Trends from the Trenches host Eleanor Howe sits down with computational biologist Eric Schadt—now an AI Scholar in Residence at Pfizer—who has spent 25 years pushing a simple idea with huge consequences: disease is a network property, not a single-gene problem. That mindset changes what “precision medicine” should mean, shifting the goal from better labels to better treatment decisions that actually improve outcomes for real patients.
They get concrete about how the AI revolution lands inside pharma and biotech. Eric explains what Pfizer’s new AI Scholar role is meant to do, what an AI-native R&D operating model looks like, and why tightly integrated teams of a few humans plus modern foundation models can outperform the old way of working. From multimodal oncology models to reasoning-first LLMs, they talk about agentic systems that can orchestrate many specialized models, not just generate text, and why interpretability and auditable traces matter when clinical decisions are on the line.
GUEST BIOs
Eric Schadt, AI Scholar, R&D, Pfizer
Eric Schadt is a scientist, founder, and R&D leader working at the intersection of artificial intelligence, systems biology, human genetics, and precision medicine. His work has focused on building computational models of human biology and translating them into new approaches to drug discovery, development, diagnostics, and patient care. He has built and led organizations across pharma, genomics, biotechnology, academia, and precision medicine, and his current focus is the development of AI-native systems capable of modeling patient state, reasoning over interventions, and continuously learning from biological and clinical outcomes.
TRANSCRIPT
Eleanor Howe
Hello
everyone,
my
guest
today
is
Eric
Schadt.
Eric
is
a
computational
biologist
who
has
spent
25
years
making
the
case
that
disease
is
a
network
property
rather
than
a
single
gene
one, work
he
started
at
Rosetta
and
Merck.
Since
then,
he's
co-founded
Sage
BioNetworks,
served
as
Chief
Scientific
Officer
at
Pacific
Biosciences,
built
the
Icahn
Institute
for
Genomics
and
Multiscale
Biology
at
Mount
Sinai,
where
he
is
still
Dean
for
Precision
Medicine,
and
founded
the
genomics
company
Sema4 and took it public.
Most
recently,
he
was
Chief
Science
Officer
at
Pathos
AI,
working
on
one
of
the
largest
multimodal
foundational
models
built
in
oncology.
And
a
few
weeks
ago,
he
joined
Pfizer's
RD
organization
as
their
first
AI
Scholar
and
started
a
new
company
that's
still
in
still.
Eric,
welcome
to
The
Trenches.
Eric Schadt
Thanks.
Thanks.
Great
to
be
here.
Eleanor Howe
All
right,
so
the
audience
has
heard
a
bit
about
you,
but
is
there
anything
that
you
want
to
add
about
the
journey
that
brought
you
here
where
you've
just
started
two
new
jobs?
Eric Schadt
Yeah,
Eric Schadt
so
well,
it's
all
indeed
about
the
journey.
So
I
would
say
starting
25
years
ago
when
I
first
transitioned
out
of
academia
into
Roche
Biosciences
to
gain
access
to
the
first,
you
know,
big
gene
chip
technologies
with
Affymetrix
and
Centenni,
you
know,
since
the
start
of
that
mission
all
the
way
until
today,
it's
been
how
do
we
both
generate
and
integrate
very
large
scales
of
data
and
then
build
predictive
models
from
those
data
to
again
predict
interesting
things?
So
I've
done
everything
from
the
diagnostic
side
to
the
drug
discovery
side
to
drug
development
to
healthcare
delivery.
And
I
would
say
the
mission
has
been
how
do
we
bring
all
the
data
can
tell
us
and
the
accelerated
modeling
and
predictive
nature
of
those
models?
How
do
we
drive
that
into
the
hands
of
the
patients
to
better
impact
their
outcomes?
And
that
mission
is
still
very
much
in
front
of
us.
And
I
would
say,
you
know,
that
medicine
generally
has
still
failed
to
benefit
at
the
level
that
we
see
the
science
and
all
other
areas
of
life,
you
know,
benefiting
from
this
big
AI
revolution.
We're
still
not
seeing
that
in
full
swing
in
patients,
in
the
clinics.
And
so
I
would
say
this
next
step
is
just
how
do
we
keep
pounding
on
that
and
getting
closer
to
the
actual
delivery
arm,
which
is
what
the
new
my
new
company
is
going
to
aim
Eric Schadt
to
do.
Eleanor Howe
Fantastic.
So
I
want,
I
want
to
just
ask
quickly
about
this
AI
scholar
position
at
Pfizer.
What
is
an
AI
Scholar?
Pfizer
just
created
this
role,
I
believe.
So
presumably
you
had
some
influence
on
designing
it.
Eric Schadt
Yeah.
Well,
as
I
was,
you
know,
building
this
new
company,
again,
focused
very
much
on
more
direct
link
to
patients
and
being
closer
to
the
provider
and
delivering
both
both
insights
and
care
plans
and
so
on
directly
to
patients.
I
was
also
talking
with
Pfizer
about
what
they
were
up
to
in
terms
of
this
AI
revolution
and
their
push
to
move
to
a
more
native
AI
operating
model.
And
of
course,
they
are
a
company
with
a
big
wealth
of
data,
a
great
diversity
of
interests.
So
as
I
was
talking
to
them,
I
saw
we
both
saw
the
potential
for
me
to
help
help
them
on
that
strategy
front
and
help
drive
a
few
big
programs.
And
that's,
in
my
view,
synergistic
with
the
efforts
I
have
at
the
New Co.
And
so
this
AI
Scholar
was
sort
of
a
role
that
was
created
as
a
consultant
in
residence
to,
you
know,
kind
of
spend
some
real
time
there
working
with
the
teams
to
see
what
could
be
on
the
big
AI
front.
Eleanor Howe
And
do
you
have
a
vision
for
what
an
AI
native
RD
organization
would
look
like?
Like
what
do
today's
organizations
not
have
yet?
Eric Schadt
Yeah,
I
think
it
it's
sort
of
the
transition
from
so
we've
all
been,
you
know,
are
benefiting
to
some
extent
from
what's
happening
in
AI
generally,
especially
with
deep
learning
and
large
language
models.
We're
all
querying
Chat
GPT
and
Claude
and
Llama
and
all
the
other,
you
know,
about
an
increasing
number
of
these
large
language
models
being
made
available
to
everybody.
And
they
have
extraordinary
capabilities.
And
so,
like
you're
seeing
lots
of
tools.
We
all
use
them
to
help
us
review
papers,
write
papers,
kind
of
evaluate
opportunities.
We
make
agents
to
do
coding
and
so
on.
So
we
are
creating
lots
of
tools,
but
to
me,
it's
more
about
how
do
you
sort
of
harmonize,
unify
all
of
those
efforts
in
more
concentrated
teams
that
together
can
do
the
work
of
what
classically
would
have
been
tens
to
even
more
than
tens
of
individuals.
So
seeing
teams
of
four
or
five
humans
where
you
have
an
AI
engineering
expert
in
there
who
knows
how
to
work
with
scaled
infrastructures
and
do
various
kinds
of
trainings
on
these
models,
and
AI
scientists
who
are
very
savvy
with
the
application
of
those
models,
working
in
tandem
with
deep
experts
on
how
to
guide
those
developments
to
solve
very
hard
problems.
And
so
today,
those
teams,
if
done
correctly,
literally
can
do
the
work
of
40
to
50
people
of
what
it
would
have
taken
to
do
that
a
short
number
of
years
ago.
So
and
it's
in
it's
embracing
all
of
that
at
a
at
a
level
that's
not
just
one
tool
at
a
time,
but
rather
trying
to
leverage
the
synergies
that
come
by
integrating
everything
and
having
a
an
operating
model
where
everybody's
kind
of
adhering
to
that.
So
I
would
say,
like
many
other
companies
and
efforts,
we're
we're
like
trying
to
figure
that
out
and
doing
experiments
and
see
what
works
and
what
doesn't.
Eleanor Howe
Okay,
so
let's
hear
a
little
bit
more
about
this
new
co
of
yours.
You
recently
launched
it,
it's
in
stealth
mode,
but
what
can
you
tell
us
about
it?
Eric Schadt
Yeah,
I
would
say
what
I
could
say
about
it
is
it's
an
effort
to,
again,
how
do
we
deliver
precision
medicine
solutions
into
the
hands
of
patients
where
today
there's,
I
wouldn't
say
it's
active
resistance,
but
many
difficulties
and
why
that's
been
hard
to
get
advanced
multi-feature
data
algorithm
informed
in
play,
in
the
hands
of
physicians,
making
decisions
on
what's
happening
to
a
patient
and
what's
the
best
way
to
treat
them.
And
so
in
thinking
through
all
of
the
great
myriad
of
issues
that
make
that
a
hard
problem,
the
new co
is
really
built
up
around
how
do
we
help
address
those
issues
and
become
part
of
the
solution
that
gets
driven
into
the
hands
of
physicians
directly
to
the
patient.
So
think
about
leveraging
AI
and
these
deep
learning
models,
large
language
models,
including
those
that
are
very
advanced
at
reasoning.
So
the
kinds
of
models
we
use
to
generate
computer
code
and
solve
very
complex
pipeline,
very
complicated
pipelines
that
are
comprised
of
many
different
software
components,
all
kind
of
brought
together
to
solve
a
big
problem,
like
the
kind
of
autonomous
agentic
systems
that
these
LLMs
have
that
kind
of
capability
are
now
kind
of
being
re-routinely
used
for
that.
And
it's
just
harnessing
that
kind
of
reasoning
and
orchestration
capability
to
evaluate
data
around
patients.
And
so
kind
of
doing
all
of
the
work
required
to
get
reasoning-based
language
models
to
think
like
that,
to
reason
like
that
and
make
those
predictions
and
decisions,
but
importantly
to
also
generate
mechanistically
interpretable
traces,
auditable
traces
that
rationalize
why
that
prediction
was
made,
all
to
aid
the
physician
in
making
a
better,
a
better
decision
around
the
patient.
Eleanor Howe
So
if
I
was
going
to
be
super
reductionistic
about
describing
this,
would
it
you
call
it
decision
support,
just
if
you
wanted
to
really
paint
it
with
a
broad
brush?
Eric Schadt
Yeah,
I
think
it
it
probably
starts
as
more
decision
support,
but
increasingly,
you
know,
goes
as
opposed
to
simply
supporting
the
decisions.
So
if
you
think
about
if
you're
going
to
deviate
from
some
guideline
standard
of
care,
you
need
to
understand
what's
the
evidence
behind
that.
And
then
you
need
to
build
up
support
for
that
evidence,
and
you
need
to
build
that
support
up
to
a
point
where
it
de
facto
becomes
the
standard
of
care.
So
think
initially
decision
support,
building
the
evidence,
but
ultimately
you've
built
up
enough
evidence
where
those
decisions
should
just
be
considered
a
standard
of
care
and
be
followed,
just
like
the
decision
trees
that
physicians
will
follow
in
today's
guideline-driven
care.
That's
what
that
AI
Eric Schadt
would
become.
Eleanor Howe
Okay,
so
not
just
decision
support,
but
also
basically
standard
of
care
design
support
as
well.
Eric Schadt
Yeah,
standards
of
care
and
the
and
the
underlying
practice
guidelines
that
define
those.
You
know,
they're
they're
typically
built
up
from
population,
you
know,
population
averages.
And
there
is
some
attempt
to
stratify
populations
into
higher
risk
groups
versus
lower
risk
groups,
and
then
consider
that
population
under
a
specific
guideline.
But
at
the
end
of
the
day,
it's
still
kind
of
taking
a
mean
behavior
over
some
population
and
making
a
decision
in
a
decision
tree
of
what
should
be
done
with
that
patient,
as
opposed
to
getting
to
individualize
trajectories
of
those
patients
within
those
groups
to
understand
which
are
likely
going
to
benefit
from
whatever
that
standard
of
care
is
versus
those
who
are
not
going
to
benefit
and
being
able
to
make
that
prediction
and
then
have
alternatives
that
can
be
done
in
in
place
of
the
standard.
So
think
of
it
as
a
more
individualized
decision
support
guideline
than
a
population-based
one.
Eleanor Howe
Well,
I
personally
am
very
excited
about
this
because
it
sounds
like
the
kind
of
thing
that
would
maybe
enable
people
to
do
a
better
job
in,
you
know,
making
medical
decisions
for
folks
who
are,
I
don't
know,
maybe
not
particularly
close
to
the
reference
genome,
for
example.
Like
that
could
be
really
great
news
for
people
who
are
not
Caucasian.
And
frankly,
for
people
who
are
not
men,
potentially
could
be
really
great.
Eric Schadt
So
I
let
Yeah,
so
that
like
it's
just
the
like
a
perfect
reflection
of
the
sort
of
gross
population
groups
we
try
to
pluck
everybody
into,
or
the
standard
of
care
may
have
grown
out
of
one
of
those
population
groups,
and
then
you
try
to
force
it
across
all
other
groups
without
much
evidence.
And
so
again,
the
aim
here
is
not
to
throw
you
know
the
baby
out
with
a
bathwater,
as
they
say,
but
to
see
when,
you
know,
when
is
that
working
and
is
it
relevant
to
this
individual
or
to
this
other
group
of
individuals?
And
if
it
is,
great,
like
that
gets
wired
into
what
should
be
used.
But
if
not,
you're
then
to
say,
okay,
we're
gonna
deviate
from
this
for
this
reason.
We
think
this
is
gonna
work
better,
and
we're
gonna
longitudinally
track
and
assess
whether
that
outcome
was
better
or
not.
Eleanor Howe
Sounds
great.
Well,
okay.
I'm
I
think
I'm
very
interested
to
hear
more
when
you
folks
are
ready
to
talk
more
about
it.
It
sounds
really
interesting.
So
then,
okay,
moving
on
then,
because
I
don't
want
to
keep
picking
you
at
you
about
the
thing
you
can't
talk
about
yet.
When
I
hear
precision
medicine
normally,
I
don't
think
of
this
kind
of
frankly
big
involved
project
that
you're
talking
about.
I
think
about
I
think
about
diagnostics.
Usually
when
people
say
precision
medicine,
they're
what
they're
talking
about
is
diagnostics.
And
I
think
like
you're
clearly
thinking
about
a
lot
of
other
angles.
And
and
I
am
I
think
you've
got
some
opinions
I
happen
to
know
about
the
problems
with
diagnostics
and
why
they're
just
not
as
good
as
they
could
be,
and
maybe
not
as
popular
as
they
could
be.
Do
you
want
to
talk
about
that
a
little
bit?
Eric Schadt
Yeah,
Eric Schadt
absolutely.
I
do
think
that
one
of
the
bigger
limitations
and
why
we
don't
see
more
advanced
precision
medicine
solutions
being
driven
through
clinical
practice
into
patients
has
a
lot
of
its
rooting
in
diagnostics
and
and
the
difficulty
of
diagnostics
from
a
from
many
different
dimensions,
from
you
know,
payers
to
the
reductionist
methodology
and
medicine
to
the
financial
incentives
and
pharma
and
so
on.
There
are
many,
many
different
dimensions
of
why
diagnostics
can't
be
more
progressive
to
be
driven
into
the
hands
of
physicians
and
patients
more
readily.
And
so
have
you
know,
learned
many,
many
hard
lessons
in
this
road.
And,
you
know,
chief
among
them
is
the
complexity
of
taking
a
what
I'll
call
like
a
multi-feature
algorithm-informed
decision
or
test
that
you
want
to
drive
into
practice
that's
actually
treatment
predictive,
not
just
prognostic,
not
just
classifying
people
into
a
different
strata,
but
also
indicating
what
treatment
may
be
best
for
them,
that
that
that's
a
very
long
journey
with
respect
to
regulation,
complicated
and
very
expensive.
And
so
until
we
address
some
of
the
issues
around
how
do
you
get
paid
for
for
those
sorts
of
tests
and
how
do
you
reduce
the
regulatory
burden,
especially
in
those
things
which
are
clearly
going
to
be
benefiting
patients,
I
just
think
it's
gonna
be,
you
know,
continue
to
be
difficult.
Eleanor Howe
And
if
if
you
got
to
wave
a
magic
wand
and
change
the
things
that
you
needed
to
to
make
diagnostics
more,
I
don't
know,
empowered
to
help
patients,
like
what
would
you
do
at
a
like
a
policy
level
or
a
funding
level?
Do
you
have
like
a
wish
list?
Eric Schadt
Yeah,
I
think,
you
know,
chief
among
those
would
be
kind
of
setting
a
more
harmonized,
unified
evidentiary
standard
for
payers
that
if
you
go
after
and
you
take
the
time
and
money
to
develop,
you
know,
an
advanced
diagnostic,
companion
diagnostic
test,
that
you
know,
that
that's
gonna
get
reimbursed
at
a
at
a
at
a
good
enough
level.
And
today,
that
doesn't
exist.
Today,
payers
have
a
very
arbitrary
standard
that
they
can
apply
and
decide
it
will
appear
from
those
on
the
outside
almost
like
a
whim
of
whether
they're
gonna
cover
something
or
not.
And
those
hurdles,
you
know,
because
they're
not
well
defined,
you're
up
front
as
a
company
taking
on
big
risk
of
the
unknowable,
of
not
knowing
if
you
go
all
the
way
over
the
finish
line
to
develop
something
like
this,
whether
there's
gonna
be
a
channel
to
be
paid.
And
so
it
needs
to,
you
know,
that
needs
to,
you
know,
I
don't
know
if
it's
you
know,
HHS
that's
coming
out
with
policy
on
that
regulations,
or
if
it's
legislative,
like
there's
some
push
on
the
pay
on
the
payers
that
needs
to
be
had
to
kind
of
help
push
in
that
direction.
So
that's
number
one.
The
other
I
I
think
is
just
kind
of
financial
incentive,
maybe
a
harder
problem
on
the
on
the
side
of
pharma
because
of
the
expense
of
developing
a
companion
diagnostic,
especially
a
complex
companion
diagnostic.
So
you'll
see
on
the
market
plenty
of
variate
diagnostics.
You
know,
are
you
EGFR
positive,
are
you
KRAS
positive
and
cancer?
And
that
will
dictate
what
standard
of
care
you
get.
But
those
are
univariate,
and
as
we
know,
most
you
know,
biology
is
complex.
And
if
we
really
want
to
stratify
a
patient
into
more
accurate
treatment
groups,
it
needs
to
be
multi-multi-feature,
not
just
a
single
gene,
but
multiple
genes.
And
then
you
need
an
algorithm
to
help
process
the
logic
behind
that
to
make
a
decision.
So
that's
that
both
costs
a
lot
of
money
to
do,
especially
in
a
clinical
trial,
because
you
have
to
make
a
whole
new
clinical
trial
arm,
just
like
its
own
RCT
randomized
control
trial,
to
prove
that
out
in
terms
of
its
ability
to
inform
on
treatment
predictiveness.
But
then
you
also
are
stratifying
the
population
into
smaller
groups,
almost
like
a
financial
disincentive
for
farmers
to
do
that,
because
most
are
public
companies.
And
if
you're
a
public
company,
you
want
to
move
the
public
market
needle.
Like
that's
not,
you
don't
move
that
needle
with
a
hundred
million
dollar
a
year
drop.
Like
that
may
sound
great
to
most
people
and
biotechs,
but
that's
not
great
for
a
public
market
pharma
company.
So
those
sorts
of
opportunities
will
just
get
deprioritized
because
pharma
wants
to
center
on
what's
what's
the
home
run.
We
need
a
billion
dollar
a
year
plus
seller
to
move
the
needle.
And
that
sort
of
runs
counter
to
this
stratification
with
biomarkers.
But
maybe
I'll
pause
there
and
see
if
I'm
getting
at
your
question.
Eleanor Howe
No,
you
absolutely
are.
And
and
actually,
I
it,
you
know,
to
me,
it
it
always
has
felt
like
the
more
diagnostics,
the
better.
Eleanor Howe
Don't
we
want
to
know
what's
going
on
in
our
own
bodies
when
we're
sick
and
how
do
we
get
the
right
treatment?
And
so
I
wonder
for
folks
who
wish
they
could
do
something.
Do
you
have
a
recommendation?
Should
we
all
write
our
senators
demanding
new
legislation?
Eric Schadt
Yeah,
it's
a
good
question.
I
would
say,
yeah,
like,
you
know,
pressure,
I
think
pressure
from
the
patients
is
the
right
angle
because
I
think
those
who
lose
the
most
from
arbitrary
evidentiary
standards,
financial
disincentives,
and
so
on,
are
the
patients.
You
know,
patients
are
going
to
get
a
suboptimal
standard
of
care
as
a
result,
and
how
how
physicians
and
systems
get
paid,
what
CPT
codes
do
they
file,
and
what
justifies
the
filing
of
that
CPT
code
to
administer
some
intervention
or
treatment
to
a
patient,
like
that
all
starts
getting
more
and
more
removed
from
the
patient
and
more
and
more
centered
on
how
does
this
system,
this
big
system
involving
form
of
payers
and
systems,
how
does
that
sustain
itself
and
you
know,
stay
in
business
and
so
on.
But
I
think
getting
advocacy
groups,
you
know,
are
certainly
a
very
effective
way,
how
targeted,
you
know,
advocacy
groups
that
got
built
up
around
HIV
AIDS,
breast
cancer,
and
so
on,
have
been
wildly
successful
in
driving
the
kind
of
sea
change
that
we
need.
So
it's
almost
like
something
along
those
lines
needs
to
happen.
Maybe
it's
to
some
extent
education
of
patients
on
how
the
system's
actually
working
and
are
the
decisions
that
are
being
made
of
highest
interest
to
them.
I
think
part
of
the
new
co-push
is
can
we
enable
a
patient
to
take
on
more
agency
than
to
giving
that
agency
over
to
the
system,
to
the
physician,
and
trade
for
them
fixing
whatever
is
wrong
with
you.
If
they
had
your
interest
front
of
mine
in
all
different
dimensions,
that
would
work,
you
know,
that
would
be
effective,
but
that's
no
longer
the
case.
So,
kind
of
how
we
enable
patients
to
take
on
more
control,
more
agency,
better
question
what
decisions
are
being
made
around
them,
what
the
alternatives
are,
and
so
on,
may
help
drive
that
advocacy.
Eleanor Howe
Okay,
great.
All
right,
everybody,
let's
get
organized.
Clearly,
we
need
a
new
advocacy
group.
Unless
somebody
wants
to
write
in
and
tell
us
about
the
group
that
already
exists
that
we
can
go
join,
that
would
be
even
better.
Eric Schadt
Well,
I
was
just
reading
a
funny,
I
thought
it
was
funny,
cartoon.
And
it
was
a
person
going
into
the
physician
and
the
physician
giving
the
diagnosis,
and
the
patient
saying
back,
well,
I
want
to
understand
why
your
diagnosis
differs
from
ChatGPT's
diagnosis
of
my
condition.
So
I
think
there's
a
you're
already
seeing
AI
enable
a
better
level
of
understanding
that's
more
accessible
to
people
and
people
leveraging
that
to
self-diagnose
and
to
understand
what
a
physician
or
what
their
medical
record
is
saying.
So
I
do
feel
like
that
match
has
gotten
dropped
and
is,
you
know,
the
fire
is
going.
And
so
I
do
feel
like
we're
at
the
cusp
of
a
revolution
that
way.
Part
of
in
support
of
that,
I
uh
like
I've
looked
at
companies,
I
don't
know
if
you've
heard
of,
like
function
health
and
superpower.
And
what
caught
my
eye
on
these
companies
is
take
function
health.
Like
over
a
very
short
period
of
time,
they
got
like
a
half
a
million
to
a
million
subscribers
who
were
paying,
I
don't
know,
something
like
$500.
And
they
were
paying
$500
to
get
a
routine
blood
test
that
you
can
get
for
free
through
your
insurance,
through
your
primary
care
physician.
And
so
you're
like
wondering
why
are
people
paying
$500
to
get
this
blood
test,
a
standard
blood
test
from
Quest,
to
you
know,
get
some
health
insights.
And
to
me,
it's
and
they
got
to,
you
know,
it's
a
half
a
million
to
a
man,
like
like
almost
a
billion
dollars
in
revenue
in
like
a
one
to
two
years
time,
like
astonishing.
Well,
and
so
like,
why
is
that?
And
to
me,
it's
just
showing
the
appetite
of
the
patient.
Like
they
are
not
getting
what
they
need
from
standard
of
care,
they
are
not
understanding
their
conditions,
their
health,
their
wellness
at
a
deep
enough
level,
and
they're
willing
to
pay
out
of
pocket
to
have
more
data
generated
with
a
promise.
Like
I
would
say,
function
health
at
those
beginning
stages
was
more
promise
that
we're
gonna
figure
out
what
these
data
mean
and
how
to
better
inform
your
care.
You
know,
just
just
that
promise
was
enough
for
those
people
to
sign
up.
So,
again,
to
me,
just
shows
the
the
need
and
the
appetite.
And
again,
combine
that
with
AI
and
you
have
a
lot
of
ingredients
to
be
very,
very
disruptive.
Eleanor Howe
Yeah,
absolutely.
And
then
also
function
health
sounds
like
a
very
scientifically
driven
company,
but
then
there's
also
charlatans
out
there,
on
the
other
hand,
taking
advantage
of
people
who
are
basically,
you
know,
desperate
and
uninformed.
And
that
is
really
sad.
Eric Schadt
So
yeah,
and
that
again
where
is
hopefully
where
as
AI
solutions
become
more
accessible
in
these
models,
especially
the
Frontier
Lab,
big
large
language
and
reasoning
models
get
better
and
better,
you
know,
that
that
does
provide
channels
to
vet
what
you're
being
exposed
to
and
try
to
distinguish
between
the
charlatans
and
the
true,
you
know,
truth
seekers
who
are
well
grounded
in
science
and
so
on.
So,
and
I
would
say
a
lot
of
my
effort
is
around
like,
yeah,
how
do
we
help
both
brand
trusted
sources
and
help
drive
solutions
from
trusted
sources
into
the
hands
of
physicians
and
patients?
And,
you
know,
like
I
think
that's
going
to
be
the
right
kind
of
strategy.
I
I
think
it's
maybe
you
might
be
thinking,
you
know,
like
it's
maybe
more
difficult
on
the
political
side
than
on
the
health
side.
I
sure
as
I
hope
it
doesn't
get
as
complicated
as
it
is
on
the
political
side.
Eleanor Howe
So
now
that
we've
talked
a
little
bit
about
the
companies
and
the
goals
and
the
failures
in
our
market,
let's
talk
about
how
these
models
that
are
changing
the
world
should
be
built.
Do
you
have
a
philosophy
about
what
you
think
is
the
most
successful
way
to
structure,
say,
an
agentic
model
or
a
collection
of
models
or
a
single
big
one
model
to
rule
them
all
structure?
Eric Schadt
Yeah,
and
that's
something
I
would
say
the
entire
community
who
builds
these
models
and
thinks
about
their
application
and
the
combinations
of
those
models
are
all
thinking
about
and
running
experiments
to
try
to
figure
out
what
is
the
right
recipe.
I
would
say
the
last
three
years
of
my
life
spent
with
building,
you
know,
foundation
models
on
DNA,
on
RNA
sequencing,
on
H&E
images,
on
clinical
text,
and
trying
to
understand
the
behavior
of
those
individual
models
and
how
best
to
query
them
to
predict
things
that
are
high
interest
in
clinical
care,
like
what's
the,
you
know,
is
this
person
going
to
respond
to
this
type
of
medication
or
not?
What's
the
patient's
prognosis
and
those
sorts
of
predictions?
And
then
how
do
you
fuse
those
models
together
to
have
this
multimodal
fusion
model
and
whether
that's
better
at
making
these
kinds
of
predictions?
And
then
how
do
you
leverage
the
LLM
models
to
kind
of
contextualize
the
multimodal
models
that
you're
building
on
those
scaled
molecular
dimensions
of
data?
So
those
were
all
the
things
we
were
thinking
about.
And
I
would
say
from
that
work,
my
continued
work
with
Pfizer
and
on
the
NUCO
and
several
others,
is
that
there
are
you
know
a
couple
of
themes
are
coming
out.
First,
these
frontier
lab
models
and
almost
these
open
weight
models
that
are
getting
more
and
more
frontier
lab-like.
So
by
those
I
mean
so
think
of
Moonshot
with
the
Kimi
K2
and
K3,
NVIDIA
with
the
Nemotron
family
of
models,
Z.AI
with
the
GLM
5.3
and
other
five
series
of
models.
Like
those,
you
know,
models
are
in
the
hundreds
of
billions
to
trillions
of
parameters,
so
getting
comparable
to
the
Frontier
Labs,
you
know,
the
Anthropics,
the
Metas,
the
Googles,
and
the
OpenAIs.
So
all
having
this
pretty
amazing
ability
to
not
just
deep
in
broad
knowledge,
but
but
very
good
reasoning
capabilities.
And
out
of
the
good
reasoning
capabilities
comes
autonomous
agentic
system
capability,
where
you
no
longer
have
to
define
lots
of
tools
and
agents
to
do
specific
things.
You
just
supply
the
reasoning
LLM
with
a
list
of
things
it
can
have
access
to
and
the
kind
of
problem
you
want
to
solve,
and
it
goes
off
and
figures
out
the
best
way
to
solve
that.
So
that
that
sort
of
capability
you
know,
you
need
as
sort
of
a
brain
core
to
whatever
set
of
models
you're
trying
to
leverage
to
predict
things.
So
you
need
that
as
your
core
kind
of
reasoning
and
orchestration
engine.
But
I
think
those
orchestrators
are
always
going
to
be
orchestrating
over
a
wide
array
of
models
that
have
differing
levels
of
capabilities,
especially
with
respect
to
certain
types
of
data.
And
that
what
the
reasoning
models
need
to
be
trained
to
do
is
to
understand
when
is
this
model
or
that
model
better
for
this
task
in
this
context?
Like
these
models
are
very
good
at
assessing
context
around
large
scales
of
information
around
a
given
individual,
say
like
a
patient.
So
to
me,
it's
going
to
be
Orton's
like
frontier
lab
type
models
having
access
to
a
wide
array
of
foundation
models
from
multimodal
to
unimodal
to
other
types
of
LLMs
that
have
been
specifically
trained
to
do
tasks
very
well.
Like
that,
that's
more
of
the
arena,
I
think,
that
that
will
be
coming.
So
it
won't
be
one
model
rules
them
all.
It'll
be
handfuls
of
really
good
models
that
are
capable
of
orchestrating
over
a
vast
array
of
other
types
of
models,
but
the
other
types
of
models
to
do
come
up
with
the
best
solution
and
complex
tasks
or
make
the
best
predictions,
you're
gonna
have
the
ability
to,
you're
gonna
have
to
have
the
ability
to
leverage
many
other
Eric Schadt
types
of
models.
Eleanor Howe
And
then
where
do
you
think
this
lab
in
the
loop
concept
is
going
to
fit
into
this
world?
Like
where
do
you
think,
when
do
you
think
we're
gonna
I
I
have
to
assume
that
lab
in
the
loop
will
become
a
normal
thing
that
everybody
does
at
some
point.
What
do
you
think
the
timeline
is
on
that?
Eric Schadt
Yeah,
I
I
would
say,
you
know,
first
of
all,
it's
like
already
happening.
So
the
whole
physical
AI
revolution
is,
you
know,
just
uh
nearly
as
sweeping
as
the
information
AI
side
with
the
deep
learning
models.
But
it's,
you
know,
it's
like
a
it's
more
expensive
getting,
you
know,
to
good,
high-functioning
robots
that
can
be
very
diverse
and
the
tasks
they
can
carry
out
is,
you
know,
like
the
physical
side
is
still
difficult
to
solve.
But
if
you
look
at
companies
like
Amazon,
go
go
look
at
how
they
do
the
the
story
of
all
this
stuff.
Like
it's
shocking
to
me
that
you
can
order
something
on
Amazon
and
it
can
show
up
at
your
door
three
hours
later.
It's
crazy.
And
so
if
you
want
to
see
how
they
do
it,
it's
it's
lots
of
the
AI
and
it's
lots
of
physical
AI,
it's
lots
of
robots
that
are
carrying
out
all
those
tasks.
And
now
you
have
a
Jeff
Bezos
forming
this
company,
Prometheus,
that
is
centered
on
how
do
you
bring
that
kind
of
automated
physical
AI
capability
to
everything
else.
And
it
won't
be
a
surprise
that
the
health
and
life
sciences
are
front
and
center
for
that.
But
there
are
a
range
of
other
companies.
You
look
at
what
Pharma
does
on
the
manufacturing
of
drug
side,
that's
all
incredibly
automated,
robotic
with
advanced
AI
solutions
in
the
mix.
I
think
the
biggest
change
that's
coming
is
the
use.
Remember
that
core
brain
reasoning
orchestration
engine
I
was
just
talking
about,
like
basically
having
that
in
the
mix
where
it
you
no
longer
have
to
have
this
well-defined
blueprint
of
what
the
robot
does
and
when,
but
now
it
can
sort
of
adapt
at
will
with
these
reasoning
type
capabilities.
So
I
think
that's
more
what
you're
gonna
see
coming
into
play
that
makes
iteration
on
the
experiments
to
do
the
experiment,
look
at
the
outcome,
assess
the
outcome,
design
the
next
experiment,
like
more
of
that
iteration
will
increasingly
happen
in
an
automated
way.
And
it
will
center
heavily
on
these
reasoning-based
orchestration
engines
with
autonomous
agentic
system
capability
that
are
are
really
enabling
that.
Eleanor Howe
So
as
far
as,
I
mean,
this
is,
you
know,
the
world
is
changing
right
now,
right?
And
every
basically
every
bit
of
it
is
changing
because
of
these
new
technologies.
Which
camp
are
you
in?
Are
you
in
the
camp
of
we're
all
gonna
be
jobless
and
there
will
be
no
work
for
anybody?
Or
in
the
camp
of
actually
we're
just
gonna
have
different
jobs,
everybody's
gonna
be
busy
and
we're
all
gonna
be
employed
doing
completely
different
things
in
the
future.
Eric Schadt
I
would,
yeah,
I'm
mainly
an
optimistic
guy.
So
like
I
definitely
am
more
in
the
camp
that
AI
is
going
to
enable
us
to
reach
new
levels
as
opposed
to
put
us
all
out
of
a
job
with
nothing
to
do,
where
we're
that's
bad.
And
and
I
would
say
my
push
to
do
things
like
the
AI
Scholar
and
this
new
company
and
maintain
activity
at
Mount
Sinai
is
all
like
like
I
don't
really
have
to
work,
but
like
it's
a
it's
like
an
existential
anxiety,
I
think,
to
not
first
of
all
not
want
to
be
on
the
sideline,
like
want
to
be
in
the
middle
and
want
to
help
craft
like
how
can
we
be
more
purposeful
about
the
world
we
want
to
live
in?
Because
if
we
just
let
it
happen
instead
back,
then
we're
likely
gonna
end
up
with
something
we
don't,
you
know,
nobody
likes.
So
wanting
to
be
more
purposeful,
but
I
think
like
innovation,
like
the
decision
making,
accountability,
responsibility,
like
those,
those
are
still
all
very
human.
And
I
think
it
won't
be
at
the
end
of
the
day
the
robot
playing
chess
with
me
that
makes
a
move
to
checkmate
me
and
sits
back
with
its
500
megawatt
center
behind
it
to
say
we're
we're
supreme.
It's
gonna
be
as
that
robot
is
making
the
checkmate
move,
we
were
all
pretty
clever
about
how
we
can
enlist
the
500
megawatt
machines
to
learn
how
to
achieve
neuroplasticity
to
enable
our
20-watt
brain
to
achieve
the
same
computational
power
and
retention
ability
as
an
LLM.
Like
we're
gonna,
those
models
can
help
train
us
and
construct
those
maps
to
make
those
kinds
of
discovery
where
we'll
we'll
we'll
make
that
obsolete.
We'll
make
the
500
megawatt
data
center
obsolete
because
we'll
figure
out
how
our
20-watt
brain
can
can
do
it
instead.
So
I
have
that
kind
of
optimism
that
I
just
feel
like
the
level
of
innovation,
like
right
now,
we're
at
this
inflection
point,
it's
disruptive,
it's
overwhelming,
it's
like
happening
so
fast.
So
there's,
you
know,
there's
gonna
be
a
lot
to
work
out.
But
I
think
at
the
end
of
the
day,
we're
gonna,
you
know,
our
ability
to
connect
complex
array
of
dots
for
what
the
next
moves
for
uh
can
be
and
what's
the
next
level
of
innovation,
and
how
do
we
enable,
you
know,
there
was
just
a
company.
Um
who
was
it?
That
was
just
Flourish.
A
company
just
raised
a
crazy
amount
of
money.
I
want
to
say
$500
million
as
a
out
of
the
gate.
That's
kind
of
focused
on
how
do
we
uncover
what
the
GPU
20-watt
brain
is
doing
that
matches
hundreds
of
megawatt
data
centers
and
these
LLMs
to
do
the
same
thing.
Like
this
is
an
amazing
learning
engine.
If
we
can
harness
that,
like
there
can
be
that
kind
of
transformation.
So
you're
already
seeing
like
that
level
of
of
innovation
and
that
kind
of
trajectory
that
I'm
speaking
to.
And
I'm
I'm
pretty
confident
we'll
figure
it
out.
Eleanor Howe
So
you
you're
you're
you're
in
the
camp
that
we're
actually
all
going
to
get
smarter
as
people
are
smarter
and
more
capable
because
of
these
tools.
Eric Schadt
Yeah,
that
we'll
be
we'll
be
able
to
pick
up
the
abilities
that
come
to
advance
ourselves
that
start
to
rival
what
are
now
only
done
in
in
the
machines.
Yeah.
Eleanor Howe
And
do
you
think
that
also
what
I
think
some
of
what
I
heard
from
your
description
is
it
sounds
like
you're
thinking
that
humans,
all
of
us
humans
are
gonna
end
up
stronger
leaders,
I
think,
is
what
I
think
you
described.
The
the
things
you
saw
us
doing
in
this
new
world
sounded
like
leadership
to
me.
Does
that
sound
is
that
is
that
accurate
to
what
you
said?
Eric Schadt
Yeah,
all
yeah,
I
don't
know
if
I've
thought
about
it
that
way
before,
but
that
could
be
a
derivative.
I
would
say
it's
more
like
we'll
better
democratize
access
to
extreme
capability
that
will
in
know-how,
and
that
there'll
still
be
this
great
heterogeneity
and
diversity
because
what
we
do
with
that
more
extreme
capability
will
be
to
some
extent
defined
by
geography,
by
cultural
context,
and
so
on.
Like
the
contextualization
will
matter.
And
so
we'll
still
see
the
same
level
of
you
know,
interesting
heterogeneity
that
I
think
we
all
love
about
the
about
humans
and
the
world
we
live
in,
but
just
at
a
at
a
like
raising
the
level,
like
instead
of
relying
on
a
machine
to
direct
us,
that
capability
will
be
more
internal.
Eleanor Howe
Okay.
Eleanor Howe
Then
on
a
more
personal
note,
I
wanted
to
ask
about
your
your
career.
You've
worn
a
lot
of
hats.
You
know,
you've
been
a
scientist,
a
professor,
a
CEO,
a
CSO,
a
co-founder,
probably
about
six
other
things.
What's
what's
your
favorite
one?
What
do
you
like
the
best?
Eric Schadt
Yeah,
well,
I
would
say
number
one,
I
feel
like
incredibly
fortunate.
Like
I
wake
up
nearly
every
day
very,
very
thankful
for
what
I've
been
able
to
do
and
the
opportunity
of
the
context
I
was
born
into
in
the
US
at
the
time,
which
just
afforded
a
person
who
had
nothing
to
go
and
figure
things
out
and
have
access
to
lots.
So
I
would
say,
you
know,
compared
to
many
other
parts
of
the
world,
like
we
have
it
very
fortunate.
So,
like
appreciating
that,
I
would
say
I've
loved
like
my
trajectory
a
lot,
like
the
ability
to
be
able
to
focus
hardcore
on
the
science
and
solving
highly
interesting
problems
and
being
exposed
to
highly
interesting
people
all
trying
to
solve
interesting
problems
and
how
do
you
connect
the
dots?
Like
so
intellectual
reward
intellectually
rewarding.
But
there's
something
to
be
said
about
being
able
to
translate
that
kind
of
learning
in
scientific
discovery
and
research
into
practice,
like
that
sort
of
translation
I've
always
been
super
keenly
interested
in.
And
you
know,
because
otherwise
we're
just
publishing
a
bunch
of
papers
and
nobody's
benefiting.
And
my
big
eras
have
been
those
who
can
take,
you
know,
from
the
attention
is
everything
paper
with
the
Transformers
to
now
OpenAI
and
all
these
other
Frontier
Lab
companies.
Like
that
to
me
is
about
the
impact.
So
I
always
wanted
to
dedicate
myself
to
helping
translate
from
the
research
and
discovery
arena
very
cool
ideas
and
solutions
into
practice.
And
I
would
say
that's
what
I've
enjoyed
the
most
is
being
that
kind
of
builder,
being
able
to
take
ideas
and
early
solutions
and
hire
the
right
groups
of
people
to
advance
those
solutions
and
then
carry
them
out
in
practice,
where
you're
doing
it
in
the
sustainable
way,
where
you're
generating
revenue
and
come
up
with
a
model
that
looks
like
it
can
survive.
Like
figuring
that
all
out
as
the
builder,
I
think
is
what
I've
enjoyed
the
most.
And
sometimes
you
are
able
to
do
that
as
the
CEO,
which
I
always
enjoy
being
the
CEO,
as
opposed
to
having
that
kind
of
a
boss.
But
other
times
you
can
do
it
as
a
CSO
and
founder
and
have
the
same
impact.
And
what
I
know
is
once
you
once
that
starts
to
like
the
level
of
building
like
starts
to
plateau
or
starts
to
becoming
more
about
maintenance,
then
I
know
it's
time
for
me
to
there
are
other
people
who
are
way
better
at
that
job
than
me.
And
then
I
should
move
on
and
go
be
interested
in
the
next
cool
thing
and
help
build
something
towards
that
end.
Like
I
so
I
think
it's
that
kind
of
entrepreneurial
cycle
that
to
me
is
very
exciting.
Eleanor Howe
And
that
is
it
for
The
Trenches.
Unless
you
had
some
closing
remarks,
I
think
we'll
call
it
a
day
here.
And
thank
you
for
joining
me
today.
Eric Schadt
Yeah,
well,
thanks
for
the
time,
Eleanor,
and
you
know,
for
the
you
know,
having
the
opportunity
to
talk
about
all
these
cool
advances
and
kind
of
happy
to
be
in
the
mix
with
everybody
else,
both
trying
to
figure
out
where
do
we
fit
in
this
new
order,
but
importantly,
again,
being
very
purposeful
in
wanting
to
be
involved
in
that
solution,
to
come
up
with
solutions
and
a
trajectory
that
leads
to
kind
of
a
better
situation
for
everybody,
as
opposed
to
just
letting
it
happen
with
no
purpose
driven.
And
I
just
guarantee
you,
when
that
happens,
we're
not
none
of
us
are
gonna
get
what
we
want.
Eleanor Howe
So,
yeah,
that
sounds
perfect.
All
right,
thank
you
so
much.