►
From YouTube: NuPIC Office Hour, Jan 10 2014 - NuPIC 2014 Goals
Description
We'll talk about goals for the NuPIC project for 2014 based on community feedback.
A
A
You
should
have
an
invite,
but
there's
yeah
you
should
have
an
invite
in
your
email.
I
think.
A
Yeah
yeah,
I
can
hear
you
all
right,
let's
see
where
should
we
there's
probably
at
least
three
or
four
of
us
here,
oh
good,
I
would
guess,
if
they're
off
the
camera,
just.
B
I
I
could
see
you
and
it
just
went
away.
Oh
there,
it
is
okay,
I'm
just
seeing
sitting
in
different
positions
and
seeing
if
you
can
see
me
yeah,
I
can
see
you
there.
Okay,.
B
You
probably
didn't
get
a
chance
to
see
this,
but
I
did.
I
did
start
a
weekly
page
on
goals,
but
I
I
wanted
to
run
it
by
you
before
we
talked
about
it.
It
doesn't
really
matter.
We
can
just
talk
about
it.
Hopefully
yeah.
It
looked
good.
D
B
A
B
B
B
A
A
B
So
we've
got
chettin
and
jeff
full
and
matt
keith,
and
somebody
else
I
don't
recognize
already
connected
and
watching.
So
it's
good
we're
live
we're.
Oh
yeah.
We're
live.
D
B
About
the
you
know,
governor
chris
christie
and
some
of
these
past,
you
know
expletives.
A
So
hello,
everybody,
I'm
I'm
matt
taylor,
I'm
at
my
home
office,
which
is
my
bedroom,
my
home
bedroom
and
in
the
nominated
rock
offices
we've
got
jeff
hawkins
and
super
tight
ahmed
and
austin
marshall,
who's
off
camera,
and
soon
we'll
probably
have
scott
purdy
come
in.
So
we're
going
to
go
ahead
and
start
with
this.
As
is,
and
we're
going
to
talk
about
our
goals
for
new
pick
for
2014
and
should
probably
say
these
are
not
necessarily.
These
are
not
tasks.
A
These
are
really
kind
of
high
level
goals
things
we
want
to
try
and
accomplish
with
tasks
and
actions
that
will
follow,
but
I
think
it's
a
good
idea
to
subitize
idea
to
really
hash
the
goals
out.
What
do
we
want
to
have
at
the
end
of
this
year,
for
our
community
and
for
any
products
that
will
eventually
be
built?
On
top
of
it?.
B
B
I
think
the
the
overwhelming
consensus
for
the
most
asked
for
sort
of
work
that
needed
to
be
done
was
to
extract
out
some
core
c
plus
plus
code,
that
runs
the
essential
parts
of
of
new
pic
and
the
cla
with
the
core
algorithms.
C
Some
of
the
goals
for
this
year,
I'm
I'm
in
the
process
of
working
on
a
wiki
document
and
I'll
finalize
it
after
this
meeting
after
we
kind
of
come
to
a
consensus
on
it
that
contains
our
goals,
but
the
rough
draft.
A
Of
it
basically
separates
our
goals
into
three
categories.
A
lot
of
this
is
based
on
the
feedback
we
got
from
the
mailing
list.
A
The
first
category
is
usability
and
that
is
really
focused
on
creating
a
good
experience
for
new
users
of
newpic
that
don't
have
the
context
of,
or
that
may
not
necessarily
have
a
machine
learning
background
or
even
a
python
background.
A
But
hopefully
we
can
get
clear
enough
documentation
so
that
they
can
build
it
and
run
it
and
see
good,
compelling
sample
applications
and
say,
oh,
I
can
think
of
a
use
for
that,
so
that
they
can
continue
onward
with
newpic.
So
this
includes
samples
and
documentations
and
also
solidifying
an
api
and
making
sure
it's
easy
to
install
and
build
the
another.
I
think
broad
topic
is
transportability,
and-
and
that
is
what
we
were
just
talking
about.
B
That
we
can
create
different
language
bindings
around.
It
also
includes
some
serialization,
a
common
serialization
format,
but
those
are
really
more
of
the
tasks.
So
I
want
to
come
to
an
agreement
on
the
kind
of
broad
goals
of
usability,
transportability
and
also
the
last
one
that
people
were
asking
for
was
hierarchy
which
isn't
necessarily
discluded
from
our
current
state
of
new
pick.
But
it's
not
necessarily
easy
to
do
it.
It
takes.
It
would
probably
take
a
lot
of
work
and
guidance.
C
From
us
to
establish
a
a
true.
A
Hierarchical
architecture
with
cla
regents
that
are
linked
together
hierarchically,
so
I
think
that's
that
that's
my
take
on
on
what.
B
Should
be-
but
I
definitely
want
to
this-
is
a
rough
draft
like
I
said
so.
I
want
to
get
jeff
and
subutai
to
weigh
in
and
anybody
else
to.
A
Yeah
I
mean
to
me:
you
know,
if
you
think
about
the
software
looping
as
the
software
package.
You
know,
I
think
all
those
goals
make
a
lot
of
sense
to
me.
I
think
you
know,
if
I
think,
about
clas
and
where
the
technology
can
go,
there's
so
many
different
industries
that
it
can
be
applicable
to
and
there's
so
many
different
problems,
and
I
think
we
should
try
to
have
a
very
stripped
down
version
of
newpick
that,
just
like,
I
said
that,
probably
in
pure
c,
plus
plus,
that
can
is
agnostic.
B
To
different
languages
and
different
platforms
and
different
industries,
and
so
on
so
as
as
pure
and
simple
as
we
can
make
it,
and
it's
not
going
to
be
a
one
size
fits
all.
There's
no
way
we
can
have
it
do
everything
but
to
be
as
broad-based
as
possible.
I
was
really
surprised
and
happily
said
from
the
community
of
people
trying
to
get
our
stuff
running
on
raspberry
pi,
and
you
know
they
were
successful
with
it
and
various
embedded
platforms.
A
And
so
on,
so
I
think
having
something
that
makes
it
really
easy,
a
very
simple
core
that
can
be
used
everywhere.
You
know,
I
think
this
seems
like
a
really
good
role.
Can
I
ask
a
dumb
question
about
that
yeah?
So
the
reason
we
don't
have,
that
is
it
because
a
lot
of
it
is
because
of
legacy
python
is
that
yeah
yeah
we
have
c
plus
bus
and
python
all
together.
Is
that
really
the
big
issue.
A
A
A
Well,
for
the,
if
we're
talking
about
extracting
a
c
plus
plus
core
that
could
potentially
be
disruptive
to
grok
or
any
anything
else,
that's
using
newpic,
so
we
have
never
really
released.
We've
never
come
out
with
an
official
release
of
new
pic
so
kind
of
by
default.
We
don't
have
an
official
api,
that's
hardened
and
expected,
so
that
sort
of
means
that
we
can
change
anything
we
want
or
the
community
can
change
anything
we
want,
because
we've
never
drawn
any
real
lines
or
boundaries
around
what's
official.
A
What's
not
so
that
would
sort
of
be
a
part
of
that.
So
I
was
asking
more
along
the
lines
of
skill
level
or
knowledge
level
or
I
don't
know
yeah.
I
think
the
community
can
help
out
quite
a
bit
and
they
already
have
been
you
know
like.
I
think
it
was
david
ragazzi-
is
that
the
name
created
a
really
clean,
build
system
in
cement
and
it's
still
in
process,
but
that
would
make
this
a
lot
easier.
You
know
certainly
mark
bresnik
and
others.
D
D
Do
you
see
this
as
a
fork
from
what
we're
using
in
garage
or
is
this
well
the
right
now
gronk
is
using
exactly
what's
in
the
garbage,
and
so
I'd
like
to
maintain
that,
if
possible.
So
if
we
went
in
this
direction,
I
guess
I'm
saying
is
this
in
addition
to
what
we
have
now
or
is
this
an
alternate
to
it?
I
guess
we
transition
towards
transition.
D
So
we
already
have
a
couple
of
questions.
I'm
gonna
start
with
the
easy
one
first,
this
is
from
matt.
Sometimes
I
want
to
talk
about
something:
that's
not
only
this
too.
So
this
one,
okay,
I'll,
we'll
put
some
time
aside
for
you.
D
So
this
is
jeff
fole
who's
he's
been
helping
us
with
our
community,
managed
virtual
machine
from
new
pic,
and
this
question
is:
is
there
a
plan
for
versioning,
so
I've
yeah?
Yes,
there
is
a
plan,
but
so
I've
brought
up
this
in
the
past
and
I
think
we've
got
roadblocked
by
a
definition
of
a
solid
api
because
we
can't
really
come
out
with
a
version
until
we
have
solidified
what
the
api
is,
because
that's
that's
part
of
the
version.
D
D
Maybe
it
was
stuart
or
someone
else,
but
just
extracting
the
core
kind
of
the
network
api
into
its
own
pure
thing
is
a
pretty
straightforward.
The
problem
is
it's
kind
of
help
useless
because.
A
D
All
these
things
have
backwards
compatibility
problems
that
come
along
with
them.
So
as
far
as
breaking
rocks
up
that's
an
issue,
so
a
language,
independent,
realization
format
may
be
something.
A
That
you
have
to
do
first,
so
ian
brings
up.
D
D
A
concern
that
c
plus
core
as
a.
A
Primary
primary
core
library
excludes
a
much
larger
set
of
developers
than
having
python
as
a
primary.
A
So
while
I
can
sort
of
share
that
concern
a
bit,
I
think
that
the
benefits
of
having
a
c
plus
core
and
the
transportability
benefits
explicitly
are
far
outweigh
that
and
if
we,
if
we
since
grack,
is
a
user
of
would
be
a
user
of
that
c,
plus
plus
core.
And
we
have
a
lot
of
python.
D
In
the
grok
environment,
as
a
as
a
user
of
the
system,
we
should
be
able
to
help
provide
any
of
the
language
bindings
and
the
tools
and
utilities
in
python
and
those
would
be
accessible
from
any
other
developers
that
want
to
come
in.
B
D
A
Excluding
python
at
all,
but
we
would
enable
java
bindings
and
dotnet
bindings,
and
you
know
ruby,
bindings
or
whatever.
Today
we
can't
do
that
because
we're
so
python
focused
once
it's
a
pure
c
plus
bus
thing.
It
really
broadens
the
range
of
environments
that
you
can
use
it.
So
I
think
it
goes
the
other
way
actually.
Well
that's
for
users,
not
for
developers,
but
even
for
developers.
I
don't
think
we're
planning
on
getting
rid
of
any
of
the
python
everything
we
have
now
so
exactly
yeah.
A
So
I'm
excited
to
have
something
that
we
can
use
within
the
jvm,
because
I
have
a
lot
of
experience
in
the
jvm
and
I'd
love
to
write
some
tools
for
it
yeah
so
definitely
developer.
What
we
have
today
is
a
really
big
hurdle,
whereas
if
we
have
a
c
plus
core,
it
would
be
very
easy
for
our
java
developers
to
start
using
right
so
chettin
says:
can
we
adopt
the
unix-style
philosophy
of
small,
independent.
C
At
this
point,
it
feels
like
new
pic,
especially
the
opf,
is
kind
of
all
or
nothing.
So
I
think
that's
a
good
point.
Yeah
the
the
opf
is,
I
think,
is
very
much
like
that
and
I
think
that's.
A
Sorry
yeah,
I
think
it's
a
good
idea.
You
know
one
once
it's,
you
know
once
we
have
a
pure
c-plus
bus
core,
we
can
do
that
kind
of
stuff
very
easily.
D
The
opf
way
of
doing
things
it's
quite
hard
right
now,
so
I
think
the
challenge
is
it's
going
to
be
not
necessarily
well
after
we
have
the.
A
The
core
extracted
there's
a
lot
of
refactorings
that
could
be
applied
to
the
core,
but
those
are
all
very
dangerous
refactorings
that
need
a
really
solid
suite
of
I.
I
would
want
regression
tests
around
that
so
that
we
can
judge
performance
changes
that
might
occur
from.
A
A
Okay,
shane
has
a
another
question
too.
I
don't
know
how
a
good
way
is
to
get
this
on
your
guys's
screen.
Maybe
this
will
work
one
sec,
because
it's.
B
B
Okay,
so
jen
says
a
practical,
intuitive
demo
of
new
pick
would
be
very
useful
in
inspiring
non-community
members
into
getting
involved
with
the
mission.
Is
it
possible
for
us
to
start
an
official
project
in
this
direction?
Are
there
any
ideas.
D
D
So
this
is
also
an
idea
from
the
community
that
everybody
seems
to
want
a
compelling
sample
application,
ideally
something
that
people
can
come
to
the
website
and
look
at
and
see
the
results.
From
I
mean
that
would
see.
That
would
be
the
most
bang
I
think
for
for
getting
people
in.
It's
just
immediately
seeing
the
value.
D
D
So
I'm
fully
in
favor
of
this
goal,
and
it
is
a
part
of
the
2004
goals
that
I'm
writing
up
in
the
wiki
as
a
part
of
the
there's
a
whole
sample
section
with
the
compelling
use
case,
I
think
it's
a
great
idea
and
we've
asked
ourselves
in
the
past
what
this
might
be
a
good
time
to
work
on
it,
and
so
I
think,
a
couple
of
thoughts.
D
One
is
you
know
coming
up
with
that
demo
and
deciding
what
the
public
stream
is
and
so
on
is
challenging,
something
that's
kind
of
really
highlight
the
cla
and
yet
be
something
we
can
achieve,
and
you
know
it
all
has
to
work.
So
I
think
it's
a
great
idea.
I
think
a
lot
of
effort
has
to
get
spent
on
thinking
about
what
it
is.
At
least,
we
have
to
have
a
good
idea.
What
that
is.
A
D
Of
gawk
on
our
website,
the
crop
website,
rock
solutions.
D
Feeding
to
public
streams
and
detecting
anomalies,
I'm
wondering
if
you
could
do
something
like
take
the
exact
same
idea
but
put
like
a
cerebral
front,
end
on
it
or
attachment
to
it.
So
on
the
new
pick
website,
you
could
not
only
see
the
anomalies,
but
you
could
see
maybe
the
more
internal
workings
of
what's
going
on
some
sort
of
visualization,
so
you
can
have
some
real
times
thing
going
on
that
we
can
even
tie
into
what
we're
doing
a
lot
of
the
data
streams.
D
Well,
if
we
do
this,
we'll
have
the
data
stream
we'll
have
an
engine
running
well
with
bitcoin
all
the
time
if
you
could
feed
off
of
that
somehow
and
say:
okay,
in
addition
to
that,
you
can
look
at
the
internals
of
what's
going
on
here.
Here's
the
column,
here's
the
cells
that
are
predicting
well,
here's
the
actual
predictions
that
are
coming
out
that
I
just
just
a
thought
we
can
maybe
somehow
camp
on
to
what
would
probably
do
with
rock.
D
I
don't
know
the
thought
yeah
we
could
feed
in
the
exact
same
data
stream
and
yeah,
I'm
saying
we're
gonna
we're
gonna
have
a
if
we
do
this,
we'll
have
a
data
stream,
we'll
have
an
engine
running
we'll
have
an
amazon
instance
running.
It
continuously
we'll
have
a
level
in
the
facebook.
D
D
I
mean
yeah,
I
would
say,
as
long
as
there's
a
clear
boundary
between
new
pic
a
line
between
the
usage
of
newpick
and
brock,
because
what
I,
what
I
would
be
interested
in
displaying
in
this
sample,
is
how
to
get
input
into
new
pick
and
how
and
decent
output
out
of
dupik
and
what
to
do
with
that,
and
that's
not
exactly
what
chetan
was
saying.
I
think.
D
D
So
if
we
have,
if
we
can
create
it's
kind
of
hard
to
be
cool
and
also
explain
how
to
do
it
technically
at
the
same
time,
you
know
well,
if
we
make
a
project
a
sample
application,
we
come
up
with
a
compelling
sample
application
and
we
put
it
in
another
repository
that
depends
on
new
pick
being
installed
and
we
make
it
easy
enough
so
that
anybody
can
clone
that
repo
and
run
it
even
if
it's
a
web
application
that
just
that
coming
in
and
seeing
the
compelling
sample
of
how
newpic
works
and
then
saying,
oh
here's,
the
code
clone
it
and
run
it
yourself.
D
A
Know
just
a
thought:
well,
whatever
stream
we
end
up
using,
we
could
make
that
data
available
yeah
so
that
someone
could
copy
it
and
try
the
same
thing
yeah.
But
but
if
that
data
was
coming
through
brock,
it
becomes
really
off
amazon
instances
or
things
like
that.
It
wouldn't
be.
They
couldn't
recreate
that
exactly.
No,
it
wouldn't
be
exactly,
but
they
could
recreate
the
cla
part
of
it.
They
could
be
kids
yeah
right,
so
some
people
are
coming
up
with
some
ideas
about
sample
applications.
D
Someone
said,
maybe
we
could
do
something
with
google
analytics
about
live,
visitor
prediction,
which
is
actually
something
we
had
an
intern
do
while
they
were
here,
but
that
was
with
our
kind
of
first
generation
of
our
our
product.
But
I
don't.
I
don't
really
remember
exactly
how
much
time
and
effort
was
was
put
into
that,
but
I
think
that's
a
good
idea,
because
that's
that's
something
that
could
be
applied
to
any
google
analytics
stream.
You
know
if
they've
got
it
on
their
page,
it
could
be
something
someone
could
immediately
use
yeah.
D
So
I
like
that
idea
and
there's
ian
also
has
an
idea
for
a
model
of
etl
tools
would
be
a
great
model
for
that.
I'm
not
exactly
sure
what
he's
referring
to
or
what
etl
tools
are.
A
The
cool
aspect
of
it
is
more
important
because
a
targeted
audience
for
this
would
be
people
who
don't
yet
care
about
newbik,
so
that
he's
kind
of
supporting
your
opinion
on
that.
D
A
And
saying,
like
you
know,
I
just
want
to
see
it
running
in
real
time
yeah.
So
we
also
have
a
question
from
joe
timmons
along
the
same
line.
He
says
what
have
you
learned
so
far
from
grok
beta
testers
or
any
of
those
early
lessons
learned
driving
changes
to
the
overall
new
pic
2014
strategy.
A
D
A
D
A
As
much
as
we're
excited
about
it's
based
on
cla
for
the
customers
of
drock,
they
don't
they
don't
have
to
deal
with
that.
So
you
know,
I
think
most
of
the
effort
we'll
be
putting
into
the
product
over
at
least
over
the
last
few
months
and
going
forward
is
going
to
be
on
things
unrelated
to
you
know,
infrastructure
and
marketing
things
and
user
interface
things
and
graphical
things,
and
things
like
that.
Fortunately,
the
new
pick
and
the
ceiling
were
usually
a
good
shake
on
an
interest
rate
all
the
way.
A
So
we'd
still
do
a
few
things.
You
know
we
just
didn't
just
did
the
encoders
the
random
distributing
code?
That's
the
smaller!
That's.
A
It
as
you.
B
Know
we've
taken
the
state
of
the
cla
at
that
at
a
recent
point
in
time.
We've
now
built
a
product
around
it
and
probably
that
product
for
the
next
six
months
to
a
year
will
be
mostly
other
things
are
unrelated
and
therefore
we'll
be
able
to.
When
we
start
taking
our
algorithmic
time,
we'll
be
working
on
new
things.
B
Okay,
yeah,
and
maybe
one
thing-
we've
learned
from
the
early
videos
is
well
looking.
I'm
actually
reasonably
pleased
how
low
glock
works.
So
you
know,
I
think
you
know
the
fact
it
really
can
be
done
in
a
production
environment,
and
you
really
can.
A
D
A
B
A
A
So
I
have
a
goal
this
year
of
documentation,
and
specifically
I
want
to
do
some
peer-reviewed
published
articles
on
on
on
the
cla
and
those
are
those
take
a
long
time
to
do
they
take
a
lot
of
you
know,
get
past
periods
and
stuff,
I'm
going
to
put
it
a
strategy
here,
but
I
want
to
exit
2014
with
several
publications,
or
at
least
ones
that
are
accepted.
A
There's
if
anyone
has
input
on
that,
I'd
be
curious
to
hear
it,
because
you
know
one
thing
we
need:
we
could
focus
on
the
neuroscience
side
of
it.
Nothing.
You
could
talk
about
the
submerging
side
of
it
so
and
then
you're
typing
journals.
A
So
anyway,
I
think
that's
it's
important
not
for
them
necessary
for
the
people
already
in
the
nuclear
community,
but
I
think
it
will
provide
a
stronger
base
for
uk
going
forward
and
I'm
also
considering
beyond.
C
Considering
something
else,
a
little
bit
broader,
several
people,
they
asked
me
to
do
a
new
version
on
intelligence,
or
maybe
maybe
there's
a
different
type
of
book
that
come
out.
That's
more
related
to
machine
intelligence.
I
don't
know,
but
I
I
wanna,
I
feel
it's
a
time.
We've
got
so
much
work
that
really
there's
things
we
know.
B
B
Been
documented
at
all,
I
feel
like
I
need
to
sort
of
put
some
of
this
in
writing
and
get
it
done.
Well
only.
You
know
our
white
paper's
old
now
I
just
read
it
again
last
weekend
and
you
know
there's
things
I
would
change.
B
That
didn't
make
anyone's
list
here
that
didn't
show
up
in
any
way
any
of
the
requests,
but
it's
a
pretty
high
priority
one
for
me,
the
other
one
that
you
kind
of
alluded
to.
You
said
hierarchy
here,
but.
A
There's
several
things:
there's
there's
you
had
six
votes
for
real
temple,
pooling.
C
Modeling,
I
think,
there's
an
opportunity
here
to
make
a
fairly
significant
advance
on
the
theoretical
side
this
year
and
I've
started
doing
that
already
and
it's
going
to
take
me.
You
know
I
just
just
came
out
of
the
meeting
where
I'm
going
over
some
of
this
subutai
and
scott,
but
I
have
a
personal
goal
to
to
knock
off
a
couple
of
those.
I
think
we
can.
I
think
we
can
take
the
temple.
A
D
So
I
just
want
to
let
you
know
that
those
those
are
those
aren't.
That's
your
third
component.
It's
not
the
top
two.
B
I
think
I
mean
yeah
judging
from
this
list.
Let
me
see
if
I
can.
Let
me
let
me
share
this
real
quick
for
people
who
are
watching
there.
We
go
so
judging
from
this
list
of
things,
you're
talking
about
the
real
temporal
pooling
and
hierarchies,
which
are
the
number
number
two
and
number
three
which
are
kind
of
should
be
grouped
together,
I
think,
and
getting
real
hierarchies
working.
B
B
I
think
I
see
how
to
do
that
now
and
at
least
from
a
theory
point
of
view,
I
think
we
can
sort
of
make
progress
in
those
three,
those
three
areas
and
and
I'm
hoping
that
we'll
be
able
to
do
some-
some
cool,
quick
coding
projects
as
well
to
verify
that
stuff.
B
So
you
can
expect
for
me,
at
least
in
the
next
month
or
two
of
some
a
nay
sooner,
but
assuming
that
some
some
textual
and
descriptive
things.
A
About
this,
maybe
we'll
talk
or
something
like
that
and
I
think
we're
making
progress
on
that.
I
think,
just
over
the
last
couple
months,
we've
made
products
in
those
issues
yeah
several
people
have
asked
you
about
it
individually.
I
think
yeah
well,
because
I
I
mentioned
it
in
the
email
on
the
list
and
then
some
people,
you
know
say
hey
what
is
that
before
I
send
it
out
to
everybody,
I
wanted
to
run
it
by
scott
think
about
reporting
language
views
because
it
could
be
very
confusing.
B
If
I
just
didn't
describe
it
correctly,
so
great
okay,
so
is
there
anything
else
you
wanted
to
talk
about?
Jeff
me?
Yes,
no,
that's
the
thing
I
wanted
to
do.
I
just
talked
about
the
documentation
and
the
theory
progress,
and
I
think
these
other
suggestions
are
all
great.
B
A
So
I'll,
I
guess
we'll
take
a
quick
look
at
this
wiki
page,
because
I
I
kind
of
want
to
get
a
consensus
on
this
and
see
if
we
need
to
change
it.
Can
you
guys
see
this
okay?
A
Okay?
So,
like
I
said,
I
split
it
up
into
usability
and
transportability.
I
haven't
really
filled
filled
this
all
out,
but
at
least
wanted
to
get
the
high
level
stuff
down
jeff.
We
could
throw
your
your
stuff
into
the
documentation
section,
but
there's
also
a
section
on
on
hierarchy
which
you're
talking
about
too,
although
my
documentation
is
not
documentation
of
new
pick
right,
it's
more.
It's
more.
You
know
cla
cla
publications,
it's
not
going
to
tell
you
how
to
use
it.
A
It's
going
to
just
describe
the
theory
and
try
to
get
it.
You
know
published
academically
yeah.
I
still
think
it's.
That
would
be
useful,
though,
because
there's
there
are
people
coming
in.
There's
that
ask
a
lot
about
the
theory
and
and
stuff
and
want
to
understand.
What's
going
on
and
not
just
use
new
pick
for
what
it
is.
I
know
I
just
the
word.
Documentation
is
a
little.
You
know
this
is
describing
api
docs
and
all
right.
Tutorials.
B
B
The
other
thing
is
hierarchy,
so
this
these
three
things
sort
of
came
from
that
list
and
I
think
they
translate.
It
translates
pretty
well
to
these
higher
level
goals.
I
tried
not
to
put
too
many
technical
tasks
in
this
document
or
just
you
know,
ideas
about
how
things
could
be
better,
but
does
does
this
generally?
Does
this
make.
C
Sense
to
you
guys,
once
we
fill
in
some
of
the
details.
C
Okay,
so
I
like
that
first
goal
of
having
a
new
user
being
able
to
get
up
and
running
within,
I
would
say.
A
B
Yeah
agreed
that
that's
part
of
you
know
I've
kind
of
separated
into
some
people
are
going
to
want
to
build
it,
and
some
people
are
just
going
to
want
to
have
it
available
yeah.
So.
A
For
the
ones
that
just
want
to
have
it
available,
we
should
have
binary
distributions
and
all
the
common
formats
exactly.
I
think
that
would
make
it
so
much
easier,
because
a
very
small
percentage
of
people
really
care
about
building
and
changing
their
the
curves
right.
A
A
So
these
are
basically
our
goals.
I'll
I'll
fill
this
out
with
the
help
of
I'll,
probably
have
team
subtitle,
demento
team
review
it
it's
already
on
the
wiki,
but
I'll
kind
of
fill
it
in,
and
this
is
kind
of
going
to
be
our
official
goals
for
for
2014,
and
hopefully
we
can
turn
this
into
a
solid
implementation
plan
for
the
next
year.
A
So
no
am
I
missing
any
questions.
Yes,
okay.
Chutten
says:
when
studying
the
software
implementation
of
the
theory,
it
would
be
useful
to
have
a
toy
example
or
benchmark
that
we
can
all
analyze
and
communicate
over.
Apparently
it
seems
that
hot
gym
sort
of
fills
this
role.
Is
there
anything
better?
We
can
come
up
with.
A
So
there
are
a
number
of
examples
in
the
repository
most
of
the,
I
think,
all
of
them,
other
than
hot
gym
are
generative
data
which
in
the
past,
we
haven't
really
used
generated
data,
but
we
have.
We
also
had
a
bunch
of
internal
benchmarks
that
were
from
customers,
but
this
is
not
a
benchmark.
This
is
a
just
a
sample.
A
To
dynamically
add
in
comments:
do
we
not
have
that
this
time
or
those
were
people
that
actually
were
joined
into
the
hangout
so
right
now,
it's
only
you
and
me,
and
the.
D
Hangout,
okay,
so
I
agree
that
we
should
have
more
examples.
So
if
someone
wants
to
do
a
language
thing
there,
so
matt
had
created
some
language
of
campus
outside
of
dubai.
B
That's
good
super
dietary,
the
temporal
poor,
specific
example.
So
if
you
were
going
down
to
that
level
but
yeah,
we
should
definitely
increase
those.
But
I
guess
it's
a
separate
question:
if
he's
talking
about
benchmarking.
B
Yeah
I
mean
if
it
think
about
yeah,
it's
chet
and
says
benchmark
once
not
just
a
toy
example,
but
benchmark.
Oh
really,
very
different,
yeah
benchmarks
would
not
be
an
easy
thing
to
like.
You
know.
B
A
For
us,
you
know
we're
trying
to
build
something:
that's
going
to
be
a
core
piece
of
machine
intelligence,
and
we
should.
We
should
be
careful
about
what
we
benchmark
and
be
able
to
understand.
A
Intelligent
systems
and
a
mappy
error
is
not
necessarily
the
be
all
and
end
all
of
benchmarks.
You
know
we
all
agree
that
you
know
that
continuous
learning.
A
We
could
have
benchmarks
around
that.
We
can
talk
about.
You,
know,
hierarchy
or
sequence
learning
and
having
benchmarks
around.
How
well
do
we
learn
high
order,
sequences
and
so
on.
Those
are
all
like
really
important
parts
of
building
an
intelligent
system
and
I
think,
building
up
some
idea
of
what
how
to
characterize
those
things
formally
would
be
really
nice.
A
So
his
follow-up
is
to
clarify
suppose
we
make
a
small
change
to
the
temporal
pooler.
How
does
that
affect
this
benchmark
and
why
we
can
talk
about
the
effect
in
the
context
of
the
example?
We
all
understand,
so
I'd
like
to
bring
up
again
that
I
think
this
is
the
job
of
regression
tests
and
regression.
Tests
are
a
bit
like
benchmarks,
so
we're
setting
standards
for
achievement
for
the
software
and
they
should
always
be
tested
every
time.
There's
changes
to
see
if
we're
still.
A
A
B
A
Along
those
same
lines
since
we've
kind
of
identified,
one
of
our
major
goals
is
to.
A
And
trying
to
extract
out
this
sequel
supposed
core-
I
mean
this.
Is
it's
not
going
to
be
an
easy
job,
but
I
think
there's
a
little
bit
of
work
that
needs
to
be
done
beforehand
and
since
we're
on
the
subject
of
regression
tests,
it's
not
going
to
be
very
comfortable,
especially
for
me,
as
sort
of
the
monitor
of
the
project,
to
to
watch
major
things,
change
within
the
architecture
of
the
framework
without
having
regression
tests
in
place
beforehand.
D
So
I
think
that
it's
more
important
before
we
make
major
changes
to
introduce
regression
tests
in
some
in
some
way
at
least
get
some
benchmarks
established
in
these
tests,
so
that
when
we
do
start
mucking
around
with
the
architecture
that
we
know
that
they
continue
to
pass.
And
we
have
some
level
of
confidence
that
we're
not
breaking
any
long-term
capabilities
of
the
software.
D
I
don't
think
anyone
would
say
no.
I
think
we
definitely
won
that.
I
think
one
kind
of
related
thing
that
would
be
a
nice
goal.
D
Is
you
know
if
we
have
those
regression
tests
and
have
you
know
build
in
whatever
mechanism
we
need
so
that
we
can
improve
and
work
through
community
pull
requests
much
faster
than
we
are
now,
and
some
of
them
are
really
hard
for
us
to
work
through
efficiently
now,
because
if
someone
makes
a
change
to
the
core
part
of
the
algorithm
or
you
know
some
of
the
core
code,
we
it's
not
the
new
pick
regression
tests
are
not
enough,
no
matter
how
good
they
are
they're,
not
enough.
D
We
still
have
to
get
it
and
test
it
against
the
product
and
make
sure
there's
nothing
changed
there.
So,
whatever
we
can
do
to
kind
of
make
that
process
smoother,
I
think,
would
be
great
yeah.
I
think
it's
definitely
a
good
thing
and
I
think
the
hesitation
is
that
it's
going
to
be
really
difficult
to
do
it
in
an
automated
way,
because
you're
going
to
have
one
that
gets
like
slightly
better
when
it
gets
slightly
worse.
Is
it
a
good
change.
D
All
the
way
across,
so
I'm
sort
of
envisioning
a
a
separate
regression
testing
repository
that
we
hook
in
to
our
tooling
server
and
the
current
flow
so
that
when
something
gets
pushes
push
to
master
after
it
passes
the
normal,
build
and
travis
with
unit
tests
and
everything,
we
run
it
through
a
more
vigorous
testing
suite-
and
maybe
we
we
don't
mark
it
as
a
as
unmergeable
from
that
initially,
but
at
least
we
can
get
something
running
so
that
we
can
inspect
past
runs
and
and
compare
them
with
current
ones.
D
I
think
that's
really
essential
before
we
start
messing
around
with
extracting
a
core,
and
you
know
introducing
real
temporal
pulling
and
and
having
a
c
plus
temporal
cooler
and
all
of
that
stuff,
even
if
it's
not
necessarily
even
if,
even
if
they're
trivial
regression
tests,
we
should
be
able
to
get
something
in
some
framework
in
place
so
that
we
can
continue
to
write
more
complicated
ones.
D
So
that
is
a
goal
of
mine
personally,
because
I
don't
think
we,
I
I'm
not
comfortable
moving
forward
with
some
of
these
changes
until
we
have
that.
So
that's
something
that
I'm
planning
on
working
on
this
early
this
year.
D
All
right,
so
I'm
gonna,
I'm
gonna,
continue
honing.
This
document
with
with
subutai's
help
and
the
rest
of
the
team's
help
and
I'll
send
an
email
to
the
mailing
list
once
we
have
it
kind
of
in
a
decent
format.
D
I'd
like
to
know
how
how
good
a
mechanism
this
is
and
how
much
people
appreciate
it.
We
can
talk
about
that
later.
Oh
the
officer
itself,
what's
that,
are
you
talking
about
yeah
yeah?
Is
this
helpful
or
not,
or
no.
D
We
get
a
lot
more
views.
Offline,
no
well,
not
offline,
but
after
the
live
event
on
our
youtube
page.
Let's
monitor
that
to
see
how
that
goes.
I
mean
I,
like
the
format.
Actually
I
think
it's
it's
nice,
it's
relaxed,
it's
kind
of
nice
to
have
a
conversation.
It's
a
little
bit
different
than
you
know.
D
Typically,
so
I
like
it
right,
oh
matt,
I
have
another
suggestion
for
a
goal:
okay,
arbitrary
goal
of
just
increasing
the
number
of
active
users
of
we've
talked
about
this
before
just
increasing
the
number
of
active
users
in
by
a
certain
percentage,
whatever
it
takes,
whether
it's
easier,
you
know
it
that
might
imply.
We
have
easier
installation
and
more
portable.
Whatever
you
know,
whatever
it
takes,
you
know
a
book
or
by
crook,
increase
the
number
of
developers
by
a
factor
of
two
or
five
whatever.
D
And
we're
talking
about
you're
talking
about
contributors,
there's
there's
contributors,
but
there's
just
users.
If
you
will
there's
there's,
I
think
the
percentage
of
people
actually
contribute
back
to
the
code
base
will
be
fairly
small,
yeah
and
even
out
of
our
list
of
I
don't
know,
80
or
so
contributors
that
we
have
now
only
you
know,
10
or
15
have
actually
pushed
code
into
master,
but
that
it's
still
increasing,
I
mean.
D
But
but
yes,
users
increase
the
number
of
contributors
and
users
and
and
having
a
a
hard
goal
right
trying
to
reach
a
certain
number
yeah,
and
it
could
be
the
number
of
people
on
the
mailing
list.
It
could
be
the
number
of
people
who
downloaded
new
pick.
Fourth,
new
pick,
you
know,
there's
a
number
of
different
metrics.
D
D
So
this
time
I
just
sent
a
limited
number
of
invitations
which,
if
you're
a
contributor,
you
should
have
gotten
an
invitation
to
join
the
hangout.
So
this
is
the
alternative
there's.
Unless
we
do,
unless
we
make
a
go
to
meeting,
which
is
a
potential,
we
create
a
go
to
meeting
account
for
new
pick
that
might
allow
more
people
to
participate.
I've
thought
about
that
before
what
about
jrc
having
an
irc
window
open?
Well,
that's
a
good
idea
too,
having
an
rc
window
and
I
do
have
it
open,
but
there's
nobody
in
it.
D
D
Okay,
all
right
guys
that
that
is
it
thanks
for
taking
part
we're
looking
for
feedback
on
these
on
these
hangouts.
So
if
you're,
if
you're
watching
the
office
hour-
and
you
have
a
comment-
please
drop
a
line
to
the
mailing
list.
It's
newpickatlist.demento.org!
You
have
to
subscribe.
You
can
subscribe
at
nomento.org.