►
From YouTube: Promoting rigor in Blockchains energy and environmental footprint research - Ashish Sai
Description
No description was provided for this meeting.
If this is YOUR meeting, an easy way to fix this is to add a description to your video, wherever mtngs.io found it (probably YouTube).
A
A
The
goal
of
this
project
is
to
sort
of
give
you
an
academic
take
on
carbon
emission
and
electricity,
consumption
of
bitcoin
and
other
cryptocurrencies
over
the
last
day,
or
so
you
have
heard
from
industry
and
in
this
short
presentation,
I'm
going
to
give
you
an
academic
take
on
the
same
topic.
A
Okay,
so
I
want
to
start
this
presentation
by
pre-facing
that
it's
really
difficult
to
measure
electricity,
consumption
and
carbon
emission
of
information
technology,
especially
when
it's
distributed
right,
so
here
have
cited
some
studies
from
early
2000s
for
internet.
So
the
first
example
that
you
see
here
is
a
study
from
early
2000s
that
predicted
that
internet
would
consume
about
50
percent
of
us's
electricity
by
2010,
and
we
know
it's
provably
wrong.
A
It's
a
way
smaller
figure
than
that
there
were
the
studies
that
predicted
would
reach
this
landmark
by
2020
again
wrong,
and
one
of
the
reasons
why
we
get
this
wrong
at
an
early
stage
is
because
when
we
don't
understand
these
technologies
that
well,
we
don't
understand
their
evolution
that
well
and
the
models
that
we
design
in
order
to
predict
electricity
and
environmental
footprint
they're
based
on
a
lot
of
assumptions,
because
we
don't
have
data
about
these
technologies
at
this
earlier
stage.
So
we
see
a
similar
trend
repeat
itself
in
bitcoin
and
other
cryptocurrencies.
A
Here,
I've
plotted
a
graph
with
some
of
the
studies
from
the
academic
literature
on
bitcoin's
energy
consumption.
I've
plotted
their
best
cases
worst
cases
and
than
the
best
guess
estimate,
and
you
can
see,
there's
a
huge
variance
in
these
studies.
They
don't
necessarily
all
align-
and
this
is
this-
is
an
indication
of
the
underlying
assumptions.
These
assumptions
are
quite
different
and
if
you
assume
a
different
value
for
one
variable,
it
can
have
a
huge
impact
on
your
final
prediction
and
these
impacts.
A
They
might
look
smaller
on
this
graph,
but
they
can
be
as
huge
as
the
electricity
consumption
of
a
large
state
right.
So
another
important
thing
that
I
want
to
highlight-
and
this
is
something
that
I
think
is
very
important-
to
understand-
that
these
models
are
very
sensitive
to
the
variables
that
that
you
use
and
the
assumptions
that
you
make
so
here's
an
example
for
the
cambridge
model
for
bitcoin.
It's
very
widely
used
it's
most
cited
in
policy
documents
and
even
in
academic
literature,
so
towards
the
right.
A
You
can
see
this
model
when
you
assume
cost
of
electricity
to
be
five
us
cents,
and
then
I've
just
used
10
us
cents
as
the
value
and
the
drop
between
the
annual
consumption
for
bitcoin
is
greater
than
the
annual
energy
consumption
of
denmark,
and
this
is
just
one
variable
that
I'm
changing
by
a
small
margin,
and
this
is
well
within
the
bounds
of
literature.
So
you
would
see
in
the
later
slides
that
a
lot
of
literature
does
use
10
us
cents
as
their
base
value
for
cost
of
electricity
right.
A
So
the
important
thing
to
note
here
is
these
models
are
extremely
sensitive
to
very
small
variation.
In
this
case,
this
is
a
slightly
big
variation,
but
again
these
are
based
on
a
number
of
variables.
If
you
vary
one,
you
get
a
huge
variance
in
your
final
prediction
right
and
that's
model
sensitivity,
but
then
there
are
some
studies
that
are
just
outright
wrong,
based
on
bad
assumptions
or
poor
understanding
of
the
infrastructure.
A
Here's
an
example:
this
article
was
published
and
nature
climate
change,
which
is
a
very
prestigious
academic
journal.
This
article
predicted
that
bitcoin
alone
could
push
the
global
temperatures
by
two
degrees,
and
then
you
can
see
below
that.
A
This
particular
one
is
not
that
reliable,
unfortunately,
because
it's
published
in
such
a
good
venue,
it's,
it
has
been
cited
by
government
policy
documents
and
it
is
still
being
used
in
academic
literature
to
build
more
models
on
top
of
this
already
known
flawed
model
right.
So
this
is
something
that
we
are
trying
to
point
out
and
I
want
to
pre-face
this
whole
presentation.
With
this
quotation
from
british
statistician,
george
box,
all
models
are
wrong,
but
some
are
useful.
A
So
the
intention
of
this
project
is
to
help
move
these
models
to
be
more
useful
and
less
strong.
So
we
want
to
give
an
infrastructure
give
a
list
of
things
that
you
need
to
do
in
order
to
avoid
known
flaws,
so
avoid
giving
bad
predictions
by
just
avoiding
things
that
we
know
are
wrong
right,
okay,
so
the
main
goal
of
this
project
is
three
folds.
We
start
by
talking
about
different
research
methodologies
that
people
use
in
order
to
measure
electricity,
consumption
and
environmental
impact
of
different
cryptocurrencies.
It's
not
just
limited
to
one.
A
We
kept
it
as
vague
as
we
could,
so
we
captured
all
cryptocurrencies
and
all
models
that
we
could
find.
Then
we
try
and
design
a
quality
assessment
framework.
This
is
quite
difficult
talking
about
scientific
rigor
in
objective
terms.
A
It's
it's
a
very
difficult
task,
but
thankfully
we
do
have
some
literature
from
cognizant
fields
to
blockchain
from
social
energy
sciences,
information
systems
that
helps
us
talk
about
scientific
quality
and
soften
up
in
an
objective
dome
and
then
in
the
end,
one
of
the
main
outcomes
of
this
project
is
codes
of
practices,
so
these
are
designed
for
researchers
and
people
in
industry.
So
these
are
guidelines
that
you
can
follow
in
order
to
either
improve
your
models
or
acknowledge
limitations
of
your
model.
A
In
a
way,
that's
easy
to
understand
right,
so
the
going
back
to
the
cambridge
model.
If
we
acknowledge
that
this
is
a
highly
sensitive
model,
give
confidence
interval
that
this
prediction
is
only
valid
within
this
range,
then
that's
still
a
better
contribution
than
just
stating.
This
is
the
absolute
consumption
of
bitcoin's
network.
A
So
here's
a
very
like
chart
of
the
research
methodology
we
used.
I
don't
have
time
to
walk
you
through
all
of
this,
but
I
do
want
to
point
out
two
important
things:
one.
When
we
started
doing
this
research,
our
goal
was
to
make
this
research
as
reproducible
as
possible,
so
the
outcomes
we
come
up
with.
If
you
follow
this
research
methodology,
hopefully
you'd
end
up
with
the
same
outcome,
so
it's
reproducible
and
it
has
really
good
coverage.
So
one
of
the
main
goals
we
had
was
to
increase
our
coverage.
A
Almost
all
existing
studies
and
reviews
they're
focused
on
some
popular
research
methodologies,
for
example
cambridge
and
economist,
so
they
heavily
focus
on
popular
studies.
We
wanted
to
increa
increase
our
coverage.
So
what
we
did
is
we
looked
at
seven
different
research
venues,
but
we
also
tried
to
capture
grey
literature.
We
looked
at
top
200
cryptocurrencies
by
market
cap
to
see
if
they
use
any
of
these
models,
so
if
they
have
their
own
models
to
measure
their
energy
footprint,
then
we
also
looked
at
crypto
climate
accord,
supporters
and
signatories.
A
We
went
to
their
websites,
try
to
see
if
they
have
any
models
in
order
to
measure
the
environmental
footprint
again.
The
goal
here
was
to
increase
our
coverage
as
much
as
we
can,
and
another
important
thing
here
is
some
of
these
models
to
avoid
issues
that
other
models
face
right.
So
we
wanted
to
improve
our
quality
assessment
approach
iteratively
when
we
come
across
a
study
that
does
one
particular
thing
really
well
and
acknowledges
limitations.
We
try
and
make
sure
we
incorporate
that
into
our
research
quality
assessment.
So
it's
not
just
based
on
cognizant
fields.
A
It's
iteratively
improved
based
on
best
practices
indicated
within
this
field,
so
I'll
start
walking
you
through
some
of
the
results.
So
the
first
result
is
research
methodologies.
Most
of
them
are
quantitative
energy,
modeling
approaches.
So
these
models
are
technical,
social,
economical
model
that
we
use
in
order
to
estimate
energy
or
carbon
footprint.
Then
there
are
a
number
of
studies
that
are
just
literature
reviews.
A
They
give
you
a
brief
narrative
overview
of
the
field.
They
can
tell
you
if
there's
a
consensus
in
terms
of
electricity
consumption,
there
were
some
simpler
models
like
data
analysis
and
statistics,
so
these
are
very
simple
models.
You
take
one
transaction,
multiply
it
with
a
cost
of
electricity
for
that
one
transaction
and
scale
it
up,
and
then
there
were
few
case
studies
and
few
qualitative
research
in
order
to
like
I
said,
in
order
to
define
scientific
rigor,
it's
a
non-trivial
task.
We
build
on
existing
knowledge,
so
most
of
the
scientific
rigor
indicators.
A
We
have
adopted
them
from
social
energy
sciences,
the
first
article,
but
we
have
also
grounded
it
in
blockchain
specific
literature.
There
are
few
recommendations
on
how
should
you
design
these
models?
So
these
are
two
of
the
main
articles,
but
again
it's
iteratively
refined
as
we
review
the
literature
in
the
end,
we
ended
up
reviewing
about
128
articles.
We
reviewed
them
in
depth
and
all
of
these
quality
indicators
that
you
see
are
a
result
of
existing
literature
in
our
own
survey.
So
we
start
with
basic
research
design,
and
this
mostly
just
applies
to
academic
literature.
A
We
don't
apply
this
framework
to
non-academic
literature,
but
then
you
have
specific
research
methodology,
categories
and
specific
checks
for
each
of
these.
So
these
are
a
scientific
trigger
indicator
for
each
of
them.
So
I'm
going
to
give
you
a
very
brief
overview
of
basic
research,
so
I
think
a
few
important
things
to
highlight
in
this
slide
application
of
theory.
So
a
lot
of
studies
within
this
field
about
74
are
not
based
on
existing
research.
A
So
a
lot
of
research
in
this
field
is
done
in
different
silos
and
they
build
on
their
own
knowledge
and
they
don't
cite
other
works.
That's
why
the
that's
why
we
don't
see
this
field
moving
at
a
fast
pace,
because
we're
not
building
on
existing
knowledge,
we're
trying
to
reinvent
that
we
low
and
over
again,
without
necessarily
a
good
standardization.
So
that's
one
thing
that
can
really
benefit
this
field:
more
standardization
and
more
transparency
in
how
these
studies
are
conducted
and
then
about
34
of
these
studies
we
reviewed
did
not
have
a
research
design
section.
A
So
if
I
wanted
to
replicate
these
studies,
I
just
couldn't
do
it
because
there's
not
enough
information
for
me
to
reproduce
it
and
again,
keep
in
mind
they're,
claiming
that
bitcoin
and
other
cryptocurrencies
consume
a
huge
portion
of
electricity
without
necessarily
giving
us
tools
to
validate
the
approach
and
a
lot
of
them.
Don't
share
their
source
code;
they
don't
share
the
data
and
in
terms
of
source
code.
A
Most
of
these
studies
are
based
on
excel
sheets,
that
you
can
just
have
a
look
at
and
see
their
calculation
if
they
don't
share
it,
there's
no
way
for
me
to
independently
validate
their
finding
right.
So
that's
another
thing,
that's
a
big
issue
and
I
think
the
most
cv
of
these
issues
is
the
reliability
of
underlying
data.
A
lot
of
these
studies.
They
take
data
from
grey
literature
without
acknowledging
potential
validity
issues
in
that
right.
A
So
in
scientific
studies,
we
have
an
expectation
that,
if
you
use
a
data
that
is
going
to
influence
your
prediction,
you
tell
us
how
reliable
the
data
is,
and
that's
something
that's
completely
missing
about.
80
of
these
studies
do
not
acknowledge
issues
in
the
underlying
data,
so
you
could
use
flawed
data
and
give
flawed
prediction.
A
So,
no
matter
how
rigorous
your
model
is,
if
your
data
is
flawed,
you
you're
not
going
to
get
good
results
out
of
it,
and
these
are
broad
categories
in
terms
of
issues
in
quantitative
energy
modeling.
So
for
the
rest
of
this
presentation,
I'm
only
going
to
focus
on
quantitative
energy
modeling
due
to
time
constraints,
but
this
is
the
most
widely
used
form.
So
we
have
some
technical
issues
in
terms
of
hardware
distribution.
A
This
particularly
applies
to
bitcoin,
so
it's
quite
important
for
us
to
understand
which
type
of
hardware
constitutes
to
how
much
of
the
network
so
very
efficient.
Very
simplified
example
is
you
could
perform
a
computation
using
10,
raspberry,
pi
or
a
very
powerful
computer
right,
and
the
energy
footprint
of
both
of
them
is
quite
different.
So
we
need
to
understand
the
distribution
of
hardware
used
in
order
to
generate
the
hashing
power
that
bitcoin
does
again.
This
applies
to
other
proof-of-work-based
cryptocurrencies
too,
and
we
don't
have
very
transparent
or
clear
data
on
hardware
distribution.
A
So
that's
something
that
you'd
see
a
lot
of
these
studies
have
flaws
in.
They
make
a
lot
of
assumptions
that
are
not
clearly
backed
by
empirical
data.
Then
we
have
some
issues
in
agreement
over
cost
of
electricity.
It
varies
a
lot.
I
have
a
graph
coming
up
in
a
bit,
then
geographic
distribution,
that's
also
quite
difficult,
and
it's
very
important
factor
when
you
look
at
carbon
emissions.
A
If
you
take
a
global
average
factor,
then
you're
not
going
to
get
very
reliable
figures
right
and
so
geographic
distribution
is
quite
important
and
then
pue
value
again
has
a
huge
influence
on
your
final
electricity
consumption
prediction.
So
here's
a
brief
overview
of
some
of
the
technical
issues.
I'll
just
highlight
two
of
them,
I'm
running
short
on
time.
The
first
one
is
constant
and
static.
Hardware
efficiency
assumption.
So
a
few
of
these
studies.
They
just
assume
that
hardware
efficiency
is
going
to
remain
same
for
a
long
period
of
time.
A
The
mora
study
that
I
highlighted
earlier,
they
assumed
that
hardware
efficiency
is
going
to
stay
static
for
next
hundred
years,
so
their
prediction
boomed
a
lot
because
of
this
assumption,
and
if
you
look
at
the
actual
data
set
in
the
simulation,
the
miners,
they
lost
millions
of
us
dollars
just
because
they
were
using
such
an
inefficient
piece
of
hardware
and
that's
why
you
get
the
figures
like
two
percent
of
two
degree:
centigrade
increase
in
temperature
right.
So
that's
one
of
the
big
issues.
Then
we
have
filling
in
the
missing
data.
Now.
A
This
is
a
big
issue.
With
this
field,
we
don't
have
reliable
data
right.
So
if
we
want
to
design
models,
we
have
to
make
some
assumptions.
But
when
we
make
assumptions,
especially
when
it
comes
to
filling
in
the
missing
data,
it
needs
to
be
backed
by
some
sort
of
evidence
that
we
can
consider
reliable
a
lot
of
these
studies.
They
just
fill
in
missing
data
using
a
random
distribution
or
a
static
distribution.
A
Again
that
might
make
it
such
that
the
older
hardware,
that's
less
efficient,
is
still
being
used
in
these
models
when
in
reality
we
don't
use
that
particular
hardware
anymore
ballooning.
Your
final
prediction
in
terms
of
cost
of
electricity,
like
I
said,
there's
a
huge
variance
and
you
can
see
in
this
chart.
It
ranges
from
0.025
up
until
0.14
and
a
lot
of
these
studies.
They
use
cost
of
electricity,
but
don't
explicitly
state
the
value.
So
it's
really
difficult
again
for
me
to
independently
verify
what
the
model,
what
the
model
section
prediction
is.
A
Some
of
them
use
variable
cost
based
on
geographic
distribution,
but
again
keep
in
mind.
Geographic
distribution
is
really
difficult
to
obtain
reliably.
So
that's
also
an
issue,
and
in
this
there's
a
trend
I
think
in
0.05,
so
that's
five,
us
cents.
You
can
see
starting
2019,
there's
an
increasing
trend
that
people
use
this
particular
value,
and
I
would
personally
attribute
this
to
the
cambridge
model.
So
people
started
assuming
that
five
uscent
is
reliable
because
the
cambridge
model
uses
that
value
and
within
the
cambridge
model
they
state
it's
based
on
expert
interviews.
A
You
don't
have
access
to
the
transcript
of
those
expert
interviews.
You
don't
know
who
those
experts
are
right,
so
I'm
not
entirely
sure
how
reliable
the
value
of
five
u.s
centers
and
similarly
we
have
issues
in
terms
of
geographic
location.
A
lot
of
these
studies
just
use
mining
pool
server
location,
so
I
can
be
present
in
asia.
I
can
sign
up
for
a
mining
pool
in
europe
without
necessarily
impacting
my
profitability,
but
these
studies
would
assume
that
I'm
located
in
europe
and
that
again
influences
their
final
results
and
another
big
issue
is
for
bitcoin.
A
We
have
somewhat
of
a
good
idea
of
about
34
of
the
network
how
it's
geographically
distributed
based
on
a
data
set
from
cambridge,
so
people
try
and
use
this
34
data
set
and
apply
it
to
the
whole
network.
They
stretch,
34
percent
up
to
100
and
basic
statistics,
tell
you
you
cannot
do
that.
You
need
good
reasoning,
why
you
are
filling
and
missing
data,
and,
if
you're
making
prediction
for
the
whole
network
using
a
small
data
set,
then
your
final
results
they
they
vary
a
lot
in
terms
of
their
reliability.
A
So
if
you
are
going
to
do
that,
because
you
don't
have
access
to
the
data
you
need
to
compound
it
with
confidence
intervals,
you
need
to
compound
it
with
sensitivity,
analysis.
You
need
to
explicitly
tell
the
reader
that
I'm
stretching
on
this
data.
I
don't
know
how
reliable
this
is.
This
is
the
call
confidence
interval
right
and
then,
like
I
said,
bue
value
again,
a
huge
variance.
A
So
in
the
end
we
have
designed
some
quotes
of
practices.
I
don't
have
time
to
walk
you
through
all
of
them,
but
basically
what
we've
done
is
we
have
looked
at,
for
example,
information
systems,
and
how
do
you
share
your
data?
So
we
argue
that
you
should
use
version
control
in
order
to
share
your
data.
You
should
use
standardized
terminologies
and,
if
you're
going
to
share
your
source
code,
let's
say
in
form
of
excel
sheets,
there's
a
huge
literature
on
how
do
you
prepare
excel
sheets?
What's
information
quality?
A
What's
data
quality,
because
even
a
single
value
in
your
excel
sheet
can
actually
impact
your
final
result
significantly
and
there's
guidance
out
there
in
existing
literature?
And
if
you
look
at
fields
like
medical
sciences,
it's
expected
that
you
do
that
before
you
submit
something
to
an
academic
journal.
We
don't
have
similar
standards
in
this
field
right,
so
we
propose
some
guidelines
on
how
you
should
do
that,
and
I
think
I'm
out
of
time
so
I'll
just
walk
you
through
some
of
our
conclusions.
A
So
I
think
it's
safe
to
conclude
that,
generally,
these
studies
lack
in
the
scientific
rigor
and
we
caution
anybody
who's
going
to
base
policy
decisions
on
these
to
use
these
as
guesses,
rather
than
accurate
estimations
and
there's
already
quite
a
bit
guidance
on
how
we
can
improve
this,
especially
from
kumi
and
our
suggestions
align
with
him.
We
need
more
real
world
data
before
we
can
actually
make
more
progress
here.
A
Designing
more
models
based
on
more
assumptions,
isn't
really
going
to
help
anybody
in
terms
of
improving
accuracy
of
these
models,
we
do
need
to
collect
more
reliable
data,
and
then
we
need
more
standardization
in
this
field
and
the
last
suggestion-
and
I
think
that's
something-
that's
probably
most
actionable
for
us
at
this
point-
if
we
do
not
get
more
data
which
might
be
a
possibility,
it's
quite
difficult
to
get
reliable
data
sets.
In
that
case,
we
need
to
explicitly
state
the
confidence
intervals.