · 2:56:21–3:04:38
- Yes, so you're describing capitalism.
So capitalism is the present solution to the system.
And what we also noticed that I think that Marx was correct
in saying that capitalism is prone to crisis,
that capitalism is a system that in its dynamics
is not convergent, but divergent.
It's not a stable system.
And that eventually it produces an enormous potential
for productivity, but it also is systematically
misallocating resources.
So a lot of people cannot participate in the production
and consumption anymore.
And this is what we observe.
We observe that the middle class in the US is tiny.
A lot of people think that they're middle class,
but if you are still flying economy,
you're not middle class.
Every class is a magnitude smaller than the previous class.
(laughing)
- I think about classes, it's really like airline classes.
(laughing)
- A lot of people are economy class,
a lot of people are economy class.
- Have we really-
- Business class and very few are first class
and some are budget.
- I mean, I understand.
I think there is a, yeah, maybe some people,
probably I would push back against that definition
of the middle class.
It does feel like the middle class is pretty large,
but yes, there's a discrepancy in terms of wealth.
There's a big wealth gap.
- So if you think about in terms of the productivity
that our society could have,
there is no reason for anybody to fly economy.
We would be able to let everybody travel in style.
- Well, but also some people like to be frugal
even when they're billionaires.
Okay, so let's take that into account.
- Yes, but I mean, we probably don't need
to be traveling lavish,
but you also don't need to be tortured, right?
There is a difference between frugal
and subjecting yourself to torture.
- Listen, I love economy.
I don't understand why you're comparing
a flying economy to torture.
I don't, although the fight here,
there's two crying babies next to me.
So that, but that has nothing to do with the calm
as to do with crying babies.
They're very cute though.
So they kind of-
- Yeah, I have two kids and sometimes I have to go back
to visit the grandparents.
And that means going from the West Coast to Germany
and it's a long flight.
- Is it true that sort of when you're a father,
you grow immune to the crying and all that kind of stuff?
Like, because like me just not having kids,
it can be other people's kids can be quite annoying
when they're crying and screaming
and all that kind of stuff.
- When you have children and you're wired up
in the default natural way, you're lucky in this regard,
you fall in love with them.
And this falling in love with them means
that you basically start to see the world
through their eyes and you understand
that in a given situation, they cannot do anything
but being expressing despair.
And so it becomes more differentiated.
I noticed that for instance,
my son is typically acting on pure experience
of what things are like right now.
And he has to do this right now.
And you have this small child that is,
when he was a baby and so on,
where he was just immediately expressing what he felt.
And if you cannot regulate this from the outside,
there's no point to be upset about it, right?
It's like dealing with weather or something like this.
You all have to get through it.
And it's not easy for him either.
But if you also have a daughter,
maybe she is planning for that.
Maybe she understands that she's sitting
in the car behind you and she's screaming
at the top of her lungs and you're almost doing an accident
and you really don't know what to do.
What should I have done to make you stop screaming?
You could have given me candy.
(laughing)
- I think that's like a cat versus dog discussion.
I love it.
'Cause you said like a fundamental aspect
of that is love that makes it all worth it.
What in this monkey riding an elephant in a dream world,
what role does love play in the human condition?
- I think that love is the facilitator
of non-transactional interaction.
When you are observing your own purposes,
some of these purposes go beyond your ego.
They go beyond the particular organism that you are
and your local interests.
- That's what you mean by non-transactional.
- Yes, so basically when you are acting
in a transactional way,
it means that you are expecting something in return for you
from the one that you're interacting with.
You are interacting with a random stranger.
You buy something from them on eBay.
You expect a fair value for the money that you sent them
and vice versa because you don't know that person.
You don't have any kind of relationship to them.
But when you know this person a little bit better
and you know the situation that they're in
and you understand what they're trying to achieve
in their life and you approve because you realize
that they're in some sense serving
the same human sacredness as you are.
And they need to think that you have.
Maybe you give it to them as a present.
- But I mean, the feeling itself of joy
is a kind of benefit, is a kind of transaction.
- Yes, but the joy is not the point.
The joy is the signal that you get.
It's the reinforcement signal that your brain sends to you
because you are acting on the incentives
of the agent that you're a part of.
We are meant to be part of something larger.
This is the way in which we out-competed other hominins.
- Take that Neanderthals.
- Yeah, right.
And also other humans.
There was a population bottleneck for human society
that leads to an extreme lack
of genetic diversity among humans.
If you look at Bushmen in the Kalahari,
that basically tribes that are not that far distant
to each other have more genetic diversity
than exists between Europeans and Chinese.
And it's because basically the out-of-Africa population
at some point had a bottleneck
of just a few thousand individuals.
And what probably happened is not that at any time
the number of people shrunk below a few hundred thousand.
What probably happened is that there was a small group
that had a decisive mutation that produced an advantage.
And this group multiplied and killed everybody else.
And we are descendants of that group.
- Yeah, I wonder what the peculiar characteristics
of that group.
- Yeah.
- I mean, we can never know.
- Me too, and a lot of people do.
- We can only just listen to the echoes in our,
like the ripples that are still within us.
- So I suspect what eventually made a big difference
was the ability to organize at scale,
be able to program each other.
- With ideas.
- That we became programmable,
that we are willing to work in lockstep,
that we went below, above the tribal level,
that we no longer were groups of a few hundred individuals
and acted on direct reputation systems transactionally,
but that we basically evolved an adaptation
to become state building.
- Yeah.
To form collectives outside of the direct collectives.
- Yes, and that's basically a part of us
became committed to serving something outside
of what we know. - Bigger than ourselves.
Yeah, then that's kind of what love is.
- And it's terrifying because it meant
that we eradicated the others.
It's a force, it's an adaptive force
that gets us ahead in evolution,
which means we displace something else
that doesn't have that.
- Oh, so we had to murder a lot of people
that weren't about love.
So love led to destruction.
- We didn't have the same strong love as we did.
Right, that's why I mentioned this thing with fascism
and you see this, these speeches,
do you want total war?
And everybody says, yes, right?
There's this big, oh my God, be a part of something
that is more important than me
that gives meaning to my existence.
(laughing)
- Fair enough.
Do you have advice for young people today
in high school, in college,
that are thinking about what to do
with their career, with their life,
so that at the end of the whole thing,
§22Transcript · 3:04:40–3:12:15
they can be proud of what they did?
- Don't cheat.
Have integrity, aim for integrity.
- So what does integrity look like
when you're at the river or at the leaf
or the fat frog in a lake?
- It basically means that you try to figure out
what the thing is that is the most right.
And this doesn't mean that you have to look
for what other people tell you what's right,
but you have to aim for moral autonomy.
So things need to be right
independently of what other people say.
I always felt that when people told me
to listen to what others say,
like read the room, build your ideas
of what's true based on the high status people
of your in-group, that does not protect me from fascism.
The only way to protect yourself from fascism
is to decide it's the world that is being built here,
the world that I want to be in.
In some sense, try to make your behavior sustainable,
act in such a way that you would feel comfortable
on all sides of the transaction.
Realize that everybody is you in a different timeline,
but is seeing things differently and has reasons to do so.
- Yeah, I've come to realize this recently,
that there is an inner voice
that tells you what's right and wrong.
And speaking of reading the room,
there's times what integrity looks like
is there's times when a lot of people
are doing something wrong.
And what integrity looks like is not going on Twitter
and tweeting about it, but not participating quietly,
not doing, so it's not like signaling
or not all this kind of stuff,
but actually living your, what you think is right.
Like living it, not signaling.
- There's also sometimes this expectation
that others are like us.
So imagine the possibility that some of the people
around you are space aliens that only look human.
So they don't have the same prayers as you do.
They don't have the same impulses,
that's what's right and wrong.
There's a large diversity in these basic impulses
that people can have in a given situation.
And now realize that you are a space alien.
You are not actually human.
You think that you're human,
but you don't know what it means,
like what it's like to be human.
You just make it up as you go along like everybody else.
And you have to figure that out,
what it means that you are a full human being,
what it means to be human in the world
and how to connect with others on that.
And there's also something, don't be afraid
in the sense that if you do this, you're not good enough.
Because if you are acting on these incentives of integrity,
you become trustworthy.
That's the way in which you can recognize each other.
There is a particular place where you can meet
and you can figure out what that place is,
where you will give support to people
because you realize that they act with integrity
and they will also do that.
So in some sense, you are safe if you do that.
You're not always protected.
There are people which will abuse you
and that are bad actors in a way that it's hard to imagine
before you meet them.
But there is also people which will try to protect you.
- Yeah, thank you for saying that.
That's such a hopeful message
that no matter what happens to you,
there'll be a place,
there's people you'll meet
that also have what you have
and you will find happiness there and safety there.
- Yeah, but it doesn't need to end well.
It can also all go wrong.
So there's no guarantees in this life.
So you can do everything right and you still can fail
and you can still horrible things happening to you
that traumatize you and mutilate you
and you have to be grateful if it doesn't happen.
- And ultimately be grateful no matter what happens
'cause even just being alive is pretty damn nice.
- Yeah, even that, you know.
The gratefulness in some sense is also just generated
by your brain to keep you going.
It's all a trick.
- Speaking of which, Kamu said,
"I see many people die because they judge
that life is not worth living.
I see others paradoxically getting killed
for the ideas or illusions
that give them a reason for living.
What is called the reason for living
is also an excellent reason for dying.
I therefore conclude that the meaning of life
is the most urgent of questions."
So I have to ask what,
Jasha Bach, is the meaning of life?
It is an urgent question according to Kamu.
- I don't think that there's a single answer to this.
Nothing makes sense unless the mind makes it so.
So you basically have to project a purpose.
And if you zoom out far enough,
there's the heat test of the universe
and everything is meaningless.
Everything is just a blip in between.
And the question is,
do you find meaning in this blip in between?
Do you find meaning in observing squirrels?
Do you find meaning in raising children
and projecting a multi-generational organism
into the future?
Do you find meaning in projecting an aesthetic
of the world that you like into the future
and trying to serve that aesthetic?
And if you do, then life has that meaning.
And if you don't, then it doesn't.
- I kind of enjoy the idea that you just create
the most vibrant, the most weird,
the most unique kind of blip you can,
given your environment, given your set of skills.
Just be the most weird set of,
like local pocket of complexity you can be.
So that like when people study the universe,
they'll pause and be like, "Uh, that's weird."
- That looks like a useful strategy,
but of course it's still motivated reasoning.
(laughing)
You're obviously acting on your incentives here.
- It's still a story we tell ourselves within a dream
that's hardly in touch with reality.
- It's definitely a good strategy if you are a podcaster.
(laughing)
- And a human, which I'm still trying to figure out if I am.
- Yeah, there's a mutual relationship somehow.
- Somehow.
- Josh, you're one of the most incredible people I know.
I really love talking to you.
I love talking to you again.
And it's really an honor that you spend
your valuable time with me.
I hope we get to talk many times
throughout our short and meaningless lives.
- Or meaningful.
- Or meaningful.
- Thank you, Alex.
I enjoyed this conversation very much.
- Thanks for listening to this conversation with Josip Bach.
A thank you to Coinbase, Codecademy, Linode,
NetSuite, and ExpressVPN.
Check them out in the description to support this podcast.
Now, let me leave you with some words from Carl Jung.
"People will do anything, no matter how absurd,
"in order to avoid facing their own souls.
"One does not become enlightened
"by imagining figures of light,
"but by making the darkness conscious."
Thank you for listening, and hope to see you next time.
(upbeat music)
(upbeat music)