back to indexJoscha Bach: Nature of Reality, Dreams, and Consciousness | Lex Fridman Podcast #212
link |
The following is a conversation with Yosha Bach,
link |
his second time on the podcast.
link |
Yosha is one of the most fascinating minds in the world,
link |
exploring the nature of intelligence,
link |
cognition, computation, and consciousness.
link |
To support this podcast, please check out our sponsors,
link |
Coinbase, Codecademy, Linode, NetSuite, and ExpressVPN.
link |
Their links are in the description.
link |
This is the Lex Friedman podcast,
link |
and here is my conversation with Yosha Bach.
link |
Thank you for once again coming on
link |
to this particular Russian program
link |
and sticking to the theme of a Russian program.
link |
Let's start with the darkest of topics.
link |
So this is inspired by one of your tweets.
link |
You wrote that, quote,
link |
when life feels unbearable,
link |
I remind myself that I'm not a person.
link |
I am a piece of software running on the brain
link |
of a random ape for a few decades.
link |
It's not the worst brain to run on.
link |
Have you experienced low points in your life?
link |
Have you experienced depression?
link |
Of course, we all experience low points in our life,
link |
and we get appalled by the things,
link |
by the ugliness of stuff around us.
link |
We might get desperate about our lack of self regulation,
link |
and sometimes life is hard,
link |
and I suspect you don't get to your life,
link |
nobody does, to get through their life without low points
link |
and without moments where they're despairing.
link |
And I thought that let's capture this state
link |
and how to deal with that state.
link |
And I found that very often you realize
link |
that when you stop taking things personally,
link |
when you realize that this notion of a person is a fiction,
link |
similar as it is in Westworld,
link |
where the robots realize that their memories and desires
link |
are the stuff that keeps them in the loop,
link |
and they don't have to act on those memories and desires,
link |
that our memories and expectations is what make us unhappy.
link |
And the present really does.
link |
The day in which we are, for the most part, it's okay, right?
link |
When we are sitting here, right here, right now,
link |
we can choose how we feel.
link |
And the thing that affects us is the expectation
link |
that something is going to be different
link |
from what we want it to be,
link |
or the memory that something was different
link |
from what you wanted it to be.
link |
And once we basically zoom out from all this,
link |
what's left is not a person.
link |
What's left is this state of being conscious,
link |
which is a software state.
link |
And software doesn't have an identity.
link |
It's a physical law.
link |
And it's a law that acts in all of us,
link |
and it's embedded in a suitable substrate.
link |
And we didn't pick that substrate, right?
link |
We are mostly randomly instantiated on it.
link |
And they're all these individuals,
link |
and everybody has to be one of them.
link |
And eventually you're stuck on one of them,
link |
and have to deal with that.
link |
So you're like a leaf floating down the river.
link |
You just have to accept that there's a river,
link |
and you just float wherever it takes you.
link |
You don't have to do this.
link |
The thing is that the illusion that you are an agent
link |
What part of that is actually under your control?
link |
And I think that our consciousness
link |
is largely a control model for our own attention.
link |
So we notice where we are looking,
link |
and we can influence what we're looking,
link |
how we are disambiguating things,
link |
how we put things together in our mind.
link |
And the whole system that runs us
link |
is this big cybernetic motivational system.
link |
So we're basically like a little monkey
link |
sitting on top of an elephant,
link |
and we can put this elephant here and there
link |
to go this way or that way.
link |
And we might have the illusion that we are the elephant,
link |
or that we are telling it what to do.
link |
And sometimes we notice that it walks
link |
into a completely different direction.
link |
And we didn't set this thing up.
link |
It just is the situation that we find ourselves in.
link |
How much prodding can we actually do of the elephant?
link |
But I think that our consciousness
link |
cannot create the motive force.
link |
Is the elephant consciousness in this metaphor?
link |
No, the monkey is the consciousness.
link |
The monkey is the attentional system
link |
that is observing things.
link |
There is a large perceptual system
link |
combined with a motivational system
link |
that is actually providing the interface to everything
link |
and our own consciousness.
link |
I think is the tool that directs the attention
link |
of that system, which means it singles out features
link |
and performs conditional operations
link |
for which it needs an index memory.
link |
But this index memory is what we perceive
link |
as our stream of consciousness.
link |
But the consciousness is not in charge.
link |
That's an illusion.
link |
So everything outside of that consciousness
link |
So it's the physics of the universe,
link |
but it's also society that's outside of your...
link |
I would say the elephant is the agent.
link |
So there is an environment to which the agent is stomping
link |
and you are influencing a little part of that agent.
link |
So is the agent a single human being?
link |
Which object has agency?
link |
That's an interesting question.
link |
I think a way to think about an agent
link |
is that it's a controller with a set point generator.
link |
The notion of a controller comes from cybernetics
link |
and control theory.
link |
Control system consists out of a system
link |
that is regulating some value
link |
and the deviation of that value from a set point.
link |
And it has a sensor that measures the system's deviation
link |
from that set point and an effector
link |
that can be parametrized by the controller.
link |
So the controller tells the effector to do a certain thing.
link |
And the goal is to reduce the distance
link |
between the set point and the current value of the system.
link |
And there's an environment
link |
which disturbs the regulated system,
link |
which brings it away from that set point.
link |
So simplest case is a thermostat.
link |
The thermostat is really simple
link |
because it doesn't have a model.
link |
The thermostat is only trying to minimize
link |
the set point deviation in the next moment.
link |
And if you want to minimize the set point deviation
link |
over a longer time span, you need to integrate it.
link |
You need to model what is going to happen.
link |
So for instance, when you think about
link |
that your set point is to be comfortable in life,
link |
maybe you need to make yourself uncomfortable first, right?
link |
So you need to make a model of what's going to happen when.
link |
And this is task of the controller is to use its sensors
link |
to measure the state of the environment
link |
and the system that is being regulated
link |
and figure out what to do.
link |
And if the task is complex enough,
link |
the set points are complicated enough.
link |
And if the controller has enough capacity
link |
and enough sensor feedback,
link |
then the task of the controller is to make a model
link |
of the entire universe that it's in,
link |
the conditions under which it exists and of itself.
link |
And this is a very complex agent.
link |
And we are in that category.
link |
And an agent is not necessarily a thing in the universe.
link |
It's a class of models that we use
link |
to interpret aspects of the universe.
link |
And when we notice the environment around us,
link |
a lot of things only make sense
link |
at the level that should be entangled with them
link |
if we interpret them as control systems
link |
that make models of the world
link |
and try to minimize their own set points.
link |
So the models are the agents.
link |
The agent is a class of model.
link |
And we notice that we are an agent ourselves.
link |
We are the agent that is using our own control model
link |
to perform actions.
link |
We notice we produce a change in the model
link |
and things in the world change.
link |
And this is how we discover the idea that we have a body,
link |
that we are situated environment,
link |
and that we have a first person perspective.
link |
Still don't understand what's the best way to think
link |
of which object has agency with respect to human beings.
link |
Is it the contents of the brain as agency?
link |
Like what's the actuators that you're referring to?
link |
What is the controller and where does it reside?
link |
Or is it these impossible things?
link |
Because I keep trying to ground it to space time,
link |
the three dimension of space and the one dimension of time.
link |
What's the agent in that for humans?
link |
There is not just one.
link |
It depends on the way in which you're looking
link |
at this thing in which you're framing it.
link |
Imagine that you are, say Angela Merkel,
link |
and you are acting on behalf of Germany.
link |
Then you could say that Germany is the agent.
link |
And in the mind of Angela Merkel,
link |
she is Germany to some extent,
link |
because in the way in which she acts,
link |
the destiny of Germany changes.
link |
There are things that she can change
link |
that basically affect the behavior of that nation state.
link |
Okay, so it's hierarchies of,
link |
to go to another one of your tweets
link |
with I think you were playfully mocking Jeff Hawkins
link |
with saying his brain's all the way down.
link |
So it's like, it's agents all the way down.
link |
It's agents made up of agents, made up of agents.
link |
Like if Angela Merkel's Germany
link |
and Germany's made up a bunch of people
link |
and the people are themselves agents
link |
in some kind of context.
link |
And then people are made up of cells, each individual.
link |
So is it agents all the way down?
link |
I suspect that has to be like this
link |
in a world where things are self organizing.
link |
Most of the complexity that we are looking at,
link |
everything in life is about self organization.
link |
So I think up from the level of life, you have agents.
link |
And below life, you rarely have agents
link |
because sometimes you have control systems
link |
that emerge randomly in nature
link |
and try to achieve a set point,
link |
but they're not that interesting agents that make models.
link |
And because to make an interesting model of the world,
link |
you typically need a system that is true and complete.
link |
Can I ask you a personal question?
link |
What's the line between life and non life?
link |
It's personal because you're a life form.
link |
So what do you think in this emerging complexity,
link |
at which point does the things that are being living
link |
Personally, I think that the simplest answer
link |
that is that life is cells because...
link |
So it's a particular kind of principle
link |
that we have discovered to exist in nature.
link |
It's modular stuff that consists
link |
out of basically this DNA tape
link |
with a read write head on top of it,
link |
that is able to perform arbitrary computations
link |
and state transitions within the cell.
link |
And it's combined with a membrane
link |
that insulates the cell from its environment.
link |
And there are chemical reactions inside of the cell
link |
that are in disequilibrium.
link |
And the cell is running in such a way
link |
that this disequilibrium doesn't disappear.
link |
And the cell goes into an equilibrium state, it dies.
link |
And it requires something like an neck entropy extractor
link |
to maintain this disequilibrium.
link |
So it's able to harvest like entropy from its environment
link |
and keep itself running.
link |
Yeah, so there's information and there's a wall
link |
to maintain this disequilibrium.
link |
But isn't this very earth centric?
link |
Like what you're referring to as a...
link |
I'm not making a normative notion.
link |
You could say that there are probably other things
link |
in the universe that are cell like and life like,
link |
and you could also call them life,
link |
but eventually it's just a willingness
link |
to find an agreement of how to use the terms.
link |
I like cells because it's completely coextential
link |
with the way that we use the word
link |
even before we knew about cells.
link |
So people were pointing at some stuff
link |
and saying, this is somehow animate.
link |
And this is very different from the non animate stuff.
link |
And what's the difference between the living
link |
and the dead stuff.
link |
And it's mostly whether the cells are working or not.
link |
And also this boundary of life,
link |
where we say that for instance, the virus
link |
is basically an information packet
link |
that is subverting the cell and not life by itself.
link |
That makes sense to me.
link |
And it's somewhat arbitrary.
link |
You could of course say that systems
link |
that permanently maintain a disequilibrium
link |
and can self replicate are always life.
link |
And maybe that's a useful definition too,
link |
but this is eventually just how you want to use the word.
link |
Is it so useful for conversation,
link |
but is it somehow fundamental to the universe?
link |
Do you think there's a actual line
link |
to eventually be drawn between life and non life?
link |
Or is it all a kind of continuum?
link |
I don't think it's a continuum,
link |
but there's nothing magical that is happening.
link |
Living systems are a certain type of machine.
link |
What about non living systems?
link |
Is it also a machine?
link |
There are non living machines,
link |
but the question is at which point is a system
link |
able to perform arbitrary state transitions
link |
to make representations.
link |
And living things can do this.
link |
And of course we can also build non living things
link |
that can do this, but we don't know anything in nature
link |
that is not a cell and is not created by still alive
link |
that is able to do that.
link |
Not only do we not know,
link |
I don't think we have the tools to see otherwise.
link |
I always worry that we look at the world too narrowly.
link |
Like there could be life of a very different kind
link |
right under our noses that we're just not seeing
link |
because we're not either limitations
link |
of our cognitive capacity,
link |
or we're just not open minded enough
link |
either with the tools of science
link |
or just the tools of our mind.
link |
Yeah, that's possible.
link |
I find this thought very fascinating.
link |
And I suspect that many of us ask ourselves since childhood,
link |
what are the things that we are missing?
link |
What kind of systems and interconnections exist
link |
that are outside of our gaze?
link |
But we are looking for it
link |
and physics doesn't have much room at the moment
link |
for opening up something that would not violate
link |
the conservation of information as we know it.
link |
Yeah, but I wonder about time scale and scale,
link |
spatial scale, whether we just need to open up our idea
link |
of what, like how life presents itself.
link |
It could be operating in a much slower time scale,
link |
a much faster time scale.
link |
And it's almost sad to think that there's all this life
link |
around us that we're not seeing
link |
because we're just not like thinking
link |
in terms of the right scale, both time and space.
link |
What is your definition of life?
link |
What do you understand as life?
link |
Entities of sufficiently high complexity
link |
that are full of surprises.
link |
I don't know, I don't have a free will.
link |
So that just came out of my mouth.
link |
I'm not sure that even makes sense.
link |
There's certain characteristics.
link |
So complexity seems to be a necessary property of life.
link |
And I almost want to say it has ability
link |
to do something unexpected.
link |
It seems to me that life is the main source
link |
of complexity on earth.
link |
And complexity is basically a bridgehead
link |
that order builds into chaos by modeling,
link |
by processing information in such a way
link |
that you can perform reactions
link |
that would not be possible for dump systems.
link |
And this means that you can harvest neck entropy
link |
that dump systems cannot harvest.
link |
And this is what complexity is mostly about.
link |
In some sense, the purpose of life is to create complexity.
link |
I mean, there seems to be some kind of universal drive
link |
towards increasing pockets of complexity.
link |
I don't know what that is.
link |
That seems to be like a fundamental,
link |
I don't know if it's a property of the universe
link |
or it's just a consequence of the way the universe works,
link |
but there seems to be this small pockets
link |
of emergent complexity that builds on top of each other
link |
and starts having like greater and greater complexity
link |
by having like a hierarchy of complexity.
link |
Little organisms building up a little society
link |
that then operates almost as an individual organism itself.
link |
And all of a sudden you have Germany and Merkel.
link |
Well, that's not obvious to me.
link |
Everything that goes up has to come down at some point.
link |
So if you see this big exponential curve somewhere,
link |
it's usually the beginning of an S curve
link |
where something eventually reaches saturation.
link |
And the S curve is the beginning of some kind of bump
link |
that goes down again.
link |
And there is just this thing that when you are
link |
in sight of an evolution of life,
link |
you are on top of a puddle of negentropy
link |
that is being sucked dry by life.
link |
And during that happening,
link |
you see an increase in complexity
link |
because life forms are competing with each other
link |
to get more and more finer and finer corner
link |
of that negentropy extraction.
link |
I feel like that's a gradual beautiful process
link |
like that's almost follows a process akin to evolution.
link |
And the way it comes down is not the same way it came up.
link |
The way it comes down is usually harshly and quickly.
link |
So usually there's some kind of catastrophic event.
link |
The Roman Empire took a long time.
link |
But would that be,
link |
would you classify this as a decrease in complexity though?
link |
I think that this size of the cities that could be fed
link |
has decreased dramatically.
link |
And you could see that the quality of the art decreased
link |
and it did so gradually.
link |
And maybe future generations,
link |
when they look at the history of the United States
link |
in the 21st century,
link |
will also talk about the gradual decline,
link |
not something that suddenly happens.
link |
Do you have a sense of where we are?
link |
Are we on the exponential rise?
link |
Are we at the peak?
link |
Or are we at the downslope of the United States empire?
link |
It's very hard to say from a single human perspective,
link |
but it seems to me that we are probably at the peak.
link |
I think that's probably the definition of like optimism
link |
So my nature of optimism is,
link |
I think we're on the rise.
link |
I think this is just all a matter of perspective.
link |
but I do think that erring on the side of optimism,
link |
like you need a sufficient number,
link |
you need a minimum number of optimists
link |
in order to make that up thing actually work.
link |
And so I tend to be on the side of the optimists.
link |
I think that we are basically a species of grasshoppers
link |
that have turned into locusts.
link |
And when you are in that locust mode,
link |
you see an amazing rise of population numbers
link |
and of the complexity of the interactions
link |
between the individuals.
link |
But it's ultimately the question is, is it sustainable?
link |
See, I think we're a bunch of lions and tigers
link |
that have become domesticated cats,
link |
to use a different metaphor.
link |
As I'm not exactly sure we're so destructive,
link |
we're just softer and nicer and lazier.
link |
But I think we have monkeys and not the cats.
link |
And if you look at the monkeys, they are very busy.
link |
The ones that have a lot of sex, those monkeys?
link |
Not just the bonobos.
link |
I think that all the monkeys are basically
link |
a discontent species that always needs to meddle.
link |
Well, the gorillas seem to have
link |
a little bit more of a structure,
link |
but it's a different part of the tree.
link |
Okay, you mentioned the elephant
link |
and the monkey riding the elephant.
link |
And consciousness is the monkey.
link |
And there's some prodding that the monkey gets to do.
link |
And sometimes the elephant listens.
link |
I heard you got into some contentious,
link |
maybe you can correct me,
link |
but I heard you got into some contentious
link |
free will discussions.
link |
Is this with Sam Harris or something like that?
link |
Not that I know of.
link |
Some people on Clubhouse told me
link |
you made a bunch of big debate points about free will.
link |
Well, let me just then ask you where,
link |
in terms of the monkey and the elephant,
link |
do you think we land in terms of the illusion of free will?
link |
How much control does the monkey have?
link |
We have to think about what the free will is
link |
in the first place.
link |
We are not the machine.
link |
We are not the thing that is making the decisions.
link |
We are a model of that decision making process.
link |
And there is a difference between making your own decisions
link |
and predicting your own decisions.
link |
And that difference is the first person perspective.
link |
And what basically makes decision making
link |
and the conditions of free will distinct
link |
from just automatically doing the best thing is
link |
that we often don't know what the best thing is.
link |
We make decisions under uncertainty.
link |
We make informed bets using a betting algorithm
link |
that we don't yet understand
link |
because we haven't reverse engineered
link |
our own minds sufficiently.
link |
We don't know the expected rewards.
link |
We don't know the mechanism
link |
by which we estimate the rewards and so on.
link |
But there is an algorithm.
link |
We observe ourselves performing
link |
where we see that we weight facts and factors
link |
and the future, and then some kind of possibility,
link |
some motive gets raised to an intention.
link |
And that's informed bet that the system is making.
link |
And that making of the informed bet,
link |
the representation of that is what we call free will.
link |
And it seems to be paradoxical
link |
because we think that the crucial thing is
link |
about it that it's somehow indeterministic.
link |
And yet if it was indeterministic, it would be random.
link |
And it cannot be random because if it was random,
link |
if just dice were being thrown in the universe,
link |
randomly forces you to do things, it would be meaningless.
link |
So the important part of the decisions
link |
is always the deterministic stuff.
link |
But it appears to be indeterministic to you
link |
because it's unpredictable.
link |
Because if it was predictable,
link |
you wouldn't experience it as a free will decision.
link |
You would experience it as just doing
link |
the necessary right thing.
link |
And you see this continuum between the free will
link |
and the execution of automatic behavior
link |
when you're observing other people.
link |
So for instance, when you are observing your own children,
link |
if you don't understand them,
link |
you will abuse this agent model
link |
where you have an agent with a set point generator.
link |
And the agent is doing the best it can
link |
to minimize the difference to the set point.
link |
And it might be confused and sometimes impulsive or whatever,
link |
but it's acting on its own free will.
link |
And when you understand what's happens
link |
in the mind of the child, you see that it's automatic.
link |
And you can outmodel the child,
link |
you can build things around the child
link |
that will lead the child to making exactly the decision
link |
that you are predicting.
link |
And under these circumstances,
link |
like when you are a stage musician
link |
or somebody who is dealing with people
link |
that you sell a car to,
link |
and you completely understand the psychology
link |
and the impulses and the space of thoughts
link |
that this individual can have at that moment.
link |
Under these circumstances,
link |
it makes no sense to attribute free will.
link |
Because it's no longer decision making under uncertainty.
link |
You are already certain.
link |
For them, there's uncertainty,
link |
but you already know what they're doing.
link |
But what about for you?
link |
So is this akin to like systems like cellular automata
link |
where it's deterministic,
link |
but when you squint your eyes a little bit,
link |
it starts to look like there's agents making decisions
link |
at the higher sort of when you zoom out
link |
and look at the entities
link |
that are composed by the individual cells.
link |
Even though there's underlying simple rules
link |
that make the system evolve in deterministic ways,
link |
it looks like there's organisms making decisions.
link |
Is that where the illusion of free will emerges,
link |
that jump in scale?
link |
It's a particular type of model,
link |
but this jump in scale is crucial.
link |
The jump in scale happens whenever
link |
you have too many parts to count
link |
and you cannot make a model at that level
link |
and you try to find some higher level regularity.
link |
And the higher level regularity is a pattern
link |
that you project into the world to make sense of it.
link |
And agency is one of these patterns, right?
link |
You have all these cells that interact with each other
link |
and the cells in our body are set up in such a way
link |
that they benefit if their behavior is coherent,
link |
which means that they act
link |
as if they were serving a common goal.
link |
And that means that they will evolve regulation mechanisms
link |
that act as if they were serving a common goal.
link |
And now you can make sense of all these cells
link |
by projecting the common goal into them.
link |
Right, so for you then, free will is an illusion.
link |
No, it's a model and it's a construct.
link |
It's basically a model that the system is making
link |
of its own behavior.
link |
And it's the best model that it can come up with
link |
under the circumstances.
link |
And it can get replaced by a different model,
link |
which is automatic behavior,
link |
when you fully understand the mechanism
link |
under which you are acting.
link |
Yeah, but another word for model is what, story.
link |
So it's the story you're telling.
link |
I mean, do you actually have control?
link |
Is there such a thing as a you
link |
and is there such a thing as you have in control?
link |
So like, are you manifesting your evolution as an entity?
link |
In some sense, the you is the model of the system
link |
that is in control.
link |
It's a story that the system tells itself
link |
about somebody who is in control.
link |
And the contents of that model are being used
link |
to inform the behavior of the system.
link |
So the system is completely mechanical
link |
and the system creates that story like a loom.
link |
And then it uses the contents of that story
link |
to inform its actions
link |
and writes the results of that actions into the story.
link |
So how's that not an illusion?
link |
The story is written then,
link |
or rather we're not the writers of the story.
link |
Yes, but we always knew that.
link |
No, we don't know that.
link |
When did we know that?
link |
I think that's mostly a confusion about concepts.
link |
The conceptual illusion in our culture
link |
comes from the idea that we live in physical reality
link |
and that we experience physical reality
link |
and that you have ideas about it.
link |
And then you have this dualist interpretation
link |
where you have two substances, res extensa,
link |
the world that you can touch
link |
and that is made of extended things
link |
and res cogitans, which is the world of ideas.
link |
And in fact, both of them are mental representations.
link |
One is the representations of the world as a game engine
link |
that your mind generates to make sense of the perceptual data.
link |
And the other one,
link |
yes, that's what we perceive as the physical world.
link |
But we already know that the physical world
link |
is nothing like that, right?
link |
Quantum mechanics is very different
link |
from what you and me perceive as the world.
link |
The world that you and me perceive as a game engine.
link |
And there are no colors and sounds in the physical world.
link |
They only exist in the game engine generated by your brain.
link |
And then you have ideas
link |
that cannot be mapped onto extended regions, right?
link |
So the objects that have a spatial extension
link |
in the game engine, res extensa,
link |
and the objects that don't have a physical extension
link |
in the game engine are ideas.
link |
And they both interact in our mind
link |
to produce models of the world.
link |
Yep, but, you know, when you play video games,
link |
I understand that what's actually happening
link |
is zeros and ones inside of a computer,
link |
inside of a CPU and a GPU,
link |
but you're still seeing like the rendering of that.
link |
And you're still making decisions,
link |
whether to shoot, to turn left or to turn right,
link |
if you're playing a shooter,
link |
or every time I started thinking about Skyrim
link |
and Elder Scrolls and walking around in beautiful nature
link |
and swinging a sword.
link |
But it feels like you're making decisions
link |
inside that video game.
link |
So even though you don't have direct access
link |
in terms of perception to the bits,
link |
to the zeros and ones,
link |
it still feels like you're making decisions
link |
and your decisions actually feels
link |
like they're being applied all the way down
link |
to the zeros and ones.
link |
So it feels like you have control,
link |
even though you don't have direct access to reality.
link |
So there is basically a special character
link |
in the video game that is being created
link |
by the video game engine.
link |
And this character is serving the aesthetics
link |
of the video game, and that is you.
link |
Yes, but I feel like I have control inside the video game.
link |
Like all those like 12 year olds
link |
that kick my ass on the internet.
link |
So when you play the video game,
link |
it doesn't really matter that there's zeros and ones, right?
link |
You don't care about the bits of the past.
link |
You don't care about the nature of the CPU
link |
What you care about are the properties of the game
link |
that you're playing.
link |
And you hope that the CPU is good enough.
link |
And a similar thing happens when we interact with physics.
link |
The world that you and me are in is not the physical world.
link |
The world that you and me are in is a dream world.
link |
How close is it to the real world though?
link |
We know that it's not very close,
link |
but we know that the dynamics of the dream world
link |
match the dynamics of the physical world
link |
to a certain degree of resolution.
link |
But the causal structure of the dream world is different.
link |
So you see for instance waves crashing on your feet, right?
link |
But there are no waves in the ocean.
link |
There's only water molecules that have tangents
link |
between the molecules that are the result of electrons
link |
in the molecules interacting with each other.
link |
Aren't they like very consistent?
link |
We're just seeing a very crude approximation.
link |
Isn't our dream world very consistent,
link |
like to the point of being mapped directly one to one
link |
to the actual physical world
link |
as opposed to us being completely tricked?
link |
Is this is like where you have like Donald?
link |
It's not an illusion.
link |
It's a form of data compression.
link |
It's an attempt to deal with the dynamics
link |
of too many parts to count
link |
at the level at which we are entangled
link |
with the best model that you can find.
link |
Yeah, so we can act in that dream world
link |
and our actions have impact in the real world,
link |
in the physical world to which we don't have access.
link |
Yes, but it's basically like accepting the fact
link |
that the software that we live in,
link |
the dream that we live in is generated
link |
by something outside of this world that you and me are in.
link |
So is the software deterministic
link |
and do we not have any control?
link |
Do we have, so free will is having a conscious being.
link |
Free will is the monkey being able to steer the elephant.
link |
No, it's slightly different.
link |
Basically in the same way as you are modeling
link |
the water molecules in the ocean that engulf your feet
link |
when you are walking on the beach as waves
link |
and there are no waves,
link |
but only the atoms on more complicated stuff
link |
underneath the atoms and so on.
link |
And you know that, right?
link |
You would accept, yes,
link |
there is a certain abstraction that happens here.
link |
It's a simplification of what happens
link |
and the simplification that is designed
link |
in such a way that your brain can deal with it,
link |
temporarily and spatially in terms of resources
link |
and tuned for the predictive value.
link |
So you can predict with some accuracy
link |
whether your feet are going to get wet or not.
link |
But it's a really good interface and approximation.
link |
It says E equals MC squared is a good,
link |
equations are good approximation for,
link |
they're much better approximation.
link |
So to me, waves is a really nice approximation
link |
of what's all the complexity that's happening underneath.
link |
Basically it's a machine learning model
link |
that is constantly tuned to minimize surprises.
link |
So it basically tries to predict as well as it can
link |
what you're going to perceive next.
link |
Are we talking about, which is the machine learning?
link |
Our perception system or the dream world?
link |
The machine world, dream world is the result
link |
of the machine learning process of the perceptual system.
link |
That's doing the compression.
link |
And the model of you as an agent
link |
is not a different type of model or it's a different type,
link |
but not different as in its model like nature
link |
from the model of the ocean, right?
link |
Some things are oceans, some things are agents.
link |
And one of these agents is using your own control model,
link |
the output of your model,
link |
the things that you perceive yourself as doing.
link |
What about the fact that when you're standing
link |
with the water on your feet and you're looking out
link |
into the vast open water of the ocean
link |
and then there's a beautiful sunset
link |
and the fact that it's beautiful
link |
and then maybe you have friends or a loved one with you
link |
and you feel love, what is that?
link |
As the dream world or what is that?
link |
Yes, it's all happening inside of the dream.
link |
But see, the word dream makes it seem like it's not real.
link |
No, of course it's not real.
link |
The physical universe is real,
link |
but the physical universe is incomprehensible
link |
and it doesn't have any feeling of realness.
link |
The feeling of realness that you experience
link |
gets attached to certain representations
link |
where your brain assesses,
link |
this is the best model of reality that I have.
link |
So the only thing that's real to you
link |
is the thing that's happening at the very base of reality.
link |
Yeah, for something to be real, it needs to be implemented.
link |
So the model that you have of reality
link |
is real in as far as it is a model.
link |
It's an appropriate description of the world
link |
to say that there are models that are being experienced,
link |
but the world that you experience
link |
is not necessarily implemented.
link |
There is a difference between a reality,
link |
a simulation and a simulacrum.
link |
The reality that we're talking about
link |
is something that fully emerges
link |
over a causally closed lowest layer.
link |
And the idea of physicalism is that we are in that layer,
link |
that basically our world emerges over that.
link |
Every alternative to physicalism is a simulation theory,
link |
which basically says that we are
link |
in some kind of simulation universe
link |
and the real world needs to be in a parent universe of that,
link |
where the actual causal structure is, right?
link |
And when you look at the ocean and your own mind,
link |
you are looking at a simulation
link |
that explains what you're going to see next.
link |
So we are living in a simulation.
link |
Yes, but a simulation generated by our own brains.
link |
And this simulation is different from the physical reality
link |
because the causal structure that is being produced,
link |
what you are seeing is different
link |
from the causal structure of physics.
link |
Hopefully, if not, then you are going to end up
link |
in some kind of institution
link |
where people will take care of you
link |
because your behavior will be inconsistent, right?
link |
Your behavior needs to work in such a way
link |
that it's interacting with an accurately predictive
link |
And if your brain is unable to make your model
link |
of reality predictive, you will need help.
link |
So what do you think about Donald Hoffman's argument
link |
that it doesn't have to be consistent,
link |
the dream world to what he calls like the interface
link |
to the actual physical reality,
link |
where there could be evolution?
link |
I think he makes an evolutionary argument,
link |
which is like, it could be an evolutionary advantage
link |
to have the dream world drift away from physical reality.
link |
I think that only works if you have tenure.
link |
As long as you're still interacting with the ground tools,
link |
your model needs to be somewhat predictive.
link |
Well, in some sense, humans have achieved a kind of tenure
link |
in the animal kingdom.
link |
And at some point we became too big to fail,
link |
so we became postmodernist.
link |
It all makes sense now.
link |
We can just change the version of reality that we like.
link |
Yeah, but basically you can do magic.
link |
You can change your assessment of reality,
link |
but eventually reality is going to come bite you in the ass
link |
if it's not predictive.
link |
Do you have a sense of what is that base layer
link |
of physical reality?
link |
You have like, so you have these attempts
link |
at the theories of everything,
link |
the very, very small of like strength theory,
link |
or what Stephen Wolfram talks about with the hyper grass.
link |
These are these tiny, tiny, tiny, tiny objects.
link |
And then there is more like quantum mechanics
link |
that's talking about objects that are much larger,
link |
but still very, very, very tiny.
link |
Do you have a sense of where the tiniest thing is
link |
that is like at the lowest level?
link |
The turtle at the very bottom.
link |
Do you have a sense what that turtle is?
link |
I don't think that you can talk about where it is
link |
because space is emerging over the activity of these things.
link |
So space, the coordinates only exist
link |
in relation to the things, other things.
link |
And so you could, in some sense, abstract it into locations
link |
that can hold information and trajectories
link |
that the information can take
link |
between the different locations.
link |
And this is how we construct our notion of space.
link |
And physicists usually have a notion of space
link |
that is continuous.
link |
And this is a point where I tend to agree
link |
with people like Stephen Wolfram
link |
who are very skeptical of the geometric notions.
link |
I think that geometry is the dynamics
link |
of too many parts to count.
link |
And when there are no infinities,
link |
if there were two infinities,
link |
you would be running into contradictions,
link |
which is in some sense what Gödel and Turing discovered
link |
in response to Hilbert's call.
link |
So there are no infinities.
link |
There are no infinities.
link |
There is unboundedness, but if you have a language
link |
that talks about infinity, at some point,
link |
the language is going to contradict itself,
link |
which means it's no longer valid.
link |
In order to deal with infinities and mathematics,
link |
you have to postulate the existence initially.
link |
You cannot construct the infinities.
link |
And that's an issue, right?
link |
You cannot build up an infinity from zero.
link |
But in practice, you never do this, right?
link |
When you perform calculations,
link |
you only look at the dynamics of too many parts to count.
link |
And usually these numbers are not that large.
link |
They're not Googles or something.
link |
The infinities that we are dealing with in our universe
link |
are mathematically speaking, relatively small integers.
link |
And still what we're looking at is dynamics
link |
where a trillion things behave similar
link |
to a hundred trillion things
link |
or something that is very, very large
link |
because they're converging.
link |
And these convergent dynamics, these operators,
link |
this is what we deal with when we are doing the geometry.
link |
Geometry is stuff where we can pretend that it's continuous
link |
because if we subdivide the space sufficiently fine grained,
link |
these things approach a certain dynamic.
link |
And this approach dynamic, that is what we mean by it.
link |
But I don't think that infinity would work, so to speak,
link |
that you would know the last digit of pi
link |
and that you have a physical process
link |
that rests on knowing the last digit of pi.
link |
Yeah, that could be just a peculiar quirk
link |
of human cognition that we like discrete.
link |
Discrete makes sense to us.
link |
Infinity doesn't, so in terms of our intuitions.
link |
No, the issue is that everything that we think about
link |
needs to be expressed in some kind of mental language,
link |
not necessarily natural language,
link |
but some kind of mathematical language
link |
that your neurons can speak
link |
that refers to something in the world.
link |
And what we have discovered
link |
is that we cannot construct a notion of infinity
link |
without running into contradictions,
link |
which means that such a language is no longer valid.
link |
And I suspect this is what made Pythagoras so unhappy
link |
when somebody came up with the notion of irrational numbers
link |
before it was time, right?
link |
There's this myth that he had this person killed
link |
when he blabbed out the secret
link |
that not everything can be expressed
link |
as a ratio between two numbers,
link |
but there are numbers between the ratios.
link |
The world was not ready for this.
link |
And I think he was right.
link |
That has confused mathematicians very seriously
link |
because these numbers are not values, they are functions.
link |
And so you can calculate these functions
link |
to a certain degree of approximation,
link |
but you cannot pretend that pi has actually a value.
link |
Pi is a function that would approach this value
link |
but nothing in the world rests on knowing pi.
link |
How important is this distinction
link |
between discrete and continuous for you to get to the book?
link |
Because there's a, I mean, in discussion of your favorite
link |
flavor of the theory of everything,
link |
there's a few on the table.
link |
So there's string theory, there's a particular,
link |
there's a little quantum gravity,
link |
which focused on one particular unification.
link |
There's just a bunch of favorite flavors
link |
of different people trying to propose
link |
a theory of everything.
link |
Eric Weinstein and a bunch of people throughout history.
link |
And then of course, Stephen Wolfram,
link |
who I think is one of the only people doing a discrete.
link |
No, no, there's a bunch of physicists
link |
who do this right now.
link |
And like Toffoli and Tomasello.
link |
And digital physics is something
link |
that is, I think, growing in popularity.
link |
But the main reason why this is interesting
link |
is because it's important sometimes to settle disagreements.
link |
I don't think that you need infinities at all,
link |
and you never needed them.
link |
You can always deal with very large numbers
link |
and you can deal with limits, right?
link |
We are fine with doing that.
link |
You don't need any kind of infinity.
link |
You can build your computer algebra systems just as well
link |
without believing in infinity in the first place.
link |
So you're okay with limits?
link |
Yeah, so basically a limit means that something
link |
is behaving pretty much the same
link |
if you make the number large.
link |
Right, because it's converging to a certain value.
link |
And at some point the difference becomes negligible
link |
and you can no longer measure it.
link |
And in this sense, you have things
link |
that if you have an ngon which has enough corners,
link |
then it's going to behave like a circle at some point, right?
link |
And it's only going to be in some kind of esoteric thing
link |
that cannot exist in the physical universe
link |
that you would be talking about this perfect circle.
link |
And now it turns out that it also wouldn't work
link |
in mathematics because you cannot construct mathematics
link |
that has infinite resolution
link |
without running into contradictions.
link |
So that is itself not that important
link |
because we never did that, right?
link |
It's just a thing that some people thought we could.
link |
And this leads to confusion.
link |
So for instance, Roger Penrose uses this as an argument
link |
to say that there are certain things
link |
that mathematicians can do dealing with infinities
link |
and by extension our mind can do
link |
that computers cannot do.
link |
Yeah, he talks about that the human mind
link |
can do certain mathematical things
link |
that the computer as defined
link |
by the universal Turing machine cannot.
link |
So that it has to do with infinity.
link |
Yes, it's one of the things.
link |
So he is basically pointing at the fact
link |
that there are things that are possible
link |
in the mathematical mind and in pure mathematics
link |
that are not possible in machines
link |
that can be constructed in the physical universe.
link |
And because he's an honest guy,
link |
he thinks this means that present physics
link |
cannot explain operations that happen in our mind.
link |
Do you think he's right?
link |
And so let's leave his discussion
link |
of consciousness aside for the moment.
link |
Do you think he's right about just
link |
what he's basically referring to as intelligence?
link |
So is the human mind fundamentally more capable
link |
as a thinking machine than a universal Turing machine?
link |
But so he's suggesting that, right?
link |
So our mind is actually less than a Turing machine.
link |
There can be no Turing machine
link |
because it's defined as having an infinite tape.
link |
And we always only have a finite tape.
link |
But he's saying it's better.
link |
Our minds can only perform finitely many operations.
link |
Yes, he thinks so.
link |
He's saying it can do the kind of computation
link |
that the Turing machine cannot.
link |
And that's because he thinks that our minds
link |
can do operations that have infinite resolution
link |
And I don't think that's the case.
link |
Our minds are just able to discover these limit operators
link |
over too many parts to count.
link |
What about his idea that consciousness
link |
is more than a computation?
link |
So it's more than something that a Turing machine can do.
link |
So again, saying that there's something special
link |
about our mind that cannot be replicated in a machine.
link |
The issue is that I don't even know
link |
how to construct a language to express
link |
this statement correctly.
link |
the basic statement is there's a human experience
link |
that includes intelligence, that includes self awareness,
link |
that includes the hard problem of consciousness.
link |
And the question is, can that be fully simulated
link |
in the computer, in the mathematical model of the computer
link |
as we understand it today?
link |
Roger Penrose says no.
link |
So the universe of Turing machine
link |
cannot simulate the universe.
link |
So the interesting question is,
link |
and you have to ask him this is, why not?
link |
What is this specific thing that cannot be modeled?
link |
And when I looked at his writings
link |
and I haven't read all of it,
link |
but when I read, for instance,
link |
the section that he writes in the introduction
link |
to a road to infinity,
link |
the thing that he specifically refers to
link |
is the way in which human minds deal with infinities.
link |
And that itself can, I think, easily be deconstructed.
link |
A lot of people feel that our experience
link |
cannot be explained in a mechanical way.
link |
And therefore it needs to be different.
link |
And I concur, our experience is not mechanical.
link |
Our experience is simulated.
link |
It exists only in a simulation.
link |
The only simulation can be conscious.
link |
Physical systems cannot be conscious
link |
because they're only mechanical.
link |
Cells cannot be conscious.
link |
Neurons cannot be conscious.
link |
Brains cannot be conscious.
link |
People cannot be conscious
link |
as far as if you understand them as physical systems.
link |
What can be conscious is the story of the system
link |
in the world where you write all these things
link |
You have experiences for the same reason
link |
that a character novel has experiences
link |
because it's written into the story.
link |
And now the system is acting on that story.
link |
And it's not a story that is written in a natural language.
link |
It's written in a perceptual language,
link |
in this multimedia language of the game engine.
link |
And in there, you write in what kind of experience you have
link |
and what this means for the behavior of the system,
link |
for your behavior tendencies, for your focus,
link |
for your attention, for your experience of valence
link |
And this is being used to inform the behavior of the system
link |
And then the story updates with the reactions of the system
link |
and the changes in the world and so on.
link |
And you live inside of that model.
link |
You don't live inside of the physical reality.
link |
And I mean, just to linger on it, like you say, okay,
link |
it's in the perceptual language,
link |
the multimodal perceptual language.
link |
That's the experience.
link |
That's what consciousness is within that model,
link |
within that story.
link |
But do you have agency?
link |
When you play a video game, you can turn left
link |
and you can turn right in that story.
link |
So in that dream world, how much control do you have?
link |
Is there such a thing as you in that story?
link |
Like, is it right to say the main character,
link |
you know, everybody's NPCs,
link |
and then there's the main character
link |
and you're controlling the main character?
link |
Or is that an illusion?
link |
Is there a main character that you're controlling?
link |
I'm getting to the point of like the free will point.
link |
Imagine that you are building a robot that plays soccer.
link |
And you've been to MIT computer science,
link |
you basically know how to do that, right?
link |
And so you would say the robot is an agent
link |
that solves a control problem,
link |
how to get the ball into the goal.
link |
And it needs to perceive the world
link |
and the world is disturbing him in trying to do this, right?
link |
So he has to control many variables to make that happen
link |
and to project itself and the ball into the future
link |
and understand its position on the field
link |
relative to the ball and so on,
link |
and the position of its limbs
link |
or in the space around it and so on.
link |
So it needs to have an adequate model
link |
that abstracting reality in a useful way.
link |
And you could say that this robot does have agency
link |
over what it's doing in some sense.
link |
And the model is going to be a control model.
link |
And inside of that control model,
link |
you can possibly get to a point
link |
where this thing is sufficiently abstract
link |
to discover its own agency.
link |
Our current robots don't do that.
link |
They don't have a unified model of the universe,
link |
but there's not a reason why we shouldn't be getting there
link |
at some point in the not too distant future.
link |
And once that happens,
link |
you will notice that the robot tells a story
link |
about a robot playing soccer.
link |
So the robot will experience itself playing soccer
link |
in a simulation of the world that it uses
link |
to construct a model of the locations of its legs
link |
and limbs in space on the field
link |
with relationship to the ball.
link |
And it's not going to be at the level of the molecules.
link |
It will be an abstraction that is exactly at the level
link |
that is most suitable for past planning
link |
of the movements of the robot.
link |
It's going to be a high level abstraction,
link |
but a very useful one that is as predictive
link |
as we can make it.
link |
And in that side of that story,
link |
there is a model of the agency of that system.
link |
So this model can accurately predict
link |
that the contents of the model
link |
are going to be driving the behavior of the robot
link |
in the immediate future.
link |
But there's the hard problem of consciousness,
link |
which I would also,
link |
there's a subjective experience of free will as well
link |
that I'm not sure where the robot gets that,
link |
where that little leap is.
link |
Because for me right now,
link |
everything I imagine with that robot,
link |
as it gets more and more and more sophisticated,
link |
the agency comes from the programmer of the robot still,
link |
of what was programmed in.
link |
You could probably do an end to end learning system.
link |
You maybe need to give it a few priors.
link |
So you nudge the architecture in the right direction
link |
that it converges more quickly,
link |
but ultimately discovering the suitable hyperparameters
link |
of the architecture is also only a search process.
link |
And as the search process was evolution,
link |
that has informed our brain architecture
link |
so we can converge in a single lifetime
link |
on useful interaction with the world
link |
and the formation of a self model.
link |
The problem is if we define hyperparameters broadly,
link |
so it's not just the parameters that control
link |
this end to end learning system,
link |
but the entirety of the design of the robot.
link |
Like there's, you have to remove the human completely
link |
And then in order to build the robot,
link |
you have to create an entire universe.
link |
Cause you have to go, you can't just shortcut evolution.
link |
You have to go from the very beginning
link |
in order for it to have,
link |
cause I feel like there's always a human
link |
pulling the strings and that makes it seem like
link |
the robot is cheating.
link |
It's getting a shortcut to consciousness.
link |
And you are looking at the current Boston Dynamics robots.
link |
It doesn't look as if there is somebody
link |
pulling the strings.
link |
It doesn't look like cheating anymore.
link |
Okay, so let's go there.
link |
Cause I got to talk to you about this.
link |
So obviously with the case of Boston Dynamics,
link |
as you may or may not know,
link |
it's always either hard coded or remote controlled.
link |
There's no intelligence.
link |
I don't know how the current generation
link |
of Boston Dynamics robots works,
link |
but what I've been told about the previous ones
link |
was that it's basically all cybernetic control,
link |
which means you still have feedback mechanisms and so on,
link |
but it's not deep learning for the most part
link |
as it's currently done.
link |
It's for the most part,
link |
just identifying a control hierarchy
link |
that is congruent to the limbs that exist
link |
and the parameters that need to be optimized
link |
for the movement of these limbs.
link |
And then there is a convergence progress.
link |
So it's basically just regression
link |
that you would need to control this.
link |
But again, I don't know whether that's true.
link |
That's just what I've been told about how they work.
link |
We have to separate several levels of discussion here.
link |
So the only thing they do is pretty sophisticated control
link |
with no machine learning
link |
in order to maintain balance or to right itself.
link |
It's a control problem in terms of using the actuators
link |
to when it's pushed or when it steps on a thing
link |
that's uneven, how to always maintain balance.
link |
And there's a tricky set of heuristics around that,
link |
but that's the only goal.
link |
Everything you see Boston Dynamics doing
link |
in terms of that to us humans is compelling,
link |
which is any kind of higher order movement,
link |
like turning, wiggling its butt,
link |
like jumping back on its two feet, dancing.
link |
Dancing is even worse because dancing is hard coded in.
link |
It's choreographed by humans.
link |
There's choreography software.
link |
So there is no, of all that high level movement,
link |
there's no anything that you can call,
link |
certainly can't call AI,
link |
but there's no even like basic heuristics.
link |
It's all hard coded in.
link |
And yet we humans immediately project agency onto them,
link |
which is fascinating.
link |
So the gap here doesn't necessarily have agency.
link |
What it has is cybernetic control.
link |
And the cybernetic control means you have a hierarchy
link |
of feedback loops that keep the behavior
link |
in certain boundaries so the robot doesn't fall over
link |
and it's able to perform the movements.
link |
And the choreography cannot really happen
link |
with motion capture because the robot would fall over
link |
because the physics of the robot,
link |
the weight distribution and so on is different
link |
from the weight distribution in the human body.
link |
So if you were using the directly motion captured movements
link |
of a human body to project it into this robot,
link |
You can do this with a computer animation.
link |
It will look a little bit off, but who cares?
link |
But if you want to correct for the physics,
link |
you need to basically tell the robot
link |
where it should move its limbs.
link |
And then the control algorithm is going
link |
to approximate a solution that makes it possible
link |
within the physics of the robot.
link |
And you have to find the basic solution
link |
for making that happen.
link |
And there's probably going to be some regression necessary
link |
to get the control architecture to make these movements.
link |
But those two layers are separate.
link |
So the thing, the higher level instruction
link |
of how you should move and where you should move
link |
is a higher level.
link |
Yeah, so I expect that the control level
link |
of these robots at some level is dumb.
link |
This is just the physical control movement,
link |
the motor architecture.
link |
But it's a relatively smart motor architecture.
link |
It's just that there is no high level deliberation
link |
about what decisions to make necessarily, right?
link |
But see, it doesn't feel like free will or consciousness.
link |
No, no, that was not where I was trying to get to.
link |
I think that in our own body, we have that too.
link |
So we have a certain thing that is basically
link |
just a cybernetic control architecture
link |
that is moving our limbs.
link |
And deep learning can help in discovering
link |
such an architecture if you don't have it
link |
in the first place.
link |
If you already know your hardware,
link |
you can maybe handcraft it.
link |
But if you don't know your hardware,
link |
you can search for such an architecture.
link |
And this work already existed in the 80s and 90s.
link |
People were starting to search for control architectures
link |
by motor babbling and so on,
link |
and just use reinforcement learning architectures
link |
to discover such a thing.
link |
And now imagine that you have
link |
the cybernetic control architecture already inside of you.
link |
And you extend this a little bit.
link |
So you are seeking out food, for instance,
link |
or rest or and so on.
link |
And you get to have a baby at some point.
link |
And now you add more and more control layers to this.
link |
And the system is reverse engineering
link |
its own control architecture
link |
and builds a high level model to synchronize
link |
the pursuit of very different conflicting goals.
link |
And this is how I think you get to purposes.
link |
Purposes are models of your goals.
link |
The goals may be intrinsic
link |
as the result of the different set point violations
link |
hunger and thirst for very different things,
link |
and rest and pain avoidance and so on.
link |
And you put all these things together
link |
and eventually you need to come up with a strategy
link |
to synchronize them all.
link |
And you don't need just to do this alone by yourself
link |
because we are state building organisms.
link |
We cannot function as isolation
link |
the way that homo sapiens is set up.
link |
So our own behavior only makes sense
link |
when you zoom out very far into a society
link |
or even into ecosystemic intelligence on the planet
link |
and our place in it.
link |
So the individual behavior only makes sense
link |
in these larger contexts.
link |
And we have a number of priors built into us.
link |
So we are behaving as if we were acting
link |
on these high level goals pretty much right from the start.
link |
And eventually in the course of our life,
link |
we can reverse engineer the goals that we're acting on,
link |
what actually are our higher level purposes.
link |
And the more we understand that,
link |
the more our behavior makes sense.
link |
But this is all at this point,
link |
complex stories within stories
link |
that are driving our behavior.
link |
Yeah, I just don't know how big of a leap it is
link |
to start create a system
link |
that's able to tell stories within stories.
link |
Like how big of a leap that is
link |
from where currently Boston Dynamics is
link |
or any robot that's operating in the physical space.
link |
And that leap might be big
link |
if it requires to solve the hard problem of consciousness,
link |
which is telling a hell of a good story.
link |
I suspect that consciousness itself is relatively simple.
link |
What's hard is perception
link |
and the interface between perception and reasoning.
link |
That's for instance, the idea of the consciousness prior
link |
that would be built into such a system by Yoshua Bengio.
link |
And what he describes, and I think that's accurate,
link |
is that our own model of the world
link |
can be described through something like an energy function.
link |
The energy function is modeling the contradictions
link |
that exist within the model at any given point.
link |
And you try to minimize these contradictions,
link |
the tangents in the model.
link |
And to do this, you need to sometimes test things.
link |
You need to conditionally disambiguate figure and ground.
link |
You need to distinguish whether this is true
link |
or that is true, and so on.
link |
Eventually you get to an interpretation,
link |
but you will need to manually depress a few points
link |
in your model to let it snap into a state that makes sense.
link |
And this function that tries to get the biggest dip
link |
in the energy function in your model,
link |
according to Yoshua Bengio, is related to consciousness.
link |
It's a low dimensional discrete function
link |
that tries to maximize this dip in the energy function.
link |
Yeah, I think I would need to dig into details
link |
because I think the way he uses the word consciousness
link |
is more akin to like self awareness,
link |
like modeling yourself within the world,
link |
as opposed to the subjective experience, the hard problem.
link |
No, it's not even the self is in the world.
link |
The self is the agent and you don't need to be aware
link |
of yourself in order to be conscious.
link |
The self is just a particular content that you can have,
link |
but you don't have to have.
link |
But you can be conscious in, for instance, a dream at night
link |
or during a meditation state where you don't have a self.
link |
Where you're just aware of the fact that you are aware.
link |
And what we mean by consciousness in the colloquial sense
link |
is largely this reflexive self awareness,
link |
that we become aware of the fact
link |
that you're paying attention,
link |
that we are the thing that pays attention.
link |
We are the thing that pays attention, right.
link |
I don't see where the awareness that we're aware,
link |
the hard problem doesn't feel like it's solved.
link |
I mean, it's called a hard problem for a reason,
link |
because it seems like there needs to be a major leap.
link |
Yeah, I think the major leap is to understand
link |
how it is possible that a machine can dream,
link |
that a physical system is able to create a representation
link |
that the physical system is acting on,
link |
and that is spun force and so on.
link |
But once you accept the fact that you are not in physics,
link |
but that you exist inside of the story,
link |
I think the mystery disappears.
link |
Everything is possible in the story.
link |
You exist inside the story.
link |
Okay, so the machine.
link |
Your consciousness is being written into the story.
link |
The fact that you experience things
link |
is written to the side of the story.
link |
You ask yourself, is this real what I'm seeing?
link |
And your brain writes into the story, yes, it's real.
link |
So what about the perception of consciousness?
link |
So to me, you look conscious.
link |
So the illusion of consciousness,
link |
the demonstration of consciousness.
link |
I ask for the legged robot.
link |
How do we make this legged robot conscious?
link |
So there's two things,
link |
and maybe you can tell me if they're neighboring ideas.
link |
One is actually make it conscious,
link |
and the other is make it appear conscious to others.
link |
Are those related?
link |
Let's ask it from the other direction.
link |
What would it take to make you not conscious?
link |
So when you are thinking about how you perceive the world,
link |
can you decide to switch from looking at qualia
link |
to looking at representational states?
link |
And it turns out you can.
link |
There is a particular way in which you can look at the world
link |
and recognize its machine nature, including your own.
link |
And in that state,
link |
you don't have that conscious experience
link |
in this way anymore.
link |
It becomes apparent as a representation.
link |
Everything becomes opaque.
link |
And I think this thing that you recognize,
link |
everything is a representation.
link |
This is typically what we mean with enlightenment states.
link |
And it can happen on the motivational level,
link |
but you can also do this on the experiential level,
link |
on the perceptual level.
link |
See, but then I can come back to a conscious state.
link |
Okay, I particularly,
link |
I'm referring to the social aspect
link |
that the demonstration of consciousness
link |
is a really nice thing at a party
link |
when you're trying to meet a new person.
link |
It's a nice thing to know that they're conscious
link |
I don't know how fundamental consciousness
link |
is in human interaction,
link |
but it seems like to be at least an important part.
link |
And I ask that in the same kind of way for robots.
link |
In order to create a rich, compelling
link |
human robot interaction,
link |
it feels like there needs to be elements of consciousness
link |
within that interaction.
link |
My cat is obviously conscious.
link |
And so my cat can do this party trick.
link |
She also knows that I am conscious,
link |
be able to have feedback about the fact
link |
that we are both acting on models of our own awareness.
link |
The question is how hard is it for the robot,
link |
artificially created robot to achieve cat level
link |
Yes, so the issue for me is currently not so much
link |
on how to build a system that creates a story
link |
about a robot that lives in the world,
link |
but to make an adequate representation of the world.
link |
And the model that you and me have is a unified one.
link |
It's one where you basically make sense of everything
link |
that you can perceive.
link |
Every feature in the world that enters your perception
link |
can be relationally mapped to a unified model of everything.
link |
And we don't have an AI that is able to construct
link |
such a unified model yet.
link |
So you need that unified model to do the party trick?
link |
Yes, I think that it doesn't make sense
link |
if this thing is conscious,
link |
but not in the same universe as you,
link |
because you could not relate to each other.
link |
So what's the process, would you say,
link |
of engineering consciousness in the machine?
link |
Like what are the ideas here?
link |
So you probably want to have some kind of perceptual system.
link |
This perceptual system is a processing agent
link |
that is able to track sensory data
link |
and predict the next frame in the sensory data
link |
from the previous frames of the sensory data
link |
and the current state of the system.
link |
So the current state of the system is, in perception,
link |
instrumental to predicting what happens next.
link |
And this means you build lots and lots of functions
link |
that take all the blips that you feel on your skin
link |
and that you see on your retina, or that you hear,
link |
and puts them into a set of relationships
link |
that allows you to predict what kind of sensory data,
link |
what kind of sensor of blips, vector of blips,
link |
you're going to perceive in the next frame.
link |
This is tuned and it's constantly tuned
link |
until it gets as accurate as it can.
link |
You build a very accurate prediction mechanism
link |
that is step one of the perception.
link |
So first you predict, then you perceive
link |
and see the error in your prediction.
link |
And you have to do two things to make that happen.
link |
One is you have to build a network of relationships
link |
that are constraints,
link |
that take all the variants in the world
link |
and put each of the variances into a variable
link |
that is connected with relationships to other variables.
link |
And these relationships are computable functions
link |
that constrain each other.
link |
So when you see a nose
link |
that points in a certain direction in space,
link |
you have a constraint that says
link |
there should be a face nearby that has the same direction.
link |
And if that is not the case,
link |
you have some kind of contradiction
link |
that you need to resolve
link |
because it's probably not a nose what you're looking at.
link |
It just looks like one.
link |
So you have to reinterpret the data
link |
until you get to a point where your model converges.
link |
And this process of making the sensory data
link |
fit into your model structure
link |
is what Piaget calls the assimilation.
link |
And accommodation is the change of the models
link |
where you change your model in such a way
link |
that you can assimilate everything.
link |
So you're talking about building
link |
a hell of an awesome perception system
link |
that's able to do prediction and perception
link |
and correct and keep improving.
link |
No, wait, that's...
link |
Wait, there's more.
link |
Yes, there's more.
link |
So the first thing that we wanted to do
link |
is we want to minimize the contradictions in the model.
link |
And of course, it's very easy to make a model
link |
in which you minimize the contradictions
link |
just by allowing that it can be
link |
in many, many possible states, right?
link |
So if you increase degrees of freedom,
link |
you will have fewer contradictions.
link |
But you also want to reduce the degrees of freedom
link |
because degrees of freedom mean uncertainty.
link |
You want your model to reduce uncertainty
link |
as much as possible,
link |
but reducing uncertainty is expensive.
link |
So you have to have a trade off
link |
between minimizing contradictions
link |
and reducing uncertainty.
link |
And you have only a finite amount of compute
link |
and experimental time and effort
link |
available to reduce uncertainty in the world.
link |
So you need to assign value to what you observe.
link |
So you need some kind of motivational system
link |
that is estimating what you should be looking at
link |
and what you should be thinking about it,
link |
how you should be applying your resources
link |
to model what that is, right?
link |
So you need to have something like convergence links
link |
that tell you how to get from the present state
link |
of the model to the next one.
link |
You need to have these compatibility links
link |
that tell you which constraints exist
link |
and which constraint violations exist.
link |
And you need to have some kind of motivational system
link |
that tells you what to pay attention to.
link |
So now we have a second agent next to the perceptual agent.
link |
We have a motivational agent.
link |
This is a cybernetic system
link |
that is modeling what the system needs,
link |
what's important for the system,
link |
and that interacts with the perceptual system
link |
to maximize the expected reward.
link |
And you're saying the motivational system
link |
is some kind of like, what is it?
link |
A high level narrative over some lower level.
link |
No, it's just your brainstem stuff,
link |
the limbic system stuff that tells you,
link |
okay, now you should get something to eat
link |
because I've just measured your blood sugar.
link |
So you mean like motivational system,
link |
like the lower level stuff, like hungry.
link |
Yes, there's basically physiological needs
link |
and some cognitive needs and some social needs
link |
and they all interact.
link |
And they're all implemented at different parts
link |
in your nervous system as the motivational system.
link |
But they're basically cybernetic feedback loops.
link |
It's not that complicated.
link |
It's just a lot of code.
link |
And so you now have a motivational agent
link |
that makes your robot go for the ball
link |
or that makes your worm go to eat food and so on.
link |
And you have the perceptual system
link |
that lets it predict the environment
link |
so it's able to solve that control problem to some degree.
link |
And now what we learned is that it's very hard
link |
to build a machine learning system
link |
that looks at all the data simultaneously
link |
to see what kind of relationships
link |
could exist between them.
link |
So you need to selectively model the world.
link |
You need to figure out where can I make the biggest difference
link |
if I would put the following things together.
link |
Sometimes you find a gradient for that.
link |
When you have a gradient,
link |
you don't need to remember where you came from.
link |
You just follow the gradient
link |
until it doesn't get any better.
link |
But if you have a world where the problems are discontinuous
link |
and the search spaces are discontinuous,
link |
you need to retain memory of what you explored.
link |
You need to construct a plan of what to explore next.
link |
And this thing means that you have next
link |
to this perceptual construction system
link |
and the motivational cybernetics,
link |
an agent that is paying attention
link |
to what it should select at any given moment
link |
to maximize reward.
link |
And this scanning system, this attention agent,
link |
is required for consciousness
link |
and consciousness is its control model.
link |
So it's the index memories that this thing retains
link |
when it manipulates the perceptual representations
link |
to maximize the value and minimize the conflicts
link |
and to increase coherence.
link |
So the purpose of consciousness is to create coherence
link |
in your perceptual representations,
link |
remove conflicts, predict the future,
link |
construct counterfactual representations
link |
so you can coordinate your actions and so on.
link |
And in order to do this, it needs to form memories.
link |
These memories are partial binding states
link |
of the working memory contents
link |
that are being revisited later on to backtrack,
link |
to undo certain states, to look for alternatives.
link |
And these index memories that you can recall,
link |
that is what you perceive as your stream of consciousness.
link |
And being able to recall these memories,
link |
this is what makes you conscious.
link |
If you could not remember what you paid attention to,
link |
you wouldn't be conscious.
link |
So consciousness is the index in the memory database.
link |
But let me sneak up to the questions of consciousness
link |
So we usually relate suffering to consciousness.
link |
So the capacity to suffer.
link |
I think to me, that's a really strong sign of consciousness
link |
is a thing that can suffer.
link |
How is that useful?
link |
And like in your model where you just described,
link |
which is indexing of memories and what is the coherence
link |
with the perception, with this predictive thing
link |
that's going on in the perception,
link |
how does suffering relate to any of that?
link |
The higher level suffering that humans do.
link |
Basically pain is a reinforcement signal.
link |
Pain is a signal that one part of your brain
link |
sends to another part of your brain,
link |
or in an abstract sense, part of your mind
link |
sends to another part of the mind to regulate its behavior,
link |
to tell it the behavior that you're currently exhibiting
link |
should be improved.
link |
And this is the signal that I tell you to move away
link |
from what you're currently doing
link |
and push into a different direction.
link |
So pain gives you a part of you an impulse
link |
to do something differently.
link |
But sometimes this doesn't work
link |
because the training part of your brain
link |
is talking to the wrong region,
link |
or because it has the wrong model
link |
of the relationships in the world.
link |
Maybe you're mismodeling yourself
link |
or you're mismodeling the relationship of yourself
link |
or you're mismodeling the dynamics of the world.
link |
So you're trying to improve something
link |
that cannot be improved by generating more pain.
link |
But the system doesn't have any alternative.
link |
So it doesn't get better.
link |
What do you do if something doesn't get better
link |
and you want it to get better?
link |
You increase the strengths of the signal.
link |
And then the signal becomes chronic
link |
when it becomes permanent without a change inside.
link |
This is what we call suffering.
link |
And the purpose of consciousness
link |
is to deal with contradictions,
link |
with things that cannot be resolved.
link |
The purpose of consciousness,
link |
I think is similar to a conductor in an orchestra.
link |
When everything works well,
link |
the orchestra doesn't need much of a conductor
link |
as long as it's coherent.
link |
But when there is a lack of coherence
link |
or something is consistently producing
link |
disharmony and mismatches,
link |
then the conductor becomes alert and interacts with it.
link |
So suffering attracts the activity of our consciousness.
link |
And the purpose of that is ideally
link |
that we bring new layers online,
link |
new layers of modeling that are able to create
link |
a model of the dysregulation so we can deal with it.
link |
And this means that we typically get
link |
higher level consciousness, so to speak, right?
link |
We get some consciousness above our pay grade maybe
link |
if we have some suffering early in our life.
link |
Most of the interesting people
link |
had trauma early on in their childhood.
link |
And trauma means that you are suffering an injury
link |
for which the system is not prepared,
link |
which it cannot deal with,
link |
which it cannot insulate itself from.
link |
So something breaks.
link |
And this means that the behavior of the system
link |
is permanently disturbed in a way
link |
that some mismatch exists now in the regulation
link |
that just by following your impulses,
link |
by following the pain in the direction where it hurts,
link |
the situation doesn't improve but get worse.
link |
And so what needs to happen is that you grow up.
link |
And that's part that has grown up
link |
is able to deal with the part
link |
that is stuck in this earlier phase.
link |
Yeah, so at least to grow,
link |
so you're adding extra layers to your cognition.
link |
And let me ask you then,
link |
because I gotta stick on suffering,
link |
the ethics of the whole thing.
link |
So not our consciousness, but the consciousness of others.
link |
You've tweeted, one of my biggest fears
link |
is that insects could be conscious.
link |
The amount of suffering on earth would be unthinkable.
link |
So when we think of other conscious beings,
link |
is suffering a property of consciousness
link |
that we're most concerned about?
link |
So I'm still thinking about robots,
link |
how to make sense of other nonhuman things
link |
that appear to have the depth of experience
link |
And to me, that means consciousness
link |
and the darkest side of that, which is suffering,
link |
the capacity to suffer.
link |
And so I started thinking,
link |
how much responsibility do we have
link |
for those other conscious beings?
link |
That's where the definition of consciousness
link |
becomes most urgent.
link |
Like having to come up with a definition of consciousness
link |
becomes most urgent,
link |
is who should we and should we not be torturing?
link |
There's no general answer to this.
link |
Was Genghis Khan doing anything wrong?
link |
It depends right on how you look at it.
link |
Well, he drew a line somewhere
link |
where this is us and that's them.
link |
It's the circle of empathy.
link |
you don't have to use the word consciousness,
link |
but these are the things that matter to me
link |
if they suffer or not.
link |
And these are the things that don't matter to him.
link |
Yeah, but when one of his commanders failed him,
link |
he broke his spine and let him die in a horrible way.
link |
And so in some sense,
link |
I think he was indifferent to suffering
link |
or he was not different in the sense
link |
that he didn't see it as useful if he inflicted suffering,
link |
but he did not see it as something that had to be avoided.
link |
That was not the goal.
link |
The question was, how can I use suffering
link |
and the infliction of suffering to reach my goals
link |
from his perspective?
link |
So like different societies throughout history
link |
put different value on the...
link |
Different individuals, different psyches.
link |
But also even the objective of avoiding suffering,
link |
like some societies probably,
link |
I mean, this is where like religious belief really helps
link |
that afterlife, that it doesn't matter
link |
that you suffer or die,
link |
what matters is you suffer honorably, right?
link |
So that you enter the afterlife as a hero.
link |
It seems to be superstitious to me,
link |
basically beliefs that assert things
link |
for which no evidence exists
link |
are incompatible with sound epistemology.
link |
And I don't think that religion has to be superstitious,
link |
otherwise it should be condemned in all cases.
link |
You're somebody who's saying we live in a dream world,
link |
we have zero evidence for anything.
link |
That's not the case.
link |
There are limits to what languages can be constructed.
link |
Mathematics brings solid evidence for its own structure.
link |
And once we have some idea of what languages exist
link |
and how a system can learn
link |
and what learning itself is in the first place.
link |
And so we can begin to realize that our intuitions
link |
that we are able to learn about the regularities
link |
of the world and minimize surprise
link |
and understand the nature of our own agency
link |
to some degree of abstraction.
link |
That's not an illusion.
link |
So it's a useful approximation.
link |
Just because we live in a dream world
link |
doesn't mean mathematics can't give us a consistent glimpse
link |
of physical, of objective reality.
link |
We can basically distinguish useful encodings
link |
from useless encodings.
link |
And when we apply our truth seeking to the world,
link |
we know we usually cannot find out
link |
whether a certain thing is true.
link |
What we typically do is we take the state vector
link |
of the universe separated into separate objects
link |
that interact with each other through interfaces.
link |
And this distinction that we are making
link |
is not completely arbitrary.
link |
It's done to optimize the compression
link |
that we can apply to our models of the universe.
link |
So we can predict what's happening
link |
with our limited resources.
link |
In this sense, it's not arbitrary.
link |
But the separation of the world into objects
link |
that are somehow discrete and interacting with each other
link |
is not the true reality, right?
link |
The boundaries between the objects
link |
are projected into the world, not arbitrarily projected.
link |
But still, it's only an approximation
link |
of what's actually the case.
link |
And we sometimes notice that we run into contradictions
link |
when we try to understand high level things
link |
like economic aspects of the world
link |
and so on, or political aspects, or psychological aspects
link |
where we make simplifications.
link |
And the objects that we are using to separate the world
link |
are just one of many possible projections
link |
of what's going on.
link |
So it's not, in this postmodernist sense,
link |
completely arbitrary, and you're free to pick
link |
what you want or dismiss what you don't like
link |
because it's all stories.
link |
No, that's not true.
link |
You have to show for every model
link |
of how well it predicts the world.
link |
So the confidence that you should have
link |
in the entities of your models
link |
should correspond to the evidence that you have.
link |
Can I ask you on a small tangent
link |
to talk about your favorite set of ideas and people,
link |
which is postmodernism.
link |
What is postmodernism?
link |
How would you define it?
link |
And why to you is it not a useful framework of thought?
link |
Postmodernism is something that I'm really not an expert on.
link |
And postmodernism is a set of philosophical ideas
link |
that is difficult to lump together,
link |
that is characterized by some useful thinkers,
link |
some of them poststructuralists and so on.
link |
And I'm mostly not interested in it
link |
because I think that it's not leading me anywhere
link |
that I find particularly useful.
link |
It's mostly, I think, born out of the insight
link |
that the ontologies that we impose on the world
link |
are not literally true.
link |
And that we can often get to a different interpretation
link |
by the world by using a different ontology
link |
that is different separation of the world
link |
into interacting objects.
link |
But the idea that this makes the world a set of stories
link |
that are arbitrary, I think, is wrong.
link |
And the people that are engaging in this type of philosophy
link |
are working in an area that I largely don't find productive.
link |
There's nothing useful coming out of this.
link |
So this idea that truth is relative
link |
is not something that has, in some sense,
link |
informed physics or theory of relativity.
link |
And there is no feedback between those.
link |
There is no meaningful information
link |
of this type of philosophy on the sciences
link |
or on engineering or in politics.
link |
But there is a very strong information on ideology
link |
because it basically has become an ideology
link |
that is justifying itself by the notion
link |
that truth is a relative concept.
link |
And it's not being used in such a way
link |
that the philosophers or sociologists
link |
that take up these ideas say,
link |
oh, I should doubt my own ideas because maybe my separation of the world
link |
into objects is not completely valid.
link |
And I should maybe use a different one
link |
and be open to a pluralism of ideas.
link |
But it mostly exists to dismiss the ideas of other people.
link |
It becomes, yeah, it becomes a political weapon of sorts
link |
Basically, there's nothing wrong, I think,
link |
with developing a philosophy around this.
link |
But to develop a philosophy around this,
link |
to develop norms around the idea
link |
that truth is something that is completely negotiable,
link |
is incompatible with the scientific project.
link |
And I think if the academia has no defense
link |
against the ideological parts of the postmodernist movement,
link |
Right, you have to acknowledge the ideological part
link |
of any movement, actually, including postmodernism.
link |
Well, the question is what an ideology is.
link |
And to me, an ideology is basically a viral memeplex
link |
that is changing your mind in such a way that reality gets warped.
link |
It gets warped in such a way that you're being cut off
link |
from the rest of human thought space.
link |
And you cannot consider things outside of the range of ideas
link |
of your own ideology as possibly true.
link |
Right, so, I mean, there's certain properties to an ideology
link |
that make it harmful.
link |
One of them is that dogmatism of just certainty,
link |
dogged certainty in that you're right,
link |
you have the truth, and nobody else does.
link |
Yeah, but what is creating the certainty?
link |
It's very interesting to look at the type of model
link |
that is being produced.
link |
Is it basically just a strong prior, and you tell people,
link |
oh, this idea that you consider to be very true,
link |
the evidence for this is actually just much weaker
link |
than you thought, and look, here are some studies.
link |
No, this is not how it works.
link |
It's usually normative, which means some thoughts
link |
are unthinkable because they would change your identity
link |
into something that is no longer acceptable.
link |
And this cuts you off from considering an alternative.
link |
And many de facto religions use this trick
link |
to lock people into a certain mode of thought,
link |
and this removes agency over your own thoughts.
link |
And it's very ugly to me.
link |
It's basically not just a process of domestication,
link |
but it's actually an intellectual castration
link |
It's an inability to think creatively
link |
and to bring forth new thoughts.
link |
I can ask you about substances, chemical substances
link |
that affect the video game, the dream world.
link |
So psychedelics that increasingly have been getting
link |
a lot of research done on them.
link |
So in general, psychedelics, psilocybin, MDMA,
link |
but also a really interesting one, the big one, which is DMT.
link |
What and where are the places that these substances
link |
take the mind that is operating in the dream world?
link |
Do you have an interesting sense how this throws a wrinkle
link |
into the prediction model?
link |
Is it just some weird little quirk
link |
or is there some fundamental expansion
link |
of the mind going on?
link |
I suspect that a way to look at psychedelics
link |
is that they induce particular types
link |
of lucid dreaming states.
link |
So it's a state in which certain connections
link |
are being severed in your mind.
link |
They're no longer active.
link |
Your mind basically gets free to move in a certain direction
link |
because some inhibition, some particular inhibition
link |
doesn't work anymore.
link |
And as a result, you might stop having a self
link |
or you might stop perceiving the world as three dimensional.
link |
And you can explore that state.
link |
And I suppose that for every state
link |
that can be induced with psychedelics,
link |
there are people that are naturally in that state.
link |
So sometimes psychedelics to shift you
link |
through a range of possible mental states.
link |
And they can also shift you out of the range
link |
of permissible mental states
link |
that is where you can make predictive models of reality.
link |
And what I observe in people that use psychedelics a lot
link |
is that they tend to be overfitting.
link |
Overfitting means that you are using more bits
link |
for modeling the dynamics of a function than you should.
link |
And so you can fit your curve
link |
to extremely detailed things in the past,
link |
but this model is no longer predictive for the future.
link |
What is it about psychedelics that forces that?
link |
I thought it would be the opposite.
link |
I thought that it's a good mechanism
link |
for generalization, for regularization.
link |
So it feels like psychedelics expansion of the mind,
link |
like taking you outside of,
link |
like forcing your model to be non predictive
link |
Meaning like, it's almost like, okay,
link |
what I would say psychedelics are akin to
link |
is traveling to a totally different environment.
link |
Like going, if you've never been to like India
link |
or something like that from the United States,
link |
very different set of people, different culture,
link |
different food, different roads and values
link |
and all those kinds of things.
link |
Yeah, so psychedelics can, for instance,
link |
teleport people into a universe that is hyperbolic,
link |
which means that if you imagine a room that you're in,
link |
you can turn around 360 degrees
link |
and you didn't go full circle.
link |
You need to go 720 degrees to go full circle.
link |
So the things that people learn in that state
link |
cannot be easily transferred
link |
in this universe that we are in.
link |
It could be that if they're able to abstract
link |
and understand what happened to them,
link |
that they understand that some part
link |
of their spatial cognition has been desynchronized
link |
and has found a different synchronization.
link |
And this different synchronization
link |
happens to be a hyperbolic one, right?
link |
So you learn something interesting about your brain.
link |
It's difficult to understand what exactly happened,
link |
but we get a pretty good idea
link |
once we understand how the brain is representing geometry.
link |
Yeah, but doesn't it give you a fresh perspective
link |
on the physical reality?
link |
Who's making that sound?
link |
Is it inside my head or is it external?
link |
Well, there is no sound outside of your mind,
link |
but it's making sense of phenomenon physics.
link |
Yeah, in the physical reality, there's sound waves
link |
traveling through air.
link |
That's our model of what happened.
link |
That's our model of what happened, right.
link |
Doesn't Psychedelics give you a fresh perspective
link |
on this physical reality?
link |
Like, not this physical reality, but this more...
link |
What do you call the dream world that's mapped directly to...
link |
The purpose of dreaming at night, I think,
link |
is data augmentation.
link |
So that's very different.
link |
That's very similar to Psychedelics.
link |
It's changed parameters about the things that you have learned.
link |
And, for instance, when you are young,
link |
you have seen things from certain perspectives,
link |
but not from others.
link |
So your brain is generating new perspectives of objects
link |
that you already know,
link |
which means you can learn to recognize them later
link |
from different perspectives.
link |
And I suspect that's the reason that many of us
link |
remember to have flying dreams as children,
link |
because it's just different perspectives of the world
link |
that you already know,
link |
and that it starts to generate these different
link |
perspective changes,
link |
and then it fluidly turns this into a flying dream
link |
to make sense of what's happening, right?
link |
So you fill in the gaps,
link |
and suddenly you see yourself flying.
link |
And similar things can happen with semantic relationships.
link |
So it's not just spatial relationships,
link |
but it can also be the relationships between ideas
link |
that are being changed.
link |
And it seems that the mechanisms that make that happen
link |
during dreaming are interacting
link |
with these same receptors
link |
that are being stimulated by psychedelics.
link |
So I suspect that there is a thing
link |
that I haven't read really about.
link |
The way in which dreams are induced in the brain
link |
is not just that the activity of the brain gets tuned down
link |
because your eyes are closed
link |
and you no longer get enough data from your eyes,
link |
but there is a particular type of neurotransmitter
link |
that is saturating your brain during these phases,
link |
during the REM phases, and you produce
link |
controlled hallucinations.
link |
And psychedelics are linking into these mechanisms,
link |
So isn't that another trickier form of data augmentation?
link |
Yes, but it's also data augmentation
link |
that can happen outside of the specification
link |
that your brain is tuned to.
link |
So basically people are overclocking their brains
link |
and that produces states
link |
that are subjectively extremely interesting.
link |
But from the outside, very suspicious.
link |
So I think I'm over applying the metaphor
link |
of a neural network in my own mind,
link |
which I just think that doesn't lead to overfitting, right?
link |
But you were just sort of anecdotally saying
link |
my experiences with people that have done psychedelics
link |
are that kind of quality.
link |
I think it typically happens.
link |
So if you look at people like Timothy Leary,
link |
and he has written beautiful manifestos
link |
about the effect of LSD on people.
link |
He genuinely believed, he writes in these manifestos,
link |
that in the future, science and art
link |
will only be done on psychedelics
link |
because it's so much more efficient and so much better.
link |
And he gave LSD to children in this community
link |
of a few thousand people that he had near San Francisco.
link |
And basically he was losing touch with reality.
link |
He did not understand the effects
link |
that the things that he was doing
link |
would have on the reception of psychedelics
link |
by society because he was unable to think critically
link |
about what happened.
link |
What happened was that he got in a euphoric state,
link |
that euphoric state happened because he was overfitting.
link |
He was taking this sense of euphoria
link |
and translating it into a model
link |
of actual success in the world, right?
link |
He was feeling better.
link |
Limitations had disappeared,
link |
that he experienced to be existing,
link |
but he didn't get superpowers.
link |
I understand what you mean by overfitting now.
link |
There's a lot of interpretation to the term
link |
overfitting in this case, but I got you.
link |
So he was getting positive rewards
link |
from a lot of actions that he shouldn't have been doing.
link |
Yeah, but not just this.
link |
So if you take, for instance, John Lilly,
link |
who was studying dolphin languages and aliens and so on,
link |
a lot of people that use psychedelics became very loopy.
link |
And the typical thing that you notice
link |
when people are on psychedelics is that they are in a state
link |
where they feel that everything can be explained now.
link |
Everything is clear, everything is obvious.
link |
And sometimes they have indeed discovered
link |
a useful connection, but not always.
link |
Very often these connections are overinterpretations.
link |
I wonder, you know, there's a question
link |
of correlation versus causation.
link |
And also I wonder if it's the psychedelics
link |
or if it's more the social, like being the outsider
link |
and having a strong community of outside
link |
and having a leadership position in an outsider cult
link |
like community that could have a much stronger effect
link |
of overfitting than do psychedelics themselves,
link |
the actual substances, because it's a counterculture thing.
link |
So it could be that as opposed to the actual substance.
link |
If you're a boring person who wears a suit and tie
link |
and works at a bank and takes psychedelics,
link |
that could be a very different effect
link |
of psychedelics on your mind.
link |
I'm just sort of raising the point
link |
that the people you referenced are already weirdos.
link |
I'm not sure exactly.
link |
No, not necessarily.
link |
A lot of the people that tell me
link |
that they use psychedelics in a useful way
link |
started out as squares and were liberating themselves
link |
because they were stuck.
link |
They were basically stuck in local optimum
link |
of their own self model, of their relationship to the world.
link |
And suddenly they had data augmentation.
link |
They basically saw and experienced a space of possibilities.
link |
They experienced what it would be like to be another person.
link |
And they took important lessons
link |
from that experience back home.
link |
Yeah, I mean, I love the metaphor of data augmentation
link |
because that's been the primary driver
link |
of self supervised learning in the computer vision domain
link |
is data augmentation.
link |
So it's funny to think of data augmentation,
link |
like chemically induced data augmentation in the human mind.
link |
There's also a very interesting effect that I noticed.
link |
I know several people who are sphere to me
link |
that LSD has cured their migraines.
link |
So severe cluster headaches or migraines
link |
that didn't respond to standard medication
link |
that disappeared after a single dose.
link |
And I don't recommend anybody doing this,
link |
especially not in the US where it's illegal.
link |
And there are no studies on this for that reason.
link |
But it seems that anecdotally
link |
that it basically can reset the serotonergic system.
link |
So it's basically pushing them
link |
outside of their normal boundaries.
link |
And as a result, it needs to find a new equilibrium.
link |
And in some people that equilibrium is better,
link |
but it also follows that in other people it might be worse.
link |
So if you have a brain that is already teetering
link |
on the boundary to psychosis,
link |
it can be permanently pushed over that boundary.
link |
Well, that's why you have to do good science,
link |
which they're starting to do on all these different
link |
substances of how well it actually works
link |
for the different conditions like MDMA seems to help
link |
with PTSD, same with psilocybin.
link |
You need to do good science,
link |
meaning large studies of large N.
link |
Yeah, so based on the existing studies of MDMA,
link |
it seems that if you look at Rick Doblin's work
link |
and what he has published about this and talks about,
link |
MDMA seems to be a psychologically relatively safe drug.
link |
But it's physiologically not very safe.
link |
That is, there is neurotoxicity
link |
if you would use a too large dose.
link |
And if you combine this with alcohol,
link |
which a lot of kids do in party settings during raves
link |
and so on, it's very hepatotoxic.
link |
So basically you can kill your liver.
link |
And this means that it's probably something that is best
link |
and most productively used in a clinical setting
link |
by people who really know what they're doing.
link |
And I suspect that's also true for the other psychedelics
link |
that is while the other psychedelics are probably not
link |
as toxic as say alcohol,
link |
the effects on the psyche can be much more profound
link |
Yeah, well, as far as I know psilocybin,
link |
so mushrooms, magic mushrooms,
link |
as far as I know in terms of the studies they're running,
link |
I think have no, like they're allowed to do
link |
what they're calling heroic doses.
link |
So that one does not have a toxicity.
link |
So they could do like huge doses in a clinical setting
link |
when they're doing study on psilocybin,
link |
which is kind of fun.
link |
Yeah, it seems that most of the psychedelics
link |
work in extremely small doses,
link |
which means that the effect on the rest of the body
link |
is relatively low.
link |
And MDMA is probably the exception.
link |
Maybe ketamine can be dangerous in larger doses
link |
because it can depress breathing and so on.
link |
But the LSD and psilocybin work in very, very small doses,
link |
at least the active part of them,
link |
of psilocybin LSD is only the active part.
link |
And the, but the effect that it can have
link |
on your mental wiring can be very dangerous, I think.
link |
Let's talk about AI a little bit.
link |
What are your thoughts about GPT3 and language models
link |
trained with self supervised learning?
link |
It came out quite a bit ago,
link |
but I wanted to get your thoughts on it.
link |
In the nineties, I was in New Zealand
link |
and I had an amazing professor, Ian Witten,
link |
who realized I was bored in class and put me in his lab.
link |
And he gave me the task to discover grammatical structure
link |
in an unknown language.
link |
And the unknown language that I picked was English
link |
because it was the easiest one
link |
to find a corpus for construct one.
link |
And he gave me the largest computer at the whole university.
link |
It had two gigabytes of RAM, which was amazing.
link |
And I wrote everything in C
link |
with some in memory compression to do statistics
link |
over the language.
link |
And I first would create a dictionary of all the words,
link |
which basically tokenizes everything and compresses things
link |
so that I don't need to store the whole word,
link |
but just a code for every word.
link |
And then I was taking this all apart in sentences
link |
and I was trying to find all the relationships
link |
between all the words in the sentences
link |
and do statistics over them.
link |
And that proved to be impossible
link |
because the complexity is just too large.
link |
So if you want to discover the relationship
link |
between an article and a noun,
link |
and there are three adjectives in between,
link |
you cannot do ngram statistics
link |
and look at all the possibilities that can exist,
link |
at least not with the resources that we had back then.
link |
So I realized I need to make some statistics
link |
over what I need to make statistics over.
link |
So I wrote something that was pretty much a hack
link |
that did this for at least first order relationships.
link |
And I came up with some kind of mutual information graph
link |
that was indeed discovering something that looks exactly
link |
like the grammatical structure of the sentence,
link |
just by trying to encode the sentence
link |
in such a way that the words would be written
link |
in the optimal order inside of the model.
link |
And what I also found is that if we would be able
link |
to increase the resolution of that
link |
and not just use this model
link |
to reproduce grammatically correct sentences,
link |
we would also be able
link |
to correct stylistically correct sentences
link |
by just having more bits in these relationships.
link |
And if we wanted to have meaning,
link |
we would have to go much higher order.
link |
And I didn't know how to make higher order models back then
link |
without spending way more years in research
link |
on how to make the statistics
link |
over what we need to make statistics over.
link |
And this thing that we cannot look at the relationships
link |
between all the bits in your input is being solved
link |
in different domains in different ways.
link |
So in computer graphics, computer vision,
link |
standard methods for many years now
link |
is convolutional neural networks.
link |
Convolutional neural networks are hierarchies of filters
link |
that exploit the fact that neighboring pixels
link |
in images are usually semantically related
link |
and distance pixels in images
link |
are usually not semantically related.
link |
So you can just by grouping the pixels
link |
that are next to each other,
link |
hierarchically together reconstruct the shape of objects.
link |
And this is an important prior
link |
that we built into these models
link |
so they can converge quickly.
link |
But this doesn't work in language
link |
for the reason that adjacent words are often
link |
but not always related and distant words
link |
are sometimes related while the words in between are not.
link |
So how can you learn the topology of language?
link |
And I think for this reason that this difficulty existed,
link |
the transformer was invented
link |
in natural language processing, not in vision.
link |
And what the transformer is doing,
link |
it's a hierarchy of layers where every layer learns
link |
what to pay attention to in the given context
link |
in the previous layer.
link |
So what to make the statistics over.
link |
And the context is significantly larger
link |
than the adjacent word.
link |
Yes, so the context that GPT3 has been using,
link |
the transformer itself is from 2017
link |
and it wasn't using that large of a context.
link |
OpenAI has basically scaled up this idea
link |
as far as they could at the time.
link |
And the context is about 2048 symbols,
link |
tokens in the language.
link |
These symbols are not characters,
link |
but they take the words and project them
link |
into a vector space where words
link |
that are statistically co occurring a lot
link |
are neighbors already.
link |
So it's already a simplification
link |
of the problem a little bit.
link |
And so every word is basically a set of coordinates
link |
in a high dimensional space.
link |
And then they use some kind of trick
link |
to also encode the order of the words in a sentence
link |
or in the not just sentence,
link |
but 2048 tokens is about a couple of pages of text
link |
or two and a half pages of text.
link |
And so they managed to do pretty exhaustive statistics
link |
over the potential relationships
link |
between two pages of text, which is tremendous.
link |
I was just using a single sentence back then.
link |
And I was only looking for first order relationships.
link |
And they were really looking
link |
for much, much higher level relationships.
link |
And what they discover after they fed this
link |
with an enormous amount of training,
link |
they are pretty much the written internet
link |
or a subset of it that had some quality,
link |
but substantial portion of the common core
link |
that they're not only able to reproduce style,
link |
but they're also able to reproduce
link |
some pretty detailed semantics,
link |
like being able to add three digit numbers
link |
and multiply two digit numbers
link |
or to translate between programming languages
link |
and things like that.
link |
So the results that GPT3 got, I think were amazing.
link |
By the way, I actually didn't check carefully.
link |
It's funny you just mentioned
link |
how you coupled semantics to the multiplication.
link |
Is it able to do some basic math on two digit numbers?
link |
Okay, interesting.
link |
I thought there's a lot of failure cases.
link |
Yeah, it basically fails if you take larger digit numbers.
link |
So four digit numbers and so on makes carrying mistakes
link |
And if you take larger numbers,
link |
you don't get useful results at all.
link |
And this could be an issue of the training set
link |
where there are not many examples
link |
of successful long form addition
link |
and standard human written text.
link |
And humans aren't very good
link |
at doing three digit numbers either.
link |
Yeah, you're not writing a lot about it.
link |
And the other thing is that the loss function
link |
that is being used is only minimizing surprise.
link |
So it's predicting what comes next in the typical text.
link |
It's not trying to go for causal closure first
link |
But the fact that that kind of prediction works
link |
to generate text that's semantically rich
link |
and consistent is interesting.
link |
So yeah, so it's amazing that it's able
link |
to generate semantically consistent text.
link |
It's not consistent.
link |
So the problem is that it loses coherence at some point,
link |
but it's also, I think, not correct to say
link |
that GPT3 is unable to deal with semantics at all
link |
because you ask it to perform certain transformations
link |
in text and it performs these transformation in text.
link |
And the kind of additions that it's able
link |
to perform are transformations in text, right?
link |
And there are proper semantics involved.
link |
You can also do more.
link |
There was a paper that was generating lots
link |
and lots of mathematically correct text
link |
and was feeding this into a transformer.
link |
And as a result, it was able to learn
link |
how to do differentiation integration in race
link |
that according to the authors, Mathematica could not.
link |
To which some of the people in Mathematica responded
link |
that they were not using Mathematica in the right way
link |
I have not really followed the resolution of this conflict.
link |
This part, as a small tangent,
link |
I really don't like in machine learning papers,
link |
which they often do anecdotal evidence.
link |
They'll find like one example
link |
in some kind of specific use of Mathematica
link |
and demonstrate, look, here's,
link |
they'll show successes and failures,
link |
but they won't have a very clear representation
link |
of how many cases this actually represents.
link |
Yes, but I think as a first paper,
link |
this is a pretty good start.
link |
And so the take home message, I think,
link |
is that the authors could get better results
link |
from this in their experiments
link |
than they could get from the vein,
link |
which they were using computer algebra systems,
link |
which means that was not nothing.
link |
And it's able to perform substantially better
link |
than GPT's V can based on a much larger amount
link |
of training data using the same underlying algorithm.
link |
Well, let me ask, again,
link |
so I'm using your tweets as if this is like Plato, right?
link |
As if this is well thought out novels that you've written.
link |
You tweeted, GPT4 is listening to us now.
link |
This is one way of asking,
link |
what are the limitations of GPT3 when it scales?
link |
So what do you think will be the capabilities
link |
of GPT4, GPT5, and so on?
link |
What are the limits of this approach?
link |
So obviously when we are writing things right now,
link |
everything that we are writing now
link |
is going to be training data
link |
for the next generation of machine learning models.
link |
So yes, of course, GPT4 is listening to us.
link |
And I think the tweet is already a little bit older
link |
and we now have Voodao
link |
and we have a number of other systems
link |
that basically are placeholders for GPT4.
link |
Don't know what open AIS plans are in this regard.
link |
I read that tweet in several ways.
link |
So one is obviously everything you put on the internet
link |
is used as training data.
link |
But in a second way I read it is in a,
link |
we talked about agency.
link |
I read it as almost like GPT4 is intelligent enough
link |
to be choosing to listen.
link |
So not only like did a programmer tell it
link |
to collect this data and use it for training,
link |
I almost saw the humorous angle,
link |
which is like it has achieved AGI kind of thing.
link |
Well, the thing is, could we be already be living in GPT5?
link |
So GPT4 is listening and GPT5 actually constructing
link |
the entirety of the reality where we...
link |
Of course, in some sense,
link |
what everybody is trying to do right now in AI
link |
is to extend the transformer to be able to deal with video.
link |
And there are very promising extensions, right?
link |
There's a work by Google that is called Perceiver
link |
and that is overcoming some of the limitations
link |
of the transformer by letting it learn the topology
link |
of the different modalities separately.
link |
And by training it to find better input features.
link |
So basically feature abstractions that are being used
link |
by this successor to GPT3 are chosen such a way
link |
that it's able to deal with video input.
link |
And there is more to be done.
link |
So one of the limitations of GPT3 is that it's amnesiac.
link |
So it forgets everything beyond the two pages
link |
that it currently reads also during generation,
link |
not just during learning.
link |
Do you think that's fixable
link |
within the space of deep learning?
link |
Can you just make a bigger, bigger, bigger input?
link |
No, I don't think that our own working memory
link |
is infinitely large.
link |
It's probably also just a few thousand bits.
link |
But what you can do is you can structure
link |
this working memory.
link |
So instead of just force feeding this thing,
link |
a certain thing that it has to focus on,
link |
and it's not allowed to focus on anything else
link |
you allow it to construct its own working memory as we do.
link |
When we are reading a book,
link |
it's not that we are focusing our attention
link |
in such a way that we can only remember the current page.
link |
We will also try to remember other pages
link |
and try to undo what we learned from them
link |
or modify what we learned from them.
link |
We might get up and take another book from the shelf.
link |
We might go out and ask somebody,
link |
we can edit our working memory in any way that is useful
link |
to put a context together that allows us
link |
to draw the right inferences and to learn the right things.
link |
So this ability to perform experiments on the world
link |
based on an attempt to become fully coherent
link |
and to achieve causal closure,
link |
to achieve a certain aesthetic of your modeling,
link |
that is something that eventually needs to be done.
link |
And at the moment we are skirting this in some sense
link |
by building systems that are larger and faster
link |
so they can use dramatically larger resources
link |
and human beings can do and much more training data
link |
to get to models that in some sense
link |
are already way superhuman
link |
and in other ways are laughingly incoherent.
link |
So do you think sort of making the systems like,
link |
what would you say, multi resolutional?
link |
So like some of the language models
link |
are focused on two pages,
link |
some are focused on two books,
link |
some are focused on two years of reading,
link |
some are focused on a lifetime,
link |
so it's like stacks of GPT3s all the way down.
link |
You want to have gaps in between them.
link |
So it's not necessarily two years, there's no gaps.
link |
It's things out of two years or out of 20 years
link |
or 2,000 years or 2 billion years
link |
where you are just selecting those bits
link |
that are predicted to be the most useful ones
link |
to understand what you're currently doing.
link |
And this prediction itself requires a very complicated model
link |
and that's the actual model that you need to be making.
link |
It's not just that you are trying to understand
link |
the relationships between things,
link |
but what you need to make relationships,
link |
discover relationships over.
link |
I wonder what that thing looks like,
link |
what the architecture for the thing
link |
that's able to have that kind of model.
link |
I think it needs more degrees of freedom
link |
than the current models have.
link |
So it starts out with the fact that you possibly
link |
don't just want to have a feed forward model,
link |
but you want it to be fully recurrent.
link |
And to make it fully recurrent,
link |
you probably need to loop it back into itself
link |
and allow it to skip connections.
link |
when you're predicting the next frame
link |
and your internal next frame in every moment,
link |
and you are able to skip connection,
link |
it means that signals can travel from the output
link |
of the network into the middle of the network
link |
faster than the inputs do.
link |
Do you think it can still be differentiable?
link |
Do you think it still can be a neural network?
link |
Sometimes it can and sometimes it cannot.
link |
So it can still be a neural network,
link |
but not a fully differentiable one.
link |
And when you want to deal with non differentiable ones,
link |
you need to have an attention system
link |
that is discreet and two dimensional
link |
and can perform grammatical operations.
link |
You need to be able to perform program synthesis.
link |
You need to be able to backtrack
link |
in this operations that you perform on this thing.
link |
And this thing needs a model of what it's currently doing.
link |
And I think this is exactly the purpose
link |
of our own consciousness.
link |
Yeah, the program things are tricky on neural networks.
link |
So let me ask you, it's not quite program synthesis,
link |
but the application of these language models
link |
to generation, to program synthesis,
link |
but generation of programs.
link |
So if you look at GitHub OpenPilot,
link |
which is based on OpenAI's codecs,
link |
I don't know if you got a chance to look at it,
link |
but it's the system that's able to generate code
link |
once you prompt it with, what is it?
link |
Like the header of a function with some comments.
link |
And it seems to do an incredibly good job
link |
or not a perfect job, which is very important,
link |
but an incredibly good job of generating functions.
link |
What do you make of that?
link |
Are you, is this exciting
link |
or is this just a party trick, a demo?
link |
Or is this revolutionary?
link |
I haven't worked with it yet.
link |
So it's difficult for me to judge it,
link |
but I would not be surprised
link |
if it turns out to be a revolutionary.
link |
And that's because the majority of programming tasks
link |
that are being done in the industry right now
link |
People are writing code that other people have written,
link |
or they're putting things together from code fragments
link |
that others have had.
link |
And a lot of the work that programmers do in practice
link |
is to figure out how to overcome the gaps
link |
in their current knowledge
link |
and the things that people have already done.
link |
How to copy and paste from Stack Overflow, that's right.
link |
And so of course we can automate that.
link |
Yeah, to make it much faster to copy and paste
link |
from Stack Overflow.
link |
Yes, but it's not just copying and pasting.
link |
It's also basically learning which parts you need to modify
link |
to make them fit together.
link |
Yeah, like literally sometimes as simple
link |
as just changing the variable names.
link |
So it fits into the rest of your code.
link |
Yes, but this requires that you understand the semantics
link |
of what you're doing to some degree.
link |
And you can automate some of those things.
link |
The thing that makes people nervous of course
link |
is that a little bit wrong in a program
link |
can have a dramatic effect on the actual final operation
link |
So that's one little error,
link |
which in the space of language doesn't really matter,
link |
but in the space of programs can matter a lot.
link |
Yes, but this is already what is happening
link |
when humans program code.
link |
So we have a technology to deal with this.
link |
Somehow it becomes scarier when you know
link |
that a program generated code
link |
that's running a nuclear power plant.
link |
It becomes scarier.
link |
You know, humans have errors too.
link |
But it's scarier when a program is doing it
link |
I mean, there's a fear that a program,
link |
like a program may not be as good as humans
link |
to know when stuff is important to not mess up.
link |
Like there's a misalignment of priorities of values
link |
Maybe that's the source of the worry.
link |
I mean, okay, if I give you code generated
link |
by GitHub open pilot and code generated by a human
link |
and say here, use one of these,
link |
which how do you select today and in the next 10 years
link |
which code do you use?
link |
Wouldn't you still be comfortable with the human?
link |
At the moment when you go to Stanford to get an MRI,
link |
they will write a bill to the insurance over $20,000.
link |
And of this, maybe half of that gets paid by the insurance
link |
and a quarter gets paid by you.
link |
And the MRI cost them $600 to make maybe probably less.
link |
And what are the values of the person
link |
that writes the software and deploys this process?
link |
It's very difficult for me to say whether I trust people.
link |
I think that what happens there is a mixture
link |
of proper Anglo Saxon Protestant values
link |
where somebody is trying to serve an abstract radar hole
link |
and organize crime.
link |
Well, that's a very harsh,
link |
I think that's a harsh view of humanity.
link |
There's a lot of bad people, whether incompetent
link |
or just malevolent in this world, yes.
link |
But it feels like the more malevolent,
link |
so the more damage you do to the world,
link |
the more resistance you have in your own human heart.
link |
Yeah, but don't explain with malevolence or stupidity
link |
what can be explained by just people
link |
acting on their incentives.
link |
Right, so what happens in Stanford
link |
is not that somebody is evil.
link |
It's just that they do what they're being paid for.
link |
No, it's not evil.
link |
That's, I tend to, no, I see that as malevolence.
link |
I see as I, even like being a good German,
link |
as I told you offline, is some,
link |
it's not absolute malevolence,
link |
but it's a small amount, it's cowardice.
link |
I mean, when you see there's something wrong with the world,
link |
it's either incompetence and you're not able to see it,
link |
or it's cowardice that you're not able to stand up,
link |
not necessarily in a big way, but in a small way.
link |
So I do think that is a bit of malevolence.
link |
I'm not sure the example you're describing
link |
is a good example of that.
link |
So the question is, what is it that you are aiming for?
link |
And if you don't believe in the future,
link |
if you, for instance, think that the dollar is going to crash,
link |
why would you try to save dollars?
link |
If you don't think that humanity will be around
link |
in a hundred years from now,
link |
because global warming will wipe out civilization,
link |
why would you need to act as if it were?
link |
Right, so the question is,
link |
is there an overarching aesthetics
link |
that is projecting you and the world into the future,
link |
which I think is the basic idea of religion,
link |
that you understand the interactions
link |
that we have with each other
link |
as some kind of civilization level agent
link |
that is projecting itself into the future.
link |
If you don't have that shared purpose,
link |
what is there to be ethical for?
link |
So I think when we talk about ethics and AI,
link |
we need to go beyond the insane bias discussions and so on,
link |
where people are just measuring the distance
link |
between a statistic to their preferred current world model.
link |
The optimism, wait, wait, wait,
link |
I was a little confused by the previous thing,
link |
There is a kind of underlying morality
link |
to having an optimism that human civilization
link |
will persist for longer than a hundred years.
link |
Like I think a lot of people believe
link |
that it's a good thing for us to keep living.
link |
This morality itself is not an end to itself.
link |
It's instrumental to people living in a hundred years
link |
from now or 500 years from now.
link |
So it's only justifiable if you actually think
link |
that it will lead to people or increase the probability
link |
of people being around in that timeframe.
link |
And a lot of people don't actually believe that,
link |
at least not actively.
link |
But believe what exactly?
link |
Most people don't believe
link |
that they can afford to act on such a model.
link |
Basically what happens in the US
link |
is I think that the healthcare system
link |
is for a lot of people no longer sustainable,
link |
which means that if they need the help
link |
of the healthcare system,
link |
they're often not able to afford it.
link |
And when they cannot help it,
link |
they are often going bankrupt.
link |
I think the leading cause of personal bankruptcy
link |
in the US is the healthcare system.
link |
And that would not be necessary.
link |
It's not because people are consuming
link |
more and more medical services
link |
and are achieving a much, much longer life as a result.
link |
That's not actually the story that is happening
link |
because you can compare it to other countries.
link |
And life expectancy in the US is currently not increasing
link |
and it's not as high as in all the other
link |
industrialized countries.
link |
So some industrialized countries are doing better
link |
with a much cheaper healthcare system.
link |
And what you can see is for instance,
link |
administrative bloat.
link |
The healthcare system has maybe to some degree
link |
deliberately set up as a job placement program
link |
to allow people to continue living
link |
in middle class existence,
link |
despite not having useful use case in productivity.
link |
So they are being paid to push paper around.
link |
And the number of administrator in the healthcare system
link |
has been increasing much faster
link |
than the number of practitioners.
link |
And this is something that you have to pay for.
link |
And also the revenues that are being generated
link |
in the healthcare system are relatively large
link |
and somebody has to pay for them.
link |
And the result why they are so large
link |
is because market mechanisms are not working.
link |
The FDA is largely not protecting people
link |
from malpractice of healthcare providers.
link |
The FDA is protecting healthcare providers
link |
So this is a thing that has to do with values.
link |
And this is not because people are malicious on all levels.
link |
It's because they are not incentivized
link |
to act on a greater whole on this idea
link |
that you treat somebody who comes to you as a patient,
link |
like you would treat a family member.
link |
Yeah, but we're trying, I mean,
link |
you're highlighting a lot of the flaws
link |
of the different institutions,
link |
the systems we're operating under,
link |
but I think there's a continued throughout history
link |
mechanism design of trying to design incentives
link |
in such a way that these systems behave
link |
better and better and better.
link |
I mean, it's a very difficult thing
link |
to operate a society of hundreds of millions of people
link |
Yes, so do we live in a society that is ever correcting?
link |
Is this, do we observe that our models
link |
of what we are doing are predictive of the future
link |
and when they are not, we improve them.
link |
Are our laws adjudicated with clauses
link |
that you put into every law,
link |
what is meant to be achieved by that law
link |
and the law will be automatically repealed
link |
if it's not achieving that, right?
link |
If you are optimizing your own laws,
link |
if you're writing your own source code,
link |
you probably make an estimate of what is this thing
link |
that's currently wrong in my life?
link |
What is it that I should change about my own policies?
link |
What is the expected outcome?
link |
And if that outcome doesn't manifest,
link |
I will change the policy back, right?
link |
Or I would change it to something different.
link |
Are we doing this on a societal level?
link |
I think it's easy to sort of highlight the,
link |
I think we're doing it in the way that,
link |
like I operate my current life.
link |
I didn't sleep much last night.
link |
You would say that Lex,
link |
the way you need to operate your life
link |
is you need to always get sleep.
link |
The fact that you didn't sleep last night
link |
is totally the wrong way to operate in your life.
link |
Like you should have gotten all your shit done in time
link |
and gotten to sleep because sleep is very important
link |
for health and you're highlighting,
link |
look, this person is not sleeping.
link |
Look, the medical, the healthcare system is operating poor.
link |
But the point is we just,
link |
it seems like this is the way,
link |
especially in the capitalist society, we operate.
link |
We keep running into trouble and last minute,
link |
we try to get our way out through innovation
link |
and it seems to work.
link |
You have a lot of people that ultimately are trying
link |
to build a better world and get urgency about them
link |
when the problem becomes more and more imminent.
link |
And that's the way this operates.
link |
But if you look at the long arc of history,
link |
it seems like that operating on deadlines
link |
produces progress and builds better and better systems.
link |
You probably agree with me that the US
link |
should have engaged in mask production in January 2020
link |
and that we should have shut down the airports early on
link |
and that we should have made it mandatory
link |
that the people that work in nursing homes
link |
are living on campus rather than living at home
link |
and then coming in and infecting people in the nursing homes
link |
that had no immune response to COVID.
link |
And that is something that was, I think, visible back then.
link |
The correct decisions haven't been made.
link |
We would have the same situation again.
link |
How do we know that these wrong decisions
link |
are not being made again?
link |
Have the people that made the decisions
link |
to not protect the nursing homes been punished?
link |
Have the people that made the wrong decisions
link |
with respect to testing that prevented the development
link |
of testing by startup companies and the importing
link |
of tests from countries that already had them,
link |
have these people been held responsible?
link |
First of all, so what do you wanna put
link |
before the firing squad?
link |
I think they are being held responsible.
link |
No, just make sure that this doesn't happen again.
link |
No, but it's not that, yes, they're being held responsible
link |
by many voices, by people being frustrated.
link |
There's new leaders being born now
link |
that we're going to see rise to the top in 10 years.
link |
This moves slower than, there's obviously
link |
a lot of older incompetence and bureaucracy
link |
and these systems move slowly.
link |
They move like science, one death at a time.
link |
So yes, I think the pain that's been felt
link |
in the previous year is reverberating throughout the world.
link |
Maybe I'm getting old, I suspect that every generation
link |
in the US after the war has lost the plot even more.
link |
I don't see this development.
link |
The war, World War II?
link |
Yes, so basically there was a time when we were modernist
link |
and in this modernist time, the US felt actively threatened
link |
by the things that happened in the world.
link |
The US was worried about possibility of failure
link |
and this imminence of possible failure led to decisions.
link |
There was a time when the government would listen
link |
to physicists about how to do things
link |
and the physicists were actually concerned
link |
about what the government should be doing.
link |
So they would be writing letters to the government
link |
and so for instance, the decision for the Manhattan Project
link |
was something that was driven in a conversation
link |
between physicists and the government.
link |
I don't think such a discussion would take place today.
link |
I disagree, I think if the virus was much deadlier,
link |
we would see a very different response.
link |
I think the virus was not sufficiently deadly
link |
and instead because it wasn't very deadly,
link |
what happened is the current system
link |
started to politicize it.
link |
The mask, this is what I realized with masks early on,
link |
they were not, very quickly became not as a solution
link |
but they became a thing that politicians used
link |
to divide the country.
link |
So the same things happened with vaccines, same thing.
link |
So like nobody's really,
link |
people weren't talking about solutions to this problem
link |
because I don't think the problem was bad enough.
link |
When you talk about the war,
link |
I think our lives are too comfortable.
link |
I think in the developed world, things are too good
link |
and we have not faced severe dangers.
link |
When the danger, the severe dangers,
link |
existential threats are faced, that's when we step up
link |
on a small scale and a large scale.
link |
Now, I don't, that's sort of my argument here
link |
but I did think the virus is, I was hoping
link |
that it was actually sufficiently dangerous
link |
for us to step up because especially in the early days,
link |
it was unclear, it still is unclear because of mutations,
link |
how bad it might be, right?
link |
And so I thought we would step up and even,
link |
so the masks point is a tricky one because to me,
link |
the manufacture of masks isn't even the problem.
link |
I'm still to this day and I was involved
link |
with a bunch of this work, have not seen good science done
link |
on whether masks work or not.
link |
Like there still has not been a large scale study.
link |
To me, that should be, there should be large scale studies
link |
and every possible solution, like aggressive
link |
in the same way that the vaccine development
link |
There should be masks, which tests,
link |
what kind of tests work really well, what kind of,
link |
like even the question of how the virus spreads.
link |
There should be aggressive studies on that to understand.
link |
I'm still, as far as I know, there's still a lot
link |
of uncertainty about that.
link |
Nobody wants to see this as an engineering problem
link |
that needs to be solved.
link |
It's that I was surprised about, but I wouldn't.
link |
So I find that our views are largely convergent
link |
but not completely.
link |
So I agree with the thing that because our society
link |
in some sense perceives itself as too big to fail.
link |
The virus did not alert people to the fact
link |
that we are facing possible failure
link |
that basically put us into the postmodernist mode.
link |
And I don't mean in a philosophical sense
link |
but in a societal sense.
link |
The difference between the postmodern society
link |
and the modern society is that the modernist society
link |
has to deal with the ground truth
link |
and the postmodernist society has to deal with appearances.
link |
Politics becomes a performance
link |
and the performance is done for an audience
link |
and the organized audience is the media.
link |
And the media evaluates itself via other media, right?
link |
So you have an audience of critics that evaluate themselves.
link |
And I don't think it's so much the failure
link |
of the politicians because to get in power
link |
and to stay in power, you need to be able
link |
to deal with the published opinion.
link |
Well, I think it goes in cycles
link |
because what's going to happen is all
link |
of the small business owners, all the people
link |
who truly are suffering and will suffer more
link |
because the effects of the closure of the economy
link |
and the lack of solutions to the virus,
link |
they're going to apprise.
link |
And hopefully, I mean, this is where charismatic leaders
link |
can get the world in trouble
link |
but hopefully will elect great leaders
link |
that will break through this postmodernist idea
link |
of the media and the perception
link |
and the drama on Twitter and all that kind of stuff.
link |
But you know, this can go either way.
link |
When the Weimar Republic was unable to deal
link |
with the economic crisis that Germany was facing,
link |
there was an option to go back.
link |
But there were people which thought,
link |
let's get back to a constitutional monarchy
link |
and let's get this to work because democracy doesn't work.
link |
And eventually, there was no way back.
link |
People decided there was no way back.
link |
They needed to go forward.
link |
And the only options for going forward
link |
was to become Stalinist communist,
link |
basically an option to completely expropriate
link |
the factories and so on and nationalize them
link |
and to reorganize Germany in communist terms
link |
and ally itself with Stalin and fascism.
link |
And both options were obviously very bad.
link |
And the one that the Germans picked
link |
led to a catastrophe that devastated Europe.
link |
And I'm not sure if the US has an immune response
link |
I think that the far right is currently very weak in the US,
link |
but this can easily change.
link |
Do you think from a historical perspective,
link |
Hitler could have been stopped
link |
from within Germany or from outside?
link |
Or this, well, depends on who you wanna focus,
link |
whether you wanna focus on Stalin or Hitler,
link |
but it feels like Hitler was the one
link |
as a political movement that could have been stopped.
link |
I think that the point was that a lot of people
link |
wanted Hitler, so he got support from a lot of quarters.
link |
There was a number of industrialists who supported him
link |
because they thought that the democracy
link |
is obviously not working and unstable
link |
and you need a strong man.
link |
And he was willing to play that part.
link |
There were also people in the US who thought
link |
that Hitler would stop Stalin
link |
and would act as a bulwark against Bolshevism,
link |
which he probably would have done, right?
link |
But at which cost?
link |
And then many of the things that he was going to do,
link |
like the Holocaust, was something where people thought
link |
this is rhetoric, he's not actually going to do this.
link |
Especially many of the Jews themselves, which were humanists.
link |
And for them, this was outside of the scope
link |
that was thinkable.
link |
I mean, I wonder if Hitler is uniquely,
link |
I wanna carefully use this term, but uniquely evil.
link |
So if Hitler was never born,
link |
if somebody else would come in this place.
link |
So like, just thinking about the progress of history,
link |
how important are those singular figures
link |
that lead to mass destruction and cruelty?
link |
Because my sense is Hitler was unique.
link |
It wasn't just about the environment
link |
and the context that gave him,
link |
like another person would not come in his place
link |
to do as destructive of the things that he did.
link |
There was a combination of charisma, of madness,
link |
of psychopathy, of just ego, all those things,
link |
which are very unlikely to come together
link |
in one person in the right time.
link |
It also depends on the context of the country
link |
that you're operating in.
link |
If you tell the Germans that they have a historical destiny
link |
in this romantic country,
link |
the effect is probably different
link |
than it is in other countries.
link |
But Stalin has killed a few more people than Hitler did.
link |
And if you look at the probability
link |
that you survived under Stalin,
link |
Hitler killed people if he thought
link |
they were not worth living,
link |
or if they were harmful to his racist project.
link |
He basically felt that the Jews would be too cosmopolitan
link |
and would not be willing to participate
link |
in the racist redefinition of society
link |
and the value of society,
link |
and there is no state in this way
link |
that he wanted to have it.
link |
So he saw them as harmful danger,
link |
especially since they played such an important role
link |
in the economy and culture of Germany.
link |
And so basically he had some radical
link |
but rational reason to murder them.
link |
And Stalin just killed everyone.
link |
Basically the Stalinist purges were such a random thing
link |
where he said that there's a certain possibility
link |
that this particular part of the population
link |
has a number of German collaborators or something,
link |
and we just kill them all, right?
link |
Or if you look at what Mao did,
link |
the number of people that were killed
link |
in absolute numbers were much higher under Mao
link |
than they were under Stalin.
link |
So it's super hard to say.
link |
The other thing is that you look at Genghis Khan and so on,
link |
how many people he killed.
link |
When you see there are a number of things
link |
that happen in human history
link |
that actually really put a substantial dent
link |
in the existing population, or Napoleon.
link |
And it's very difficult to eventually measure it
link |
because what's happening is basically evolution
link |
on a human scale where one monkey figures out
link |
a way to become viral and is using this viral technology
link |
to change the patterns of society
link |
at the very, very large scale.
link |
And what we find so abhorrent about these changes
link |
is the complexity that is being destroyed by this.
link |
That's basically like a big fire that burns out
link |
a lot of the existing culture and structure
link |
that existed before.
link |
Yeah, and it all just starts with one monkey.
link |
One charismatic ape.
link |
And there's a bunch of them throughout history.
link |
Yeah, but it's in a given environment.
link |
It's basically similar to wildfires in California, right?
link |
The temperature is rising.
link |
There is less rain falling.
link |
And then suddenly a single spark can have an effect
link |
that in other times would be contained.
link |
Okay, speaking of which, I love how we went
link |
to Hitler and Stalin from 20, 30 minutes ago,
link |
GPT3 generating, doing programs that this is.
link |
The argument was about morality of AI versus human.
link |
And specifically in the context of writing programs,
link |
specifically in the context of programs
link |
that can be destructive.
link |
So running nuclear power plants
link |
or autonomous weapons systems, for example.
link |
And I think your inclination was to say that
link |
it's not so obvious that AI would be less moral than humans
link |
or less effective at making a world
link |
that would make humans happy.
link |
So I'm not talking about self directed systems
link |
that are making their own goals at a global scale.
link |
If you just talk about the deployment
link |
of technological systems that are able to see order
link |
and patterns and use this as control models
link |
to act on the goals that we give them,
link |
then if we have the correct incentives
link |
to set the correct incentives for these systems,
link |
I'm quite optimistic.
link |
So humans versus AI, let me give you an example.
link |
Autonomous weapon system.
link |
Let's say there's a city somewhere in the Middle East
link |
that has a number of terrorists.
link |
And the question is,
link |
what's currently done with drone technologies,
link |
you have information about the location
link |
of a particular terrorist and you have a targeted attack,
link |
you have a bombing of that particular building.
link |
And that's all directed by humans