►
Description
Fun Friday! 2D object recognition, brain / tech chat, HTM Forum discussion -- Watch live at https://www.twitch.tv/rhyolight_
A
A
A
A
A
A
A
Possibly
the
best
example
of
how
I
already
read
this
I
think
these
answers.
What
is
three
learning
rules?
What
are
these
three
learning
rules
suit?
I
was
talking
about
learning,
growing,
new
connections,
yeah
I.
Think
that
the
answer
is
consistency
basically
I
mean
you
can't
learn
something
if
you
can't
rely
on
similar
stimulations
causing
similar
activations
in
the
input.
I
think
that's
the
gist
of
it.
A
These
are
yeah.
These
are
both
good
videos
to
watch.
If
you
want
to
understand
STRs
and
that
I
have
not
read
this
whole
theory
or
this
whole
paper
on
semantic
folding.
But
it
looks
interesting.
It
says:
yep
I
understand
that
the
same
input
should
output
the
same
output.
Otherwise
you
could
never
learn
anything.
My
question
is
why
two
inputs
that
are
similar
should
generate
similar,
but
not
identical.
Strs
you've
read
the
semantic
folding
paper.
A
That's
interesting
question:
why
why
two
inputs
that
are
similar
to
generate
similar
but
not
identical,
SP
R's
well
they'd
only
generate
identical
SBR
as
if
they
were
the
exact
same
sensory
input,
but
if,
but
if
you
feel
something
just
barely
different,
it's
got
a
encoded
there
that
difference.
It
has
to
encode
that
difference
right,
it's
going
to
it
has
to
be
close
to
the
previous
one,
but
not
exactly.
A
A
Think
this
is
true
for
categorical
data.
Okay,
wait
so
will
says:
I,
don't
understand
the
reason
for
requiring
semantic
similarity.
Wouldn't
spatial
pooling
learning
happen
just
as
reliably
on
randomly
random
Lee
but
deterministically
generated
STRs
yeah?
Definitely,
we've
got
a
random
distributed,
scaler
encoder,
for
example,
but
it
has
to
be
consistent
and
it
has
to
still
adhere
to
those
same
rules.
It's
not
really
random.
You
can
pick
so
you
can
start
randomly,
but
then
every
other
thing
that's
like
that
has
to
be
consistent
with
those
semantics
that
you
would
randomly
chose.
A
A
And
I
haven't
read
this
version:
I
have
read
the
previous
version
and
it's
been
updated
since
then
a
bit
so
I
still
have
yet
to
read.
This
I
might
actually
read
through
this
on
Twitch
live
next
week,
because
I
I
need
to
read
it.
I
want
to
go
over
it
and
I
could
comment
on
it
as
I
read
it.
If
that's
interesting,
anyone,
that's
something.
I
could
do.
A
B
B
A
Me
get
to
a
post
from
mark
here,
okay,
so
here's
just
for
the
context
trying
to
understand
basically
HTM
and
the
model
of
cortical
columns
in
HTM,
a
single
neuron
and
a
given
mini
column,
gets
inputs
and
outputs
of
every
neurons
and
mini
columns
of
the
given
cortical
column.
That's
sort
of
it
could
I
mean
the
the
key
thing
about
this
is
that
it's
configurable,
like
every
layer
can
can
potentially
be
input
to
one
or
or
many
other
layers.
A
Okay,
how
can
a
pattern
share?
It
in
different
cortical
columns
be
predicted,
while
cortical
columns
do
not
send
information
to
each
other.
First
of
all,
they
do.
They
do
send
information
to
each
other.
There
are
lateral
connections
in
layer,
two
that
are
sharing
information
to
each
other.
I
know,
Marc
knows
about
these
I
mean
maybe
my
first
letter
is
encoded
in
a
cortical
column.
In
the
second
letter
in
another
cortical
column,
no
see
they're
all
going
to
be
encoded
in
all
the
columns.
A
When
you
look
at
a
letter
a
and
then
you
look
at
the
letter,
B
the
same
cortical
columns
are
looking
at
that
letter.
It's
not
like
the
same
cortical
columns
in
v1
are
looking
at
it.
Mark
says
this
is
an
excellent
and
perceptive
questions
due
to
how
humans
like
to
breakdown
problems
and
combine
them
back
together
and
disciplines
such
as
engineering
or
physics,
we're
used
to
seeing
hierarchy
in
a
way
that
combines
and
condenses
information
as
you
ascend
the
levels
right.
That
makes
sense.
That's
that's
very
that's
a
very
classical
view
of
hierarchy.
A
It's
very
structured
view.
The
lower
levels
feed
into
the
higher
levels,
those
higher
levels,
get
information
from
many
children
and
somehow
make
sense
of
them
and
create
their
own
picture
of
the
world
and
send
it
up
to
another
higher
level
which
does
the
same
thing.
So
the
same
process,
I
Falco,
the
same
processes
is
happening
at
every
level
of
the
hierarchy.
Essentially,
here
we
got
here
Falco
and
bit
King
thanks
for
joining
and.
A
So
anyway,
I'm
going
through
your
poster
mark
agreeing
with
everything
so
far,
so
we're
used
to
seeing
hierarchy
like
what
you
just
said.
This
leads
to
a
central
command
and
control
node
somewhere
towards
the
top
of
the
logical
structure
right.
This
is
a
classical
view.
Much
of
the
classical
literature
assumes
that
this
is
what
is
going
on
in
the
brain.
It
presents
this
as
a
fact
without
any
actual
support
from
the
known
wiring
of
the
brain.
A
When
we
try
to
apply
this
model
to
the
brain,
we
can
see
in
case
anybody's
reading,
along
with
me,
when
we
try
and
play
this
model
to
the
brain
we
can
see.
There
are
layers
of
processing,
but
the
wiring
just
does
not
support
the
concept
of
information
merging
to
some
central
node.
Absolutely
jeff
has
talked
about
this
for
ages.
He
always
shows
the
picture
of
the
monkey
monkey
hierarchy
right.
A
A
Where
is
the
picture
I?
Always
it's
it
always
pops
up.
Maybe
somebody
could
help
me
out
here
if
they
find
it,
it's
like
a
jumble,
see
some
goop
little
stuff.
It's
just
a
jumble
of
gosh
darn
it
it's
in
it's
always
in
his
presentations.
Anyway,
it's
a
jumble,
it's
hard
to
say
that's
hard
to
say.
Ok,
it's
hard
to
grasp!
Oh
wait!
Where
did
it?
Where
was
I
right?
A
It
seems
to
stay
mostly
in
a
doesn't
support
the
concept
right,
it
seems
to
say
mostly
in
a
parallel
format,
as
it
courses
from
area
to
area
in
the
brain.
I
struggled
this
for
the
longest
time
it
is
hard
to
grasp,
but
it
seems
that
the
recognition
is
distributed
as
a
cooperative
effort
using
short-range
lateral
connections
to
each
area
of
the
brain.
A
This
allows
the
individual
column
to
recognize
it's
possible
bit
of
the
overall
picture
and
vote
with
its
neighbor,
on
which
of
many
possible
larger
scale,
things
that
it
may
be
a
part
of
all
the
computations
are
local,
but
they
build
to
a
global
picture.
It's
exactly
the
right,
that's
perfect,
with
sequential
recognition,
so
we're
talking
about
temporal
right.
You
know
like
over
time
right.
This
is
requiring
movement,
I!
Think!
A
Well,
in
from
the
mental
standpoint,
sequential
object,
recognition
requires
movement
of
sensors
through
the
space
movement
through
the
sensory
space,
so
with
sequential
recognition
in
the
bit
that
is
being
voted
on
is
also
the
transition
between
this
current
pattern
and
the
pattern.
This
adds
temporal
recognition
to
the
spatial
recognition,
so
I
think
when
learning
when
we're
learning
new
objects,
I
think
this
is
correct.
The
suit
that's
the
bit
we're
gonna
be
predicting,
given
our
experience
in
that
out
in
an
object
space
just
in
space.
A
Iii
really
think,
there's
probably
two
two
spaces
like
two
spaces
in
your
brain.
One
space
where
you
model
your
egocentric
surroundings,
I
think
I've
talked
about
this
before
and
another
space
where
you
model
objects,
ll,
eccentric
ly
and
yes,
as
you're
moving
through
object
space.
So
there's
there's
always
an
intersection
of
these
spaces.
Right
I
think
this
is.
This
has
to
be
happening
somewhere.
A
This
intersection
between
these
spaces,
where
you've
got
you're
at
egocentric
space
and
the
little
Peaks
little
the
like
the
allocentric
objects,
can
pop
and
and
render
and
exist
in
that
egocentric
space
and
and
we
can
sort
of
seamlessly
merge
the
two
spaces.
That's
like
a
core
thing,
our
brains,
doing
seamlessly
merging
spaces,
and
so,
as
you
move
your
senses
through,
what
you
perceive
is
egocentric
space.
A
You
you're
also
moving
through
allocentric
space,
so
you're
learning
up
object
models
in
its
in
another
space,
not
the
space,
not
your
egocentric
space
when
you're
learning
object
models
in
a
different
space,
I
think,
but
that,
but
you
have
to
use
the
sequential
sequential
recognition
to
build
up
those
models,
I
think
and
so
yeah.
That
makes
sense
to
me
and
then
once
you
have
built
those
models,
then
you
can
use
them
to
recognize.
A
Bring
up.
You
know
trigger
these.
These
objects
to
based
on
a
single
sensory
feature.
Then
you
can
imagine
not
descending.
Imagine
it's
not
like.
It
consciously
happens.
It's
just
it's
when
you
touch
something
in
a
space,
no
matter
how
house
even
do
it
with
just
one
touch
you
can
fill
in
the
rest
of
it
if
you
have
enough
experience
with
those
sensations
in
those
points
in
space
like
I
know,
this
is
some
kind
of
cup.
If
I
accidentally
touch
it
like
this
I
don't
know
exactly
it
could
be
a
coffee
cup.
A
It
could
be
a
long
cup
of
could
be
a
beer
glass
I,
don't
know,
but
there's
a
ton
of
things
that
I
would
drink.
Out
of
that
would
match
that
that
sensory,
you
know,
input
that
I
get
in
object,
space
and
some
object
space,
but
but
there's
this
okay
I'll
keep
going
and
the
example
you
provide
the
eyes
move
around
and
keep
placing
a
group
of
letters
in
the
center
of
the
visual
field
there.
A
A
Second,
here
and
I
look
real
close
and-
and
if
you
put
your
thumb
out
to
in
front
of
you
like
this
and
and
and
you
look
at
your
thumb
that
you're
basically
the
size
of
your
thumb
that
is
v1
like
that
field
of
view,
the
size
of
your
thumb
projects
to
to
v1
back
here
and
that's
like
the
most
the
most
densely
packed
area
of
processing,
that's
going
on
is
in
that
space.
So
when
you
look
at
that
word
the
word
second
over
here
and
I'm,
looking
at
it
to
my
hole,
v1
my
hole.
A
V1
is
right
there
like
that.
That's
that's
about
perfect,
that
this
size
that
I've
highlighted
right
there.
That's
my
v1,
where
I'm
standing
right
now
so
I
can
scan
my
v1
all
over
and
there's
a
bunch
of
macro
columns
in
v1
right.
So
all
of
them
are
working
together
to
recognize
the
individual
letters,
but
also
the
words
I
mean
because
that
that's
part
of
it
too,
you
have
column,
oh
god
hold
on
hold
on
you.
Gotta
meet
my.
A
A
Audio
separated
out
so
I
can
have
desktop
audio
and
and
mic
audio
anyway,
what
I
was
saying
about
yeah
I'll
just
keep
reading,
because
I
think
you're
I
think
you're
on
the
right
track
there.
This
macro
column
recognizes
a
letter
and
the
next
macro
column.
The
neck
recognizes
is
a
different
letter.
A
They
are
forming
a
guess
that
this
is
part
of
a
pattern,
and
this
depends
because
you
can
focus
on
a
letter
or
a
word,
usually
when
we're
reading
we're
folk,
not
not
letters
but
they're,
forming
a
guess
that
this
is
part
of
a
pattern,
in
this
case,
a
word
that
they
have
learned.
So
these
two
columns
are
voting
on
a
two
letter.
Diagraph,
the
the
second
and
third
macro
columns
are
likewise
voting
on
a
different
two
letter
diagrams.
This
process
is
happening
all
over
the
entire
visual
field.
A
At
the
same
time,
none
of
the
macro
columns
know
they're
part
of
a
particular
word
or
phrase
just
a
little
bit
that
they
see
right.
So
if,
for
example,
one
one
of
the
macro
columns
is
just
covering
is
just
covering
one
letter,
it
might
only
be
able
to
recognize
that
one
letter
and
there's
another
beside
it
that
might
be
recognizing
a
different
letter
and
now,
in
that
case,
there
they're
gonna
be
operating
on
different
objects.
A
C
A
Why
I
would
play
it,
but
whenever
we're
close
to
AG
I
will
play
it?
Okay,
they're
forming
a
guess
at
this
part
of
the
pattern,
a
word
that
they've
learned.
So
these
two
columns
are
voting
on
a
two
letter
diagram.
Second
third
map
of
columns,
likewise
voting
on
a
different
two
letter
diagraph.
This
is
this
process
happening
all
over
the
whole
visual
field.
None
of
the
macro
columns
know
they're
part
of
a
particular
word
or
phrase
just
a
little
bit.
They
can
see.
A
Jeff
Hawkins
described
this
as
looking
at
the
world
through
a
straw
yeah.
The
larger
local
group
of
macro
columns
rapidly
settle
on
some
pattern
that
we
might
consider
a
representation
of
a
word
or
phrase.
I
think
that's
a
pretty
fair
representation
of
of
what
Jeff
thinks
yeah
for
them
for
them.
I'm
not.
C
A
Sure
about
this
exact
part,
the
second
advert
you
know
the
first
and
second
versus
versus
the
second
and
third
excuse
me:
cuz
I,
keep
thinking
you
know.
First,
there's
there's
word
recognition,
then
there's
letter
recognition
do
I,
have
a
Marvin
t-shirt,
no
I,
don't
I
only
have
Skynet
and
how
I
had
I
used
to
have
another
one,
but
I
think
I
wore
it
out.
The.
A
I'm,
just
thinking
about
the
idea,
like
imagine
you
saw
the
word
second
here
and
and
imagine
that
this
see
was
really
an
A
or
something
like
that,
and
so
you,
you
might
just
breeze
over
it
and
not
even
notice
it
right.
That's
just
stuff,
that's
something
that
your
brain
does
you've
seen
those
you
know
those
those
things
where
they
say.
Can
you
read
this
and
it's
all
jumbled
up,
but
you
can
everybody
can
still
read
it
just
because
it's
they
replace
things
with
things
that
look
somewhat
like
what
they
they
should
look
like.
A
So
that
tells
you
there's
obviously
word
recognition
going
on
and
and
maybe
when
you're
in
that
mode,
you're,
maybe
you're
operating
it'll
lay
a
level
up
I'm
talking.
This
is
I'm
not
talking
for
the
research
team
here
I'm
talking
about
how
I
think
this,
but
maybe
when
you're
doing
that
type
of
level
reading
and
you're
not
focusing
on
the
detail,
you
are
sort
of
operating
in
a
mode.
That's
a
little
bit
up
up
the
hierarchy,
I'm,
not
sure,
but
so
I'm,
not
certain.
But
overall
this
totally
makes
sense.
A
That
you're
moving
through
an
object
and
so
you're
recognizing
that,
as
you
move
through
the
space
of
a
word
or
whatever,
that
it's
a
consistent
pattern,
the
stable
constellation
pattern
could
persist
for
several
I
or
hand
finger
movements,
building
up
longer,
word
groupings
as
you
learn
more
patterns,
yeah
keep
in
mind
that
each
macro
column
can
Center
on
a
different
mini
column
that
was
the
winner
and
recognizing
the
local
pattern.
This
is
where
I'm,
not
so
sure
about
this
is
sort
of
so
I
think
this
is
again.
This
is
an
interesting
idea.
A
This
is
one
of
the
things
that
that
mark
Brown
has
been
one
of
the
ideas.
That's
crucial,
I,
think
central
to
like
the
idea
of
Calvin
text
grids
right,
mark
I
think
that's.
This
is
like
the
thing
that
links
it
to
HTM.
If
you
can
think
of
a
mini
column
as
sort
of
an
originator
based
on
lateral
connections
of
of
the
hex
grid,
pattern
or
or
a
cortical
column
could
be
based
around
any
mini
column
may
be
something
that
moves
or
isn't
so
discretely
defined
in
flesh
right.
A
So
I'm
not
sure
about
this
I
think
it's
an
interesting
idea,
but
I
know
that
Jeff
doesn't
think
this
is
true.
I
think
he
really
thinks
there
are
these
physical,
these
physical
macro
columns,
and
that
many
columns
operate
within
them
in
the
million
cortex,
this
macro
column,
neighborhood
yeah
to
20
to
50,
microns,
Mike,
Michael,
yeah
micro
columns.
This
means
the
center
of
a
macro
column
is
not
a
fixed
place.
It
depends
on
the
pattern
that
is
sensed.
A
A
We
I
think
it
I
think
it's
very
possible
that
the
grid
cell
mechanism,
that
we're
always
talking
about
and
the
bumps
that
we're
talking
about
grid
cell
bumps
it's
very
possible
that
those
are
constantly
moving
through
the
cellular
space.
That
hexagonal
pattern
is
definitely
moving
and
I
think
it's
moving
in
response
to
either
motion
or
the
environment
moving
or
you
moving
or
something
like
that.
A
It's
so
so.
I've
been
missing
chat,
sir,
about
that
at
least
they
haven't
been
popping
up.
But
what
is
the
relation
to
many?
What
is
the
relation
to
many
columns
in
HTM
Mini,
two
columns,
there's
nothing
that
fixes
a
comment.
No
there's
nothing
in
our
theory!
No
I,
don't
think
so.
So
I
mean
it's
not
something
that
is
really
worth
arguing
about
right,
I,
just
I
can't
I
can't
resolve
it
in
my
head
with
with
our
theory,
but
there's
nothing
that
says.
C
A
A
It's
I'm
gonna
assume
this,
so
the
center
of
a
macro
column
is
not
a
fixed
place.
It
depends
on
the
pattern
that
is
sensed.
This
is
true
for
all
macro
columns,
so
the
Alpha
pattern
is
not
fixed
to
a
rigid
location
or
grid.
This
is
true
for
okay,
okay,
the
output
is
the
constellation
of
macro
columns
that
one.
C
A
This
competetive
process,
so
the
so
you're
describing
a
macro
column
competition.
It
is
this
similar
to
is
this
through
voting
the
macro
column,
competition,
each
global
input
state
would
result
in
a
collection
of
local
output
pattern
of
bits.
I
have
described
this
as
a
constellation
of
stars.
I
think
that's
an
apt
description.
Every
learned,
input,
pattern
or
sequence
results
in
a
different,
stable
constellation.
A
Each
global
input
state
would
result
in
a
collection
of
local
output
pattern
of
bits,
so
I'm
gonna,
I'm
gonna,
keep
rereading
this
in
the
keep
you
reading
this
until
like
until
I
understand
it.
So
I
get
the
center
of
a
macro
problems,
not
a
fixed
place.
It
depends
on
the
pattern
that
is
sent
so
in
response,
so
yes,
and
they
formed
from
the
winning
mini
columns
and
local
inhibition.
Okay.
A
So
in
response
to
some
raw
sensory
perception,
many
columns
activate
and
you're
saying
that
from
those
men
and
winning
mini
columns
in
some
form
of
local
inhibition.
These
macro
columns
can
then
then
emerge
from
at
least
one
of
them.
I
mean
you
only
need
one
mini
column
right
and
then
you
apply
the
hex
pattern,
like
you
like
you've,
written
about
in
other
posts,
I
think
so
so
you're
saying
a
sensation
can
cause
a
different
pattern
of
macro
columns
to
emerge
base.
A
A
What
would
the
macro
column
mean
again?
Okay,
that
I
can
I
can
I
can
visualize
that
occurring.
I
can
visualize
that
occurring
each
global
input
state.
When
you
say
global
input.
State
you
mean
the
whole
sensory
field.
Does
that
mean
like
contextually
everything
that's
happening
around
you
at
once
would
result
in
a
local
output
pattern
of
bits
which,
which
represents
that
global
input,
State
right,
okay,
okay,
yeah
I
can
see
that
I
can
see
that.
A
See
it
says
so:
here's
I
haven't
I
know
you
do
you
have
I
gotten
to
the
main
point,
because
I
get
this
I
get
this
now
now
I'm,
trying
to
figure
out
how
this
matches
with
my
understanding
of
how
we're
constructing
a
cortical
column,
because
we're
like
building
it
up.
No,
because
I'm
going
to
I
mean
we
want
to
understand
all
the
different
layers
and,
and
so,
if
we're
gonna
say:
okay,
it's
not
okay,
I
would
assume.
So
you
get
a
constellation
of
bits.
That's
an
object
right.
A
That
makes
sense
if,
assuming
all
the
sensory
input
is
focusing
on
one
thing
and
that
that
we're
taking
that
huge,
dense
representation
and
we're
narrowing
it
down
to
mini
column,
activations
narrowing
it
down
to
macro
column,
activations,
and
there
you
have
a
connection.
Essentially
that
represents
the
objects
across
all
of
your
macro
columns,
each
one
having
different
mini
column,
activations.
That
makes
sense
to
me
totally
whether
or
not
they're
fixed
in
place.
Does
it
matter.
Are
you
saying
that
that
may
be
the
mechanism
by
which
they
emerge?
I?
A
This
general
process
is
happening
all
levels,
so
higher
levels
are
perceiving
and
voting
on.
The
groups.
Groupings
formed
by
lower
levels
keep
in
mind
that
this
is
not
strictly
a
pipeline,
as
there
are
huge
numbers
of
fiber
tracks
crossing
up
and
down
these
hierarchies
and
between
processing
streams.
It's
not
strictly
a
pipeline.
A
A
So
aren't
these!
So
if
this
okay
I
know
I've
been
just
rambling,
but
something
just
popped
in
my
head,
if
we,
if
all
my
senses,
converge
on
an
object
at
once
for
whatever
reason-
and
so
this
object
is
somehow
now
projected
across
my
cortex
and
in
a
sense
because
of
these
sensory
sensory
inputs
that
have
activated
many
columns
I'm,
assuming
that
the
same
input
wins
the
same
alpha
pattern
yeah,
you
have
to
assume
that
that's
it's
crucial.
A
So
let
me
let
me
think
about
something
here:
how
do
these
different
hex
patterns
that
might
emerge
in
different
places
in
your
brain
say,
for
example,
I'm
going
to
touch
a
cup
on
this
side
and
then
a
couple
on
this
side?
I've
got
activations
happening
on
both
sides
of
my
brain
and
and
in
this
with
thinking
along
these
lines.
A
Mark
each
side
is
going
to
have
an
emergent
pattern
emerge
from
sensory
input
to
mini
column,
activations,
to
columns
macro
columns,
there's
more
keep
going,
okay,
you're
talking
to
me
right,
but
let
me
just
talk
through
this
because
I
think
I'm
either
understanding
something
better
or
I
have
a
question
that
I
need
to
answer.
I
have
the
answer
it's
so.
If
I've
got
one
activation
of
hex
grid
over
here
and
another
one
over
here,
how
can
they
converge
to
the
same
object.
B
A
A
Question
in
and
and
in
an
hour
I
guess
you
could
just
vote
I
guess
you
could
apply
the
voting.
You
just
apply
the
lateral
voting
to
converge
on
them
on
the
object,
and
we
could
do
this
using
this
model
that
you're
you're
talking
about
yeah.
Okay,
we
could
use
that
MA.
It
doesn't
really
matter
how
we're
describing
the
column
necessarily
I
mean
this.
A
This
this
could
be
I
still
keep
thinking.
I'm
went
in
the
wrong
direction,
except
I'm,
pointing
to
whoops
I
keep
pointing
over
here.
This
could
be
useful,
I
feel
like
once.
We
understand
how
this
circuit
works,
because
then
we
can
emerge
the
circuit.
You
know
what
I
mean
we
can.
We
can
figure
out
a
way
to
take
a
big
substrate
of
cells,
a
huge
population
of
cells
and
and
then
have
a
way
to
to
cut
it
up
into
these
cortical
processing
units.
A
Is
that
one
of
your
points
mark,
because
that
makes
sense
that
would
that
seems
like
something
that
could
be
applicable
down
the
road?
Okay
he's
saying
keep
reading,
so
that's
general,
that's!
Okay!
Great!
This
general
process
is
happening
at
all
level
levels,
so
higher
levels
are
perceiving
and
voting
on
groups
for
my
lower
levels.
This
is
not
a
pipeline.
A
huge
number
of
fiber
tracks
crossing
up
and
down
the
hierarchy
is
between
them.
Yeah.
That's!
Oh
I,
get
what
you're
saying
it's
not
a
pipeline!
Yeah
messy!
It's
messy!
A
A
Is
what
makes
it
beautiful
the
messiness
is,
what
makes
it
redundant
and
resilient?
You
know,
you
don't
know,
you
know
you
don't
know
where
your
input
might
be.
Coming
from
me,
you
have
to
you,
have
to
generalize
everything
is
forced
to
generalize.
It
feels
like
since
you're
perceiving
your
entire
environment.
At
the
same
time,
these
perceptions
are
likely
different
aspects
of
the
thing
you
are
perceiving
you
think
of
it
as
like
the
room,
the
room
that
you're
in
or
the
object
that
you're
interacting
with
right.
A
A
Yes,
we're
talking
about
I'm
personally,
pursuing
a
slightly
different,
take
mark
I
thought
that
was
a
great
explanation
by
the
way
I
think
you're
I
think
you're
on
track.
This
makes
sense
to
me
and
I
know
we
have
differences,
especially
in
resolving
the
the
calven
stuff
and
the
hex
grids,
but
I
like
this
idea
and
and
again
I
think
we
came.
I
talked
about
this
at
the
very
end
of
the
last
hex
grid
post.
That
I,
like.
A
Know
how
to
apply
it
yet
I
think
this
could
really
become
applicable
when
we
finally
get
to
the
when
we
have
the
compute
power
to
simulate
a
big
substrate
of
cells,
and
how
can
we
split
them
up
essentially
in
real
time
live,
you
know
into
computing
units,
because
you
don't
necessarily
want
to
predefined
your
computing
units
right.
You
want
him
to
be
super
flexible,
so
it
would
make
sense
to
me
if
these
really
weren't
predefined
and
it
went
just
copied
it
was.
A
B
A
This
is
a
big
post
if
you're
interested
in
this,
and
this
idea
of
hexagonal
grids
at
a
sort
of
macro
column
level,
read
this
post,
it's
very
it's
it's
good,
there's
a
lot
of
good
visuals,
some
of
them
most
of
them
come
from
this.
It's
an
old
book,
the
Cerebral
code
by
William
Calvin,
but
then
mark
really
illustrates
it.
Well,
as
you
go
along
with
his
own
stuff,
I,
don't
know
what
you
used
to
do
your
illustrations
mark.
Let
me
find
some
of
them.
Oh
coin
is
like
nothing
loading.
This
was
chalked
full.
A
It
used
to
be
chalked
up
there.
Oh
there,
we
go.
Here's
some
so
here
here
he's
talking
about
these
lateral
connections
between
you
know
where
a
mini
column
might
originate
how
far
it
can
reach
and
where
other,
where
it
might
reach
potentially
others
that
are
also
sort
of
resonating
because
of
that
input
and
making
links
with
them.
So
he's
got
all
this
really
drafted
out.
A
Well,
so,
if
you're
interested
in
that
take
a
look
there,
oh
I'm
sorry
good
grief,
not
even
showing
my
test
I'm
trying
to
show
these
great
pictures
that
marked
it
and
I'm,
not
even
showing
my
desktop.
So
this
is
it
the
HTM
many
columns
to
hexagonal
grids
I
will
put
a
link
to
it
and
chat
like
I
said:
I
think
this
is
really
interesting
and
and
I
think
it
jives
with
the
stuff
that
we've
been
doing
to
at
Numenta,
I.
A
Think
I
think
it
really
does
I,
don't
think,
there's
really
anything
that
conflicts
with
it
I
think
they
work
together.
Well
and
I
know.
Mark
has
been
trying
to
convince
Jeff
to
pursue
this
idea
and
I've
pitched
it
to
him
occasionally.
But
Jeff
is
really
focused
on.
He
wants
to
know
what
that
circuit.
Does
he
wants
to
know
what
that
circuit
he's
not
gonna.
A
Some
things
I
had
not
thought
about
at
all
and
then
I
talked
to
sue
batai
again
about
the
thalamus
and
his
co-sign
poster
on
relay
cells
in
the
thalamus,
so
very
technical,
so
I'm
gonna,
edit
those
podcasts
and
you'll,
see
those
soon
hey,
Tatiana
I
do
appreciate
that
you
changed
your
name,
that's
very
cool
of
you.
Jeff
mentioned
he
was
thinking
about
1
dimensional
grid
cells,
yeah
he's
always
thinking
about
that
anytime.
We
have
a
problem.
That's
and
the
answer
is
two
dimensional
grid
cells.
We
always
think
can
we
do
it
with
one
dimension?
A
And
the
answer
is
always:
yes,
it's
it
seems
like
you
can
solve
any
dimensional
problem
in
space
with
a
combination
of
one
dimensional
grid
cells.
So
it's
still
on
the
table.
The
idea
of
one
dimensional
grid
cells.
Okay,
thank
you
mark
sorry,
okay,
oh
yeah,
and
this
we
also
have
this
long
thread
where
I
expressed
a
lot
of
my
thoughts
about
this
idea.
This
will
probably
blow
your
mind.
But
how
does
this
information
come
together
and
make
decisions
and
initiate
actions?
I
maintain
that
the
cortex
contents
are
shared
with
these
subcortical
structures,
the
lizard
brain?
A
Let's
get
the
brain
out
here,
get
the
brain
out
the
subcortical
structures
in
the
midbrain
I
assume
there
we
go
I'm
going
to
leave
this
cerebellum
out.
It's
boring!
It's
not
boring!
I'm!
Sorry,
it's
the
BRIT!
No,
no
parts
of
the
brain
are
boring.
I'm
gonna
finish
this
paragraph,
so
these
older
parts,
in
an
evolutionary
sense
and-
and
you
know,
I
asked
Jeff
specifically-
is
the
thalamus
older
than
the
neocortex
in
this.
In
this
past
podcast,
and
he
said
you
know,
I
really
don't
know
it's
a
hard
question
to
answer.
A
A
The
cortex
contents
are
shared
with
the
subject:
cortical
structures,
certainly
through
the
valve
through
the
thalamus,
and
then
these
older
parts
decides
and
directs
actions
through
projections
to
the
forebrain
okay
through
the
thalamus
I
bet.
That's
coming
from
the
balance,
I
suppose
that
if
you
could
say
that
there
was
an
executive
spot
in
the
brain,
it
would
turn
out
to
be
the
flam
ik
nucleus.
Absolutely
I
think
that
modeling
this
area
will
end
up
being
the
part
that
finally
makes
a
I
react
in
a
way
that
we
consider
having
intelligence
hi-five
mark.
That's
absolutely
right!
A
That
is
exactly
the
track
that
we're
heading.
We
are
really
focused
right
now
on
the
thalamus
and
trying
to
figure
out
how
it
is
controlling
attention,
specifically
how
how
does
it
shed
light
on
in
within
the
cortex?
How
does
it
put?
How
does
it
direct
the
cells
and
so
you're
gonna
like
these
next
two
podcasts
cuz
I
just
talked
about
this
exact
stuff,
with
sue
Bataille
and
Jeff.
A
A
So
the
problem
with
distributed
network
solutions
is
that
we
need
local
connectivity,
so
you'd
have
to
have
all
your
local
connectivity
within
each
node
of
the
distribution
and-
and
if
we're
talking
about
this
type
of
system,
where
the
structure
of
the
processing
units
is
inherent
there
or
is
somehow
invoked
by
the
input,
it
seems
impossible
to
distribute
it
across
nodes
in
a
physical
computer
network
unless
we
have
very
high
elasticity
or
plasticity
in
the
system.
So
we're.
What
we
really
need
is
plastic
chips.
A
We
need,
we
need
neuromorphic
chips,
chips
that
can
grow
somehow
or
somehow
simulate
the
growth
of
permanence
--is
between
synapses
of
neurons.
That's
that's
what
we
really
need
as
far
as
hardware,
so
the
limitation
in
size
that
we
have
right
now
is
absolutely
do
to
compute
power
and
hardware,
and
also
just
that
we
don't
have
the
right
type
of
hardware.
So
that's
a
big
thing
too
you're,
looking
forward
to
the
podcasts
cool
I
need
to
edit
them
they're
on
my
list
of
things.
To
do.
A
A
A
It's
very
exciting
and
and
I'm
it's
taken
so
long
to
come
up
with
a
picture
of
what's
happening
in
one
cortical
column
like
what
are
these
layers
doing
and-
and
it
feels
like
we're
so
so
close,
because
the
grid
cells
was
a
huge
huge,
unlocking
of
so
many
things
and
understanding
that
that
that
grid
cells
and
how
they
map
space
and
how
the
cortex
could
be
using
it
to
compute
and
and
model
was
just
opened
up.
So
many
possibilities,
and
it
just
feels
like
it's
going
very
quickly.
Now
the.
B
A
Is
going
very
quickly,
so
falco
used
to
have
a
friend
who
designed
cpus
for
a
living
I
I
didn't
so
that
still
sounds
as
though
software
solution
would
be
able
to
simulate,
given
at
a
much
slower
pace
to
prove.
Oh
yeah.
Definitely
we're
constantly
building
software
simulations
at
Numenta
to
test
these
theories.
All
the
theories
like
if
we
have
a
paper,
that's
out,
we
have
software
theories.
We
have
software
running
simulations
to
test
that
the
theories
do
what
we
think
they're
they
do
bit.
Kings
got
a
lot
of
hardware
experienced
bit.
A
A
Where
do
the
layers
doing
and
this
guy
we
had
in
the
office
just
totally
was
infatuated
the
same
questions
that
we
have
been
researching
for
years,
and
so
it
was
a
great
meeting
and
I
was
so
exciting
to
see
somebody
that
was
so
excited
to
see
what
we
were
working
on,
because
we
just
spoke
the
same
language,
so
I
just
sort
of
observed
that
I
was
on
the
double.
The
best
thing
is
that
this
does
not
need
hex
grids
to
be
useful
to
HTM.
That's
true,
yeah,
that's
true!
It's
that's
true!
I!
A
Don't
think
about
that.
The
hex
grids,
if
that
works
out
like
a
bonus
on
top
of
everything
that
could
help
just
connect
everything
together,
right
I
did
not
interview
him
or
her
for
the
series.
I
didn't
have
time
to
it
was
like
he
popped
in
the
office
once,
but
I
have
a
feeling
that
you'll
you'll
hear
more
from
this
particular
neuroscientist.
A
A
Yeah
ask
an
off
topic
question:
if
it's
fun
Friday
and
let
me
let
me
check
yeah,
it
is
it's
fun
Friday
you
can
add
you
can
we
can
go
off-topic
whatever
you
want,
I'm
only
going
to
be
strict
with
off-topic
chat
on
on
work
sessions,
which
is
like
Tuesday
and
Thursday?
Okay.
Well,
then,
no,
if
it's
a
boring
question,
I,
don't
want
to
hear
it,
but
you're
gonna.
Ask
it
anyway
right.
A
It
may
be
an
interesting
answer.
Some
boring
questions
have
really
interesting
answers
like
what
did
you
do
last
night,
mine
was
not
I
mean
mines,
I
didn't
have
a
I
didn't
have
an
exciting
night.
Last
night,
I
sat
here's.
What
I
did
I
worked
on
I
tried
to
work
on
that
research
map
that
was
written
yesterday?
I,
don't
think
I'm
gonna
put
you
guys
through
that
anymore.
That
was
not.
That
was
a
boring,
it's
kind
of
a
boring
project
and
it
has
nothing
to
do
with
HTM.
It's
just
like
a
web
post.
A
When
will
truly
intelligent
machines
be
a
reality
and
with
the
winning
vote
is
25
years?
My
answer
to
your
question
is
no
no
and
it's
more
like
hell.
No,
not
no,
but
hell.
No
I
would
be
so
surprised
if
we
had
anything
near
AGI.
I
do
think
that
when
we
sort
of
slip
past
that
that
border
things
will
happen
pretty
quickly
but
I,
don't
think
we're
close
to
having
it
I'm
thinking
more
like
20
years,
20
years,
maybe
15
20
years.
A
Something
like
that,
but
who
can
say
I
mean
who
can
say
it's
hard,
this
hard
stuff
and
I
can
tell
you
one
thing:
it's
not
gonna
the
it's
not
gonna,
be
neural
networks.
Oh
no!
No!
No!
No!
It's
definitely
going
to
be
neural
networks.
It's
not
gonna,
be
deep
learning,
it's
not
gonna,
be
deep
learning
and
because
deep
learning
doesn't
give
you
this
rich
layers
of
this,
this
layer
doing
that
this
layer
doing
that
and
all
the
different
interconnectivity
that
you
actually
have
in
your
brain.
A
We
know
it
doesn't
like
there's,
certainly
Bayesian
properties
of
statistics
and
stuff
that
are
that
are
affecting
everything
in
your
brain
all
the
time
you
know
it's
a
population
effect
of
having
millions
and
millions
of
neurons
and
stuff
operating
at
the
same
time,
you're
going
to
be
able
to
look
at
it
from
a
Bayesian
lens,
but
you
can
look
at
anything
with
a
bee
lenz
and
that
doesn't
necessarily
tell
you
how
it
works.
You
have
to
know
how
it
works.
A
So
deep
learning
doesn't
tell
you
how
it
works,
and
even
some
of
the
people
that
are
deep
into
deep
learning
have
a
hard
time
explaining
how
it
works
it
with
HTM
and
all
of
our
theory,
and
you
look
at
the
videos
and
the
papers
and
everything
we're
starting
from.
How
does
it
work?
We
want
to
know
how
it
works
and,
from
that
point
on
once
we
understand
how
it
works,
we'll
build
something
with
it.
Everything
we've
built
is,
after
understanding
how
it
works.
A
15
says
hex
grids
and
thalamus
for
the
AI
win.
Definitely
definitely
that
direction.
Yes,
you
mark
says
he's
working
on
this
right
now
so
five
years.
That's
awesome.
Look
good!
Good
luck!
I
would
be
happy
if
it
happened
in
five
years,
so
you
saw
the
poll
I
wanted
to
hear
my
opinion.
Yeah,
it's
I
think
it's.
If
I
think
it's
2015
twenty
years
out.
We
will
achieve
it,
though,
because
I
mean
it's
knowledge
and
it's
there
it's
accessible.
It
obviously
works.
Somehow
it's
not
magic,
so
we'll
figure
it
out.
Okay,
your
boring
question.
A
I
when
you
say
you
guys
it's
really
just
me,
I'm
the
one
jumping
between
JavaScript
and
Python,
which
I
should
learn
if
I
want
to
understand.
Oh,
if
you
want
to
understand,
learn,
Python
learn,
Python,
yes,
learn!
Python
JavaScript
is
nice,
but
it's
not
if
you
want
to
be
a
machine
learning
and
if
you
want
to
learn
the
language
of
the
field,
learn
Python
for
sure.
A
A
Be
in
Python,
with
C++
optimizations,
all
right,
hey!
Well,
nice
to
see
you,
let's
go
back
to
the
project
at
hand
now
that
we
finally
have
or
passed
the
the
forum
stuff.
Let
me
check
my
my
viewer
things.
This
work
we're
supposed
to
be
able
to
press
a
button.
It'll
tell
me
how
many
people
are
watching,
but
it
doesn't
work
all
the
time.
A
I'm
only
writing
in
JavaScript
when
well
here's
the
thing.
Falco,
you're,
gonna,
learn
the
logic,
no
matter
which
language
I,
use,
writing
and
reading,
or
two
different
things.
You
can
easily
read:
Python
or
JavaScript
that
shouldn't
be
a
big
problem
once
you
know
the
syntax
of
the
it's,
it's
easy
once
you
know,
you
know
how
programming
works,
they're,
basically,
pseudocode,
JavaScript
and
Python
are
dynamic
enough
that
you
can
read
through
them
and
and
if
the
codes
written
well
enough,
that
you
can
read
it
like
pseudocode,
so
I,
don't
think
it
matters.
A
What
I
write
it
matters,
what
it
matters,
what
you
choose
to
write
and
I
would
suggest
Python
if
you
want
to
work
in
the
machine
learning
space
for
the
AI
space,
so
yeah,
Python
and
JavaScript
are
pretty
close
in
style,
different
spelling,
that's
just
something
become
a
polyglot
programmer.
As
my
my
advice
to
you
here
comes
in
a
beta
with
the
emotes
how's,
it
going
good
to
see
you
it's
going
better.
Today,
it's
fun
Friday,
so
I
didn't
put
any
productivity
restrictions
on
me
or
anything.
So
we're
just
chatting
talking
about
what
we
were.
A
A
You
helped
me
out
with
python
at
some
point,
I
think
he
says,
he's
fluent
and
c,
plus,
plus
and
QT,
but
I
know
yeah.
I
can
understand
all
that
stuff.
Yeah.
Those
languages
are
pretty
easy
to
understand,
I
think!
What's
on.
What's
on
a
higher
priority
list
on
Dementors
research,
the
thalamus
or
place
cells?
A
That's
a
good
question.
We've
been
researching
place
cells
for
a
long
time,
so
I
think
play
cells.
Probably
those
thalamus
is
sort
of
weird
just
now
moving
towards
thalamus,
but
we've
definitely
been
thinking
about
play
cells
for
a
long
ever
since
we've
thought
about
grid
cells,
because
when
you
think
about
grid
cells,
a
lot
of
stuff
comes
along
with
that
head.
A
Direction
cells,
band
cells,
border
cells,
there's
a
bunch
of
different
cells
that
are
conjunctive
that
detect
certain
aspects
of
space
that
seem
to
be
conjunctive
of
other
cells,
working
together,
other
layers
of
cells
or
something
working
together.
It's
not
fully
understood,
but
play
cells
being
one
of
those.
That's
like
a
landmark,
you
know,
helps
identify
things
by
a
particular
landmark
or
something
film.
Tire
says:
what
do
you
think
about
consciousness?
A
It's
like
free!
It's
like
freewill,
a
bit.
It's
a
hard.
It's
a
slippery
slope
once
you're
talking
up
once
you
start
talking
about
it,
I
think
I
think
consciousness
may
end
up
to
be
nothing
special.
Maybe
I
want
to
take
that.
Hey!
Yes,
do
that
post
me,
the
forum
mark!
You
want
to
talk
about
consciousness.
Okay,
if
I
got
a
place
for
you,
Mark's
gonna,
post
it
into
chat.
Unless
you
want
me
to
find
it,
you
might
be
on
your
phone
or
something
but
there's
a
there's,
a
lot
of
forum
posts.
A
B
A
Or
goes
some
yeah
I'm
gonna
find
that
I'm
gonna
find
that
post.
It's
really
easy,
I
think
I.
Think
here
just
do
this
consciousness
in
general,
you
can
just
search
for
consciousness
on
HTM
form
and
you'll.
Find
something
interesting
and
but
I
think
mark
has
something
as
a
specific
post
and
put
this
by
bit.
A
King
there
he
is
Serge,
but
that's
it
it's
the
thought
experiment
one
or
is
it
the
not
what
you
think
either
one
either
way
you'll
find
you'll
find
some
interesting
conversation
on
consciousness
from
Mark
Brown
right
right
here
on
HTM
forum,
okay,
sorry,
cookie,
dough!
That's
what
Descartes
said:
coke
coal
Guiteau
ergo
sum.
What
does
that
mean?
I
am
Not
sure
that
I
that
I
exist
I'm,
not
sure
I'm,
conscious
or
I'm,
not
conscious.
A
All
right,
okay,
film
tire
can't
say
for
myself.
I
believe
consciousness
is
simply
a
useful
contract.
Construct
is
about
Union
and
that
helped
us
survive,
not
a
real
and
unique
mechanism.
It's
definitely
something
that
helped
us
survive.
Everything
that's
going
on
up
here
is
was
evolved
to
help
you
survive
everything.
A
Now
it
might
not
play
the
same
role
as
it
used
to
anymore,
because
society
has
changed
drastically
in
the
past
few
hundred
years
and
a
lot
of
those
mechanisms
don't
return
the
same
benefits
as
they
use.
I.
Think
therefore,
I
am
of
course.
Oh
here
we
go.
Here's
here's
the
forum,
post
consciousness
versus
intelligence,
now
I
said
I'm
not
gonna,
get
into
it,
so
I'm,
not
I'm
gonna.
Let
you
guys
read
that,
because
I
don't
want
to
talk
about
consciousness
on
fun,
Friday,
it's
just
it's
not
fun!
For
me,
I!
Don't
it's
too
much!
A
A
A
It's
the
same
question,
it's
hard,
it's
hard
to
say,
ask
it,
you
can't
ask
it
that's
the
thing:
how
can
you
know
how
can
you
know
hold
on
I
gotta
start
using
these
buttons?
How
can
you
know
something
is
conscious
unless
it
can
declare
that
it's
conscious,
I,
don't
know
I
I
am
gonna,
have
trouble
declaring
anything
conscious
until
it
gives
me
some
indication
that
it's
conscious,
I
guess
a
cow,
can't
declare
that
it's
conscious.
So
is
it
that's
a
question?
Do
we
have
to
have
a
common
language?
A
A
Merge
jeaious,
object
project
with
main
okay,
so
I'm
gonna
do
this
I'm
gonna
merge
the
JavaScript
object
project
with
main
and
then
we're
gonna
go
code,
yeah,
I'm,
gonna,
start
I'm,
gonna,
start
coding
and
getting
some
fun
stuff
done
here,
because
I'd
like
to
see
this
work,
let's
go
to
my
let's
just
dump
it.
Let's
just
move
it
right
over.
Oh
I'm,
actually
running
it
right
now.
Here's
the
here's,
the
JavaScript
project
that
we're
running-
and
this
is
basically
just
displaying
a
couple
of
sample-
objects
that
we're
creating
as
a
part
of
an
object,
schema.
A
That's
going
to
be
a
part
of
an
object,
recognition
tasks
that
we're
gonna
build
in
gherkin.
Oh
man,
oh
yeah,
yeah
I
forgot
about
this
I
forgot
about
this
thanks
for
reminding
me.
This
is
a
will
he's
been
helping
me
with
this
helping
on
on
both
of
these
projects.
So
so
in
this
one
looks
like
you
have
jQuery
matches
the
dependency,
we're
adding
jQuery
the
proper
way.
All
right
so
we're
yeah
yeah,
makes
sense
to
me.
A
Okay
and
then
just
parcel
static
just
want
to
make
sure
that
it
runs
for
me
and
I'm
really
only
doing
this,
because
every
once
in
a
while
I
forget
to
check
in
a
dependency
that
should
work
and
so
everything
works,
and
so
it's
always
good
to
download
it
and
run
it
before
you
merge
something
just
in
case,
although
it's
not
like
we've
got
a
big
production
team.
It's
only
a
couple
lines.
Yeah,
so
I'll
merge
that
no
big
deal
I,
don't
think
I
need
to
review
it.
A
I'm
just
going
to
merge
it
merge
it
I
should
get
it.
I
should
have
emerged.
A
merge,
sound
to
merge,
pyaare,
merge,
sound
I'm
thinking
about
ways
of
enhancing
my
twitch
stream,
so
I'm
thinking
about
the
different
sounds
and
things
I'm
gonna
end
up
adding
at
some
point.
Okay,
so
back
to
the
code
check
out
master.
A
B
A
B
A
A
A
B
A
A
A
A
She
moves
in
mysterious
ways:
I
got
it,
although
I
might
get
I
know
this
video
be
demonetised
because
of
this,
and
it
might
not
even
be
you
know,
I
better
turn.
This
off
I
got
to
be
careful
with
this.
If
I'm
playing
like
like
top
40
type
stuff,
the
record
companies
will
will
make
this
video
they
won't.
Even
let
me
show
it
on
YouTube,
so
I
think
I,
better
rethink
that
I
start
playing
the
more
eclectic
things
like
mellow
beats.
A
A
Okay?
So
now,
I'm
gonna
have
to
reconfigure
this
Python
project,
because
this
is
no
longer
valid
because
I
moved
it.
So
that's
fine
project
interpret
invalid,
so
I
need
to
delete
it.
Somehow
this
is
invalid,
so
I'm
going
to
delete
it
now,
add
another
one.
Add
local
and
it's
gonna
be
in
Python
the
env.
A
D
A
A
A
A
I
should
probably
also
deactivate
which
Wow
what's
gonna
happen
now
I
think
I
just
removed
that
still
works.
Okay,
it
must
have
done
it
tricky
in
the
tricky
way,
all
right:
new
environment,
Python,
ve,
env,
Python
interpreter
Python,
3
yep.
That's
what
I
want!
Okay
and
now
I
need
to
go
into
that
environment
and
run
the
setup
yeah
that
one
okay,
all
right
so
now,
I
need
to
go
into
that
environment
there.
It
is
source,
ven-ven,
activate
and
now
I
can
say:
pip
install
requirements,
text.
B
B
A
Missus
some
chat
stuff.
How
much
noise
would
you
need
to
add
the
channel
to
defeat
their
deep
learning?
Quite
a
bit.
I
bet
I
mean
because
it
recognizes
it
when
it's
playing
in
the
background,
even
with
mine,
even
when
I'm
not
streaming
it
through
desktop
audio.
So
what
they,
what
they
do
is
they
they
sample
portions
of
it,
and
then
they
do
that
with
all
their
songs.
You
know
they
turn
on
all
their
songs
by
sampling,
portions
of
it
and
then
creating
like
some
type
of
spectrogram
analysis.
So
they
so
they
create.
A
They
take
temporal
sequences
of
songs
and
basically
just
do
frequency
analysis
on
them
and
put
them
into
a
spatial
context,
and
then
you've
got
little
bits
of
spatial
context.
They're
like
little
bits
of
songs
that
you
can
compare
other
little
bits
of
songs
to
and
if
you
get
close,
then
close
enough,
you
know
within
the
threshold
it
seems
to
work
pretty
well
and
I.
Think
like.
If
you've
got
someone
talking
in
it,
it
doesn't
affect
the
signature
of
the
spectral
analysis
of
the
music,
apparently
in
a
way
that
would
alter
the
spectrogram
enough.
A
A
From
from
what
I
understand,
okay
requirements,
dev
text
and
how
much
noise
you
need
to
add
I,
don't
know
a
lot,
although
it's
it's
I
could
probably,
if
you
added
white
noise
it
might
it
might
really
mess
it
up.
So,
if
I
just
like
Shh,
because
additive
noise
seems
seems
to
have
effect
but
deep
networks
at
these,
this
thing
called
dropout
to
enforce
sparsity
and
that
helps
the
counter.
But
that
being
said,
I
don't
know.
Some
level
of
noise
would
probably
mess
it
up,
especially
if
it
was
tonic
in
some
way.
A
What
do
I
think
about
more
modern
music
I,
like
I
like
more
modern
music,
two
words
on
a
virtual
environment
for
a
newbie
like
me,
please
sure
so.
A
virtual
environment
is
a
Python
thing
for
the
most
part.
I,
don't
know
if
I
think
there's
something
similar
enough,
there's
probably
something
similar
in
other
languages,
but
you
have
a
system
Python
installed.
Let
me
let
me
use
my
buttons
here.
You.
A
Python
installed
it
at
the
root
level
of
your
computer,
that's
available,
basically
at
the
system
level,
for
all
programs
or
everything
to
use,
and
this
one
this
system
place
is.
Is
you
don't
necessarily
want
to
mess
it
up
mess
with
it
too
much,
because
in
Python
you
can
pip,
install
X
and
pip,
install
Y
and
pip
install
anti-gravity,
and
when
you
do
that
it
affects
the
Python
that
it's
installed
on
to.
A
A
Wait:
hi
hi
cash
I,
just
navigated
into
PI
cash
just
by
typing
PI
cash.
When
did
this
happen
in
bash
I've?
Never
I've,
never
known
this
could
happen.
Is
this
something
that's
always
been
there
and
I
just
never
knew
about
it.
So
I'm
in
my
root
directory
I
just
type
in
TA
I
go
into
my
NCA
directory
I
type.
A
A
Oh
I've
installed
the
in
virtual
environment
now
with
everything
it
needs,
so
they
can
run
the
tests.
This
is
all
just
because
I
moved
my
virtual
environment,
folder
or
I
wanted
to
move
it
out
of
the
root
directory.
So
now,
if
I
run
these
tests,
it
should
they
should
all
Pappas
great.
So
this
is
these
are
all
of
my
environment
tests.
So
far,
oh
thank
you
for
telling
me
to
switch
screen,
so
I
got
everything
working
in
Python
and
so
that,
so
this
is
a
pain.
A
A
You
have
to
use
an
alternate
Python
editor
and
alternate
webstorm
editor
javascript,
evident
editor
I'm,
going
to
use
Adam,
I
guess
because
I
don't
know
why
I
don't
know
why
so
I'm
have
one
in
one
directory:
HTM
2d
to
2d
and
JavaScript,
and
in
here
I
should
be
able
to
say
parcel
static
index
that
should
work,
and
that
should
be
fine
too.
So
both
projects
are
now
working
in
the.
B
B
A
A
B
B
C
D
A
B
A
D
C
A
A
Nobody
said
it
said
anything
about
the
Hal
9000
shirt
except
turbo,
and
he
liked
it
so
then
we
made
then
we
had
many
space,
odyssey'
jokes,
all
right.
All
right.
Nothing
has
been
done
basically
except
okay,
the
if
we
look
on
the
Trello
board.
Oh,
let
me
see
what
my
wife
says
later.
Whenever
you're
done
with
work,
great
all
right,
I've
been
you've
been
paid
to
sleep,
you
mean
the
band
sleep.
You've
been
paid
to
sleep,
oh
paid
to
sleep,
yeah
can
I
can
imagine
that
was
it
for
some
type
of
scientific
study.
A
A
Mission
tests
a
2d
environment,
so
then
we
have
to
define
first.
So
here's
some
topics
we
want
to
go
through.
You
want
to
talk
about
the
object,
schema
right
and
then
we
want
to
talk
about
the
Python
part.
I,
don't
have
very
good
names
for
this
stuff,
yet
you
know
and
then
the
JavaScript
part
and
it's
window
management
here
you
got
paid
for
slow
for
sleeping
law
because
you
were
doing
long
hours.
Yeah
I've
done
that
I
was
in
the
military.
A
Okay
object
schema!
Let's
put
that
first,
that's
the
important
thing.
So
so
we
want
to
define
I'm,
not
even
that
that
concerned
about
the
Python
and
JavaScript
parts
of
it.
Those
are
just
like
for
people
to
use
to
understand.
What's
going
on
and
and
maybe
their
libraries
there,
that
they
might
use
to
like
load
the
objects
or
create
an
environment,
you
know
to
run
their
their
AI
on
right,
so
the
object
schema
for.
First
we
got
to
talk
about
an
object
environment.
A
There
you
go
that's
better.
The
terminology
is
hard
to
parse,
because
I
want
to
talk
about
sort
of
the
object
environment
and
the
schema
at
the
raw
and
the
same
time,
and
when
I
say
the
word
environment,
it's
it's
it's
a
term.
That's
it
doesn't
really
mean
environment,
it's
the
place
in
which
the
objects
exist.
Let's,
let's
I
want
to
look
at
this
because
this
so
here's
an
object.
A
A
If
we're
talking
about
such
abstract
terms,
already,
it's
really
hard
to
create
a
more
elegant
term
for
like
object
for
a
thing
or
whatever,
but
I
do
want
to
maybe
visually
show
that
an
object,
a
2d
space
has
a
width
and
a
height
okay.
Let's
talk
about
this.
Let's,
let's
try
it.
Let's
focus
not
so
much
on
terms.
I
know:
I'm,
gonna
use.
The
word
object,
schema
it's
just
environment
that
I'm
having
the
problem
using
yeah
I,
know,
I
mean
I
can
do
synonyms
all
day.
A
Long
I
wanted
to
I
want
to
I
want
to
put
together
a
string
of
words
that
describe
what
I'm
saying
instead
of
using
the
word
environment,
I'm,
gonna,
say:
I'm
gonna,
get
rid
of
this
part,
I'm
gonna
say
objects,
2d
or
2d,
that's
sort
of
assumed,
so
objects
can
exist
within
a
2d
space
which
so
that's
true
objects
can
exist
with
it.
Let's
just
say:
objects
exist
within
a
2d
space.
A
Objects
that
well
I,
don't
know
I
think
maybe
that's
all
we
need
to
say
about
about
that
at
each
each
location
in
the
space
can
be
identified
with
an
x
and
y
integer
coordinate,
so
make
sense,
so
I'm
just
defining
the
environment
within
which
these
objects
exist.
I've
got
to
get
rid
of
this
thing
over
here.
It's
just
over.
Okay,
that's
better!
A
With
the
height
it's
gonna
be
different,
always
the
same,
they
can
be
different,
doesn't
matter
it
shouldn't
matter
so
yeah
I
would
say
just
the
fact
that
I
am
calling
out
that
it
has
a
width
and
a
height.
You
can
make
the
assumption
that
can
be
different.
All
these
examples
are
squares,
but
I,
don't
think
it
I,
don't
think
it
matters
I.
A
A
Think
when
you're,
when
you're,
defining
a
test
or
a
schema
or
something
like
that,
I
feel
like
everything
you
say
matters,
and
you
have
to
be
explicit
only
about
the
things
that
are
explicit.
You
know
what
I
mean
so
each
location
the
space
can
be
identified
with
x
and
y
coordinate.
Again,
each
location
may
have
a
feature:
okay,
I'm
gonna,
I'm,
gonna
capitalize,
the
important
terms
feature
so
then.
Let
then
we'll
talk
about
features.
Okay,
now
I'm
breaking
this
down
a
little
I'm
starting
to
feel
a
little
bit
better
about
about
this.
A
B
A
Is
a
sense?
Is
I'm
okay,
I'm
not
even
going
to
find
that
a
feature.
A
A
2D
space:
okay,
that's
pretty
clear
objects.
We
defined
as
collections
of
features
in
2d
space
features
always
contain
data.
A
feature
is
just
one
of
the
A's
in
a
box
of
a
so
it's
so
this,
for
example,
this
one.
This
is
a
feature,
that's
a
feature,
that's
a
feature,
and
this
is
just
these
features
happen
to
lay
out
in
this
shape
in
the
object
space
in
the
objects
yeah
in
the
object
space
right
so
we're
gonna,
say
every
single,
because
this
is
an
XY
coordinate.
This
is
a
different
XY
coordinate.
A
Okay,
that
feature
I'm
gonna,
say
always
contain
data,
because
why
would
you
have
a
feature
if
there
weren't
something
to
to
sense
right
so
features
always
contain
data
features?
Data
is
used
by
an
agent
to
identify
objects.
A
A
Want
to
make
I
want
I
want
to
make
features
extensible,
so
I
don't
want
to
define
anything
about
the
data
type
okay
feature
initially
I'm
just
going
to
put
initially
all
features
consist
of
a
simple
data
type,
but
should
be
extensible
to
contain
any
data
type
I'm,
gonna
I
think
I'll.
Leave
it
at
that
I.
A
Think
that
defines
it
because
right
now
we're
talking
about
A's
and
B's
and
characters
or
whatever
is
being
the
feature
data,
but
it
should
be
able
to
be
anything
we
want
it
should
you
should
be
able
to
go
to
you
know
X
4
y
4
and
not
just
get
an
a
get
a
rich
piece
of
data,
not
just
it
could
be
anything
like
a
it
could
represent
red
or
17
or
whatever
I
don't
know
so.
Features
should
be
extensible
but
should
be
extensible
to
obtain
any
type
of
data
type.
Okay,
I'm,
just
I'm
gonna.
A
B
A
A
Now
agency,
so
so
here
comes
the
challenge.
Oh,
let's
set.
Let's
talk
about
the
object
library,
the
object
schema
features.
Now
we
should
say
the
object
library
object,
library
which
we
haven't
really
created.
Yet
we
have
it.
We
have
them
here
in
objects,
but
they
really
shouldn't
be
here
right.
They
should
be.
They
should
be
up
here,
I'll
go
ahead
and
move
them,
and
then
we
can
then
I
can
say:
split
out,
object,
library
from
folder,
that's
I'm
in
progress,
I
like
to
keep
on
top
of
my
tools.
You
know
spread
this
back
over.
A
A
B
A
A
So
what
we
should
do
is
have
a
training
set
right.
We
create
some
objects
to
read
yeah,
that's
probably
what
I
should
do
so
the
attack
is
so
okay,
so
I
had
I
had
in
my
head
several
different
types
of
tasks
that
we
could
do
with
this,
but
let's,
let's
nail
down
like
the
simplest
one
first,
what
we
should
be
able
to
do
is
have,
let's
just
say,
ten
different
objects
in
the
object
library
we
could
just
label
them
0
through
9
or
whatever,
and
each
one
is
distinctly
different.
A
So
we
have
to
make
some
decisions.
Should
we
try
this
first
with
just
one
feature:
should
we
try
it
with
multiple
features?
If
we're
gonna
start
simple,
we
should
try
it
with
one
feature.
I
know
this
sounds
like
simple,
but
why
make
the
test
any
harder
than
someone
could
actually
make
it
pass?
So
what
I
want
is
like
the
most
basic
ground
level
test
for
out
2d
object,
recognition
that
I
can
create
so
I'm
thinking.
A
These
objects
are
gonna,
be
really
simple,
so
maybe
I
should
just
to
two
objects,
one
with
one
set
of
one
feature
spread
in
some
fashion.
Another
with
another
set
in
a
different
fashion,
yeah
test-
definitely
test
one
thing
at
a
time,
but
it's
almost
impossible
in
this
case
because
I'm
writing
like
this
broad
functional
test,
I'm
starting
from
the
functional
test
unit
tests,
absolutely
test
one
thing
at
a
time,
but
when
you're
running
it
when
you're
creating
a
functional
test,
you
have
to
test
lots
of
stuff.
A
So
this
is
going
to
test
like
the
whole
system
and
it
hasn't
even
been
built
yet,
but
still
we
have
to
start
really
simple,
because
I
don't
want
somebody
to
try
and
build
something,
overly
complicated,
just
to
make
a
test
pass.
We
want
to
take
baby
steps,
so
let's
have
the
test
also
build
it
up
with
baby
steps.
A
So
so,
while
I'm
thinking
about
that,
let's
start
talking
about
our
object
library,
one
first,
ten
objects
have
one
single
feature
at
10
positions:
I'm
thinking
about
something
like
that:
I
wonder
if
we
could
just
here's
I'd
like
them
to
be
differentiable
by
humans.
First
of
all,
so
you
can
look
at
it
and
say:
yeah,
that's
object!
1!
A
A
Let's
start
with
1
what
I
would
it
would
be
nice
if
I
had
an
editor
but
I'm,
but
I
don't
want
to
write
an
editor
for
these
objects.
I
think
that's
a
little
silly.
So
what
I'll
do
right
now
is
I've
got
the
I'm
at
a
place
where
I
want
to
create
an
object
model
to
do
great
object,
library,
and
then
this
says
two'd.
What
is
some
of
that?
Dude?
A
Okay,
so
one
thing
I
gotta
get
this
thing
working
again
because
I
know
it's
not
gonna
work
anymore,
because
I
moved
two
objects.
So
if
I
move
the
objects
I'm
going
to
have
to
change
where
I'm
finding
them
so
I
hope
this
ooh,
this
might
not
work
because
objects
have
to
be
in
static
right.
That's
where
they're
expected
to
be
in
the
static
folder
I
could
put
a
soft
link
to
them.
I
guess.
A
I
forget
how
soft
links
work
and
get
if
I
do
a
relative
soft
link
within
the
same
repository
I
think
it
might
work,
we're
gonna,
try
it
we're
gonna,
try
it
okay,
so
statics
everybody
understand
what
a
soft
link
is
in
operating
systems.
It's
a
like
you
can
you
can
link
to
a
file
or
directory'
down
one
like
this
from
the
director
you're
in
and
it
will
create
you
a
nice
link
to
it.
So
it
pretends
like
it's
actually
there,
so
you
can
run.
A
I
can
run
this
web
server
now
and
it's
as
if
that
objects,
folder
is
actually
there
where
I
want
it
to
be,
but
it's
really
not
it's
down
lower.
So
if
we
run
it
now,
parcel
Oh
am
I
in
the
wrong
place.
Yes,
parcel
static
index,
let's
see
all
right,
so
that's
still
work.
So
now
we
can
go
and-
and
our
objects
are
right
here
and
we'll
just
let's
just
keep
it
with
a
and
B.
A
D
A
Why
can't
what's
the
deal
place
off?
Okay,
hey
I'm,
gonna,
replace
all
the
A's
and
B's
with
exes?
No,
not
the
little
ones,
not
the
little
ones
case
sensitive
place
off.
Okay,
Hey!
Look
it
refreshed
itself,
nice
parcel
refreshed
itself.
Will
these
could
be
my
legacy?
I'm
JC
I.
Don't
think
so!
Hey
check
this
out
a
changed
ox
ox
ox
ox
checked,
object
X.
Now
this
is
object.
A
well
just
object,
a
and
object
B,
and
here
we'll
do
that.
A
A
Okay,
all
right
pop.
This
is
again
object.
B,
not
a
I'm
liking
parcel
will,
if
you're
still
listening
its
he's
gone,
no
he's
still
there
like
it,
you
can
look
and
see
who's
in
chat,
so
I'm
just
checking
still
bit
King
code,
all
the
things
Falco
Freeman,
hey,
Freeman,
I,
didn't
know,
you're
listening
and
I.
Don't
know
some
of
you
guys.
Some
of
you
guys
were
always
in
here,
I
wonder
if
your
BOTS,
not
not
Arab,
Yusef
I,
know
you
in
film
tire
thanks.
A
You
guys
for
watching
tacky
on
Ted
turba,
appreciate
you
guys
being
here
anyway
onward
and
upward.
So
I've
got
two
objects.
Let's
just
make
this
this.
These
are
the
two
objects
we're
gonna
compare.
We
should
be
able
to
compare
so
first
of
all,
I
feel
like
we
should
give
the
agent
unfettered
access
I
mean
where
I'm
not
going
to
restrict
how
they're
trained
on
this
on
this
data.
A
A
Yeah,
we
could
totally
do
this
with
just
two
objects
at
first,
because
we
could
say,
here's
the
objects
and
then
we
give
them
one
of
them
and
allow
their
agent.
How
many
touches
will
it
take?
So
this
is
the
tricky
part
because
it
might
be
hard
with
just
two
objects,
but
let's,
let's
try
it
anyway,
let's
just
let's
just
start
simple,
we'll
start
simple
with
this
we'll
start
simple
with
this:
that's
fine,
okay,
a
soft
spot
I'll
go
ahead
and
specify
the
format
you
put
it
real
quickly
and
the
readme.
A
A
A
A
Object
a
it
doesn't
matter
what
they're
named.
They
are
completely
different,
so
there's
no
there's
no
like
overlapping
features,
but
I,
don't
think
I'm
gonna
call
it
out
right
now,
there's
a
reason
why
you
want
to
call
your
features
by
a
letter:
Y,
not
just
number
them
I.
Don't
know
it
was
just
it
doesn't
matter
right.
It
doesn't
matter.
I
could
change
them
to
numbers
if
you've
got
a
strong
feeling
about
it,
but
I
don't
really
care
I,
guess
I'm.
A
By
the
way,
if
you
want
to
know
more
about
HTM,
you
can
go
here
and
I've
got
more
commands.
You
can
try
out
if
you
want
to
know
more
about
the
channel
or
anything
associated
with
it,
I'm
trying
out
my
stream
deck
thing.
This
is
what
you're
thinking
in
here.
Oh,
it's,
not
pointing
the
right
way
anymore.
I
didn't
want
to
waste.
I
didn't
want
to
waste
time.
You
know
going
over
a
bunch
of
weird
twitch
stuff,
but
this
is
a
very
cool
little
thing.
A
A
So
what
are
the
original
nupoc
demo
objects
was
a
duck,
a
cell
phone.
Oh
man,
you
remember
that
from
a
long
that
was
like
Zeta
1
I,
wasn't
there
I
wasn't
at
Numenta,
then
I've.
That
was
a
that
was
an
old
version
of
new
pic
that
doesn't
even
exist
in
the
code
anymore,
like
I,
think
all
that
stuff
got
thrown
out
that
was
before
sparse,
distributed
representations
well,
yeah
I
know
we
had
a
whole
vision
system
and
it
was
it.
A
It
seemed
to
work
really
well,
but
they
just
gave
up
on
it,
and
it's
because
this
is
crazy.
There's
a
there's
a
paper
on
this
there's
a
paper
written
by
dilip,
George
and
Jeff
Hawkins,
and
it
says
hierarchical
temporal
memory
in
it,
but
it
was
like
it's
like
a
bipolar
paper:
it's
like
half
written
from
the
Bayesian
standpoint,
that's
dilip,
George
and
half
written
from
a
biological
standpoint,
which
is
obviously
Jeff
and
is
doing.
Why
is
it
doing
that?
That's
so
weird.
A
A
So
anyway,
this
paper,
it's
it's
like
bipolar,
right,
half
of
its
beige
and
half
of
its
basically
leaning
towards
HTM,
so
that
old
model
used
used,
Bayesian
inference
well,
I,
don't
they
didn't
call
it
Bayesian
inference
back
then
I,
don't
even
know
what
I
wasn't
even
here.
So
I
don't
really
know
well
how
it
worked.
A
Okay,
but
my
my
understanding
is
that
there
was
there
were
Bayesian
models
heavily
in
play
in
new
mental
world
back
before
Dilip
George
left
and
Dilip
George
wanted
to
pursue
the
Bayesian
aspect
of
it
and
he
went
and
formed
vicarious,
and
that
is
that's
all
Bayesian
stuff
as
far
as
I
know,
so
that
we
just
sort
of
stopped
work
on
that
on
that
vision,
thing
that
you
that
got
you
into
it
and
focused
on
sparse
representations,
because
that
early
stuff
wasn't
sparse,
didn't
have
that
type
of
stuff.
At
all.
A
A
A
A
If
I
get
it
the
right
place
this
right
here,
this
is
the
agent
and
so
it's
going
to
have
a
location
in
the
space.
So
so
now
this
brings
up
a
couple
questions
right,
if
you
think
about.
If
this
is
going
to
be
our
agent,
we
immediately
have
some
restrictions.
What
happens
if
the
agent
is
at
zero
zero?
A
We
have
to
make
a
decision
about
how
we
feed
the
mist.
What
do
we
do
with
the
missing
input
from
those
cell,
those
those
sensors
that
are
not
on
the
map?
If
we're
right
here?
If
the
agents
right
here,
there's
gonna
be
missing,
input,
so
I
feel
like
we
should
I,
don't
want
to
wrap
around
the
object.
I,
don't
think
I
want
to
wrap
around
the
object.
I
want
this
to
be
a
distinct.
Well,
it
probably
won't
matter,
probably
won't
matter.
A
It
shouldn't
matter,
we
could
depth,
we
could
wrap
around
the
object
or
we
could
just
say
you
just
don't
get
anything.
If
your
sensor
goes
because
gets
off
the
object
space,
you
don't
get
anything
I,
think
that's
simpler.
I
think
this
is
even
simpler.
I'm
just
saying
if
your
sensor
is
outside
of
the
bounds
of
the
object,
there
is
no
feature
there.
A
Then
you
don't
have
to
worry
about
adding
and
subtracting
indices
and
it's
simpler
for
people
who
are
trying
to
wrap
their
heads
around
it
because
they
don't
have
this
additional
concept
of
wrapping
around
and
coming
to
the
other
way.
It's
just
bounded
by
one
thing,
so
I'm
gonna,
err
on
the
side
of
simplicity,
I,
think
yeah.
A
But
I,
like
your
suggestions,
don't
don't
be
discouraged
from
making
suggestions,
it's
always
good
to
think
about
things
in
different
ways
before
moving
forward,
so
so
I
think
I'm
going
to
try
and
put
this
in
words
on
in
the
readme
here
for
the
agents.
Okay
and
agent
capitalized
this
an
agent
can
okay.
This
is
a
tricky
thing,
so
an
agent
can
sense
objects
by
positioning
itself
in
a
location
on
the
object
and.
C
D
A
Receive
features
in
space
according
to
the
orientation
of
its
sensors
orientations,
a
bad
word,
location
of
its
sensors
agent,
observing
an
object,
will
receive
features
and
space
according
to
the
location
of
a
sensor.
Ok,
let's
do
some:
let's
do
some
passkey,
no
I
don't
want
to
do
the
rescue.
What
I'll
do
is
I
mean
pictures
are
good
here,
so
maybe
I
should
just
refer
to
this.
A
So
it's
not
really
it's
it's
sort
of
on
the
object
I
mean
when
you
think
about
agent.
You
can
also
think
about
finger
and
when
you're
touching
something
your
finger
does
exist
on
that
object.
So
it's
it's
in
the
object
space,
it's
a
reference
frame,
but
if
it's
it's
sort
of
different
with
your
eyes
anyway,
I
think
it's
a
little
bit
better.
Okay,
okay,
an
agent
observing
an
object,
will
see
features
in
space
according
to
the
location
of
its
sensors.
A
B
A
In
the
right
place,
you're
not
seeing
what
I'm
doing
here,
but
it's
it's
real
stupid,
I'm,
dealing
with
OBS,
stuff
and
I
have
to
otherwise
the
chats
not
going
to
show
up
properly
so
give
me
just
a
sec.
I
can
fix
it
so
that
or
the
chat
Keith
will
keep
disappearing.
I
don't
want
the
chance
to
disappear.
It's
not
fun,
it's
more
fun
to
have
chat
that
doesn't
disappear.
Obviously,
alright!
Here
we
go
okay,
okay,
we're
getting
there.
We
got
there.
We
got
there.
A
A
A
A
So
this
is
where
it
gets
tricky
because
I
want
to
start
using
HTM
terms.
I
want
to
start
using
it's
not
a
big
deal.
Pal
girl,
like
pg-13,
pg-13,
spine
testicle,
testicle,
technical,
testicle,
tits
and
boobs
whatever,
as
long
as
we're
not
just
being
rude
for
just
rude
to
each
other.
I
want
to
say
proximal
input
here,
but
but
I
know,
but
I
also
don't
want
to
say
proximal
I
play
frustrating
though
twitch
chat
is,
if
you
actually
refresh
the
page
at
loses
all
its
history.
A
You
know
what
you
can
do
is
I
know
you
can
you
can
have
a
pair
motion
show
you
something
you
like
this.
If
you,
if
you
are
concerned
about
that,
there
is
a
program
called
what
is
it
twitch,
there's
a
bunch
of
twitch
chat,
clients,
twitch
chat,
client,
I,
I
found
one
there's.
One
called
chatty
I
have
T's
T
C.
A
So
if
you
get
this,
which
I
I'll
download
it
right
now
because
I
downloaded
it
on
my
Mac
and
I
and
I
used
it
several
times
and
it
works
great.
So
I
would
recommend
this
TC.
It's
a
little
chat.
Client.
You
can
have
chat
open
all
the
time
to
this
twitch
channel
or
any
of
your
other
trip
channels
that
you
watch
and
just
always
be
connected
to
them
and
you'll
never
lose
the
history
so
even
like.
Sometimes
people
come
and
chat
in
the
chat
room
when
I'm
offline.
Usually
it's
me
but
people.
A
Okay,
I
almost
just
shaved
a
yak,
so
you
can
go
get
that
chat.
Client
if
you
want,
or
another
type
of
chat
client
I'll
set
up
that
later.
If
I
had
the
yak
shaving
but
and
I
almost
have
it
first
adjust
that
right,
where
we're
eating
we're
worried
for
sensors,
each
sensor
gets
and
I'm
not
going
to
say
proximal
input,
I'm,
not
gonna,
say
proximal
input.
I
do.
C
A
A
A
So
I'm
struggling
right
now,
I
want
to
just
define
the
fact
that
every
sensor
gets
a
unique
feature
from
the
object
space
relate
in
relation
to
the
location
of
the
agent
I.
Guess
that's
what
I
say.
Each
sensor
has
access
to
the
feature
beneath
it.
Maybe
is
that
is
that
a
clear
enough
term
future
beneath
it.
A
A
One
space
up
or
actually
just
say,
y
-1
right
I,
mean
just
to
really
definitely
definitively
define.
This
South
is
is,
is
a
agent
actually
agents
agent
y
-1?
Is
that
clear
South
is
going
to
be
agent.
Why
actually
I
got
a
backwards?
Didn't
it
plus
one
minus
one
I
think
that's
clear
and
will
do
agent
X,
plus
one
agent
X
minus
one,
okay,
just
to
be
declaring
where
the
sensors
are
each
sensor
is
access
to
the
feature
beneath
it.
D
A
A
Is
north
is
South
is
plus
right,
X
is
plus,
East
is
plus,
West
is
I.
Think
that's
right,
n
is
ease,
these
East
is
not
East
is
East,
is
gonna,
be
plus
West
is
gonna
be,
but
you
but
the
wise
needed
to
be
flipped.
That's
that's
correct.
Cuz
I
was
thinking
about
the
graph.
Instead
of
the
you
know,
it
all
depends
on
the
space,
so
we're
gonna
do
this
top
is
well.
Maybe
we
should
do.
Maybe
we
should
be
like
this
agent
plus
instead
of
doing
Y
and
X
will
just
say
agent.
No,
no.
A
This
is
fine.
I
think
this
is
right,
I,
don't
know
what
you
mean
left
is
zero
left
is
not
zero
left
is
minus
one
from
where
you're
at
right
is
plus
one
from
where
you're
at
I'm.
Looking
at
looking
at
this,
if
I'm
at,
if
I'm
at
X
is
10
y
is
9,
north
would
be
y
-1
right
south
would
be
y
plus
1,
which
is
what
we
have
here.
East
would
be
x,
+
1
west
would
be
X
minus
1.
A
A
Why
did
it
do
that
I
forget
how
to
copy/paste?
Sometimes
each
sensor
has
access
to
the
feature
beneath
it.
A
Agents
should
use
their
sensors
to
to
attempt
to
identify
the
object
under
test
under
observation
at
each
time,
step
that
makes
sense,
so
the
agents
will
be
will
move
to
a
location
first
yeah,
the
times
time,
step
I
haven't
talked
about
time.
Yeah,
that's
going
to
be
in
the
challenge,
I
think
that's
good.
Well,
we
will
define
steps
in
the
challenge.
So
agents
should
use
their
sensors
to
attempt
to
identify
the
object
under
observation
each
time
step,
so
they
have
an
opportunity
to
identify
an
object
at
each
time.
Step.
A
A
A
A
D
B
A
A
B
A
Okay,
oh
you
probably
put
that
there
will.
Okay,
that's
fine,
so
cat
this
good,
ignore
onto
that
good
ignore,
so
we
can
combine
our
to
get
ignores
together
and
then
remove
get
our
M.
This
get
ignore!
Actually
this
one
I'll
probably
put
things
like
D
s
store
so
blah
blah
I'll
leave
that
D
s
store:
whoops,
whoops,
whoops,
okay,
okay,
okay
into
the
Python
I'm.
Just
this
is
just
all
clean
up:
I'm,
just
cleaning
up
before
I
push
cash
fight;
okay,
V
and
V
vs
code,
PI
cache.
A
A
A
What
about
package
lock,
you
guys
do
guys
that
are
node
node
engineers,
you
know
how
to
do.
Node
I
noticed
that
recently
packaged
lock
started
appearing
in
my
in
my
node
files.
Is
that
something
that
typically
gets
checked
in
or
is
that
something
that
goes
like
only
for
production
to
lock
things
in
production
commit
it?
Okay,
thank
you
just
making
sure
okay
and
the
dist
I
want
to
get
rid
of
that
dist.
D
A
Get
ad
get
ignore
me:
okay,
get
ignore,
read
me
that
stupid
cache,
so
I'm
gonna
reset
all
this
stuff
get
reset
head
cache
all
right.
So
I've
got
my
JavaScript,
eternal
javascript
readme
package
like
package
static
index
main
objects.
No,
no!
That
shouldn't
exist.
That
should
not
good.
So,
let's
figure
that
get
remove
objects
force.
A
One
question:
what
stopped
you
from
starting
from
a
text
environment
in
the
building
up
to
a
JavaScript
or
PI
game,
environment?
I,
don't
know
I,
just
I
didn't
think
about
it.
I,
don't
I've,
never
used
PI
game.
Actually,
I
used
2
pi
game
once
but
I,
because
when
you
say
text
environment,
do
you
mean
test
I,
don't
understand
what
you're
asking
I'm
trying
to
write
an
agnostic
functional
test,
essentially
that
anybody
could
create
a
solution
as
and
solve
it
in
whatever
language
they
want.
Oh.
A
An
ASCII
environment
you
mean
so
that
I
could
define
the
objects
in
ASCII
I.
Don't
think
that's
a
good
data
format.
It's
not
flexible!
You
I
want
something.
I,
don't
know
it's
it's
it's
just
it's
not
a
very
good.
It's
not
a
good
flexible
data
format,
building
blocks
in
unity
yeah,
so
these
could
definitely
be.
These
objects
could
be
rendered
in
unities
for
sure
for
sure,
yeah
and
I
I
did
do
a
little
rendering
in
unity,
but
I
won't
be
doing
that
for
a
while,
but
I
probably
will
come
back
to
it.
A
D
A
Know
I
feel
like
the
Yamhill
schema
is
just
a
much
better
data
format
than
ASCII.
It's
it's.
It's
just
flexible.
It's
easy
to
edit
you
its
small.
It's
a
lot
smaller.
It's
not
that
that's
a
big
deal,
but
I
just
I,
definitely
prefer
something.
That's
that
looks
like
this.
Then
then
ask
ya:
it's
not
important
for
the
project.
A
Hey
the
thing
worked:
awesome,
hey
ho,
hosts
ptosis,
interesting
name,
thanks
for
following
I
appreciate
it
ain't,
no
new
noise
works;
I,
don't
know
your
first
name,
but
I
appreciate
you
coming
in
here,
you're
actually
rating.
You
brought
a
party
awesome
thanks
for
coming.
You
know
it's
going
a
lot
better
today
than
it
was
going
yesterday.
That's
for
sure,
if
you're
new
to
the
channel,
obviously
you're
all
new
to
the
channel,
here's
here's
some
some
background
and
some
commands.
If
you,
if
you
want
to
oh
wait,
I
didn't
do
that
right.
A
Commands
I
got
a
chat
bot.
So
if
you
want
to
know
anything
about
what
I'm
doing
right
now,
since
you
guys
are
all
new
I'll
kind
of
give
you
an
idea
of
what
I'm
trying
to
do
here,
I
am
trying
to
define
a
2d
object,
recognition,
challenge
or
test
in
a
language
agnostic
way
and
in
a
format
so
that
we
can
apply
like
a
biologically
correct
solution
to
it.
So
the
idea
is,
you
have
an
object
space
like
this.
That's
just
two-dimensional!
This
is
a
very
simple.
D
A
Yeah,
so
so
each
one
of
these
spaces
in
here,
so
this
is
just
X&Y
and
they're
and
they're
integers.
You
can
imagine
this
being
as
large
or
as
small
as
you
wanted.
Thanks
for
the
follow
bit
code
appreciate
it
thanks
for
the
raid,
that's
awesome,
so
so
there's
like
two
different
types
of
features
in
the
space.
So
the
idea
being
that
you
get
an
agent
and
the
agent
has
four
sensors
north
south
east
west
and
each
one
of
those
sensors
gets
a
different,
has
access
to
observe
a
different
feature
in
that
object.
A
Space
and
as
the
agent
moves
through
the
object
space
it
attempts
to
identify
the
object.
It
does
sort
of
look
like
minesweeper.
Doesn't
it
this
it's
not,
and
so
the
challenge
would
be
to
create
and
the
way
this
is
structured
with
the
four
sensors.
It
relates
to
a
biological
neural
network
model
called
hierarchical
to
portal
memory,
and
so
that's
the
community
that
that
I'm
working
with
to
try
and
build
this
project
the
HTM
forum.
This
is
that's
the
form
that
we're
on
here.
A
If
you,
if
you
want
to
check
it
out
HTM
forum
and
that's
where
this
blog
post
is
at,
but
each
one
of
these
sensors,
as
you
can
see
here,
is,
has
a
complete
cortical
column
structure
like
you're,
going
to
model
a
cortical
column.
Cortical
column
is
a
structure
in
the
brain
that
has
many
different
layers
and
so
we're
trying
to
like
nail
down
and
try
and
figure
out
how
how
intelligence
works
in
the
neocortex.
So
this
is
a
simulation
of
layers
in
the
neocortex,
so
we're
gonna
simulate
four
layers
in
the
neocortex.
A
Each
one
is
a
sensor
around
this
agent
and
we're
gonna.
Let
the
agent
move
through
the
space
and
try
and
feel
out
these
objects
and
then
tell
it.
This
is
object,
one
right
and
it'll
and
it'll
randomly
pop
through
the
space
and
and
get
a
feeling
for
these
objects,
building
up
a
representation
of
them
through
space
over
time
in
these
cortical
columns
and
all
these
were
these
cortical
columns
and
there's
each
one
of
them
will
have
an
object
pool,
what's
called
a
where
they're
trying
to
do
active
recognition.
Sack
of
joy.
A
Thank
you
for
following
I
appreciate.
It
I'm
talking
about
neuroscience
and
the
AI
stuff
right
now,
so
this
is
like
one
sensor.
This
would
be
maybe
the
East
sensor,
the
South
sensor.
They
all
see
different
things
in
space,
but
they
all
are
trying
to
identify,
not
the
features
they're
seeing
but
the
the
objects,
so
the
objects
are
not
in
this
space
like
this
is
an
object.
This
whole
box
is
an
object
and
each
object
has
a
bunch
of
features
so
we're
gonna.
A
Let
the
agent
move
through
the
object,
space
and
learn
an
object
and
then
say:
okay,
here's
another
one.
Learn
it
and
it'll
go
to
learn
that
object,
and
then
the
challenge
is
we're
going
to
give
it
another
object
and
not
tell
it
what
it
is
and
let
it
move
through
the
space
and
see
how
many,
how
much
it
has
to
move
before
it
can
identify
the
object.
A
That's
the
challenge,
it
seems
really
simple
and
the
challenge
is
really
simple
but
to
implement
it,
as
you
can
see
from
the
the
moving
parts
of
a
cortical
column,
it's
pretty
complicated,
so
this
is
like
HTM
Theory,
hierarchical
temporal
memory
theory.
This
is
like
the
culmination
of
several
years
of
work
of
Numenta
coming
together
in
this
project.
There's
two
different
papers
involved
here:
there's
a
theory
of
how
columns
in
the
neocortex
enable
learning
in
the
structure
of
the
world
and
the
second
one
which
is
locations
in
the
neocortex.
A
A
theory
of
sensory
motor
object,
recognition
using
cortical
grid
cells
grid
cells
are
amazing.
If
you
don't
know,
a
grid
cells
are
go.
Google
grid
cells
because
they're
super
cool,
so
yeah.
This
is
for
the
new
person
coming
to
this
topic.
It
it
is
really
overwhelming,
but
honestly
it's
not
any
harder
than
Bayesian
neural
network.
Sorry
I
mean
our
calculus.
It's
it's
it's
something
you
have
to
spend
some
time
thinking
about,
because
it
involves
a
lot
of
intricacies
of
neural.
A
You
know
aspects
of
neurons
and
connections
and
synapses
and
all
its,
but
it
just
takes
time.
You
can
learn
anything
given
enough
time
and
practice
and
motivation.
I
I've
certainly
come
to
that
conclusion.
Based
on
my
experience
in
the
field,
all
right.
Okay,
so
thanks
for
the
words
of
encouragement,
you
guys
are
being
nice
to
each
other.
This
is
one
the
team
that
will
one
day
get
the
Nobel
Prize
I
hope
that
would
be
awesome,
we'll
see
the
grid
cells,
which
is
we're
basing
a
lot
of
our
work
on,
did
get
the
Nobel
Prize.
A
B
A
They
won
the
Nobel
Prize
because
and
grid
cells
are
cool
grid
cells
are
super
cool.
If
you
don't
know
what
grid
cells
are,
go,
look
them
up.
Go
google
check
this
out.
I
know
I.
Do
this
all
the
time
Google
grid
grid
cells
and
the
first
video
that
pops
up
is
my
smiling
face
because
I
did
a
whole
video
on
grid
cells
and
with
a
ton
of
visualizations
and
everything.
A
But
like
look
at
this,
if
you
wanted
to
see
what
grid
cells
are
and
how
they
work
check
out
this
video,
it's
pretty
neat
I'll
paste
it
for
you.
If
you
want
to
go
watch
it
offline,
but
this
is
like
crucial
understanding.
Grid
cells
is
crucial
to
understanding
sensory
motor
integration
in
HDM
3.
So
if
this
piques
your
interest
that
grid
cell
video
is
really
visual
and
it
might,
it
might
sort
of
give
you
an
entry
entry
entrance
pathway
into
HTM
theory,
so
it's
not
any
harder
than
Bayes
and
neural
networks.
Yeah,
that's
true.
A
It
took
me.
It
was
easier
for
me
to
understand
HTM
than
it
was
to
understand
Bayesian
Bayesian
stuff,
because
you
really
have
to
understand
derivatives
and
calculus
to
truly
get
that
and
I
never
took
calc
2
or
calc
or
calc
2
in
college.
I
am
NOT
a
mathematician,
so
I
skipped
that
stuff,
but
the
HTM
stuff,
the
bio
of
stuff
based
on
biology
is,
is
definitely
definitely
better
and
you
know
not
better
I,
don't
say
better.
Better
is
not
the
right
word.
It's
more!
It's
it's!
It's!
A
It's
moving
in
the
direction
towards
real
intelligence
right,
real,
real
AG
I.
What
you
might
call
a
GI
single
order.
Memory
mat
can
even
understand
yeah.
That's
an
inside
joke,
I
think
I
took
guitar
classes,
I
took
a
few,
but
I
think
I
took
three
guitar
classes.
Our
data
structures
applied
to
this
data
structures.
Let's
see
well
I
mean
that's
a
really
generic
term.
What
do
you?
What
do
you
mean
our
data
structure
is
applied
to
this?
A
There
are
data
structures
inherent
in
the
system
and
I
mean
that
I
can
I
can
explain
to
you,
but
I'd
have
to
draw
out
the
whole
I
mean
here's.
The
deal
if
this
is
a
cortical
column
in
your
brain
in
your
neocortex,
it's
split
up
into
a
bunch
of
layers.
More
than
six
I
mean
they
say,
there's
six
layers,
but
there's
there's
a
bunch
and
each
one
of
those
is
like
processing,
something
and
doing
something
specific
each
one
of
those
has
connections
to
other
layers
and
in
different
places.
A
Some
of
them
are
different
types
of
connections.
There's
proximal
connections,
distal
connections,
apical
connections-
some
of
them
are
actually
spreading
out
laterally
and
sharing
their
information
with
other
cortical
columns
that
are
nearby
that
are
doing
the
exact
same
thing
and
they're
all
getting
feed-forward
sensory
input
from
the
senses
and
doing
object,
recognition
simultaneously
and
the
gist
of
it.
A
That's
HTM
theory,
sensory
motor
integration,
theory
and
our
job
as
Numenta
has
been
to
understand
this
circuit
and
what
it
does,
because
it's
the
key
common
circuit
in
in
the
cortex
in
the
brain,
I,
think
and
if
we
can
understand
this
circuit,
we
can
understand
intelligence
in
the
brain
and
we
can
build
real.
Intelligent
things
understand
the
world
because
they
model
the
world
through
movement
and
and
sensory
perception
and
and.
A
A
Constantly
it's
getting
sensory
input,
it's
getting
efference
copies
of
its
own
movement
commands,
so
they
can
tell
how
it
is
changing
its
environment
when
it
does
make
the
moves,
so
it's
kind
of
deep
stuff,
but
it's,
but
it's
super
important
and
and
there's
also
the
thalamus
over
here
each
one
of
these
also
like
loops
loops
to
the
thalamus
to
so
a
part
of
this
circuit.
This
would
be
like
this
is.
A
This
is
all
neocortex
and
then
the
the
thalamus
is
a
super
important
part
of
this,
and
it's
part
of
this
cortical
circuit
as
well,
and
we've
recently
been
when,
when
thinking
a
lot
about
the
thalamus,
you
know
what
I
got
more
look
check
this
out.
So
here's
your
brain.
This
is
your
bread.
This
is
your
brain
on
on
Twitch.
No,
so
I
get
rid
of
this
part.
It's
not
that
important.
So
your
neocortex
is
all
this
jazz
and
well.
Some
of
this
is
just
normal
core
textures
hippocampus
back
there
in
Taranto
cortex
right
around
here.
A
This
is
your
thalamus,
this
football-shaped
thing,
okay,
so
every
this
is
like
a
sheet
of
cells.
You
can
stretch
it
out
like
like
a
dinner
napkin,
if
you,
if
you
wanted
to
it,
wouldn't
be
very
comfortable,
but
you
could
you
can
stretch
it
out
like
a
dinner
napkin,
because
it
has
this
laminar
structure
and
the
thalamus
fits
right
up
in
there
like
it's.
A
They
work
so
closely
together
that
we
have
to
understand
what
the
thalamus
is
doing
too,
and
this
is
what
is
controlling
attention.
This
is
probably
the
control.
This
is
the
place.
This
is
in
a
perfect
location
to
control
what
this
does
to
control.
What
you're
thinking
about?
What
are
you
attending
to?
How
am
I
imagining
the
objects
that
I've
stored
in
my
brain
I
also
think
well.
Jeff
also
thinks
that
we
just
talked
about
this
the
other
day.
That
is
also
controlling
the
dilation
of
time
and
space.
A
So
it
has
it
as
an
effect
through
oscillatory
control
like
by
controlling
the
oscillations
across
the
neocortex.
How
objects
are
warped
in
time
and
how
movements
are
warped
in
space,
so
it
is
like
the
culmination
of
time
and
space
happens
right
here
in
these
little
footballs
in
your
brain
right
there.
This
this
is
where
time
and
space
merge
yeah.
So
I
keep
talking
about.
Jeff
Hawkins
he's
my
boss.
A
A
You
know,
I,
think
that
there's
no
such
thing
as
talam
I
I
think
that
it's
the
thalamus
and
there's
two
parts
to
it
anyway,
thanks
for
the
interest,
so
that's
how
you
first
found
me
film
tire
right
after
the
deep
mind,
blog
post
about
grid
cells,
oh
cool,
I'm,
glad!
That's
awesome!
I'm
glad!
You
found
that
because
I
put
a
lot
of
work
into
that
grid
cells,
video
I'm,
really
glad
it
turned
out
it's
one
of
them,
one
of
my
most
successful
video
productions.
So
yeah
there's
a
bunch
of
HTML
videos,
no
problem!
A
A
This
twitch
stuff
is
fun,
but
it's
it
can
be
super
exhausting,
sometimes
here's
here's
the
book
coming
to
fix
this
pawned
intelligence.
I,
don't
know
if
you've
seen
this
I
read
this
book
in
2006
before
I
even
knew
Jeff
I
picked
it
up
in
st.
Louis
that,
like
Barnes
and
Noble,
or
something
or
it
was
a
Borders
I
think,
and
it
was
really
enlightening.
Like
I
read
it
and
I
was
like
wow
like
it
was
a
mind.
Blown
moment.
A
I
was
like
this
is
this:
is
it
like
this
guy's
on
the
right
track
and
now
I
work
for
him?
So
he
put
your
mind
to
it.
You'd
do
anything
because
I
didn't
I,
didn't
like
try
and
is
it
a
matter,
never
mind?
Okay,
back
to
oh
yeah,
this
all
right
object.
Recognition
stuff
we
were
trying.
I
was
just
about
to
commit
a
bunch
of
changes
and
I
was
going
to
add
something.
I
was
gonna.
Add
JavaScript
I
was
gonna,
add
a
cache
the
DEF
cache
file
to
this.
Let's
put
it
at
the
end.
A
D
A
All
right
so
I'm
ready
to
commit
this
took
me
all
day
all
day
to
get
to
this
point
where
he
can
finally
commit.
This
is
what
this
is
the
point
where
I
would
hit
the
button
that
would
do
some
type
of
celebration
or
something
so
yes,
I
am
addicted
to
twitch.
Already,
let
me
tell
you
a
few
things
about
twitch
I
was
gonna,
add
something
too
yeah.
I
was
but
I'm
not
going
to
do
that.
Yet
I
down
here,
I
am
I'm
missing
this
part
in
this
test.
A
C
A
A
A
A
It's
a
Space
Odyssey
reference
space
on
its
feet:
yeah
2001,
a
Space
Odyssey
that
looks
pretty
good.
Okay,
so
see
the
origin
in
the
forum
post,
which
goes
right
here.
Purpose
of
this
is
to
define
two
2d
object.
Recognition
tests
now
BIC
schema
exist
when
2d
space
height
with
at
fourteen
five
I'm
missing
a
period
no
big
deal,
there's
an
object.
Schema
features
always
contain
data,
here's
the
object,
libraries
and
objects,
and
it
is
just
double-checking
everything
agency,
an
agent
exists
of
location,
object,
space,
Oh,
open
the
pod
door.
A
New
I've
got
a
Python
area,
a
Python
place
in
here
I'm
just
going
to
keep
yeah,
so
my
Python
place
simply
has
a
library
with
an
agent
and
an
environment,
and
I
haven't
really
done
much
with
this
I've
just
created
just
enough
of
this
to
write
some
tests,
so
we
have
a
two
dimensional
environment
and
create
with
this
width
and
height,
and
then
we
have
one
of
these
objects
that
we
could
load
into
the
environment,
which
sets
its
features.
Basically
and
that's
super
simple
code.
All
the
things
did
this
actually
and
chat.
A
Let's
see
and
then
the
Python
or
the
JavaScript
part
is
just
a
visualization
of
the
objects.
There's
is
that
the
only
code
we
have
yeah
it's
not
even
all
it
does
is
just
render
them
to
a
canvas.
So
we
can
see
them.
So
it's
that
simple
I
guess
so
so
I'll
just
I'll
just
describe
that.
Okay,
the
tricky
thing
is
tricky
thing.
What
is
the
tricky
thing?
The
these,
these
two
different
code
bases,
are
serving
different
purposes.
That's
sort
of
bothers
me
and
I
almost.
A
Objects,
yeah
objects
in
the
object
library,
so
currently
it's
just
yeah
so
I'm
describing
that
now,
because
I
didn't
want
to
write
the
visualization
code
in
Python
honestly,
because
I
haven't
found
a
good
visualization
I
guess,
maybe
I
could
do
it
in
PI
game
I
could
totally
do
it
in
PI
game
good
night.
Somebody
mentioned
PI
game
earlier,
but
right
now
the
only
reason
we
have
to
two
places
for
the
code
is
so
that
one
of
them
looks
like
this.
One
of
this,
the
Java
Script
one
is
showing
the
objects.
A
That's
the
only
reason,
because
I
didn't
want
to
hey
you're
a
tachyon.
You've
done
PI
game.
So
is
how
easy
is
it
to
basically
create?
It
should
be
really
super
easy
to
do.
This,
like
I've,
got
two
spaces
here.
These
are
just
two
canvases
very
easy
yeah.
So
maybe
how
are
interested?
Are
you
in
recreating
this
visualization
and
PI
game?
It's
easy!
It's
Python,
because
I
don't
necessarily
want
this
visualization
in
JavaScript.
It
could
be
easy
in
Python
I'm,
just
throwing
it
out
there,
no
pressure,
no
pressure,
Friday
night.
A
Let's
go
wild
break
out
the
IDs
and
the
die
cook
in
the
Mountain
Dew,
no
pressure,
I'm,
not
I'm,
not
putting
any
pressure
on
you,
but
that
sounds
like
a
great
idea:
cuz
I,
don't
really
want
to
javis.
Don't
really
want
javascript
in
here.
I
just
want
an
example.
I
just
want
some
type
of
code
example
and
the
J
s
was
just
so.
We
could
define
the
objects
and
see
them
and
make
sure
that
they
make
sense.
Oh
man,
if
you
could
twitch
that
that
would
be
awesome.
A
A
A
D
A
Thing
to
add
I
think
whether
or
not
tachyon
ends
up
doing
it.
That's,
let's
add
it
up
there,
because
that
makes
sense.
So
I
haven't
created
a
test.
Spec
I
have
updated
that
I
have
emerged,
that
I
have
split
out,
that
I
have
created
and
not
a
library
of
two
objects
for
testing
yeah
you're
not
gonna,
get
far
with
a
weak
internet
connection
and
a
bad
graphics
card.
Oh
my
sub
notification!
Oh
you
did
you
seriously
subscribe.
I,
didn't
I,
didn't
get
a
notification,
that's
for
sure
I
don't
know.
A
What's
going
on,
this
looks
like
it's
on
now.
I
have
to
go
check
well.
Thank
you.
I
really
appreciate
that.
That's
awesome,
I
I,
don't
know.
What's
going
on
with
myself
notification,
I'll
have
to
fix
it
off
to
figure
it
out.
Thank
you
very
much.
You've
got
a
brain
by
your
name
now,
which
Anna
and
ass
nope.
A
A
Yeah,
not
let's
not
I'd
like
to,
but
not
yet
I,
don't
necessarily
want
to
use
new
pick.
If
we're
going
to
use
anything
I
want
it
to
be
Python
3.
If
we're
going
to
use
Python
and
that
sort
of
means
new
picks
out,
we
could
use
the
community
fork,
we
could
use,
we
got,
we
don't
need
to
use
Python.
Also,
like
I
said,
I
want
to
know.
A
No.
We
can't
that
it's
not
easy.
If
we
just
did
an
investigation
of
this,
did
you
see
on
the
forum?
Let
me
show
you
on
the
forum
I
added
a
new
pick:
Python
3.
Should
we
make
new
potatoes?
They
pick
Python
3
end
of
life,
so
we
are.
We've
basically
announced
that
we
are
not
so
you
saw
that
okay,
you've
watched
all
that
we
are
not
going
to
update
5
on
3.
We
still
need
to
write
research
code
yeah.
A
Well,
the
divide
between
Python
2,
&,
3
shouldn't
still
be
a
thing,
but
people
with
people
with
legacy
code
bases.
It
is
still
really
still
a
thing,
and
it
is
it's
not
a
fun
upgrade
for
for
those
of
us
who
the
then
the
problem
with
new
pick
is
not
just
Python
3.
It's
that
it
needs
to
be
refactored
really
badly
like
there's,
there's
yeah,
there's
even
in
the
Python
code,
the
stuff
that
we
have
to
take
out
to
make
Python
through
the
transition
to
Python
3
easier
is
like
you
have
to
it's
like
brain
surgery.
A
You
know
it's
like
you've
got
to
take
this
out,
which
means
you
have
to
take
this
out,
which
effects
this
and
it's
tied
to
that
it's
kind
of
spaghetti.
It's
it's,
not
the
best
software
engineered
project.
You
have
to
understand
all
of
this
code
and
nupoc
came
from
scientific
research.
So
it's
it's
all,
based
on
research
code,
that
sort
of
evolved
into
being
productized
in
different
in
different
formats,
and
so
it's
hard.
So
there's
there's
definitely
still
a
divide,
so
we
said
we're
not
even
gonna.
A
Do
we
have
new
pet
core,
which
is
C++
and
has
swig
bindings
to
Python
too,
and
that
new
pet
core?
We
might
try
and
create
Python
3
bindings.
On
top
of
as
a
so,
we
can
write
research
over
it.
That's
one
option,
however:
the
the
community,
the
nupoc
community,
have
already
a
fork
of
new
pet
core
called
new
pic
CPP
and
they've
done
a
bunch
of
work
on
here
recently
to
update
this,
so
that
it
it
doesn't
have
some
of
the
issues
like
that
doesn't
have
suik
anymore,
they
upgraded
it.
A
So
it
has
a
PI
bind,
and
so
it's
much
easier
to
write
the
Python
3
bindings
against
this.
So
we
may
end
up
using
this
instead,
so
we'll
see,
ok,
yeah,
all
the
C
by
is
so
so
I
think
this
is.
This
would
actually
be
simpler.
I've
talked
to
Louis
about
this
too,
and
he
sort
of
feels
the
same
way
like
if
we
really
want
to
move
fast.
A
We
might
just
take
the
community
fork
and
start
writing
research
against
the
community
for
believe
it
or
not,
and
I
know
that
the
guys
behind
the
community
fork
would
love
that,
but
we're
still
debating
it,
but
everything
that
is
a
new
pic
will
eventually
be
integrated
into
this
project.
No
there's
a
lot
of
stuff
in
new
pic
that
we
don't
need.
For
example,
again,
I'm
gonna
go
back
to
recent
foreign
posts,
because
this
has
come
up
recently
or
is
the
latest
where's?
A
My
latest,
we
were
just
someone
was
here:
we
go
biological
constraints,
so
I
was
just
talking
about
this
earlier.
This
is
in
the
fork
these
dirt.
These
are
guys.
D,
Mac
and
Bresnik
are
guys
that
have
been
working
on
the
fork
for
a
long
time
and
so
they're
talking
about
should
I
take
out
this
from
new
pic
backtrack
and
TM
should
I
take
out
a
top-down
computing
decode,
and
should
they
take
out
the
temporal
Pooler?
My
answer
is:
yes,
remove
them
all
we're
not
going
to
be
using
them.
Numenta
does
not
need
this
stuff.
A
This
is
this
is
old
legacy
code
we're
not
going
to
use
it,
so
the
core
of
HTM
still
is
a
new
pit.
It's
in
the
network
API
and
the
algorithm
the
spatial
pulling
algorithm
temp
for
M.
You
know
the
boosting
and
even
the
encoders.
That's
not
core
code.
Encoders
are
nice
to
have
anybody
can
make
their
own
encoders,
so
there
will
be
stuff
in
nupoc.
That
does
not
that
we
don't
use,
but
to
solve
this
problem
that
we're
doing
you
don't
need
any
of
that
stuff.
You
don't
need
that
stuff.
A
A
Ok,
where
am
I
and
your
work?
Do
you
deal
with
low-level
stuff
like
assembly
stuff
or
not?
No,
no.
We
don't.
We
write
research
code
in
Python
and
then
optimize
it
with
c
plus
plus
so
typically,
there's
there's
not
there's
a
wrap.
A
Python
wrapper
around
a
c
project.
Do
you
want
me
to
work
with
coded
images?
Are
imported
pngs
like
the
grid
and
all
the
symbols?
Oh
I.
A
Don't
think
you
need
to
work
with
any
images,
don't
work
with
any
images.
Can
you
not
just
draw
it
and
I
mean
and
just
use
the
string
definition,
because
this
is
the
object,
are
going
to
be
visualizing.
The
only
thing
you
have
to
visualize
is
an
X
at
you
know
four
or
five,
so
you
could
make
an
image
of
an
X
if
you
want,
but
it
seems
like
it
would
be
just
as
easy
to
draw
a
text.
A
X
can
the
community
for
it
be
compiled
down
to
something
that
will
run
on
the
GPU,
yeah
I'm
sure
it
can.
It's
been
done
with
nupoc
core,
and
so
it's
probably
easier
to
do
with
with
the
community
fork,
because
they
seem
to
have
cleaned
up
some
of
the
hard
things
like
the
cat
and
proto
serialization
stuff.
They
pulled
all
that
stuff
out.
So
that
should
be
easier.
A
But
if
you
don't
have
to
deal
with
that,
it's
not
no
Falco
I.
Don't
you
don't
quite
understand
it's
not
that
only
the
latest
they're.
All
the
theory
is
in
this
project
and
when
I
say
all
the
theory,
I
really
mean.
Let
me
go
back
to
Newman,
says
comm,
slash
papers,
everything
in
this
section
all
the
theory.
All
four
pages
of
papers
is
in
that
will
be
in
the
project.
We
don't
need
backtracking,
TM,
actually
for
anomaly
detection.
The
only
thing
might
be
for
anomaly
detection.
A
You
might
want
to
use
the
backtracking
TM,
but
we
were
moving
away
from
anomaly
detection,
so
you're
right
about
that.
The
anomaly
detection
stuff
was
not
most
of
the
the
product.
Ization
of
that
included.
Non-Biological
changes
that
were
throwing
out.
We
did.
The
backtracking
thing
is
non-biological,
so
we're
threw
it
out
and
we're
moving
forward
with
just
the
biological,
the
biological
stuff.
A
So
it's
not
just
the
latest.
It's
everything,
because
all
the
latest
theory
is
built.
On
top
of
you
know
the
idea
of
thousands
of
synapses
and
what
are
they
doing
in
many
columns
in
spatial
bullying
and
temporal
memory?
The
newer
stuff
just
adds
this
idea
of
motor
place
location
object,
spaces
all
that's
all
that
stuff.
On
top
of
it
and
the
layers
of
the
of
the
cortical
column.
D
A
A
A
B
B
A
A
A
That
would
be
cool
one
of
the
things
I
wanted
to
do
with
this
was
it
would
be
nice
if
I
could
you
know,
click
it
and
then
move
the
feature
to
add
a
feature
change
a
feature,
so
you
could
sort
of
edit
it
through
the
UI
instead
of
touching
the
amyl,
but
you
know
that's
not
totally
necessary.
Just
bonus.
A
D
D
B
A
A
There
you
are
I,
think
you're
on
the
team
on
zero
boards,
so
I
need
to
add
you
to
a
board
somehow
or
add
the
team
to
the
board
or
something
about
this
board
change.
Permissions
changed
now:
twitch,
Team,
Twitter,
so
I
have
to
add
you
to
the
twitch
team.
So
let's
go
twitch
team
view
team
page
members
now.
B
A
A
A
Okay,
team
members
settings
this
team
is
private
or
visible
to
those
outside
slack
team
linking
Trello
business
class
now
members
and
back
to
board.
So
we
should
have
one
board
on
here
and
I
hope
you
guys
have
full
access
to
this,
like
you
should
be
able
to
create
a
new
Guardian
I
got
to
get
back
to
side.
Hustle,
see
you
all
the
next
one,
thanks
bit
code
for
joining,
and
thanks
for
the
subscription,
that's
awesome,
you're
the
first
person
to
subscribe
to
my
channel
so
kudos.
A
A
Right,
what's
so,
cool
I
thought
you're
leaving
the
code
all
right,
so
we're
all
we're
all
collaborating
on
the
same
code
on
the
same
well,
the
same
yeah,
the
same
code
in
the
same
Trello
board.
So
this
is
cool
and
honestly
you
guys,
if
you
have
any
ideas
or
something
throw
them
up
there
and
I'll
sort
them
out
the
next
time
I'm
online.
We
could
talk
about
them
or
whatever,
oh
dude.
Thank
you.
That's
so
cool!
A
A
A
D
A
A
A
A
I'm
gonna:
stop
you
don't
want
it,
you
don't
want
a
subscription.
Oh
you
wanted
to
do
it
right.
You're
gonna,
subscribe
yourself.
It's
up
to
you.
Okay
I
mean
it's
always
more
authentic
if
somebody
else
subscribes
himself,
but
if
I've
been
sort
of
gifting
out
subscriptions
every
once
in
a
while,
so
I'm
more
than
happy
to
accept
a
subscription
from
anybody.
A
Okay
back
to
where
we
were
at,
you
know
what
I
should
do,
though
this
board
is
so
this
is
gonna,
be
my
favorite
board
show
menu.
You
can't
do
anything
to
the
Trello
Oh.
You
can't.
You
can't
give
comments.
Okay,
what
is
the
problem,
then?
Can
you
create
a
new
one?
Okay,
something's
wrong,
then
what
should
it?
Oops
whoops
whoops?
Not
this
about
this
Bourgh
change.
Permissions
commenting
permissions
team
members,
add
remove
missions,
all
members
team
members.
Okay,
does
that
help?
Hopefully
you
can
comment
now.
A
A
Let's
go
back
to
the
team's
page,
maybe
a
team
organization.
Public,
oh
I,
should
make
this
public
yeah,
make
it
public
I,
don't
mind.
Publics.
Fine,
but
I
do
want
to
view
the
team
page
check.
Something
I
can
give
comments
and
check.
Something
is
solved.
I
can't
I
never
subscribe
to
any
one
twitch
before,
but
you
are
the
right
person
to
start
thanks.
That's
nice,
film,
tire
I'd
like
to
what's
your:
what's
your
real
name,
it
can't
really
be
film
tired,.
A
A
It's
better.
If
we
make
suggestions.
Okay,
all
right,
that's
fine,
yeah
I
could
make
you
guys
admins.
Let's
see
how
it
goes
this
way
because
I'd
love
it.
If
you
guys
comment
comment
all
day
long
and
then
maybe
during
my
twitch
sessions,
I'll
I'll
update
it
and
with
with
feedback
and
we'll
see
if
it
works.
That
way,
and
if
that's
not
working
quick
enough,
if
you
guys
are
moving
on
stuff
while
I'm,
not
live-streaming,
then
we'll
deal
with
that.
Then
that
makes
sense
to
me:
okay,
cool.
A
Your
parents
are
weird
on
names:
yeah.
We
have
to
update
the
instructions.
I'm
gonna
put
this
off
I
think
for
now,
because
we
might
not
even
have
a
JavaScript
prop
portion
of
this
at
some
point
and
I'm
really
interested
in
this
gherkin
thing.
I
think
this
might
be
the
meat
of
this
project
really
is
to
define
functional
tests
in
this
agnostic
way,
especially
since
we've
got
our
object.
Files
in
yam.
Also
anybody
can
read
those
in
the
ammo
and
then
and
yeah.
So
I'm
gonna
look
into
that,
probably
next
before
a
big
chunk
of
work.
A
A
A
A
A
B
A
Just
going
to
reply
here
not
specifically
to
Falco,
just
like
a
general
reply:
yeah
Monday,
so
you
guys,
if
you're
interested
in
figuring
out
what
we're
working
on
in
research,
I,
think
Jeff's
Jeff's
gonna,
do
a
paper
review
or
something
on
Monday
at
the
research
meeting.
He's
definitely
has
a
topic
and
I'm
sure
it
probably
has
to
do
is
with
the
thalamus
so
Monday,
it's
it's
like
10,
10,
15,
ish
or
so
I
will
be
live-streaming.
A
A
A
B
A
Will
says
I
I
might
have
a
go
at
adding
a
build
system
of
the
validates
code,
style
and
PRS
for
each
of
the
sub
languages.
I
may
also
look
at
bringing
in
the
Java
HTM
port
to
hook
up
an
agent
that
that
would
be
cool.
That
would
be
cool,
I
mean
I
would
think
it
would
be
cool
to
have
JavaScript
directory
Python
directory
Java
directory
and
then
because
we've
got
HTML
3
of
those
languages
and
it
would
be
neat
to
DES.
A
Let
people
work
in
whatever
language
that
they
prefer,
because
people
are
going
to
be
more
productive
and
in
their
their
original
programming,
language
or
their
comfort
or
comfortable
chronic
programming
language.
So
that
I'm
all
for
that.
But
but
just
know
that
I
mean
this.
The
code
we
wrote
in
the
JavaScript
part
might
all
go
away
if
we're
gonna,
replace
it
with
pygame
and
and
it
which
is
fine,
I
mean
the
only
reason
we
wrote
this
code
and
it
wasn't
much
code
right.
It's
really.
A
Barely
any
code
was
just
to
visualize
the
objects,
so
we
could
make
sure
that
they
visually
made
sense,
because
I
can't
look
at
this
and
visualize
it
at
all.
So
that's
really
the
only
reason
we
did
that,
but
but
that's
a
good
idea.
Well,
if
I
fully
support,
it
have
some
object:
visualization
code
and
a
Java
Script,
my
change
soon
but
anyway,
and
we
have
some
Python
code
with
an
example.
A
A
A
Alright,
that's
it
that's
it
folks!
That's
it
folks!
Okay,
let's,
let's
wrap
up
with
the
raid.
This
is
cool.
This
is
the
first
time
this
is
a
couple.
I
had
a
couple
of
nice
firsts
today.
First
time,
I've
ever
been
rated
by
new
noise
networks.
This
guy
is
interesting
because
he's
writing
a
game
like
stardew
valley,
but
it
has
like
a
climate
change
context
to
it.
So
I
go
and
watch
them
and
it's
interesting
to
see
the
game
get
better
and
better
as
you
go
and
watch
them.
A
A
D
A
A
All
right,
I
think
it
just
happened.
I
think
it
happened.
It's
happening
right
now.
There
you
go
you're
too
slow
all
right.
So
now,
I
gotta
figure
out
how
to
turn
this
on.
My
chat
screen
just
died
because
of
that
somehow,
oh
well,
all
right,
I'm
glad
you
guys,
like
the
stream.
Let's
go
raid,
somebody
let's
go
raid.
Somebody,
let's
see
who's
streaming.
A
This
was
a
fun
one.
Yesterday,
it
just
wasn't
I
just
couldn't
I've
already
I've
already
raided.
All
these
guys,
I've
haven't
raided
talk
to
me,
goose
man,
I
like
that
guy
he's
fun.
He
looks
like
he's
playing
a
game,
though
I'd
rather
read
somebody
in
science
and
tech.
Its
market
research,
Vista,
forgotten,
gamedev,
I've,
seen
him
before
he's
fun,
I
could
go,
watch
gets,
Bo
gets
Bo,
I
gotta,
see
him
a
little
bit.
A
Former
Triple
A
gameplay
programmer
gone
indie,
so
he's
working
on
games,
Oh
radical
hi,
it's
farcry,
4,
farcry,
3,
Ghost
Recon
are
these
the
games.
You've
worked
on
or
the
games.
You
played
oh
worked
on
me,
so
this
guy's
live
now
and
he's
gay
Minh.
Let's
give
him
a
few
more.
Oh,
that's
fun!
Look
at
that,
that's
fun!
Ok,
let's
get
him
of
a
few
followers
or
a
few.
Let's
go,
let's
go
rate
them!
That's
what
I'm
trying
to
say.
Let's
go
raid,
the
guy,
ok,
I
gotta
get
in
my
room
now
raid
gets
phone.
A
One
viewer
is
ready
to
raid.
Do
I
still
have
few
viewers,
seven?
Okay,
all
right!
You
guys
ready
thanks
for
watching
the
stream
again,
I
will
be
online.
Monday
I
hope.
You'll
join
me
for
the
new
mental
research
meeting,
because
it
should
be
interesting
if
you're
interested
in
neuroscience
join
me.
There
take
care.