►
From YouTube: AI Chat Episode 1: machine learning, deep learning, ANNs, AI in games, biological systems
Description
Broadcasted live on Twitch -- Watch live at https://www.twitch.tv/rhyolight_
An overview of today's AI landscape. How AI might apply to games.
A
A
So
it
seems
like
when
I
start
right
on
time
it
takes
a
while
for
people
to
start
flooding
in
so
if
I'm
going
to
have
sort
of
a
scheduled
show
with
content,
I
thought
I
might
start
the
stream
a
little
bit
earlier,
and
hopefully
that
doesn't
bug
anybody.
So
let
me
know
if
there's
a
better
way
to
do
it.
Otherwise,
I'll
do
something
like
this
and
I'll
just
keep
something
like
this
on
the
screen.
That's
me
doing
that.
A
A
A
A
Yes,
we
are
starting
soon:
I'm
gonna,
I'm
gonna,
wait
until
one
o'clock,
I
got
about
four
more
minutes
so
and
I
even
I
have
an
agenda
and
everything
and
I
might
nobody
might
even
watch
this,
because,
when
I
put
it
to
you
table,
probably
turn
it
so
I'm.
Basically,
just
talking
to
myself,
it's
practicing
in
front
of
the
mirror,
a
very
elaborate
expensive
mirror.
A
A
B
A
A
A
Eventually,
I'll
just
animate
this
there's
so
many
things
you
can
do
on
Twitch,
it's
very
cool
aside
from
like
things
popping
up
on
the
screen.
If
anybody
follows
me,
I,
like
pops
up
on
the
screen,
which
is
cool,
chat
pops
up
on
the
screen
too
I
know
most
of
you
guys
know
this,
but
this
is
all
new
for
me.
Yeah
nice
name,
I've,
never
thought
in
the
midst.
It
would
use
twitch
for
something
serious.
This
account
is
for
trolling
great,
so
your
Tachyon
on
on
the
forums,
great
yeah
that
that
that
name
not
not
good.
A
Anyway,
there's
a
discord:
if
you
want
to
chat
in
discord,
you
know
where
to
find
it
I,
don't
I,
don't
know
how
to
say
your
name
die.
Rick.
Eight
eight!
Is
that
right,
a
couple
more
minutes
and
we'll
start
in
the
meantime.
I'm
just
messing
around
yeah
I
go
to
discord.
Where
is
it
a
go
to
HTM
forum?
There's
a
link
there
and
there's
also
on
my
twitch
page
there's
a
link
down
there
that
says
discord
and
you
can
click
it,
but
I'm
not
I.
A
A
Dee
Eric's
Dee
Eric's,
alright,
alright,
let's
get
started
because
it's
about
that
time.
It
is
one
o'clock,
so
I'm
gonna
remove,
wait
still
trying
to
figure
out
the
twitching
her
face
here
we
go.
I
must
turn
this
off.
I
still
have
chat
up
and
I'm
going
to
talk
about
AI
stuff,
so
I
wanted
to
start
real
basic.
So
I
wanted
to
talk
about
terms
because
the
terms
can
be
confusing
honestly.
What
is
artificial
intelligence
versus
machine
intelligence
versus
machine
learning,
because
they
all
honestly
feel
like
different
things
to
me.
First
of
all,
artificial
intelligence.
A
The
term
I
really
dislike
I've,
never
felt
comfortable
with
it,
because
what
is
artificial,
but
if
we
create
intelligence,
what
is
artificial
about
it?
It's
not
artificial.
It's
like
saying
artificial
life.
If
you
create
something
that
we
categorize
as
life,
it's
no
longer.
It's
not
artificial.
Neither
is
life.
Where
isn't?
It
is
intelligent
where
it
isn't.
There's
no
artificial
in
any
aspect
in
my
head,
so
I
hate
the
term
artificial
intelligence.
A
I
think
you
should
call
it
biological
intelligence
or
non-biological
intelligence,
which
is
why
I
like
the
term
machine
intelligence,
because
that's
the
closest
thing
to
non-biological
intelligence.
That's
what
we're
working
on
machine
intelligence!
That's
what
people
say:
artificial
intelligence
equals
equals
human-made,
but
the
terms
are
on:
let's
call
it
what
it
is.
A
It's
it's
machine
intelligence,
but
that's
what
I
like
to
call
it
so
I
try
you
have
to
use
the
term
AI
and
today
as
well,
because
every
it's
a
huge
hype
term,
but
I
don't
like
to
use
it
I
would
rather
call
it
what
it
is:
we're
working
on
non-biological
intelligence
systems
and
we
specifically
are
looking
towards
biological
systems
to
better
understand
them.
So
machine
learning
now
I
think
is,
is
I.
Think
of
as
a
field
of
artificial
intelligence
research
and
it's
all
obviously
machine
intelligence.
A
Let's
just
ignore
this
term:
let's,
let's
call
it
machine
intelligence,
Ted
of
artificial
intelligence.
I'll
just
do
that
and
then
we
can
but
I'm
gonna
change.
I'm
gonna
keep
the
name
of
the
of
the
channel
or
the
the
chat,
of
course,
because
that's
how
people
recognize
it
nobody's
going
to
go
to
machine
intelligence
chat
right
so
anyway,
when
we're
talking
about
machine
intelligence,
machine
learning
is
a
type
of
machine
intelligence
and
it's
sort
of
is
everything.
We've
got
today
can
be
sort
of
be
categorized
as
machine
learning.
People
use
this.
A
These
two
terms
also
interchangeably
but
I
like
to
refer
to
machine
learning,
as
encompassing
all
the
artificial
neural
networks
that
were
building
today.
I
think,
if
you
talk
to
someone
that
was
doing
deep
learning,
they
would
say
they're
working
on
machine
machine
learning
and
and
if
anyone
is
using
artificial
neural
networks,
they
probably
tell
you
they're
working
on
machine
learning,
so
that
term
sort
of
exists
in
a
more
technical
way
for
the
engineers
or
and
the
implementers
that
are
working
on
it,
that
they
like
to
use
that
term.
A
When
they're
talking
about
specific
implementations,
I,
never
refer
to
HTM
as
machine
aren't
learning
necessarily,
but
you
could.
You
could
call
it
that
I
like
to
use
the
term
machine
intelligence,
because
it's
more
descriptive.
It's
not.
We
want
it's
the
intelligence
bit
we're
trying
to
crack.
We
can
make
things,
learn
that's
easier
than
making
something
intelligent.
A
You
have
to
learn
to
be
intelligent,
so
we're
partially
there
but
okay,
so
that's
sort
of
the
terminology
bit
I
wanted
to
get
past
and
by
the
way,
if
anyone
has
questions
at
any
time,
just
pipe
in
on
chat
at
some
point,
I'll
get
the
voice
thing
working
in
discord,
but
not
today.
So
in
addition
to
this
terminology,
I
also
want
to
talk
about
strong
versus
weak
AI
and
I
want
to
be
really
clear
everything
that
humans
have
created
up.
Until
this
point
is
we
chaotic?
A
A
There's
the
Turing
test,
which
I
think
nobody
really
uses
much
anymore
and
there's
other
methods
of
trying
to
decide
if
something
as
intelligent,
I,
don't
necessarily
want
to
talk
about
those
methods,
but
there's
a
difference
between
things
that
are
sort
of
intelligent
or
locally
intelligent,
or
you
know
able
to
apply
learning
and
interesting
ways
and
in
very
grant
and
various
specific
fields
right
that
the
weak,
AI
stuff
that
we
have
today
like
I,
said
all
of
it.
Everything
we
have
today's
weak
AI
is
super
impressive.
There's,
there's
a
lot
of
stuff.
A
You
can
do
with
weak
AI
and
I'll
talk
a
bit
about
that
later,
as
we
talk
about
sort
of
different
forms
of
AI,
but
what
we're
really
striving
towards
is
strong
AI
what
we
really
want.
What
we
really
need
is
strong
AI
the
reason
self-driving
cars
don't
work
today
is
because
we
don't
have
strong
AI
everything
we
have
is
weak
and
you
can't
shove
a
bunch
of
weak
AI
systems
together
and
and
create
strong
AI
and
AI,
and
I
really
don't
think
that
these
systems
will
naturally
evolve
into
strong
AI
systems.
A
I
think
there
must
be
some
form
of
paradigm
shift
for
today's
weak
AI
systems
to
make
the
jump
to
strong
AI.
Now
that's
a
topic
of
conversation
on
this
channel
which
I'll
talk
about
later,
so
you
can
say
AGI
instead
of
strong
AI.
If
you
want
to
use
that
term,
it's
pretty
much
the
same
thing,
but
we
in
a
way
I
like
to
think
about
it
is
if
you're
strong,
AI,
you
understand
and
things
and
that's
a
that's
a
loose
term,
but
but
I
think
it's
an
appropriate
term.
B
A
Looks
at
all
your
photos
and
says
how
there's
Steve,
there's
Steve,
that's
Steve,
that's
Steve!
That's
Jill!
There's
Steve!
That's
Jill!
That's
Muhammad,
bla,
bla,
bla,
bla,
bla!
All
through
your
your
photos
do.
Does
that
a
I?
Does
that
system
understand
your
your
relationships
with
your
friends?
Does
it
understand
that
these
are
humans?
Does
it
understand
what
a
human
life
is?
Does
it
understand
the
environments
that
these
things
exist
in
it
doesn't
understand
anything?
B
A
So
that
when
you
predict,
when
you
make
an
action
against
the
world,
you
understand
it
well
enough
to
predict
what
will
happen
in
that
environment
that
strong
AI?
That's
what
we're
trying
to
strive
for
I,
say
we
I
sort
of
mean
humanity
in
general.
I
also
mean
the
company
that
I
work
for
our
goal
is
understand:
intelligence,
how
it
works
in
the
brain
and
to
create
intelligent
systems
in
non
non-biological
ways.
Okay,
so
weak
versus
strong
went
over
any
questions
feel
free
to
add
them
in
chat.
I.
A
Think
most
of
you
guys,
probably
this
is
like
a
review,
but
I
was
hoping
that
some,
some
folks
from
from
twitch
that
have
been
interacting
with
it
might
be
interested
in
learning
a
bit
about
AI
from
the
ground
up
would
join
into.
So,
if
you
have
any
questions,
let
me
know,
even
if
they
seem
dumb
I,
don't
mind.
There
is
no
stupid
question
alright.
A
So
let's
talk
about
some
different
types
of
AI
and
I'm
still
using
the
term
AI,
because
you
have
to
so
I
like
to
talk
about
old
AI
and
this
stuff,
you
might
call
it
classic
AI,
but
this
stuff
has
been
around
for
decades
and
decades.
It's
a
lot
of
this
is
rules
rule-based
systems
like
the
like
lookups
lookup,
table
type
systems
and
then
just
expert
systems.
When
we
talk
about
expert
systems,
I
can't
spell
this:
it's.
A
A
subject
matter,
expert
or
a
team
of
experts,
custom
wrote
some
software
to
understand
something,
but
basically
to
two
hundred
to
understand
the
rules
well
enough
to
make
decisions
so,
but
these
aren't
really
intelligent
systems,
they're,
very
fragile
and
brittle
they're,
they're
custom
created
specifically
for
problem
domains
and
they've
been
around
for
so
long
and
they're
everywhere.
Honestly,
if
you're
doing
gaming
stuff,
this
is
a
lot
of
the
AI
that
you
will
use
or
that
you
might
create.
A
A
A
Like
what
Freeman
said,
good
old-fashioned,
AI
sort
of
before
you
get
into
the
the
Bayesian
probabilities
and
statistics
and
stuff
that
come
along
with
neural
networks,
that
you
know
that
that
were
all
those
big
discoveries
and
in
the
end
of
this
last
century,
that
is
the
sort
of
the
next
thing,
and
that
is
I
would
call
a
n
n
simple,
an
tens.
Maybe
point
neurons.
Let's
call
it
simple,
because.
A
B
A
This
is
where
you
have
to.
This
is
tricky
for
me,
because
I
am
NOT.
A
mathematician
I
did
not
go
through
calculus
in
high
school
I
had
to
understand
derivatives
and
stuff
to
understand
how
these
sort
of
neural
networks
worked.
So
you
might
have
to
do
a
bit
of
self-study,
and
once
you
get
it,
though,
it
makes
sense
how
how
it
works,
but
essentially
it's
like
population
effects.
A
If
you,
if
you
get
a
whole
whole
bunch
of
neurons
and
you
can
create
millions,
thousands
who
knows
it
just
a
ton
ton
of
neurons
and
and
then
you
you
connect
them
together
in
and
we're
going
to
have
like
simulated.
This
is
like
a
simulated
layer
of
cells,
so
this
is
really
inspired
by
neurobiology
in
the
late
late,
80s
and
80s,
or
something
was
I,
think
roof
started
to
write
a
bit
earlier
than
that
the
perceptron
was
even
earlier
than
that
I
believe,
but
this
was
originally
I
think
called
the
perceptron.
A
This
idea
of
a
point
of
a
point
neuron
and
essentially
it
gets
it
can
be
connected
to
all
these
different
inputs
and,
at
the
neuron
level,
dependent
on
all
these
different
inputs
in
their
states
right.
Each
one
of
these
can
have
a
weight,
so
you
could
have
one
weighted
really
high
one
weighted
really
low.
So
there's
sort
of
this
thing
is
monitoring
all
this
thing's
it's
connected
to
and
it's
connections
have
the
states
to
and
and
it
can
use
that
to
decide
what
its
state
is
going
to
be.
A
B
A
Keep
in
mind,
though
this
is
a
simple
neuron
model.
The
neuron
model
is
basically
I,
have
an
put,
which
is
an
axon
and
I
can
have
a
whole
bunch
of
inputs,
which
are
all
basically
the
same
and
they
can
have
weights
and
I
decide
what
what
I'm
doing
based
on
those
inputs.
So
what
you
can
do
with
this,
this
type
of
model
is
take
an
image.
For
example,
let's
say:
there's
someone
we're
trying
to
recognize
digits.
This
is
like
the
classic
artificial
intelligence,
hello
world
program.
It's
called
amnesty
M
NIST.
A
If
you
want
to
look
it
up
in
EM
in
ist
and
it's
just
basically
a
huge
collection
of
handwritten
digits
and
they're
all
labeled.
So
you
know
here's
one
and
it's
a
nine
and
here's
one.
It's
a
5,
etcetera,
and
so
in
this
case
you
might
have
you
might
split
this
up
into
pixels
right
and
then
turn
it
put
it
into
an
array.
You
know
so
each
pixel
gets
sort
of
unrolled
down
this
array
and
and
then
we'll
create
a
bunch
of
neurons.
A
B
A
This
is
we're
in
the
realm
of
deep
learning
right
now,
there's
there's
lots
of
different
tricks.
You
can
do
if
you
want
to
learn
more
about
the
different
tricks.
There's
there's
convolutional
neural
networks,
which
I
don't
think
I'll
describe
right
now,
but
maybe
that
could
be
a
topic
of
another,
a
a
chat
in
the
future.
There's
a
generative
networks
or
gans,
which
was
a
big
big
thing
right
now.
A
I,
don't
have
any
experience
with
those
there's
reinforcement,
learning
which
is
very
interesting
because
with
reinforcement,
learning
and
and
I
guess,
I
should
call
it
deep
reinforcement
learning.
Therefore,
if
we're
relating
it
to
neural
networks,
because
you
can
do
reinforcement,
learning
without
the
deep
part,
yeah
I
think
tachyon
is
saying:
I'll
use
your
other
name
yeah,
it's
it's
all
probabilistic,
absolutely
the
the
deep
learning
approach
is
is
relies
heavily
on
on
Bayesian
dynamics
in
probabilities
that
that
exists
there,
which
is
really
cool
stuff.
A
Let's
talk
quickly
about
reinforcement,
learning,
because
I
think
it's
really
cool
reinforcement
learning
is
not
necessarily
an
artificial
neural
network.
Reinforcement
learning
is
when
you
have
an
actor,
and
you
have
some
environment
that
the
actor
wants
to
act
upon
so
with
reinforcement,
learning
the
environment.
Has
a
state
and
the
state
changes
over
time
something
over
here,
something
or
whatever
and
I
can
assume
like
this
is
reality,
and
this
is
an
actor
that
I
was
acting
upon
reality
when
the
actor
acts
upon
reality.
A
He
gets
a
state
back
and
there's
some
connection
so
that
you
can
tell
whether
you
should
be
rewarded
or
punished
for
that
action
based
upon
the
state.
So
there's
a
there's
a
goal
reward
system
here,
that's
really
interesting
because
with
h0
thank
you,
tachyon
I
really
appreciate
it.
You
guys
are
great.
A
A
The
current
deep
reinforcement
learning
systems
have
a
deep
learning
model,
but
the
problem
with
deep
learning
models
and
I'll
get
into
this
in
a
minute
is
that
their
best
doing
spatial
pattern,
recognition
and
in
the
world
where
you're
taking
action
and
the
state
of
the
world
is
constantly
changing.
You
need
to
be
good
at
temporal
pattern.
Recognition
and
HTM
is
good
at
temporal
pattern,
recognition
so,
but
I'll
get
into
that
later.
I.
A
A
I
would
call
biological
neural
networks
I'm
sure
that
there
are
more
than
just
HTM,
but
HTM
is
obviously
the
one
that
I'm
going
to
talk
about
a
lot
there,
and
the
main
difference
here
is
remember
and
I
told
you
in
a
point:
neuron
there's
like
there's
an
output
and
then
there's
a
bunch
of
weights,
and
then
it
decides
based
on
the
state
of
all
these
weights
and
all
these
things
that's
connected
to
whether
it's
going
to
be
active
or
not.
The
a
real
neuron.
B
A
A
So
the
thing
is
the
there's
different
receptivity
zones
here,
there's.
If
this
neuron
is
stimulated
in
different
places,
it
does
different
things.
So
if
it
gets
stimulation
on
these
synapses
on
these
apical
dendrites,
it
might
mean
one
thing
and
signals
get
sent
to
the
soma
that
indicate
in
some
cases,
indicate
that
something's
going
on
there.
If
you
get
stimulation
close
to
the
cell
body,
this
is
proximal.
A
This
is
a
proximal.
This
is
apical
and
then,
if
you
get
stimulation
close
to
the
cell
body,
this
causes
cell
fire
right
away
this.
This
is
what
causes
cells
to
the
cell,
to
fire
and
and
for
years.
We
didn't
even
know
why
these
others,
these
other
dendrites,
were
there.
So
all
the
rest
that
are
way
out,
you
know
beyond
some-some
range.
These
are
distant
and
and
until
recently
we
really
didn't
know
what
they
did
or
what
the
point
was.
A
The
point
is
anyway
in
the
HTM
neuron
model,
and-
and
we
think
this
is
continually
being
backed
up
by
experimental,
neuroscience
research-
we're
constantly
reading
papers
about
this,
and
this
is
called
an
N
DMA
spike,
okay
in
DMA
spike.
If
you
want
to
look
it
up
in
the
neuroscience
literature,
if
one
of
these
segments
distal
segments
gets
a
whole
bunch
of
activity
on
it,
it
will
send
a
spike
down
the
dendrite
okay
to
the
so
on,
but
the
soma
won't
fire.
It
will
just
increase
its
voltage
enough.
A
So
it's
really
close
to
firing
and
we
call
this
a
predictive
state.
Then
the
cell
is
closer
to
firing
than
all
of
its
neighbors,
so
when
it
does
get
proximal
stimulus,
it
fires
first
first
because
it's
almost
to
the
threshold
of
firing.
This
NDMA
spike
causes
pyramidal
neurons
to
go
into
it's.
This
predictive
state,
the
point
neuron
in
deep
learning,
networks
and
other
artificial.
The
simple
and
ends
don't
have
this
at
all:
they've
got.
A
There
is
no
predictive
state,
it's
either
on
or
off,
and
then
you
know,
or
you
can
have
scalar
or
value
associated
with
it.
So
the
computation
occurring
here
is
at
its
core
different.
Like
you,
you
can't
take
and
deep
learning
network
right
and
just
install
this,
because
the
deep
learning
architecture
is
dependent
on
the
architecture
of
the
perceptron
or
the
the
point
neuron.
So
you're
changing
the
functionality
of
the
point.
Neuron
itself.
You
have
to
change
the
architectures
of
all
your
networks
in
order
to
do
that
and
all
the
frameworks
that
are
running
them.
A
So
this
is
not
easy
to
just
swap
and
say
well,
we'll
just
upgrade
our
deep
learning
networks
to
use
the
HTM
your
own
model.
It
just
doesn't
work
that
way.
Okay,
okay,
so
we
have
to
talk
about
spatial
versus
temporal
pattern,
recognition
next,
okay,
what
happens
when
the
neuron
fires
after
the
first
one
that
fired
first?
So
there's
a
chain
of
events?
If
so,
you've
got
a
real
you've
got
to
think
of
any
layer
of
cortex.
Let's
say
here's
a
layer
of
cortex
I:
don't
care
what
it
does.
It
gets
some
proximal
input.
A
Okay,
it'll
it'll
get
some
some
some
layers
get
apical
input,
some
don't,
but
some
layers
get
proximal
input
and
some
will
get
distal
input
as
well,
and
you
don't
know
where
it's
coming
from.
So
if
you're,
if
you're
a
neuron
in
this
layer
of
cortex,
you
don't
know
what
your
input
represents
in
any
way,
which
is
tricky
right,
so
you
have
an
output.
A
So
so
one
neuron
in
here
might
have
a
proximal
connection
that
comes
from
over
here
and
and
distal
connections
that
come
from
wherever
over
there
when
it
fires
it
fires,
it
has
an
output.
It's
a
part
of
this
whole
layers
output.
So
it
will.
It
will
be
a
part
of
that
layers
of
representation.
Each
layer
represents
something
if
you
look
at
the
whole
population
of
cells
and
what's
on
any
point
in
time,
that's
representing
something
right.
A
So
this
neuron,
the
fact
that
it's
firing
at
some
point
in
time
represents
something
and
that
big
population
of
cells-
and
we
it's
really
hard
to
tell
what
it
represents.
There
could
be
a
part
of
a
big
whole.
You
know
a
feature
that
maybe
a
thousand
neurons
together
will
represent.
You
know
what
I
mean
so
to
your
question
when
it
fires
after
the
one
that
fired
first,
it's
just
another
in
the
chain
of
events,
so
it
would
go.
A
A
It
potentially
could
have
apical
input
and
if
you,
if
you
block
them
in
different
ways,
you
can
you
can
connect
one
to
the
other
and
that
this
way
to
that
way-
and
you
know,
there's
a
million
things-
you
might
do
knowing
that
at
some
point
you
need
to
have
sensory
input
from
the
world
right,
that's
that's
the
thing
sensory
input
has
to
come
from
somewhere
and
then
you
create
some
form
of
architecture
to
understand
that
sensory
input.
That's
the
key
point
understand
it.
A
A
A
This
is
this
is
getting
kind
of
deep,
but
in
our
comms
plus
paper.
I
can
remember
this
properly.
So
we
have
this
three
3-mile
three
I
don't
want
to
get
into
this
right
now,
but
I'm
gonna
skip
that
question.
Falco
asked
it
on
the
form.
Okay,
it's
sorry
to
do
that.
It's
too
deep!
It's
too
deep
for
chat
session.
That's
like
a
very
specific
HTM
question.
A
Note
no
worries
spatial
analysis
versus
temporal
analysis,
so
most
of
the
things
that
current
weak
AI
systems
like
deep
learning
pattern,
video,
sorry,
deep
learning,
image
classification
systems
like
your
Google
photos
or
your
Facebook,
you
know
face
recognition,
voice,
recognition.
All
of
that
stuff
uses
deep
learning.
B
A
It
uses
spatial
analysis
tools
so
and
what
I
mean
by
that
is,
if
you've
got
a
corpus
of
images,
let's
say:
you've
got
like
image:
net
is
one
that
has
just
I,
don't
know
how
many
millions
and
millions
of
images
and
each
image
it's
got
a
list
of
labels.
So
this
one
has
a
cat
and
a
dog
and
a
house
and
a
badger
I,
don't
know
it
just
it
lists
all
the
things
water
tree.
A
A
A
So
we
can
train
deep
learning
systems,
because
if
you've
got
a
million
photos
of
dogs,
then-
and
you
run
each
one
of
those
like
you-
can
run
whole
batch-
its
thousands
of
pictures
right
and
there's
there's
a
way
that
you
can
map
it
out.
Usually
they
do
this
with
convolutional
neural
networks,
which
I
don't
want
to
get
into
hugely,
but
you
can
take
you
know
every
image
you
get
and
you
create.
A
You
know
these
a
whole
bunch
of
kernels
convolutions
that
look
at
different
sections
and
each
one
looks
at
a
different
section:
they're
all
trying
to
do
feature
recognition,
some
portion
of
the
of
the
space
you
can
talk
about
channels
and
how
much
information
is
embedded
in.
He
picked
each
pixel
there's
there's
a
ton
of
ways
to
extract
information
or
to
encode
information
in
images
in
a
way
that
they
eventually
are
going
to
be
represented
by
a
big
ole
array
of
point.
A
B
A
A
one
of
these
networks
that
you've
created
and
telling
it
at
the
end
every
time
you
show
it
something
you
allow
it
to
know
if
it
was
right
or
wrong
and
what
it
classified
it
to
be.
And
then
you
take
that
error
and
you
do
some
form
of
credit
assignment
to
identify
which
neurons
contributed
to
the
accuracy
of
that
error
and
your
accuracy
of
that
classification
and
you
back
propagate
the
error.
A
A
A
So
the
combination
of
these,
when
you
start
adding,
you
know
a
bunch,
a
bunch
of
hidden
layers,
because
you
can
add
as
many
hidden
layers
as
you
want
is
you're.
Just
constantly
refining
these
into
sort
of
you
can
you
can
decrease
the
size
of
the
features
you're
getting
at
so
at
the
end,
you
can
have
a
classification
of
the
most
commonly
seen
things
like
dog
or
cat
right,
the
things
that
people
are
actually
labeling.
A
A
Dissin,
though
they'll
tune
everything
and
increase
the
the
quality
of
all
the
parameters
and
then
run
it
again
and
then
train
again
and
every
time
they
do
that
their
refining
the
parameters
and
training
it
again,
and
then
they
do
it
again
and
you
get
to
a
point
where
you're
happy
with
the
classification
performance
and
you
put
that
model
into
production.
That
model
will
no
longer
learn
ever
again.
It
does
not
learn
it
learned
when
you
trained
it
and
it
learned
on
all
the
spatial
stuff
that
it
saw.
A
But
it
has
no
perception
of
what
came
after
the
next
thing,
because
it's
just
doing
batch
processing
of
spatial
input.
So
it
has
no
idea
what
time
is
which
is
totally
different
from
you
and
me,
and
every
other
thing
that
lives
and
thinks
and
moves
on
this
planet.
Our
brains
all
evolved
specifically
to
react
to
move
through
space
and
time
is
a
key
component
of
that.
A
So
we
need
a
temporal
pattern,
recognition
and
we're
try
and
do
it
the
way
the
brain
does
it
and
the
problem
with
current
deep
learning
systems.
And
yes,
there
are
some
types
of
like
deep
learning
like
LST
m'lord,
long,
short-term
memory,
which
is
a
funny
name
and
something
else
is
I
read
recently
the
that
are
trying
to
do
temporal
pattern,
recognition
but
they're
all
just
doing
it
in
batches.
A
You
know
it's:
it's
hacking,
deep
learning,
spatial
pet,
what
it's
good
at,
which
is
a
spatial
pattern,
recognition
and
it's
trying
to
hack
it
it
seems
like
I,
have
to
me.
You
have
to
run
everything
in
batches.
What
you
need
is
an
online
learning
system.
Online
learning
means
that
every
new
piece
of
information
you
get
you've
learned
from
it
right
away.
You
don't
have
to
you
know
like
like,
for
instance,
photo
classification.
If
you
go
to
google
photos,
they
are
constantly
in
the
background.
A
A
A
A
So
that's
one
one
thing
Oh,
so
you
might
ask:
how
do
you
do
voice
right?
How
do
you
do
voice
recognition
because
or
like
Shazam,
that
that
service,
that
you
listen
to
a
part
of
a
song
and
it'll
say?
Oh,
that's,
that's
you
know
whatever
song
is
so
in
all
those
cases
that
I've
ever
investigated
what
you
end
up.
A
Taking
a
sample
of
some
temporal
stream
of
data
and
then
running
a
spatial
pattern,
recognition
analysis
on
it
so
like
you'll,
do
a
Fourier
transform
on
like
an
audio
file
to
break
it
up
into
frequency,
bins
or
you'll.
Do
a
spectrogram
analysis
or
something
like
that
of
a
voice
file,
and
then
what
that
essentially
does
is
it
gives
you
a
spatial
representation
of
it?
A
Temporally
it'll,
actually,
you
know,
draw
it
out
over
time
and
you'll
get
a
little
image,
and
so
you
can
do
spatial
image,
recognition
and
pattern
pattern
matching
on
that
image,
because
what
you
can
do
like
forces
am,
for
example,
you
can
have
different
signatures
for
different
windows
in
the
song
and,
if
you
do
that
off
ahead
of
time-
and
you
have
those
easily
indexed
and
whenever
anybody
sends
you
one
all
you
have
to
do
is
match
it
against
those
and
so
I
think
that's
the
type
of
stuff.
It's
doing.
A
A
A
Some
of
the
I'll
just
throw
out
some
ideas,
and
if
anyone
has
something
interesting,
they
want
to
to
bring
up
then
go
for
it,
but
so
some
of
the
obvious
places
I
think
that
you
could
do.
F2V
FFT
could
be
used
in
the
encoding
process.
Yeah
yeah
yeah
I've,
always
used
FFT,
see
I've
used
FFT
to
encode
input
from
audio
there's
this
video
of
me
on
YouTube
doing
it
I
got
a
whole
bunch
of
views
from
it,
because
the
band
that
I
was
playing
in
the
background
posted
it
out
on
Facebook
anyway,.
A
A
Yeah
I
can
find
a
link
here.
Let's
do
it
I'll
help.
You
find
a
link,
it's
gonna
be
on
the
forum.
Oh
no,
it's
called
new
pic
critic,
sound
and
analysis.
There
is
new
pic
critic,
yeah
and
there's
there's
the
video
okay,
but
this
is
old
stuff.
This
is
years
ago.
I
did
this
like
five
years
ago,
so
it
might
not
be
the
best
quality
stuff.
A
A
You
can
make
a
prediction:
you
can
get
a
prediction
of
what
you
think
is
coming
next,
but
going
into
the
cell,
the
cell
state,
and
trying
to
extract
what
sequences
could
I
currently
be
in.
What's
contributing
to
this
prediction,
that's
hard
to
do
yeah,
so
you
guys
chat
on
the
forum
connect
on
the
forum.
There's
private
messages
there.
This
would
be
a
great
place
to
do
it.
A
A
If
you
search
grid
cells,
go
to
videos,
here's
a
video
of
grid
cells-
this
is
me
so,
but
I'll
paste
this
in
chat
here.
This
is
super
exciting
for
those
of
you
who
aren't
following
the
HTM
community.
Everybody
else
knows
how
crazy
I
am
about
this.
This
is
so
cool,
so
the
scientists
won
Nobel
Prize
for
this
I
think
in
2014,
2014,
right,
yeah,
I,
think
so
so
they
found
cells
and
I'll.
A
So
when
you
think
about
enemy
control,
and
you
think
about
how
would
you
model
the
enemy?
What
is
that
enemies
model
and
current
AI
systems
for
game?
Devs
I?
Don't
think
those
enemies
really
have
models,
I
think
they
just
have
rules
I.
Think
they're,
mostly
like
expert
systems
that
have
like
a
canned
enemy
behavior,
but
they
don't
learn
from
you
know
the
environments.
They
don't
learn
over
time
with
a
proper
HTM
model
with
that,
so
you
know
a
temporal.
B
A
That
updates
as
an
enemy
moves
through
the
space
of
the
game.
You
would
have
something
that
could
potentially
learn
learn
the
environment
as
it
goes
along
even
other
people's
movements
or
other
enemy's
movements.
I,
don't
know
what
pathfinding
is.
Are
you
talking
about
path?
Integration,
flippin
positions,
use
HTM,
the
game,
master
game
platform
that
controls
the
rules
of
the
world
depending
on
what
kind
of
rules
we
pre
taught
them?
I
you've
lost
me
there.
Maybe
that's
that's,
but
over
my
head,
I'm,
not
sure.
Maybe.
A
You
could
use
HTM
to
model
a
whole
environment,
but
the
thing
is-
and
this
is
a
tricky
thing:
if
you're
gonna
go
this
route,
you
have
to
have
movement
baked
into
the
system
like
you
have
to,
if
you're
gonna
create
an
every
intelligent
thing
that
has
ever
existed,
has
had
the
ability
to
move
through
this
environment.
So
we
need
to
know.
We
need
to
understand
how
movement
contributes
to
intelligence.
A
That's
one
of
like
the
core
research
things
that
we
do
in
my
company
is
try
to
understand
how
movement
yeah
modeling
the
environment
is
a
hard
problem.
That's
that's.
Deep
learning
is
not
necessarily
a
model
heavy
and
it's
model
heavy
in
some
ways,
but
not
it
doesn't
like
create
a
model
of
the
world.
You
need
a
model
of
the
world
of
reality,
not
just
like
an
input
space.
You
know
anyway,
enemy
control,
okay,
environment
control.
That
was
another
thing.
A
But
the
thing
that
we're
also
this
would
go
pair
well
with
reinforcement
learning,
because
what
we're
missing
in
HTM
is
is
an
age
is
agency.
Essentially,
HTM
has
a
creates
a
model,
it's
a
model
of
the
neocortex,
so
it
creates
a
sensory
motor
model
of
the
world
based
upon
movements
through
space.
It
creates
representations
of
the
objects
we've
interacted
with
in
the
space
that
you've
interacted
with,
but
it
doesn't
generate
the
movements
necessarily.
We
still
have
to
figure
out
how
to
how
to
incorporate
this,
with
reinforcement,
learning
systems
that
generate
movements.
A
B
A
Use
already
made
open-world
games
and
generate
quests
yeah
yeah.
You
could
use
any
anytime
so
just
for
deep
learning
systems
anytime,
you
have
of
just
a
buttload
of
data
like
if
you
have
like
10,000
levels
in
a
game
or
in
an
old
game,
or
something
like
that.
You
could
potentially
figure
out
how
to
train
a
model
on
that
on
classifying
something
on
that,
so
that
you
can
generate
new
levels,
so
it
may
be
level
generation.
But
yes,
it's
a
reinforcement.
B
A
Sure
so,
no
so
does
it.
So
this
is
theoretically
okay,
but
theoretically
you
do
not
need
to
know
where
you
are
in
the
world
to
learn
where
you
are
or
to
learn
your
environment.
You
build
it
around
you
and
then
once
you
recognize
think
about.
Like
think
about,
like
you
open
your
eyes
in
a
strange
room,
you
don't
have
to
know
where
you
are
to
create
a
representation
of
that
room
in
your
brain.
A
Let's
say
you
walk
out
the
door
and
you
recognize
oh
I'm,
on
10th,
Street
and
I'm
on
that
building
that
I
never
walked
in
before
you
there's
a
there's,
a
reoccurring
that
occurs
here.
Okay,
you
have
to
you
have
to
understand
something
called
place:
cells
in
addition
to
grid
cells
and
I,
haven't
talked
about
play
cells
at
all
and
I
barely
talked
about
grid
cells,
but
place
cells.
Our
cells
that
respond
to
like
certain
landmarks,
so
they're
like
broader
and
scope.
A
So
when
you
walk
out
the
door,
you
might
see
the
cell
phone
tower
across
the
street,
which
is
a
landmark
that
says
oh
I'm,
on
Fifth
Street,
whatever
that
just
anchored
you
to
a
place
that
you've
been
before
your
grid
cells
operate
within
that
place
area,
so
they
map
space
out
normally,
no
matter
where
you
are
so
it
needs
a
certain
level
of
detail
and
input.
Absolutely
you
have
to
that's
pattern
matching
and-
and
it's
not
just
spatial,
it's
temporal
spatial
temporal
pattern
matching
and
it's
the
same
thing
with
objects
as
it
is
with
environments.
A
A
A
This
is
probably
one
of
a
few
things
and
then
what
if
I
use
my
whole
hand,
even
without
opening
my
eyes,
I
know
exactly
what
it
is,
because
I've
held
this
a
million
times
before
not
a
million,
but
but
you
do
that
by
moving
your
sensors
through
an
object
space
and
then
matching
on
all
the
things
all
the
different
things
that
have
felt
like
that
in
that
space
before
right,
that's
sort
of
a
key
to
HTM,
3
and
new
mentis
theory.
We
have
a
bunch
of
videos
about
this.
A
A
B
B
A
Other,
like
you,
can,
if
you
put
two
you're
going
to
recognize
something
faster
by
grabbing
it
all
at
once
with
all
four
or
five
fingers,
then
then
touch
one
at
a
time.
You
know
you
know
what
I
mean
doing
like
like
that,
that
doesn't
accept,
but
when
they
all
inform
each
other,
then
you
immediately
know
it's
a
cylinder.
If
you
were
to
go
like
this,
it
might
not
be
so
obvious.
A
A
That
that's
deeper
topic
so
either
a
high
level
of
detail,
five
fingers
or
long
term
memory
to
be
able
to
combine
a
lot
of
observations.
Yes,
yes,
you
got
it
okay,
so
that
where
was
I
Oh
game
development
right,
we
modeling
the
players
is
another
way
to
do
it.
You
could
create
a
model
that
just
for
the
current
player
and
it
could
evolve
over
time.
You
know
you
created.
It
follows
the
player
along
on
their
journey
or
whatever.
A
So
you
can
learn,
potentially
things
like
I,
don't
know,
I,
guess
preferences,
but
it's
not
just
spatial
preferences.
It's
like
Oh
they've
done
these
few
actions.
Usually
they're
gonna
do
this
next.
So
if
you
can
get
an
indication
of
they're
likely
to
do
this
next,
their
that
next,
that
might
be
really
useful
in
a
game
because
then
you
can
shock
the
hell
out
of
them
and
they
won't
know.
What's
coming
yeah.
A
The
type
of
player
it
is
here's
one
thing
that
here's
something
that
we
do
that
I,
that
I've
done
quite
a
few
times
so
there's
something
you
can
do
in
HTM,
which
is
create
an
anomaly
model,
and
let
me
maybe
I
should
turn
our
turn
they're,
not
that
one
this
one
so
so,
as
this
is
kind
of
cool
I've,
actually
done
this
in
Minecraft.
So
let's
say
this
is
like
a
Minecraft
world
and
I've
got
a
house
over
here
and
a
tree
or
whatever
and
I've
got
this
track
and
I
go
over
here.
A
A
The
same,
we
use
the
same
tactic
for
other
things,
not
just
position
I'll
get
to
your
question
in
a
minute
in
cognition
we.
What
we
will
do
is
is,
we
might
have
a
stream
of
data
and
and
that's
coming
in,
and
there
might
be
different
patterns
of
activity
that
we
know
we
want
to
recognize.
So
here's
one
pattern:
here's
one
pattern,
I,
don't
know.
There's
one
pattern:
something
like
that,
and-
and
we
know
this
is
some
type
of
classification,
but
it's
never
exact.
You
know
it's
just
like
we
know
this
is
one
pattern
of
activity.
A
It's
another
pattern
of
activity,
and
if
we
want
to
do
that
type
of
temporal
classification,
we
could
do
it
if
there's
a
finite
number
of
these
patterns,
because
we
could
train
one
model
on
this
pattern
only.
We
can
train
one
model
on
this
pattern
only
and
one
model
on
this
pattern.
Only
so
the
only
things
that
these
have
seen
or
is
just
that
class
of
input
and
then
we'll
run
when
we
have
an
actual
input
like
like
that
when.
A
A
A
A
A
We
make
all
of
our
code
open-source
under
an
a
GPL
license,
so
you
can
do
whatever
you
want
with
it
and
research
settings
scientific
settings,
but
if
you
do
end
up
creating
products
with
that
that
code,
we
there's
a
license
involved,
but
but
we
don't
actually
build
anything,
we're
just
we're
just
doing
the
research
and
trying
to
figure
out
the
core
science
of
it,
and
that's
that's
really.
What
Jeff
wants
to
do?
Jeff
Hawkins
is
our
founder.
He
he
founded
palm
and
handspring.
He
invented
the
Palm
Pilot.
You
know
this
is
claim
to
fame,
which.
A
Big
deal
of
the
time-
and
he
you
know,
invented
graffiti-
some
of
the
old
heads
will
probably
remember
graffiti.
Some
people
just
loved
it
so
much
anyway,
he
pretty
much
funds
our
company
in
our
research,
so
that
is
his
goal
is
to
understand
intelligence,
yeah,
the
godfather
of
smartphones
right
read
on
intelligence.
If
you're
interested
he
wrote
that
book.
A
A
Ok,
anyway,
back
to
the
the
topic,
we're
still
talking
about
AI
and
game
development.
We
talked
about
enemy
control,
modeling,
the
environment
and
environmental
actions,
and
modeling
different
players
talked
about
generating
levels.
So
another
way
you
could
use
it
that
you
could
do
today
with
today's
deep
learning.
Technology
is
creating
graphics,
especially
if
you've
got
like
NPCs.
Have
you
seen
you
can
go,
look
on
youtuber
or
or
whatever
there's
a
lot
of
memes
about
it
now,
but
they've
got
AI
systems
that
can
just
create
a
face.
A
What
is
the
URL
of
that
there
there's
a
it
says
this
face
is
not
real
or
this
person
is
not
real,
calm
or
whatever
somebody
will
find
it
for
me,
but
it
will
just
generate
a
face
for
you
and
it'll.
Look
like
a
real
person.
Unless
you
look
really
hard,
it'll,
look
like
a
real
person
and
you
can't
tell
it
it's
not
a
real
person
like
it
looks
really
good
and
that's
cool
for
games
right,
because
you
don't
have
to
do
that.
Just
for
people.
You
could
do
that
for
buildings.
A
You
could
do
that
for
grass
to
do
for
trees.
You
could
create
all
kinds
of
artifacts
in
the
game
that
are
sort
of
pseudo
there
there
it
is.
This
person
does
not
exist,
calm.
Let's
look
at
this
I
bring
it
over
here
there
we
go,
there's
a
kid
that
does
not
exist
and
every
time
you
refresh
it,
let's
see
there's
another
person
that
does
not
exist.
Now,
if
you
start
looking
really
closely
at
these
you'll
recognize
some
interesting
things
like
this
woman
has
a
tooth
in
the
middle
of
her
face,
how's
that
get
there.
A
Oh,
that
I
got
I.
Gotta
move
the
chat
there.
It
is
a
watch
at
my
chest
not
working
on
this
whatever
and
and
if
you
keep
doing
it
like
you'll,
see
some
weird
really
weird
stuff
like
eyes
will
be
different
like
these
eyes.
There's
something
wrong
about
those
eyes,
and
especially
with
hair
and
earrings.
Sometimes
you'll
get
one
earring
that
looks
totally
different
and
can't
this
is
a
pretty
good
one,
but
and
then
another
one
will
like
look
like
it
came
from
outer
space.
Here's
something
like
what
is
this
right
here.
A
A
It'll,
it
doesn't
quite
know
it
doesn't
know
what
an
earring
is.
It
doesn't
know
what
a
scarf
is,
but
if
it
sees
like,
but
it
might
put
together
pieces
of
each
and
think
that
it
makes
sense
right
and
that
that's
what
you
see
and
stuff
like
this,
it's
not
really
understanding
people
or
objects.
It's
just
like
just
generating
things
based
on
all
of
the
different
observations
and
the
probabilities
of
having
those
observations
together
in
the
same
place
before
interesting
stuff
I
mean
this
is
really
cool.
You
could
do
a
lot
with
this
I
think.
A
So
that's
so
that's
something
you
could
do
for
like
creating
NPCs
tons
of
stuff,
and
you
could
do
that
right
now
with
deep
learning
models
and
you
could
train
the
model
sort
of
while
you're
producing
the
software
and
serialize
it
and
you
know,
deploy
it
with
the
system.
It
would
not
be
able
to
learn
in
real
time.
I.
Don't
think
any
of
this
stuff
is
gonna,
be
able
to
learn
in
real
time,
because
you
have
to
retrain
it
and
read
to
pull
it
on
your
running
software
and
I.
A
A
A
B
A
Somebody
showed
me
the
picture
of
the
cat
yeah,
there's
some
really
wacky
stuff
coming
out
of
that.
Okay,
so
I'm
done
with
this
stream
I'm
coming
up
on
an
hour,
I
think
I'll
take
a
few
more
questions.
If
you
want
to
what
what
I
would
like
to
try
and
do
is
raid,
another
channel
and
I've
never
done
this
before
so
I
have
to
figure
out
how
it
even
works.
I
think
I
can
do
it
with
the
chat
and
but
now
I
have
to
figure
out
like
who's
online.
A
Hey
thanks
for
the
for
the
compliment.
I'm
gonna
do
this
every
week,
every
Monday
we'll
talk
about
something
else
and
it
may
not
be
I
might
not
have
a
huge
Genda,
but
especially
if
I
get
voice
working.
Maybe
it
could
just
be
a
time
for
me
to
chat
with
the
community.
Whoever
is
interested
in
talking
about
machine
intelligence
or
AI
or
whatever
you
want
to
call
it
and,
let's
see
so
I'm
gonna
go.
A
A
Yeah
bye
de
da
de
Rick
Derek,
good
luck
on
your
test.
Let's
send
I'm
gonna
send
some.
We
could
so
when
you,
when
you
raid
another
channel,
it's
like
you,
send
all
of
the
people
that
are
currently
watching
your
channel
to
another
channel,
so
I'm
thinking
about
going
to
this
guy
named
Ryu
because
he
seemed
nice
and
he
signed
up
to
two
for
my
channel
and
he's
he's,
got
some
really
cool,
really
clean
code,
so
I
respect
clean
code,
he's
got
really
well
commented
code
and
he
does
all
the
TDD
stuff,
hey
nerd
fever.
A
Thank
you
for
the
follow-up
reciate.
It
I'm
trying
to
figure
out
how
to
raid
another
channel
so
I
think
is
it?
Is
it
I
really
don't
know
if
I'm
good
somebody
know
how
to
do
this?
I'm
gonna
have
to
look
it
up
how
to
raid
twitch
channel
how
to
raid
a
twitch
channel,
okay,
how
to
use
raids?
Oh
it's
just
slash
raid
channel
name
raid.
A
B
B
From
scratch,
inside
building
my
own
platform
for
only
a
game
on
top
that
server
based
and
today
the
whole
focus
is
refactoring
the
server
code,
because
you
know
that
components
have
the
coordinator
which
you
can
see
in
this
diagram.
It's
this
part
of
my
server.
It's
pretty
complicated
and
messy
and
I'm
trying
to
clean
it
up
so
the
overall
architecture,
a
multiplayer
server
game.
You
are
controlling
you're
in,
like
the
clear
movie
controlling
it
through
a
client.