Kind: captions
Language: en
it turns out that if you train a
planarian and then cut their heads off
the tail will regenerate a brand new
brain that still remembers the original
information I think planaria hold the
answer to pretty much every deep
question of life for one thing they're
similar to our ancestors so they have
true symmetry they have a true brain
they're not like earthworms they're you
know they're much more advanced life
form they have lots of different
internal organs but they're these little
um they're about you know maybe two
centimeters in in the centimeter to two
in size
I have a head in the tail and the first
thing is plenary are Immortal so they do
not age there's no such thing as an old
planarian so that right there tells you
that these theories of thermodynamic
limitations of on lifespan are wrong
it's not it's not that well over time of
everything degrades no planaria can keep
it going for uh probably you know how
long have they've been around 400
million years right so these are the
actual so the planaria in our lab are
actually in physical continuity with
planaria that we're here 400 million
years ago
the following is a conversation with
Michael Levin one of the most
fascinating and Brilliant biologists
I've ever talked to he and his lab at
Tufts University works on novel ways to
understand and control complex pattern
formation in biological systems Andre
carpathi a world-class AI researcher is
the person who first introduced me to
Michael Levin's work
I bring this up because these two people
make me realize that biology has a lot
to teach us about Ai and AI might have a
lot to teach us about biology
this is Alex Friedman podcast to support
it please check out our sponsors in the
description and now dear friends here's
Michael Levin
embryogenesis is the process of building
the human body from a single cell I
think it's one of the most incredible
things that exists on Earth from a
single embryo so how does this process
work yeah it is it is an incredible
process uh I think it's maybe the most
magical process there is and I think one
of the most fundamentally interesting
things about it is that it shows that
each of us takes the journey from
so-called just physics to mind right
because we all Start Life as a single
quiescent unfertilized oocyte and it's
basically a bag of chemicals and you
look at that and you say okay this is
chemistry and physics and then nine
months and some years later you have an
organism with high level cognition and
preferences and an inner life and so on
and what embryogenesis tells us is that
that transformation from physics to mind
is gradual it's smooth there is no
special place where you know a lightning
bolt says boom now you've gone from from
physics to True cognition that doesn't
happen and so we can see in this process
that the whole mystery you know the
biggest mystery of the you of the
universe basically how you get mind from
matter from just physics in quotes yeah
so where's the magic into the thing how
do we get from information encoded in
DNA
and make physical reality out of that
information so one of the things that I
think is really important if we're going
to bring in DNA into this picture is to
think about the fact that what DNA
encodes is the hardware of Life DNA
contains the instructions for the kind
of micro level Hardware that every cell
gets to play with so all the proteins
all the signaling factors the ion
channels all the cool little pieces of
Hardware that cells have that's what's
in the DNA the rest of it is in
so-called generic laws and these are
laws of mathematics these are laws of
computation these are laws of um of
physics of all all kinds of interesting
things that are not directly in the DNA
and that that process you know I think I
think the reason the reason I always put
just physics in quotes is because I
don't think there is such a thing as
just physics I think that thinking about
these things in binary categories like
this is physics this is true cognition
this is as if it's only faking other
these kinds of things I think that's
what gets us in trouble I think that we
really have to understand that it's a
Continuum and we have to work up the
scaling the laws of scaling and we can
and we can certainly talk about that
there's a lot of really interesting
thoughts to be had there so the physics
is deeply integrated with the
information so the DNA doesn't exist on
its own the DNA is a integrated as in
some sense in response to the the laws
of physics at every scale the laws
of the environment it exists in yeah the
environment and also the laws of the
Universe I mean the thing about the
thing about the the DNA is that it's um
once Evolution discovers a certain kind
of machine that if if the physical
implementation is appropriate it's sort
of uh and this is hard to talk about
because we don't have a good vocabulary
for this yet but it's a very kind of a
platonic notion that if the machine is
there it pulls down interesting uh
interesting things that you do not have
to evolve from scratch because the laws
of physics give it to you for free so
just as a really stupid example if
you're trying to evolve a particular
triangle you can evolve the first angle
and you evolve the second angle but you
don't need to evolve the third you know
what it is already now why do you know
that's that's a gift for free from
geometry in a particular space you know
what that angle has to be and if you
evolve an ion Channel which is Ion
channels are basically transistors right
they're voltage-gated current
conductances if you evolve that ION
channel you immediately get to use
things like truth tables you get logic
functions you don't have to evolve the
logic function you don't have to evolve
a truth table doesn't have to be in the
DNA it's you get it for free right and
the fact that if you have nand Gates you
can build anything you want you get that
for free all you have to evolve is that
that first step a first little machine
that that enables you to couple to those
laws and there's laws of adhesion and
and many other things and this is all um
that interplay between the the hardware
that's set up by the genetics and the
software that's paid right the
physiological software that basically
does all the computation and the
cognition and everything else is a real
interplay between the information and
the DNA and the laws of physics of
computation and so on so is it fair to
say just like this idea that the laws of
mathematics are discovered
they're Laden within the fabric of the
universe in that same way the laws of
biology are kind of discovered yeah I
think that's absolutely and it's
probably not a popular view but I think
that's right on the money yeah well I
think that's a really deep idea
then embryogenesis is the process of
revealing
of um
embodying of manifesting these laws
you're not building the laws you're just
creating the capacity to reveal yes I
think again not the standard view of
molecular biology by any means but I
think that's right on the money I'll
give you a simple example you know some
of our latest work with these xenobots
right so what we've done is to take some
skin cells off of an early frog embryo
and basically ask about their plasticity
if we give you a chance to sort of
reboot your multicellularity in a
different context what would you do
because what you might assume by the
thing about embryogenesis is that it's
super reliable right it's very robust
and that really obscures some of its
most interesting features we get used to
it we get used to the fact that acorns
make oak trees and frog eggs make frogs
and we say well what else is it going to
make that's what it you know that's what
it makes that's a standard story but the
reality is and so and so you look at
these um these skin cells you say well
what do they know how to do well they
know how to be a passive boring
two-dimensional outer layer keeping the
bacteria from getting into the embryo
that's what they know how to do well it
turns out that if you take these skin
cells and you remove the rest of the
embryo so you remove all of the rest of
the cells and you say well you're by
yourself now what do you want to do so
what they do is they form this little um
this multi-little creature that runs
around the dish they have all kinds of
incredible capacities they navigate
through mazes they have various
behaviors that they do both
independently and and together they uh
they have a uh but they basically they
Implement Von Neumann's dream of
self-replication because if you sprinkle
a bunch of loose cells into the dish
what they do is they run around they
collect those cells into little piles
they they sort of mush them together
until those little piles become the next
generation of xenobots so you've got
this machine that builds copies of
itself from loose material in its
environment none of this are
things that you would have expected from
the Frog genome in fact there's wild
type the genome was wild type there's
nothing wrong with their genetics
nothing has been added no Nano materials
no genomic editing nothing and so what
we have done there is engineer by
subtraction which you've done is you've
removed the other cells that normally
basically bully these cells into being
skin cells and you find out that what
they really want to do is is to be this
they want their default behaviors to be
a xenobot but in Vivo in the embryo they
get told to be skinned by these other
cell types and so so now so now here
comes this this really interesting
question that you just posed
when you ask where does the form of the
tadpole and the Frog come from the
standard answer as well it's it's it's a
selection so over over millions of years
right it's been shaped to to produce the
specific body with that's fit for froggy
environments where does the shape of the
xenobot come from there's never been any
zenobots there's never been selection to
be a good xenobot these cells find
themselves in the new environment in 48
hours they figure out how to be an
entirely different uh proto-organism
with new capacities like kinematic
self-replication that's not how frogs or
tadpoles replicate we've made it
impossible for them to replicate their
normal way within a couple days these
guys find a new way of doing it that's
not done anywhere else in the biosphere
well actually let's step back and Define
what are xenobots
so a xenobod is uh self-assembling
little proto-organism it's also a
biological robot those things are not um
distinct it's a member of both classes
how much is a biology how much is it
robot
at this point most of it is biology
because what we're doing is we're
discovering natural uh behaviors of
these uh of these of the cells and also
of the cell collectives now one of the
really important parts of this was that
we're working together with Josh
bongard's group at University of Vermont
their computer scientists do Ai and
they've basically been able to use an
evolutionary a simulated Evolution
approach to ask how can we manipulate
these cells give them signals not rewire
their DNA so not Hardware but experience
as signals so can we remove some cells
can we add some cells can we poke them
in different ways to get them to do
other things so in the future there's
going to be you know we're now and this
is this is future on published work but
we're doing all sorts of interesting
ways to reprogram them to new behaviors
but before you can start to reprogram
these things you have to understand what
their innate capacities are okay so that
means
engineering programming you're
engineering them in in the future and in
some sense the the definition of a robot
is something you impart engineer yeah
and first versus evolve I mean um
it's such a fuzzy definition anyway in
some sense many of the organisms within
our body are kinds of robots yes yes and
I think robots is a weird line because
it's we tend to see robots as the other
I think there will be a time in the
future when there's going to be
something akin to the Civil Rights
movements for robots but we'll talk
about that later perhaps sure anyway
um
so how do you can we just Linger on it
how do you build a zenobot what are we
talking about here
from from whence does it start and how
does it become
the Glorious zenobot yeah so just to
take one step back one of the things
that a lot of people uh get stuck on is
they say well uh you know engineering
requires new DNA circuits or it requires
new nanomaterials you know what the
thing is we are now moving from Old
School engineering which use passive
materials right that things like you
know wood metal things like this that
basically the only thing you could
depend on is that they were going to
keep their shape that's it they don't do
anything else you it's on you as an
engineer to make them do everything
they're going to do and then there were
active materials and now computationals
this is a whole new era these are
agential materials this is your you're
now collaborating with your substrate
because your material has an agenda
these cells have you know billions of
years of evolution they have goals they
have preferences they're not just going
to sit where you put them that's
hilarious that you have to talk your
material and to keep your kitchen that's
it that is exactly right that is exactly
right stay there it's like getting a
bunch of cats or something and trying to
organize the shape out of them it's
funny we're on the same page here
because in a paper this is this is
currently um just been accepted in
nature by engineering one of the figures
I have is building a tower out of Legos
versus dogs right yeah so think about
the difference right if you build out of
Legos you have full control over where
it's going to go but if somebody knocks
it over it's game over with the dogs you
cannot just come and stack them they're
not going to stay that way but the good
news is that if you train them then
somebody knocks it over they'll get
right back up so it's all right so as an
engineer what you really want to know is
what can they depend on this thing to do
right that's really you know a lot of
people have definitions of robots as far
as what they're made of or how they got
here you know design versus evolve
whatever I don't think any of that is
useful I think I think as an engineer
what you want to know is how much can I
depend on this thing to do when I'm not
around to micromanage it what level of
uh what level of dependency can I can I
give this thing how much agency does it
have which then tells you what
techniques do you use so do you use
micromanagement like you put everything
where it goes do you train it do you
give it signals do you try to convince
it to do things right how much you know
how intelligent is your substrate and so
now we're moving into this into this
area where you're you're working with
agential materials that's a
collaboration that's not that's not old
old style what's the word you're using a
gentle a gentle what's that mean agency
if it comes from the word agency so so
basically the material has agency
meaning that it has some some level of
obviously not human level but some level
of uh preferences goals memories ability
to remember things to compute into the
future meaning anticipate uh you know
when you're working with cells they have
all of that to some to various degrees
is that empowering or limiting having
material as a mind of its own literally
I think it's both right so it raises
difficulties because it means that it if
you if you're using the old mindset
which is a linear
um kind of extrapolation of what's going
to happen you're going to be surprised
and shocked all the time because biology
uh does not do what we linearly expect
materials to do on the other hand it's
massively liberating and so in the
following way I've argued that advances
in regenerative medicine require us to
take advantage of this because what it
means is that you can get the material
to do things that you don't know how to
micromanage so just as a simple example
right if you if you you had a rat and uh
you wanted this rat to do a circus trick
put a ball in the little hoop you can do
it the micromanagement way which is try
to control every neuron and try to play
the thing like a puppet right and maybe
someday that'll be possible maybe or you
can train the rat and this is why
Humanity for thousands of years before
we knew any Neuroscience we had no idea
what's behind what's between the ears of
any animal we were able to train these
animals because once you recognize the
level of agency of a certain system you
can use appropriate techniques if you
know the currency of motivation reward
and Punishment you know how smart it is
you know what kinds of things it likes
to do you are searching a much more much
smoother much nicer problem space than
if you try to micromanage the thing and
then regenerative medicine when you're
trying to get let's say an arm to grow
back or an eye to repair so birth defect
or something do you really want to be
controlling tens of thousands of genes
at each point to try to micromanage it
or do you want to find the high level
modular control roles let's say build an
arm here you already know how to build
an arm you did it before do it again so
that's I I think it's it's both it's
both the difficult and it challenges us
to develop new ways of engineering and
it's it's hugely empowering okay so how
do you do I mean maybe sticking with the
metaphor of dogs and cats
I presume you have to figure out the
find the dogs and uh dispose of the cats
um because you know it's like the old
herding cats is an issue so you may be
able to train dogs
I suspect you will not be able to train
cats
or if you do you're never going to be
able to trust them so is there a way to
figure out which material is amenable
to hurting is it in the lab work or is
it in simulation right now it's largely
in the lab because we are our
simulations do not capture yet the most
uh interesting and Powerful things about
biology so the simulation what what
we're pretty good at simulating are feed
forward emergent types of things right
so cellular automata if you have simple
rules and you sort of roll those forward
for every every agent or every cell in
the simulation then complex things
happen you know ant colony or algorithms
things like that we're we're good at
that and that's and that's fine the
difficulty with all of that is that it's
incredibly hard to reverse so this is a
really hard inverse problem right if you
look at a bunch of termites and they
make a you know a thing with a single
chimney and you say well I like it but
I'd like two chimneys how do you change
the rules of behavior for each termite
so they make two chimneys right or or if
you say hear a bunch of cells that are
creating this kind of organism I I don't
think that's optimal I'd like to to
repair that birth defect how do you
control all the all the individual
low-level rules right all the protein
interactions and everything else rolling
it back from the anatomy that you want
to the low-level Hardware rules is in
general intractable it's a it's an
inverse problem that's generally not
soluble so
um right now it's mostly in the lab
because what we need to do is we need to
understand how biology uses top-down
controls so the idea is not not
bottom-up emergence but the idea of
things like gold directed uh test
operate exit kinds of Loops where where
it's basically an error minimization
function over a new space it's not a
space of gene expression but for example
a space of anatomy so just as a simple
example if you have you have a
salamander it's got an arm you can you
can amputate that arm anywhere along the
length it will grow exactly what's
needed and then it stops that's the most
amazing thing about regeneration is that
it stops it knows when to stop when does
it stop it stops when a correct
salamander arm has been completed so
that tells you that's the right that's a
that's a uh a mean Zen's kind of
analysis where it has to know what the
correct limb is supposed to look like
right so it has a way to ascertain the
current shape it has a way to measure
that Delta from from what shape it's
supposed to be and then we'll keep
taking actions meaning Remodeling and
growing and everything else until that's
complete so once you know that and we've
taken advantage of this in the lab to do
some some really wild things with with
both planaria and frog embryos and so on
once you know that
um you can start playing with that uh
with that homeostatic cycle you can ask
for example well how does it remember
what the correct shape is and can we
mess with that memory can we give it a
false memory of what the shape should be
and let the cells build something else
or can we mess with the measurement
apparatus right so it gives you it gives
you those kinds of so so the idea is to
basically appropriate a lot of the um
approaches and Concepts from cognitive
neuroscience and Behavioral Science into
things that uh previously were taken to
be dumb materials and you know you get
yelled at in class if you if you for
being anthropomorphic if you said well
my cells want to do this and my cells
want to do that and I think I think
that's a that's a major mistake that
leaves a ton of capabilities on the
table so thinking about biologic systems
is things that have memory have almost
something like cognitive ability
but
I mean
how incredible is it you know that the
salamander arm is being
rebuilt not with a dictator
it's kind of like the cellular automata
system all the individual workers are
doing their own thing
so where's that oh wait top down signal
that doesn't control coming from like
how can you find it yeah like why does
it stop growing how does it know the
shape how does it have memory of the
shape and how does it tell everybody to
be like whoa whoa slow down we're done
so the first thing to to think about I
think is that there are no examples
anywhere of of a central dictator
because in in this in this kind of
science because everything is made of
parts and so we we even though we we
feel as a unified Central sort of
intelligence and kind of point of of
cognition we are a bag of neurons right
we all intelligence is collective
intelligence there's this this is
important to kind of um and think about
because a lot of people think okay
there's real intelligence like me and
then there's collective intelligence
which is the ants and flocks of birds
and you know termites and things like
that and and you know and and maybe it's
appropriate to think of them as a as a
as an individual and maybe it's not a
lot of people are skeptical about about
that and so on but you've got to realize
that we are not there's no such thing as
this like indivisible Diamond of
intelligence that's like this one
Central thing that's not made of parts
we are all made of parts and so if if
you believe that which I think is is
hard to uh to get around that that we in
fact have a centralized um set of goals
and preferences and we plan and we do
things and so on you are already
committed to the fact that a collection
of cells is able to do this because we
are a collection of cells there's no
getting around that in our case what we
do is we navigate the three-dimensional
world and we have Behavior this is
blowing my mind right now because we are
just a collection of stuff oh yeah yeah
so when I'm moving this arm
I feel like I'm the central dictator of
that action but there's a lot of stuff
going on like every all all the cells
here collaborating in some interesting
way they're getting signal from the
central nervous system well even the
central nervous system is is
misleadingly named because it isn't
really Central again it's it's what it's
just a bunch of cells I mean all of the
right there are no you there are no
singular indivisible intelligences
anywhere we are all every every example
that we've ever seen is is a collective
of some of something it's just that
we're used to it we're used to that you
know we're used to okay this thing is
kind of a single thing but it's really
not you zoom in you know what you see
you see a bunch of cells running around
and so is there some unifying I mean
we're just jumping around but that
something that you look as the the
biological signal versus the biochemical
the
um the chemistry the electricity
maybe the life isn't that
versus the cells
it's the uh there's there's an orchestra
playing and uh the resulting music is
the dictator that's not bad um Dennis
that's Dennis Nobles uh kind of view of
things he has two really good books
where he talks about this musical
analogy right so so I think that's
that's I like it um I like it is it
wrong though I don't think it's no I
don't think it's wrong
um I don't I don't think it's wrong I
think I think the important thing about
it is that we have to come to grips with
the fact that a true a a true proper uh
cognitive intelligence can still be made
of Parts those things are and in fact it
has to be and I I think it's a real
shame but I see this all the time when
you have uh when you have a collective
like this whether it be uh a group of
robots or a you know a collection of
cells or neurons or whatever as soon as
as soon as we gain some insight into how
it works right meaning that oh I see in
order to take this action here's the
information that got processed via this
camera mechanism or whatever immediately
people say oh well then that's not real
cognition that's just physics I think
this is this is fundamentally flawed
because if you zoom into anything what
are you going to see of course you're
just going to see physics what else
could be underneath right that's not
going to be fairy dust it's going to be
physics and chemistry but that doesn't
take away from the magic of the fact
that there are certain ways to arrange
that physics and chemistry and in
particular the bioelectricity which I
like a lot uh to give you an emergent uh
Collective with goals and preferences
and memories and anticipations that do
not belong to any of the subunits so I
think what we're getting into here and
we can talk about how how this happens
during embryogenesis and so on what
we're getting into is the origin of the
of a self yeah with a big with a capital
S so we ourselves there are many other
kinds of selves and we can tell some
really interesting stories about where
selves come from and how they become
unified yeah is this the first
or at least humans tend to think that
this is the the level at which the self
with the capital s is first born
but uh and we really don't want to see
um human civilization or Earth itself as
one living organism yeah that's very
uncomfortable to us it is yeah but is um
yeah where's the self born we have to
grow up past that so what I like to do
is uh I'll tell you two quick stories
about that I like to roll backwards so
so as opposed to so if you start and you
say okay here's a paramecium and you see
it um you know it's a single cell
organism you see it doing various things
and people will say okay I'm sure
there's some chemical story to be told
about how it's doing it so that's not
true cognition right and people will
argue about that I I like to work it
backwards I said let's let's agree that
you and I as as we sit here are examples
of true cognition if anything is if
there's anything that's true cognition
we are we are examples of it now let's
just roll back slowly right so you roll
back to the time when you're a small
child and used to doing whatever and
then just sort of day by day you roll
you roll back and eventually you become
more or less that paramecium and then
and then you sort of even below that
right as an unfertilized Osa so it's no
one has to my knowledge no one has come
up with any convincing discreet Step At
which my cognitive Powers disappear
right it just doesn't the biology
doesn't offer any specific step it's com
it's incredibly smooth and slow and
continuous and so I think this idea that
it just sort of magically shows up uh at
one point and then and then uh you know
humans have true selves that don't exist
elsewhere I think it runs against
everything we know about Evolution
everything we know about developmental
biology these are all slow continua and
the other really important story I want
to tell is where embryos come from so
think about this for a second amniot
embryo so this is humans birds and so on
mammals and birds and so on imagine a
flat disc of cells so there's maybe 50
000 cells and in that so when you get an
egg from a from a fertilizer let's say
you buy a fertilized egg from a farm
right that that egg uh will will have
about 50 000 cells in um in a flat disc
it looks like a little little tiny
little Frisbee and in that flat disc
what will happen is there'll be uh one
one set of cells will uh becomes will
become special and it will tell all the
other cells I'm I'm going to be the head
you guys don't be the head and so it'll
amplify symmetry breaking amplification
you get one embryo there's a there's a
you know there's some neural tissue and
some other stuff forms now now you say
okay I had one egg and one embryo and
then there you go what else could it be
well the reality is and I used to I I
did all of this as a grad student if you
um if you take a little needle and you
make a scratch in that blasted room in
that in that disc such that the cells
can't talk to each other for a while it
heals up but for a while they can't talk
to each other what will happen is that
uh both regions will decide that they
can be the embryo and there will be two
of them and then when they heal up they
become conjoint Twins and you can make
two you can make three you can make lots
so the question of how many selves are
in there cannot be answered until it's
actually played all the way through it
isn't necessarily that there's just one
there can be many so what you have is
you have this medium this this
undifferentiated I'm sure there's a
there's a psychological
um version of this somewhere that I
don't know the proper terminology but
you have this you have this list like
put ocean of potentiality you have these
thousands of cells and some number of
individuals are going to be formed out
of it usually one sometimes zero
sometimes several and they they form out
of these cells because a region of these
cells organizes into a collective that
will have goals goals that individual
cells don't have for example make a limb
make an eye how many eyes well exactly
two so individual cells don't know what
an eye is they don't know how many eyes
you're supposed to have but the
collective does the collective has goals
and memories and anticipations that the
individual cells don't and that that the
establishment of that boundary with its
own ability to maintain to to pursue
certain goals that's the origin of of
selfhood
but I is that goal
in there somewhere but they always
destined like are they discovering that
goal like where the hell did Evolution
um discover this when you went from the
prokaryotes to you you carry out excels
and then they started making groups and
when you make a certain group you make a
you you make it sound
that's such a tricky thing to try to
understand you make it sound like this
cells didn't get together and came up
with a goal but the very Act of them
getting together
revealed the goal that was always there
there was always that potential for that
goal so the first thing to say is that
there are way more questions here than
than certainties okay so everything I'm
telling you is is Cutting Edge
developing you know stuff so so it's not
as if any of us know the answer to this
but but here's here's here's my opinion
on this I think what evolution I I don't
think that Evolution produces solutions
to specific problems in other words
specific environments like here's a frog
that can live well in a froggy
environment I think what evolution
produces is problem-solving machines and
that that will that will solve problems
in different spaces so not just
three-dimensional space this goes back
to what we were talking about before we
the the brain is a evolutionarily a late
development it's a system that is able
to to pursue goals in three-dimensional
Space by giving commands to muscles
where did that system come from that
system evolved from a much more ancient
evolutionarily much more ancient system
where collections of cells gave
instructions to for cell behaviors
meaning cells move to to divide to to
die to change into different cell types
to navigate morphe space the space of
anatomies the space of all possible
anatomies and before that cells were
navigating transcriptional space which
is a space of all possible Gene
expressions and before that metabolic
space so what evolution has done I think
is is is produced Hardware that is very
good at navigating different spaces
using a bag of tricks right which which
I'm sure many of them we can steal for
autonomous vehicles and Robotics and
various things and what happens is that
they navigate these spaces without a
whole lot of commitment to what the
space is in fact they don't know what
the space is right we are all brains in
a vat so to speak every cell does not
know right every cell is some other some
other cells external environment right
so where does the with that border
between you you and the outside world
you don't really know where that is
right every every collection of cell has
to figure that out from scratch
and the fact that Evolution requires all
of these things to figure out what they
are what effectors they have what
sensors they have where does it make
sense to draw a boundary between me and
the outside world the fact that you have
to build all that from scratch this
autopoiesis is what defines uh the
border of a self now biology uses like a
um a multi a multi-scale competency
architecture meaning that every level
has goals so so molecular networks have
goals cells have goals tissues organs
colonies uh and and it's the interplay
of all of those that uh that enable
biology to solve problems in new ways
for example and xenobots and various
other things
um this is
you know uh it's it's exactly as you
said in many ways the cells are
discovering new ways of being but at the
same time Evolution certainly shapes all
this so so evolution is very good at
this agential bioengineering right when
evolution is uh discovering a new way of
being an animal yet one animal or a
plant or something sometimes it's by
changing the hardware you know protein
changing proteins protein structure and
so on but much of the time it's not by
changing the hardware it's by changing
the signals that the cells give to each
other it's doing what we as Engineers do
which is try to convince the cells to do
various things by using signals
experiences stimuli that's what biology
does it has to because it's not dealing
with a blank slate every time as you
know if you're a Evolution and you're
trying to uh uh make make a make an
organism you're not dealing with a
passive material that is is fresh and
you have to specify it already wants to
do certain things so the easiest way to
do that search to find whatever is going
to be adaptive is to find the signals
that are going to um convince cells to
do various things right
your sense is that Evolution operates
both in the software and the hardware
and it's just easier more efficient to
operate in the software yes and I should
also say I I don't think the distinction
is sharp in other words I think it's a
Continuum but I think we can but I think
it's a meaningful distinction where you
can make changes to a particular protein
and now the enzymatic function is
different and it metabolizes differently
and whatever and that will have
implications for Fitness or you can
change the huge
um amount of information in the genome
that isn't structural at all it's it's
uh it's signaling it's when and how do
cells say certain things to each other
and that can have massive changes as far
as how it's going to solve problems I
mean this idea of multi-hierarchical
competence architecture which is
incredible to think about so this
hierarchy that Evolution builds I don't
know who's responsible for this
I also see the incompetence of
bureaucracies of humans when they get
together
so how the hell does evolution build
this where
at every level only the best get to
stick around they somehow figure out how
to do their job without knowing the
bigger picture
and then there's like the bosses that do
the bigger thing
somehow or that you can now abstract
away the small group of cells as a as an
organ or something and then that organ
does something bigger
in the context of the full body or
something like this
how is that built is there some
intuition you can kind of provide of how
that's constructed that that
hierarchical confidence architecture
I love that confidence just the word
confidence is pretty cool in this
context because everybody's good at
their job somehow yeah no it's really
key and the other nice thing about
competency is that so so my my central
belief in all of this is that
engineering is the right perspective on
all of this stuff because it gets you
away from uh subjective uh terms you
know people talk about sentience and
this and that those things very hard to
define or people argue about them
philosophically I think that engineering
terms like competency like um you know
pursuit of goals right all of these
things are uh are empirically incredibly
useful because you know it when you see
it and if it helps you build right if I
if I can pick the right level I say uh
this thing has I believe this is X level
of like con if you competency I think
it's like a thermostat or I think it's
like a a better thermostat or I think
it's a you know a a various other kinds
of you know many many different kinds of
complex systems if that helps me to
control and and predict and build such
system then that's all there is to say
there's no more philosophy to argue
about so so I like competency in that
way because you can quantify you could
you have to in fact you have to you have
to make a claim competent at what and
then or if I say if I tell you it has a
goal the question is what's the goal and
how do you know and I say well because
every time I deviated from this
particular State that's what it spends
energy to get back to that's the goal
and we can quantify and we can be
objective about it so so so the the when
we're not used to thinking about this I
I give a talk sometimes called why don't
robots get cancer right and the reason
robots don't get cancer is because
generally speaking with a few exceptions
are our architectures have been you've
got a bunch of dumb parts and you hope
that if you put them together the the
the the overlying machine will have some
intelligence and do something rather
right but the individual Parts don't
don't care they don't have an agenda
biology isn't like that every level has
an agenda and the final outcome is the
result of cooperation and competition
both within and across levels so for
example during embryogenesis your
tissues and organs are competing with
each other and it's actually a really
important part of development there's a
reason they compete with each other
they're not all just uh you know sort of
helping each other they're also
competing for for information for
metabolic for limited metabolic
constraints
but to get back to your your other point
which is you know which is which is this
seems like really efficient and and good
and and so on compared to some of our
human efforts we also have to keep in
mind that what happens here is that each
level
bends the option space for the level
beneath so that your parts basically
they don't see the the geometry so so
I'm using um and I think I I take this
this seriously uh terminology from from
like um from like relativity right where
the space is literally bent so the
option space is deformed by the higher
level so that the lower levels all they
really have to do is go down their
concentration gradient they don't have
to in fact they don't they can't know
what the big picture is but if you bend
the space just right if they do what
locally seems right they end up doing
your bidding they end up doing things
that are optimal in the in the higher
space conversely because the components
are good at getting their job done you
as the higher level don't need to to try
to compute all the low level controls
all you're doing is bending the space
you don't know or care how they're going
to do it give you a super simple example
in the um in the tadpole we found that
okay so so tadpoles need to become frogs
and to become to go from a tadpole head
to a frog head you have to rearrange the
face so the eyes have to move forward
the Jaws have to to come out the
nostrils move like everything moves it
used to be thought that because all
tadpoles look the same and all frogs
look the same if you just remember if
every piece just moves in the right
direction the right amount then you get
your you get your frog right so we
decided to test we I had this hypothesis
that I thought I thought actually the
system is probably more intelligent than
that so what did we do we made what we
call Picasso tadpoles so these are so
everything is scramble so the eyes are
on the back of the head their jaws are
off to the side everything is scrambled
well guess what they make they make
pretty normal frogs because all the
different things move around in novel
paths configurations until they get to
the correct froggy sort of frog face
configuration then they stop so so the
thing about that is now imagine
Evolution right so so you make some sort
of mutation and it does like every
mutation it does many things so so
something good comes of it but also it
moves your mouth off to the side right
now if if if there wasn't this
multi-scale companies you can see where
this is going if there wasn't this
multi-scale competency the organism
would be dead your Fitness is zero
because you can't eat and you would
never get to explore the other
beneficial consequences of that mutation
you'd have to wait until you find some
other way of doing it without moving
them out that's really hard so so the
fitness landscape would be incredibly
rugged Evolution would take forever the
reason it works one of the reasons it
works so well is because you do that no
worries the mouth will find its way
where where it belongs right so now you
get to explore so so what that means is
that all these mutations that otherwise
would be deleterious are now neutral
because the competency of the parts
make up for all kinds of things so all
the noise of development all the the
variability in the environment all these
things the companies do the parts makes
up for it so the so so that's all that's
all fantastic right that's all that's
all great the only other thing to
remember when we compare this to human
efforts is this every component has its
own goals in various spaces usually with
very little regard for the welfare of
the other levels so so as a simple
example you know
um you as a as a complex system
um you will go out and you will do you
know Jiu Jitsu or whatever you'll have
some to go rock climbing scrape a bunch
of cells off your hands and then you're
happy as a system right you come back
and you've you've accomplished some
goals and you're really happy those
cells are dead they're gone right did
you think about those cells not really
right you had some you had some bruising
out selfish SLB that's it and so and so
that's the thing to remember is that
um you know and we know this from from
history is that is that just being a
collective isn't enough because uh what
the goals of that Collective will be
relative to the welfare of the
individual Parts is a massively open
place justify the means I'm telling you
Stalin was on to something no that's the
danger but we can exactly that's the
danger of uh
for us humans we have to construct
ethical systems
under which we don't take seriously the
full mechanism of biology and apply it
to the way the world functions which is
which is an interesting line we've drawn
the world that built us
is the one we reject in some sense when
we construct human societies the idea
that this country was founded on that
all men are created equal that's such a
fascinating idea it's like uh you're
fighting against
nature
and you're saying well there's something
bigger here than um yeah a hierarchical
competency architecture yeah uh but
there's so many interesting things you
said so from an algorithmic perspective
the act of bending the option space
that's really that's really profound
because if you look at the way AI
systems are built today there's a big
system like I said with robots and as a
goal and he gets better and better at
optimizing that goal at accomplishing
that goal but if biology built a
hierarchical system where everything is
doing computation
and everything is accomplishing the goal
not only that
it's kind of dumb
you know with the uh with the limited
with the bent option space it's just
doing the thing that's the easiest thing
for it in some sense and somehow that
allows you to have
um Turtles on top of turtles literally
dump systems on top of dump systems that
as a whole create something incredibly
smart yeah I mean every system is has
some degree of intelligence in its own
problem domain so so cells will have
problems they're trying to solve in
physiological space and transcriptional
space and then I could give you some
some cool examples of that but the
collective is trying to solve problems
in anatomical space right and forming a
you know a creature and growing your
blood vessels and so on and then the
collect the the the the whole body is
solving yet other problems they may be
in Social space and linguistic space in
three-dimensional space and and who
knows you know the group might be
solving problems and and um you know I
don't know some sort of financial space
or something so one of the major
differences with with most
um uh with most AIS today is is a the
the kind of flatness of the architecture
but also of the fact that they're
constructed
from outside their their borders and
their you know so so if you're so to a
large extent and of course there are
counter examples now but but to a large
extent our technology has been such that
you create a machine or a robot it knows
what its sensors are it knows what its
effectors are it knows the boundary
between it and the outside world all
this is given from the outside
biology constructs this from scratch now
the best example of this that that
originally uh in in robotics was
actually Josh bongard's work in 2006
where he made these these robots that
did not know their shape to start with
so like a baby these are floundered
around they made some hypotheses well I
did this and I moved in this way well
maybe I'm a whatever maybe I have wheels
or maybe I have six legs or whatever
right and they would make a model and
eventually they would crawl around so
that's I mean that's really good that's
part of the autopoiesis but we can go a
step further and some people are doing
this and then we're sort of working on
some of this too is this idea that let's
even go back further you don't even know
what sensors you have you don't know
where you end and the outside world
begins all you have is is certain things
like active inference meaning you're
trying to minimize surprise right you
have some metabolic constraints you
don't have all the energy you need you
don't have all the time in the world to
to think about everything you want to
think about so that means that you can't
afford to be a micro
um reductionist you know all this data
coming in you have to coarse grain it
and say I'm going to take all this stuff
and I'm going to call that a cat I'm
gonna take all this I'm going to call
that the edge of the table I don't want
to follow off of and I don't want to
know anything about the microstates what
I want to know is what is the optimal
way to cut up my world and by the way
this thing over here that's me and the
reason that's me is because I have more
control over this than I have over any
of this other stuff and so now you can
begin to write so that's
self-construction that that figuring out
making models of the outside world and
then turning that inwards and starting
to make a model of yourself right which
immediately starts to get into issues of
agency and control because
in order to if if you are under
metabolic constraints meaning you don't
have the energy right that all the
energy in the world you have to be
efficient that immediately forces you to
start telling stories about course
grained agents that do things right you
don't have the energy to like laplaces
demon you know calculate every every
possible uh State that's going to happen
you have to you have to coarse grain and
you have to say that is the kind of
creature that does things either things
that I avoid or things that I will go
towards that's a mate or food or
whatever it's going to be and so right
at the base of uh simple very simple
organisms starting to make
models of Agents doing things that is
the origin of uh models of of Free Will
basically right because you see the
world around you as having agency and
then you turn that on yourself and you
say wait I have agency too I can I do
things right and and then you make
decisions about what you're going to do
so all of this one one model is to view
all of those kinds of things as
being driven by that early need to
determine what you are and to do so and
to then take actions in the most
energetically efficient space possible
right so free will emerges when you try
to simplify tell a nice narrative about
your environment I think that's very
possible yeah do you think free was an
illusion
so so you're kind of implying that it's
a useful hack
well I'll say two things the first thing
is I think I think it's very plausible
to say that a