►
From YouTube: DevoWorm (2021, Meeting 30): GSoC #11, Size Scaling of Cells, Contingent Tapes of Life, LCB-II
Description
Update on Summer of Code activities (Mainak Deb, Improving DevoLearn), Size Scaling, Timing, and Division Dynamics in Single-celled contexts. Liquid Crystal Biology (LCB). Historical contingencies and replaying the tape of life, Virtual Environments for training computational developmental agents. Attendees: Bradly Alicea, Krishna Katyal, Jesse Parent, Akshay Nair, and Mainak Deb.
A
Hello,
welcome
to
the
meeting.
No,
I
don't
think
anyone's
going
to
show
up
today.
Maybe
a
couple
people
later
but
I'll
get
started
for
those
of
you
on
youtube
and
you
can
follow
along.
I
can
provide
you
with
the
links
in
our
slack
and
if
you
need
an
invite
to
our
slack,
just
contact
me
and
I
can
give
you
the
link
into
the
slack.
A
B
A
C
C
C
I'll
just
go
over
it
quickly,
so,
okay,
so
this
presentation
is
actually
about
the
g-shock
project
for
this
year,
which
was
upgrading
so
I'll
just
get
started
with
it.
Okay,
so
the
first
thing
that
oh
yeah
yeah,
the
first
one
that
I
actually
had
to
build
from
scratch
was
actually
in
the
model
that
could
segment.
Then
it
came
from
the
cl
against
every
uhd
data.
C
C
C
C
So
they
seem
pretty
close
now
and
I
think
the
results
are
pretty
much
acceptable.
So
that's
so,
okay,
so
I'll
just
go
up
to
the
next
okay,
okay,
so
this
section
with
the
upgradation
of
the
self-membrane
segmentation
model.
So
this
was
the
one
that
was
basically
already
present
in
the
regular
library
and
the
first
thing
that
I
looked
at
when
I
started
out
this
expedition
for
upgrading
this
thing.
C
B
C
C
B
C
C
C
C
C
C
C
C
C
B
C
C
C
B
A
A
A
D
A
B
G
G
G
G
B
G
G
B
G
G
A
Well,
yeah
I
mean
like
people
will
measure
divisions
like
so
it's
usually
a
function
of
the
lineage
tree.
So
when
there's
a
branch
in
the
lineage
tree,
there's
this
you
know
division
and
you
can
time
the
divisions
out
and
the
timing
is
not
always
uniform.
You
know
it
has
some
variation
in
the
interval
between.
G
A
G
G
A
F
G
G
G
G
A
G
A
G
A
Okay,
so
this
is
the
paper
and
it
looks
like
they
have
so
they're
using
a
neural
network.
So
what
they're
doing
and
we've
done
this
sort
of
before
kind
of
modeling
lineage
trees
and
that?
But
they
want
to
understand
like
how
cell
size
differs
by
division.
So
sometimes
when
cells
divide,
they
usually
are
about
half
the
size
of
their
mother
cells.
And
then
you
know
there's
this:
they
can
grow
in
size
a
little
bit
or
they're.
You
know,
there's
an
asymmetry
in
division.
A
So
if
you
take
a
mother
cell
and
you
divide
it
into
two,
each
daughter
cell
will
be
half
the
size
of
its
mother
cell
and
but
sometimes
there's
an
asymmetry
in
that
size.
So
sometimes
maybe
the
daughter
one
daughter
cells,
like
sixty
percent
of
the
mother
cell,
one
other
cells,
forty
percent,
and
so
that
that
you
know,
depending
on
the
organism
that
holds
or
there's
some
sort
of
scaling
between
the
two.
A
And
then
this
goes
on
and
you
end
up
with
an
organism-
that's
much
bigger
than
the
embryo.
So
you
get
this,
but
you
have
to
there's
this
with
a
cell's
divide.
There's
this
sort
of
control
that
controls
the
size
of
the
cells
and
ultimately,
the
size
of
the
organ
that
it's
going
to
form
so
there's
different
non-size
control
mechanisms
that
actually
depend
on
history.
A
So
it
says,
there's
a
conditional
dependence
on
the
size
distributions
of
history.
We
had
a
discussion
about
this
in
our
group
on
saturday,
a
rather
lengthy
discussion
about
like
the
role
of
historical
contingency
and
neural
networks,
and
so
you
know
the
or
markov
models
actually
where
markov
models
don't
have
a
history,
and
so
there
are
ways
at
which
you
can
like
incorporate
a
history
into
your
markov
model,
which
is
basically
where
you're
transitioning
between
states
and
so
that
would
we
had
a
discussion
on
that.
A
But
it
looks
like
that's
what
they're
doing
here
as
well
they're,
in
this
case
it's
a
neural
network
but
they're
actually
trying
to
incorporate
a
historical
aspect
to
it.
So
it's
it's
drawing
from
previous
data.
You
know
maybe
like
the
last
20
or
30
examples
and
it's
kind
of
using
that,
as
not
as
like
a
just
a
training
like
a
general
training
model
but
like
as
a
time
dependence.
A
So
it's
a
bit
different
because
the
data
instead
of
just
sampling,
generally
speaking,
it's
you
know,
there's
this
time
to
it's
picking
up
this
pattern
in
time
and,
as
you
know,
things
that
happen
in
time
are
temporally
dependent.
So
you
know
your
state
at
one
point
in
time
correlates
with
your
next
point.
You
know
your
next
date
and
time,
but
over
long
distances
it
may
not
correlate
at
all.
So
there's
this
there's
this
dependence
on
history.
A
You
know
they
have
the
size
that
they,
you
know
when
they
divide,
they
have
a
size.
They
can't
get
any
bigger
than
something
they
can't
get
any
smaller
than
something
and
in
single
cell
organisms
like
this.
This
is
important
in
terms
of
like
metabolism
and
so
forth,
so
they
don't,
but
they
don't
model,
they
don't
use
like
a
metabolic
model.
They
don't
use
an
internal
model.
They
looks
like
they
use
this
sort
of
model
of
the
kind
of
you
know
assume
certain
things
about
the
size.
A
They
just
use
a
scaling
law,
which
is
where
you
have
this
trend
in
the
data,
and
then
you
can
get
a
range
of
outcomes
and
it's
it's
a
stochastic
generation
which
means
it's
sort
of
drawing
from
a
distribution
at
random.
So
you
know
again,
you
have
these
distributions,
like
a
gaussian
distribution.
Just
drawing
from
that.
G
A
So
you
know,
usually
it's
like
this,
where
you
have
a
branching
pattern
and
then
let's
say
that
this
lineage
tree
is
asymmetrical,
so
this
this
other
cell
here
will
divide
at
a
certain
time
and
then
you'll
have
two
daughter
cells.
Those
daughter
cells,
though,
will
divide
at
different
times
yeah
and
then
that's
so
you
can
see
that
they
don't
always
divide
at
the
same
time.
Sometimes
they
divide
synchronously,
but
sometimes
it
doesn't
and
there's
you
know
the
different
reasons
for
this.
It
could
be
in
a
multicellular
organism.
A
Sometimes
like
in
c
elegans,
there's
a
sort
of
a
program
time
for
this.
So,
like
you
know,
if,
if
this,
this
lineage
is
going
to
go
off
and
form
some,
you
know
organ
that
needs
to
be
formed
in
a
certain
amount
of
time.
Then
that's
going
to
be.
You
know
something
that
is
you
know
it's
going
to
divide
at
a
certain
time
scale.
A
This
thing
might
be
the
germ
cell,
so
it
divides
at
a
different
time
scale
I
mean
so
their
differences
in
terms
of
the
organ
now,
with
like
the
organisms
they're
talking
about
yeasts
and
bacteria,
it
doesn't,
I
don't
think,
there's
any
sort
of
like
this.
A
strong
asymmetry
like
this,
it's
just
that
they
divide
stochastically
over
time,
which
means,
it's
just
like
say,
there's
a
gaussian
distribution
with
a
mean
of
like
10
minutes.
A
This
is
just
a
hypothetical
because
it's
not
biologically
plausible,
but
then
it
would
just
sample
around
that
10
minute
interval,
maybe
like
anywhere
from
8
to
12
minutes.
You
get
to
have
another
round
of
division,
and
so
this
just
happens.
You
know
over
you
know,
that's
a
very
slow.
You
know
it
just
will
draw
over
that
time.
There's
no
real
pattern
to
it,
but
it
could
be
asymmetrical
that
way.
A
A
So
the
the
mother
cell
is
coming
through
here
and
then
there's
a
flow
that
shear
a
sheer
force
that
pushes
the
daughter
cells
out.
I
think,
and
then
there's
this
birth
size
and
generation
time.
So
you
can
see
what
I
was
talking
about
before.
What
I
think
krishna
is
talking
about
where
you
start
at
a
size
and
then
it
the
cell
grows
and
then
it
divides,
and
so
then
it
goes
back
to
a
sort
of
a
characteristic,
so
division
size,
and
then
it
grows
again
and
it
goes
back.
A
I
think
roughly
twofold,
because
you
have
every
rounded
division.
You
have
this
doubling
of
the
size,
just
by
virtue
of
the
cells
growing
and
dividing
growing
and
dividing
like
that
now,
so
they
can
observe
that
in
microfluidics,
and
then
I
think
this
stuff
here
is
just
shows
kind
of
yeah,
so
they're
using
the
medians
of
the
content
conditional
probability
density
function,
which
is
just
like
a
function
like
a
gaussian
distribution.
A
A
Drawing
from
it
I
mean
they're
necessary,
drawn
from
our
probability
function
they're,
just
drawing
from
like
you
know
their
local
neighborhood
like
if
the
division
time,
I
guess,
is
10
minutes
it
could
be
like
you
know,
either
side
of
that
the
timing
isn't
precise,
but
it's
it's
generally
within
that
window,
and
then
temperature
will
affect
this
time.
So
if
you
have
a
lower
temperature,
your
time
will
be
longer
because
the
cell's
metabolism
is
lower.
A
A
E
A
Yeah,
we
do
actually
we
have.
I
think
we
have
a
number
of
data
sets
and
they're
in
the
csv
format,
and
they
have
a
lot
of
division
times
and
subdivi.
So
like
we've,
sampled,
the
division
times
and
gotten
derivative
division
times,
and
all
that
I'll
have
to
pull
that
up
and
put
that
in
the
slack
yeah.
A
So
this
is
forked
from
d9w.
Who
actually
is
someone
who
he
submitted
a
paper
to
this
workshop
last
summer
and
it
was
run.
He
was
one
of
the
organizers
of
it.
This
was
at
the
a
life
conference,
and
so
you
know
they
they're
interested
in
computational
development,
developmental
biology
and
then
agents
developmental
agents,
so
they're
doing
this
project,
pi
grn,
which
is
where
you
can
model
these
artificial
gene
regulatory
networks.
A
So
gene
regulatory
network
is
where
you
have
a
model
of
a
gene.
So
it's
like
something
that,
like
a
binary
sequence,
that
encodes
something
like
a
phenotype,
so
you
could
create
a
a
quote,
unquote:
gene
for
maybe
like
a
pencil
all
right,
you
could
describe
a
pencil
using
a
binary
string
or
a
series
of
binary
strings,
so
you
could
have
like
multiple
jeans
describing
a
pencil
and
that
pencil
is
going
to
have
you
know
it.
A
We
can
observe
regulatory
elements
that
regulate
genes
in
their
expression,
so
sometimes
genes
are
expressed.
You
know
at
a
very
high
level,
sometimes
at
a
very
low
level.
Sometimes
certain
parts
of
the
genes
are
expressed
and
not
others
given
the
environmental
context,
and
so
these
are
and
then,
of
course,
you
can
use
this
to
model
objects
in
the
real
world
like
pencils
or
car,
you
know,
building
car
bodies
or
whatever
you
want
it.
A
A
They
just
have
the
setup
and
all
this,
but
I
wanted
to
bring
this
to
people's
attention
that
this
exists,
because
this
is
maybe
something
that's
useful
to
people
if
you're
interested
in
say,
genetic
algorithms
or
you're
interested
in
generative
modeling.
This
is
a
good
way
to
go
and
we
haven't
really
talked
too
much
well.
We
haven't
really
gotten
too
much
into
like
how
these
artificial
models
work,
but
you
know
they're
they're
going
to
be.
You
can
use
a
lot
of
computational
tricks
to
get
it
to
work
and
work
in
different
ways.
A
So
this
is
something
that
you
might
want
to
check
out.
Another
thing
I
want
to
share
is
that
we
have
our
google
summer
of
code
stub
on
the
website.
So
this
is
the
website
for
diva
worm
and
let
me
put
in
the
chat
and
our
summer
of
code.
Stub
has
our
people
who
are
the
students
in
every
year
the
students
who
are
in
in
the
program
and
accepted
into
the
program.
A
So
this
is
my
not
care
class
of
2021.
My
knock.
If
you
have
a
better
picture
yourself,
let's
send
it
to
me,
but
I
just
took
a
screen
grab
off
of
one
of
our
sessions
we
have.
Last
year
we
had
mayukdev
and
oswal
singh
who
did
open
devo
cell
integration
and
then
the
beginnings
of
diva
learn
so
and
also
did
something
sort
of
a
sub
project
under
that
called
diva
worm
ai.
So
if
you
weren't
here
last
year,
there's
this
deworm
ai
repository,
which
has
a
bunch
of
models.
A
So
you
know
at
this
point:
we've
have
a
lot
of
ai
models
for
looking
at
cells
and
analyzing
cells
cell
shape
cell
size,
and
things
like
that.
So
this
is
like
a
nice
collection
of
models,
and
maybe
this
fall
will
kind
of
reevaluate
where
we
are
with
that
and
what
we
can
use
it
for,
because
we
have
all
this.
These
tools,
we
can't
we
haven't
really
used
them
for
a
lot
yet
and
we
need
to
maybe
use
it
a
little
bit
more.
A
So
then,
the
year
before
we
had
this
open
devo
cell
project
by
vinay
verma,
and
that
was
a
nice
segment,
a
semantic
segmentation
package.
Rnab
was
the
year
before
and
he
did
this
developmental
networks
and
xml
frameworks
for
the
c
elegans
embryo.
This
is
actually
where
we
looked
at
some
of
the
time:
timing,
data
or
refine
that
data
a
bit.
So
we
were
doing
that
a
while
back
and
then
in
2017
siddhartha
did
image
processing
with
image
j,
so
we
actually
created
an
image
j
module
as
well.
A
So
we
have
all
these
tools-
and
I
don't
know
exactly
where
we're
going
with
this:
it's
something
that
we'll
need
to
sort
of
focus
on
how
to
implement
it.
You
know
in
a
way,
that's
you
know.
Maybe
we
can
create
some
sort
of
nice
analyses
from
this,
but
there's
a
way.
We
really
have
a
lot
of
tool.
You
know
we
have
a
lot
of
tools
in
place
now.
A
I
think
that's
going
to
be
an
exciting
time
coming
up
for
for
the
group
and
then
go
to
the
submissions
document,
and
there
isn't
much
here
new
I'll,
reiterate
this
neurops
workshops
list.
So,
if
you're
interested
in
going
submitting
something
to
an
europe's
workshop,
the
schedule
is
now
online
and
what
you
do
is
you
go
find
one
of
these
sessions
and
you
try
to
find
it
on
on
your
favorite
search
engine
and
you
can
maybe
find
their
website
where
they
will
tell
you
what
the
deadline
is
for
submission.
A
The
other
thing
we
had
this
mathematics
at
diva,
where
we
talked
about
that
last
week,
that's
still
ongoing,
like
I
showed
a
couple
of
additions
to
that.
I
know.
Jesse
didn't
see
it
west
last
time,
but
we
can
go
over
that.
Maybe
at
some
point
I
can
put
that
go
over
that
more
detail
in
another
meeting,
and
so
we
have
a
couple.
We
have
the
kindle
book
and
and
the
boring
billion,
so
the
kindle
book
is
something
that
krishna
was
working
on.
G
A
A
A
A
Probably
the
life
will
die
off
before
that
because
it'll
get
pretty
hot.
You
know
in
terms
of
like
the
sun
being
getting
hotter
and
brighter
and
larger
and
then
it'll
actually
engulf
earth,
and
so
that
means
that
our
sort
of
any
solar
system,
in
fact
not
just
earth
any
solar
system
where
there's
life
there's
this
finite
window
for
life
to
exist,
and
so,
if
life
starts
at
a
certain
point
after
a
planet
is
formed,
then
you
know
how
long
does
it?
How
long
can
life
sort
of
sit
on
a
planet
and
evolve?
A
A
If
we
were
to
rerun
the
tape
from
the
starting
point,
do
you
think
a
technological
spacefaring
civilization
would
re-emerge
on
earth
before
our
planet
becomes
inhospitable
and
people
said
no
by
a
small
majority?
And
he
says
I
don't
have
a
good
answer
for
this,
because
the
start
time
above
is
somewhat
arbitrary
and
causally
connected
to
prior
history.
Okay,
so
this.
A
In
other
words,
if
I
just
took
any,
if
I
just
went
back
and
I
just
replayed
the
tape
life,
forgetting
all
we
know
about
like
life
on
earth-
and
this
could
be
true
of
another
planet
where
there's
life,
would
you
have
primates,
for
example?
Or
would
you
have
dinosaurs
or
anything
you
know?
Would
you
have
life
on
land
or
maybe
life
would
look
a
lot
different?
A
Maybe
it
wouldn't
even
be
anything
that
resembles
life
on
earth
now
or
life
on
that
other
planet,
and
so
the
question
is:
is
basically
replaying
this
process
of
evolution
and
recombination,
and
all
this
and
reproduction
and
you're
getting
this
result.
So
people
have
done
this
experiment,
but
they've
never
really
done
it
with
an
end
point
in
mind,
which
I
think
is
interesting,
that
there's
this
time
window
for
life
and
people
have
talked
about
like
you
know,
if
you
start
off
with
a
planet,
that's
hospitable
to
life,
but
the
conditions
aren't
right.
A
You
know
it
may
not
give
life
a
long
time
to
sort
of
intelligent
life
a
long
time
to
emerge,
and
so
these
are
all
questions
you
can
ask
by
replaying
the
tape
of
life,
and
so
this
actually
came
from
steven
j
gould
about
30
years
ago.
He
asked
he
used
this
metaphor,
replaying
life's
tape,
and
so
this
is
a
gould.
I
believe
think
thought
that,
like
everything,
would
be
dramatically
different
every
time
you
rerun
the
tape.
A
Things
are
different,
and
so
we
don't
really
know,
but
this
relies
on
the
sort
of
stochasticity
of
the
evolutionary
process.
So
that
means
that
if
you
rerun
the
tape
of
life,
things
really
can't
be
the
same,
because
life
is
stochastic
and
you
have
enough
events,
stochastic
events
that
you
end
up
in
a
much
much
different
place,
or
it's
very
likely
to
end
up
in
a
much
different
place,
and
so
a
direct
approach
to
quantifying
evolutionary
predictability
requires
massively
parallel,
carefully
controlled
and
fully
characterized
evolution.
A
Experiments
and
some
people
have
done
this
experimental
evolution
with
bacteria
with
yeast
with
other
organisms.
C
elegans.
Usually
you
know.
If
you
have
a
short
generation
time,
you
can
do
large
scale
evolution
experiments
where
you
can
do.
Maybe
you
know
one
thousand
five
thousand,
even
like
a
hundred
thousand
generations
over
a
number
of
years,
and
so
you
can
get
a
lot
of
data
on
what
life,
at
least
in
these
organisms
will
look
like
if
you
replay
that
tape-
and
so
you
know
you
get
different
mutants,
but
this
isn't
really
the
question.
A
The
question
is:
is
that
you
know?
Can
you
start
off
and
then
end
up
with
this
great
complexity,
and
so
people
have
done
this
sort
of
work,
and
this
is
a
review
of
this
area
of
work,
and
so
but
people
aren't
really
close
to
answering
it.
It's
still
one
of
these
things
that's
kind
of
a
mystery,
and
so
this
paper
replaying
the
tape
of
life
in
the
21st
century
kind
of
updates.
This
idea
a
bit.
This
is
where
you
know.
A
First
of
all,
we
need
to
have
an
experimental
evolution
where
we
can
ensure
the
sort
of
replicability,
because
then
we
can't
say
if
we
don't
have
that,
we
can't
say
whether
you
know
this,
whether
it's
just
a
statistical
artifact
or
whether
stochasticity
is
overcoming
this
repeatable
repeatability
so
but
accumulating
data
is
now
challenging
this
classic
answer
of
no.
So
we
have
this
experimental
issue
and
if
we
can
overcome
that
we
can
really
get
a
good
answer
and
but
then
you
know
you
can
look
at
other
things,
other
than
experimental
evolution.
So.
A
Of
alleles
described
in
the
literature
that
cause
non-deleterious
phenotypic
differences
among
animals,
plants
and
yeasts
that
indicates
a
similar
phenotype
can
often
evolve
in
distinct,
taxa
they're,
independent
mutations
in
the
same
genes.
So
you
sometimes
have
convergent
evolution
of
like
things
like
wings
or
actually
flight
or
wings,
where
you
know
they
have
this
functional
thing,
but
it's
done
in
many
different
ways,
so
like
insect
wings
and
bat
wings
and
those
things
are
very
different
in
their
structure,
but
they
both
achieve
flying
or
flight
and
they
both
have
sort
of
a
similar
structure.
A
So
these
things
sort
of
speak
to.
Maybe,
if
you
play
the
type
of
life
again,
maybe
you
do
end
up
with
more
or
less
the
same
outcome.
So
you
know
we
have
to
take
these
sorts
of
things
into
account
as
well,
and
so
imagining
other
possible
paths
for
evolution
runs
into
four
important
issues.
One
is
resolving
the
influence
of
contingency,
so
this
is
you
know.
If
history
gives
you
a
certain
hand,
you
have
to
play
the
hand
you
can't
get
another
hand
from
the
dealer
and
so
that
that
always
has
an
influence
on
things.
A
A
So
how
do
we
even
know
how
to
predict
evolution
and
if
we
understand
those
principles,
maybe
we
can
understand
this
problem
better
and
for
estimating
the
probability
of
occurrence
for
complex
evolutionary
events?
So
you
know
if
we
understand
that
multicellularity
is
this
fluke
or
the
stochastic
thing,
then
you
know
we
understand
that.
Maybe
it
can't
happen
twice
or
maybe
it
happens
in
a
very
different
way.
A
A
It
kind
of
goes
very
deeply
into
a
very
traditional
area
of
biology
and
there's
a
lot
that
we
can.
You
know
I
mean
people
have
done
these
studies
again
and
again.
So
you
know
if
we,
if
we
wanted
to
like
make
a
contribution
here,
would
be
very
much
like
you'd
have
to
find
a
niche
area,
but
nevertheless,
I
think
there's
some
interesting
things
to
say
here,
so
you
know
extrapolating
from
short
bouts
of
evolution
to
the
entire
span
of
life
evolution.
A
A
And
finally,
you
know
this
goes
back
to
even
the
wall
of
complexity
that
stephen
jabul
talked
about,
which
is
that
you
know
when
you
have
this,
when
you
have
something
like
bacteria,
which
are
very
simple
and
in
the
pre-cambrian
dominated
the
biota
on
earth,
and
they
still
dominate
the
biota
on
earth
by
the
way.
But.
B
A
They
were
the
only
game
in
town
back.
Then
there
was
a
lot
of
complexity
in
that
type
of
organism.
You
know
and
then,
but
then,
as
you
go
further
out
in
complexity
in
terms
of
complexity,
there's
less
and
less
frequency
of
occurrence
and
there's
less
and
less
complexity
in
terms
in
specific
lineages.
So
this
is
this
wall
of
complexity,
where
you
get
these
bacteria
that
have
a
large
number
of
constituents
and
you
get
fewer
constituents
as
you
go
through
time.
So
there
are
fewer
opportunities
to
really
make
very
large
scale
changes.
A
B
B
F
A
A
A
Yeah
I'll
put
them
in
the
slack
in
the
open
arm
slack,
so
I'm
gonna
go
through
one
more
paper.
I
did
want
to
mention
that
I
I
was
planning
on
doing
this
liquid
crystal
biology,
so
this
was
for
susan's
benefit.
I
didn't!
I
don't
want
to
do
it
because
she's
out
with
she's
getting
surgery,
I
think,
but
I
did
want
to
point
to
this
book-
the
physics
of
liquid
crystals.
A
You
know
usually,
like
bacteria
will
kind
of
flow
around
or
they'll
move
around
if
they
get
jammed
they're
too
many
of
them
in
one
place
they
get
jammed
up,
and
so
that's
a
phase
transition
in
terms
of
their
organization
across
cells,
and
so
there
are
a
lot
of
interesting
things
in
this
book.
It's
it
put
it
because
it's
a
classic
book
people
might
be
interested
in
reading
it.
For
reasons
other
than
like.
A
I
think
we
talked
about
a
paper
that
featured
some
of
these
topics
last
week,
and
so
I'm
gonna
come
back
to
this
one
seasons
back,
but
I
wanted
to
make
because
I
put
it
in
the
slack.
I
wanted
people
to
understand
why
that
was
so.
That's
that's
an
interesting
classic
book.
Finally,
I'm
going
to
go
into
this
paper.
I
think
so
this
is
from
mit
tech
review.
This
is
it's:
it's
an
ai
paper,
it's
machine
learning
and
it's
an
endlessly
challenging
playground
teaches
ais
how
to
multitask.
A
A
So
in
this
case
you
have
an
environment.
Where
there's
this
maze,
there's
geometry
and
the
agent
is
developing
or
it's
it's
behaving.
It
could
be
developing.
Maybe
not-
and
it's
interacting
with
these
geometric
shapes
such
that
it's
maybe
telling
it.
You
know
where
things
are
in
space,
so
you
can
encode
different
landmarks
or
you
know
in
in
the
case
where
you
have
to
learn
what
a
step
is.
A
You
know
you
would
make
a
step
up
from
one
layer
to
another,
and
you
know
this
agent
can
learn
what
a
step
is
or
how
to
generalize
to
other
types
of
steps,
or
you
know
a
tower.
What
a
tower
is
how
to
climb
a
tower.
So
there's
all
these
things
that
the
agent
can
learn
in
an
environment
like
this.
It
couldn't
learn
if
you're,
just
programming
it
or
alert.
You
know
teaching
it
from
pictures
or
something
like
that.
A
So
this
is
you
know
they
can
build
these
3d
environments,
put
agents
in
them
and
then
the
3d
environment
and
a
3d
environment
is
limiting
because,
like
I
said
with
embryos,
you
can
have
things
like
temperature,
which
really
influence
the
state
of
the
agent.
So
if
you
have
a
developing
agent,
that's
a
developing
group
of
cells.
You
know
you
might
have
a
temperature
chamber
here
with
some
geometry
to
it
and
you
know
other
like
light
sources
that
come
in
and
out
of
play.
A
And
so
again
you
know
this
is
much
better
than
if
you
have
like
a
bunch
of
pictures
that
you
train
the
agent
on
where
you
just
hard
coded
something
into
the
agent,
and
so
the
playground
manager
has
no
specific
goal,
but
the
general
ideas
that
improve
the
capability
of
the
player
over
time.
So
you
know
there
are
actually
platforms
where
you
can
download
the
platform
and
you
can
actually
just
put
agents
into
it,
because
you
know
these
these
cost
a
lot
to
make.
A
You
know
a
lot
of
time
and
expertise,
and
so,
if
you
have
these
pre-made
environments,
it's
a
lot
easier
to
just
pick
them
up
and
start
training
things
in
it.
So
this
is,
I
mean,
learn
to
experiment
these
bots
improvised
a
ramp.
So
here
they
learn
that
you
can
use
some
feature
of
the
environment
as
a
ramp,
and
so
that's
what
they're
doing
it's
not
necessarily
obvious
to
the
programmer.
Maybe
what
that
is,
but
that's
what
the
agents
are
learning
for,
so
so
yeah!
A
This
is
the
this
is
from
the
mit
tech
review.
So
they
have.
You
know
you:
can
they
have
these
ai
generating
algorithms
aigas,
which
are
genetic
algorithms
which
generate
ais
where
yeah
so
they're,
algorithms
that
generate
ais?
And
it's
an
interesting
area
here?
This
is
something
that
they're
working
on
in
open,
ai,
but
there's
a
lot
you
can
do
with
this.
I
think
it's
very,
very
flexible,
so
I'd
be
interested
in
here.
You
know
if
people
were
inspired
by
that
you
know.
A
B
B
A
Welcome
all
right,
yeah
there's
my
knock,
and
so
my
not
good
luck
with
your
submission
and
if
you
have
any
questions,
let
me
know,
but
I
think
it's
pretty
straightforward
and
I
look
through
your
submission
and
everything
looks
to
be
there.
I
don't
think,
there's
anything
missing,
yeah,
okay,
all
right!
Well,
thanks
for
attending
and
I'll
be
putting
that
information
in
the
slack
and
next
week
you
know
I
will
see
if
you
have
anything
you
want
to
present
like
krishna
did
this
week.
A
Let
me
know
and
I'll
make
time
for
it
and
we'll
you
know
we'll
follow
up.
I
know
next
week,
it'll
be
officially
the
end
of
gsox.
So
we'll
you
know
we'll
probably
talk
about
some
other
things,
but
maybe
we'll
have
a
discussion
about
how
to
integrate
a
lot
of
our
tools
into
one
place,
or
you
know
the
paper
that
we
have
been
promising
and
evil
learn
so.