From the MIT GenAI Summit: Building Strong Business with Generative AI

thank you all so much for joining
um this is going to be an amazing panel
hard to follow what we just listened to
but uh we're very excited we have four
amazing people we're going to talk about
how to build strong businesses with
generative AI
um I'll just start by introducing our
amazing panelists uh we have Paul here
Paul karamov is a partner software
architect at Microsoft he's working
directly on integrating large language
models primarily into the Microsoft
Office Suite so Paul has touched you
know products that have millions of
users and he's really at the Forefront
of some of the implementation around gen
AI so welcome Paul thank you for coming
next up we got Christian Christian cos
is uh amazing he founded and he's been
leading the generative AI for ads group
at meta and his amazing team you know is
comprised of Engineers scientists
working on multiple modes of gen AI text
video image audio multi modal to invent
the future of digital advertising but
most importantly Christian really wants
you to know that he doesn't host an
award-winning podcast he doesn't run
marathons in his spare time and nobody
asked him to be on the Forbes 30 under
30. foreign
Delphine is up next Delphine zukia is a
senior partner at McKinsey and Company
here in Boston she's one of the leaders
and I think recently leading now the
generative AI practice for uh Health
Tech I think if that's correct
um so Delphine has a sort of Life
Sciences Health Tech med tech
specialization and she's got a lot of
generative AI experience she's been
counseling companies in that space she's
done an AIML deployment in the space and
she's an MIT Alum so welcome back
Delphine
and last but not least we have Armin
mccridgeon who is a senior principal at
Flagship pioneering where he's leading
the pioneering intelligence initiative
so Armin is expanding the use of AI in
portfolio companies and he's applying AI
to healthcare biotech and a lot more and
most of you may know Flagship as the the
place where moderna was born so
hopefully we get to hear something
around the flagship model in a little
bit so Armin gets to see startups at
their earliest stages and we're really
excited to touch on some of those topics
here welcome Armin my name is Sid sriram
I'm a second year MBA student also part
of the organizing committee for the
summit great to see this turnout thank
you all for joining uh so we're going to
talk today about
a lot of wonderful things but over the
last six to 12 months a lot has changed
in the world of generative AI we've had
some amazing Technologies come to the
Forefront chatgpt Dolly but today we're
really going to discuss how to build
strong businesses out of those
Technologies uh we'll talk about what
success looks like what the risks are
we'll talk about startups big Tech and a
lot more so I'd like to start by asking
each of you
how the last six to 12 months have kind
of changed in the context of your AI
work so Paul can you kick us off
it's 12 months I feel like there was
like a Cambrian explosion of players in
the space everybody started racing
towards uh you know productizing open AI
models I work a lot with GPT family
models so that's what I think about
first and foremost but lots and lots of
startups as well as now large companies
getting into the game and it feels like
uh like a large race and then in terms
of just my own work we we have products
that have tens and some cases hundreds
of millions of users and
understanding how to impact those users
in positive ways becomes really
interesting and it's not uh you know
it's not a straightforward uh solution
to that so it's a huge race it's it
feels like it's going pretty quick I'm
very excited for what's going to happen
over the next six to 12 months I think
there's going to be uh some more really
interesting announcements and
revolutionary advances both in science
and in the business applications of it
thank you Paul go ahead Christian
uh well Kimber and explosion is such a
good I'm going to steal that uh that's
that's that's very good
um I I think three things have changed
uh for me the first one is just the
sheer amount of attention that this area
is getting
um you know I spent about three years at
meta evangelizing generative Ai and one
of the things that you realize when
you're kind of in this space is how
um this room is above average nerdy and
most people that you talk to uh
potential customers they're not and so
there's been a lot of fear and
trepidation and uncertainty about oh you
know is AI mature enough
um is this really the you know what we
should be investing in uh chat GPT and
stable diffusion those two things have
really put AI out there and created far
more attention even inside Facebook for
this area than I could have ever done
alone though I will claim the credit
later on and
the the second thing that's changed is I
think the the speed at which certain
problems that I expected to be hard have
been solved in large part thanks to the
open source Community things like how do
you do photorealistic uh Renditions of
people with a stable diffusion type
model how do you preserve the
proportions of a product very important
in ads
um how do you preserve perspective when
you're doing sort of digital digital ads
all these things turned out to be solved
by the thousands of people that that are
working on these things in the open
source Community now
and I think finally the third thing
that's changed is just the huge amount
of fomo that's happening in other
Industries outside of my my modest
little field where banks are thinking
about this life sciences are thinking
about this and you know maybe it was
myopic on my part but I never thought of
the field of application being quite
that large
yeah it's very interesting and I think
probably Delphine is going to touch on a
little bit of that in a bit but go ahead
Delphine about how your life has changed
over the last few years so actually I'll
start on the personal front you
mentioned I was I was an undergrad here
uh way back when uh when you could
actually draw draw a neural network on a
piece of paper and I try to predict the
stock market at the time and it didn't
quite work and I was very frustrated so
I moved on to Medical Imaging and I was
just incredibly excited about what it
could do for Physicians but I I realized
AI is just not ready to be adopted which
is you know one reason I went into
business
and what I've noticed over the last year
is that finally the business world is
willing to hear about AI in a way that
it's not just a cool technology but
there are all kinds of questions and
Eric
um mentioned them around the adoption of
AI and how you should have some design
thinking around it and so you know I
used to have a lot of AI conversations
with cios and heads of r d and now you
have them with board members now you
have them with CEOs who actually want to
understand how it works which you know
makes it very fun for and I know many of
you are excited to talk about technology
and finally we have now a platform but
there's a lot of questions and I think
you know for us it's very important to
think about the responsibility of making
sure it's being rolled out the right way
links and Armin go ahead great hi
everyone Armin mccirchen it's a pleasure
being here I was telling Delphine that I
think last time I was here probably was
12 years ago and I'm still taking
classes in this in this building so it's
a pleasure being back
um I think I'm going to Echo a lot of
the things that I think my co-panelists
said but uh just to kind of maybe
emphasize a few things I think the level
of excitement probably about AI probably
has been the highest I've seen in my
lifetime
maybe I'm just going to use an analogy
maybe on all of you since I have a
four-year-old he was asking me today
about Orchestra and positioning of the
different instruments
um if you if you for a second imagine
that the in an orchestra all of the
different instruments are the tools that
companies have digital tools to address
different problems uh maybe for an
average company AI would be French horns
nothing against French or horns but
what's happening now I think is a lot of
companies are actually either just
bringing too many French horns
they may not need it but they are or
pushing some of his French horns to the
front of the line even though violins
are supposed to be front of the line
again nothing nothing bad about it
except that's I think what's also
happening at the same time I think just
the kind of the level of excitement also
is lowering the barrier for some of the
companies to adopt ml tools well they
generally probably would not adopt those
tools which is a great thing and then
and then the last thing is I think there
are problems as maybe Delfino saying
there are problems that at least in life
sciences now we are seeing what we can
solve that before we'd not even think of
tackling and mostly due to generative AI
I think those are probably the three
things that at least we have been seeing
in the last six to 12 months Christian
and I are French so this is going to be
about friends not China I I just learned
about this French Roots common French
Roots yesterday so if they start
spontaneously French people in Disguise
I think we're artificial intelligence
was invented right yeah
wonderful
um so Christian if if you can I think
everybody
um is kind of wondering this so and
curious about this can you can you tell
us about a couple of uh maybe one or two
cool products or features
um that your team has been working on
um and you know what kind of Market
needed solving and why your team decided
to work on that thank you uh that's uh
that's a that's a big question but and I
have to be a little bit cagey in how I
answer it for for obvious reasons but
let me let me maybe start by what are
the big problems in digital advertising
today and sort of how generative AI is
going to lead us but also other
companies like Google Amazon tiktok
probably to want to solve the same
problems there's essentially two things
um that are really really hard in
digital ads today one is a purely
workload kind of problem and the other
one is how do you personalize digital
advertising and they're they're largely
intertwined today if you ask any
marketer or any creative strategist what
one of their major pain points are they
will tell you it's just the sheer amount
of stuff that I need to create because
you and I will think
um I want to advertise a bottle of water
so I'll just take a picture and sort of
throw that out there at no because in
fact every time you want to create one
ad you need to create different versions
of it for you know Facebook for
Instagram for Instagram stories Facebook
feed different things like that so in
fact it's never just one asset it's
something like 47. and then your manager
reads an article on LinkedIn about a b
testing so now it's like oh well now
it's 47 times 2 because we want to make
sure that we have the best one
um and then you learn that oh but ads
have a very short lifetime so you need
to redo everything three times a month
and then by the way it's not just
Facebook you're also on Tick Tock you're
also on all these other things so your
life is horrible and generative AI is
um
generative AI the promise of it is
really the ability to sort of do what
what uh
sort of in a Java kind of way you know
you make it once and then you kind of
run it anywhere because we can do things
like General generative outpainting we
can create variations on a theme
automatically all of that is what we're
building towards in the very long run
what you want to do is create ads that
are not only liquid across these
different systems but also ads that are
personalized to different people so
obviously all these large companies
including meta but also Google know a
lot about us as consumers so how do I
use that knowledge to generate an ad
that is pertinent to you right off the
bat and so in the very long term those
are the kinds of problems that you
should expect these companies to solve
yeah that's that's a really interesting
sort of uh goal or Target but it helps
us move faster but I think
there's there's a little bit more of
sort of risk when it comes to bias when
you have these sort of higher Stakes
um things when you get personalized ads
and you're putting a little bit of trust
in sort of the machine so how do you
weigh risk versus
um sort of customer engagement or speed
to Market or some of those
uh I'm glad you used the word trust
because I I think that's kind of the big
risk for me as a product person
obviously in generative Ai and AI in
general there's different kinds of risk
you know legal societal uh technical
risk and there's frankly there's better
people that talk about those than than
me so you know from a product
perspective when you're in AI it's the
biggest trade-off is between your own
um the speed of your ambition
and the level of trust that you're that
your consumer that your customer sorry
has in you and again this is one of
these situations where we're all above
average nerdy so we kind of like Ai and
we're you know we're eager to try it
most people are not like us most people
uh you know don't particularly trust AI
and if you work depending on who your
employer is sometimes you'll discover
your your customer doesn't particularly
trust you
um and so when you're faced with that
you know the trade-off is well I can't
show up with the solution that sort of
says look I'm going to automate and make
your life perfect because nobody will
buy that and that you know is a recipe
for failure
um it's very much how do you adopt a a
an assist augment automate kind of
framework to your product where you
begin by earning the right to speak to
your customer by saying let me find the
the sort of boring tactical but but
painful parts of your job and let me use
AI to kind of assist you in that process
over time you were in the right to talk
about how do I augment the productivity
uh of of your job and and in the very
long run you know maybe you'll get to
the point where you can say well look
let me automate this for you but the
biggest risk is the loss of trust and if
you lose the trust of your consumer then
it's it's really really hard to get back
yeah yeah that's uh it's really
interesting and Delphine I'm sure you
are seeing companies that are on a much
sort of broader scale outside the tech
industry perhaps trying to understand
what these Technologies are what
generally I even is and how they can
implement it how do you and if you if
you take say medical as an example or
biotechnology risk is really important
there and assessing trust and risk so
how do you sort of convince the people
that you counsel that the application is
safe while still
bringing value
yeah and Healthcare is an area that's
highly regulated so risk is obviously
one that sucks it's an important
conversation
you know one thing I have to say about
Healthcare and life sciences and Armin
will will refer to that as well I'm sure
is that generative AI if you
if you define it more as large
pre-trained models the Transformer
models that are the underlying
technology that's been around for a
while actually and um
and it's been around where there's
already workflows that exist where
whatever it's generating can actually be
caught by a scientist and evaluated and
then they decide you know what to do
with it so an example is drug Discovery
you know these models are known for
large language models but actually you
can do very exciting things treating a
molecule like a sentence and or
molecular compound and then you know
words or different molecules what it
allows humans to do is to essentially do
something that organic chemist was not
able to do which is understand the
structure the relationship between the
molecules
um and that's been used now for you know
at least three four three to four years
by large pharmaceutical companies to
significantly accelerate drug
development only about 12 percent of
what starts in the lab actually makes it
to be FDA approved so it's a lot of cost
and unfortunately a lot of unmet need
um that area there's less risk
conversations now we're getting to more
the chat bot craze uh which is you know
initially people think oh great we're
going to start communicating with our
hcps with you know AI generated emails
and of course the risk Department says
well hang on a second are you going to
read that email before it gets sent so
these are the kinds of you know I'd say
fun conversations that are happening but
my my prediction is where there is a
workflow and there is good design
thinking you're going to see it take off
like we have where it's you have to
invent new workflows especially risk
needing to look at things I think that's
where you're going to see a little bit
of a Slowdown yeah that's that's super
interesting um Armin you know feel free
to continue pulling on that thread but
I'm really curious if you can just kind
of explain the the flagship process to
start with just really briefly so that
everybody understands kind of how
startups are born from Technologies at
Flagship and how you're implementing
that or how it might have changed for
generative AI
thank you thanks for giving me an
opportunity to talk about Flagship so
Flagship was founded in about 2000 it's
a company that creates companies and by
creating it starts from scratch it
starts from literally from ideation from
exploration uh we don't necessarily go
find IP and say hey come actually we're
going to incubate we're going to
accelerate your development we don't do
that
we have mostly phds actually and then
this we work on ideation and coming up
with about 150 Explorations every year
six to eight of those converting to
companies we call them protocols these
are not real companies protocols stands
for product companies so we start before
they become real companies we are
product companies and we have a very
well defined process of managing those
protocols after about a year some of
those protocols will become new cost
many new companies and then become
growth costs meaning growth companies
the idea in our case is that if we trust
the process if we have enough people who
know what they're doing greatest ideas
will emerge we don't necessarily have to
start with the greatest idea
in the very beginning it's not that it's
the one day I'm going to wake up and I'm
going to have the best idea ever but the
best idea over a period of time will
emerge and once it emerges we'll know
what that idea is and I can take it
forward
so then afterwards we do invest Capital
we actually will hire CEO Partners to
actually run our companies uh who are
going to be Flagship by employees in
addition to obviously running the
company as well
now relate to generative AI uh about a
year ago we started this initiative and
I'm heading now called partnering
intelligence the idea is actually to
build Central capability in AI at
Flagship to help our companies but also
start new companies that are more
generative
uh in terms of the kind of using the AI
computational tools and techniques
we are also at the same time I think
Eric was mentioning about kind of maybe
associations and causations when talking
about AI so a lot of time I think what
we do is uh as probably everyone else
uses uh AI in terms of defining what our
causal what are associative and try to
project them forward in our case as
delfino's eluding there is a lot of
workers that has been done in Pharma in
figuring out how we can actually use
generative models in drug Discovery for
example now we can use this diffusion
models that Dal imagery stable
diffusional use to actually create
completely new proteins that the world
has never seen for example
and we have a company called generate
biomedicines that has shown it actually
published the paper it's available for
everyone to go and have a look at at the
same time something that we leverage a
ton I think many people probably don't
yet but we actually leveraged the
hallucination aspect of these large
language models uh partly because it's
very useful for imagination so these are
things that probably are not true but
these are things that actually can force
us to imagine much further when we
ourselves could sit at our desk and
imagine so we are we are very
intentionally prompting these models to
hallucinate and hallucinate a lot more
than in many cases they are actually
allowed to and then we are leveraging it
in our company creation company ideation
process as well
yeah and you know you have these uh sort
of experiments or or prototypes at the
at the initial stages but do you have a
process in place that is thinking about
how these I mean I'm sure you do is how
these actually become viable profitable
businesses and uh can you tell us a
little bit more about how it's kind of
like a staged process and how you
eliminate the ones that you think might
not work and move on the ones that do
um and they result in successes like
moderna yeah so it's initially we in the
very beginning we we don't focus on the
business models we don't focus on
necessarily on the market opportunity a
lot of it is driven by science and
technology meaning can we actually
enable vet Science and Technology
assuming that if we are able to do it
then we'll probably hopefully find ways
to also commercialize it it's not 100
always true but I think if you start
from the very beginning trying to
actually run due diligence on a market
that doesn't exist it's not necessarily
the most uh
productive place to to spend your time
on so we just start mostly with the with
the idea and with with the technology
what happens over time and see though is
is that as the idea grows when we
started actually narrowing down and
trying to figure out what is the market
we are going to go after how we're going
to go after the market what are the new
business models that potentially we
actually need to think about to go after
those markets but it mostly happens it
mostly emerges rather than us having a
stage gate process saying hey now it's
time January 1st to talk about the
business model or talk about the
commercial uh viability of this it just
happens over time yeah I like that
fluidity
um great thank you so so Paul
um Armin has talked quite a bit about
how startups are born from from these
Technologies and from generative AI
specifically how do things work in the
big text like Microsoft
[Music]
um
can you talk talk us through a little
bit about uh building your own large
language models versus uh integrating
existing ones into say the Microsoft
Office Suite how you weigh those
trade-offs
um and if you if you're able to um I'm
interested in a discussion around the
different modalities of generative AI so
text image video audio whatever you can
share there that would be interesting as
well
yeah sure um so there's uh when you work
at a large company uh there is uh good
and bad in terms of support for
Investments uh such as you know large
language models getting them into the
products
some advantages are there's lots of
people so lots of ideas and so you end
up benefiting from effective exploration
of the possible solution space of the
different applications of these models
of the stack and you don't have to
figure out all of it yourself you can
talk to others you learn what they have
done you apply it you experiment your
own you contribute back to this
community and that's a fantastic
advantage
another Advantage is that uh just
there's just a lot of supporting
structure right there you need to do
user research there user researchers you
need to have a to build a user base to
do the experimentation well chances are
there are already products that have
pretty large user bases and you can do
that experimentation uh so so that is
awesome
um at the same time the challenge is
that well you have large user bases and
large products and there's a lot of
stakeholders and you know unlike in a
startup it's a little bit more difficult
to make sure everybody's got the buy-off
everybody's comfortable with the risk
you know we talk a little bit about
hallucination ethical aspects of these
things and uh you know there's a lot of
users and not everybody's comfortable
with those you can't just throw things
over the fence and see what works what
does it you have to apply a little bit
more control a little bit more process
um
in terms of what we do
um
you know you asked about training your
own large language models it is not a
small investment right so training of
this model takes a lot of money uh if
you want power and so that's one of
those things where it's there but you
have to get enough of a buy off for that
to actually have but you can
um I think that it is a little bit of a
question of where you are in that
Journey uh because anybody can create a
I'm going to call it a vanity startup on
top of open AI apis from the comfort of
their home office in probably 30 minutes
throw it out there see what happens it
is probably not going to be super
interesting so you have to figure out
what is the unique value that you're
going to bring and then in case of uh
Microsoft Office in case of large
products how does it apply to everybody
who is using your product because you
have different user bases from edu to
from education to Enterprises to
governments consumers
Etc once you figure out that maybe you
find an application for okay now I gotta
train my own and training is a load
right it could be your distilling could
be your fine tuning for a specific
application things like that
um and I think maybe after that you're
going to okay now I need my own like
really large language models then you're
in the business of competing with open
AI or different things and that's I I
don't know that we're there just at the
moment different modalities become
really interesting we talk a lot about
text there's obviously a lot of
investments into images into videos into
voice I think real power comes when
there's a combined when you create a
document in Word text is nice it pops
when you add images right you work in
teams
um you might again text is nice images
are great but the large component of
these videos and augmenting that with AI
becomes really cool so I feel like
there's a lot of power in the
combination of the different modalities
uh and I'm very excited to see what's
going to happen with that Christian do
you do you sort of agree with that
because I think you've mentioned before
that you think that text will sort of
reign supreme and but you've also
mentioned that your team has worked on
these multimodal models how do you see
that changing or evolving towards the
future
uh yes I did mention that that I I
thought text was ultimately where things
where things will go but maybe not for
not for that not for a particularly
obvious reasons so I'll come back to
that
um our team sort of is is forced to work
on multimodal things because that's the
nature of of digital advertising you
know when we use things like Tick Tock
Instagram reels it's a reflection of the
fact that consumers are using these
surfaces therefore you need ads on them
guess what you know Tick Tock is a video
first platform Instagram reels is a
video first platform therefore if you
want to solve The Advertiser problems
that I mentioned earlier then you know
what now we have to do video and guess
what video is the multimodal format by
definition it's got visual aspects it
got audio it has content which is
usually text and so you're kind of
dragged into this whether or not you
want to do that
um uh
the reason why I think that text is
ultimately a more interesting modality
doesn't necessarily have to do with that
it has to do with the fact that and that
if you mentioned this
um you can text doesn't just mean this
is human understandable words what it
actually means is that there's many
things that you can abstractly call a
language that large language models can
learn the grammar of implicitly and a
language can be anything it can be
English it can be it can be French it
can also be JavaScript it can be python
it can be protein sequences it could be
a building anything that you can sort of
constrain in in a way where you can
express it as code is learnable by a
large language model which really means
that text shouldn't be thought of as
text it should be thought of as a
convenient way to represent reality and
therefore to generate reality
um and and that's why I think that
ultimately the the large language model
research is more compelling but not
necessarily in the context of ads got it
and Paul you touched a little bit on
um high costs
um so for the the first question is do
you think that things will remain that
way for a while and secondly because of
these high costs now do you think that
the bigger companies like Microsoft have
sort of an advantage over startups when
trying to get to the Forefront and do
you think there will be like this Rift
that forms where the bigger companies
rise up
uh I I don't think cost will stay the
same I think everything we've we've had
in the computer industry for the last 50
more years has pointed to cost go down
all the time things get democratized
things get cheaper new technologies new
approaches all of that is going to
happen uh at the same time
we also know that every new technology
initially comes with high cost and so
just by nature of that companies with
deeper Pockets have a little bit more of
an advantage over just a startup
but startups have a history of
disruption they're going to happen
they're going to be the ones bringing
some of these new technologies
monetizing it and so I don't think it'll
be that the Big Deck always has this big
Advantage nobody else can touch it I
don't think it's going to calcify I
think we will continue to see the cycle
of innovation bringing us down
disruption new Tech High cost Innovation
bringing costs down
great
um
I wanna I wanna also ask you a little
bit about
metrics and measuring success so you're
working on several products presumably
across the office suite and some of
these products will not maybe see the
light of day what are some of these
metrics that you're measuring on to to
ensure that products become successful
Revenue lines or revenue streams for
let's say Microsoft
yeah great question so
in in our world a lot of our products
have been around for quite some time if
you think about Microsoft Word Microsoft
Outlook Etc gosh tens of years I think
ward has just celebrated its 40th
anniversary that seems a long time
um so it's like it is a successful
product it is you can argue it could be
better it could be worse it has more
user few years it is a successful
product I think a lot of what we talk
about is applications of this deck like
the way we're thinking about them today
the features we're building are they
going to be successful are they going to
be less successful and in terms of
metrics to judging it uh there are
obviously some sort of technical metrics
how fast are things right
um how much hallucination exists in this
given model output do we generate the um
sort of the right amount of tax do we
cover all the languages all that all
that is good
um but in terms of business success you
have to look at things that are
disassociated from Tech you have to look
at things like user engagement uh you
want to if you think about money and
monetizing you want to think about your
top of the funnel how many users are
attracting how much retention you're
driving and ultimately Tech any Tech
including large language models is in
service of that so whenever you're
building an application you have to
disassociate from Tech and you have to
look at your business goals your
objectives and make sure that the tech
Choice you're making the Investments
you're doing are aligned to those
business goals yeah that's that's really
a good point uh and I think you said
um sort of user engagement is an
important one Delphine I think you had
an AIML deployment that like a med tech
company recently and one of the key
metrics was I think customer engagement
or user engagement
um can you can tell us a little bit
about that but also I'm curious to hear
about how you put these metrics in place
when things are much broader and you're
touching a lot of different companies
so I think one thing that's important to
note is the deployment of AI is is not
new and you know gen AI is just a class
of AI and so is that there's some I'd
say truth that we've been finding over
the last 10 years of doing these these
deployments this is the Quantum black
arm of McKinsey uh you know first of all
it's a bit of a sobering fact but only
release
um
28 of what you would call a digital
transformation which is just being able
to put an analytical product out there
succeed
so this other 72 percent they get caught
in Pilot mode they get caught where the
users actually don't really want the
technology they get caught because it's
a great technology but it can't scale
you don't have the pipes to give you the
data for example
so a lot of time we spend up front is
educating Executives around that because
if you are excited and if you invest up
front you need to know that it is going
to be a journey
um now that said we absolutely encourage
folks to think of the art of the
possible and experiment because it's
very cheap up front and the metrics are
really no different than what you would
use for business I mean this is what
Paul was saying it's it's really I mean
at the end of the day it's your margin
profile many companies are willing to
trade that off and and get Revenue first
before and as they figure out the cost
because users are are quite important
um and then yes you you measure along
the way adoption metrics that are going
to give you some signal that the users
like like to engage with uh with your
product
what is encouraging to me with
generative AI I would say is you know if
you look at sort of out of these
transformation it's only about 15
percent of the effort is on data science
and then if you add data engineering
it's probably another 20 well that's
still 30 we can cut
um if we have some of these models
already available and if we can
democratize who's able to Tinker with
them now unfortunately the other 70
percent that is all the various jobs
that Eric was talking about this is risk
this is Design This is the business
owners who also need to be around and
and take it through so no different but
I think now finally uh everybody is
starting to ask the question how do I
deploy AI whereas before it was more
like what is AI and not even
understanding the journey
yeah that's that's really interesting
and Armin I'm curious
um about your thoughts on this 28
success metric which um which Delphine
just said because that's not what I
would have expected uh and also when
when you're
when your Technologies reach the stage
where you think that they will become
successful businesses
um how are you measuring that so it's
kind of a similar question
you know interesting I think 28 I
probably knew partly because I also
worked at uh delphin's firm in the past
uh but it's a I feel like a lot of it is
at least from what I've seen has to do
with cultural issues as well it's how
much kind of people buy into it how much
people kind of daily daily Drive the
adoption of it
um
I was just going to maybe comment very
quickly on the metrics as well and then
I'll get to your question
um it's interesting because I was
thinking uh at least in biology how do
we measure actually success of
generative Ai and in a way it's probably
no different than anything else at the
end of the day we are making we are
creating proteins molecules at the end
they are gonna take them and test them
you've got to have a wet lab to go and
see duvet bind the very dark or not
what's Affinity or not so it's it's not
necessarily very different except now
you can actually accelerate that process
and at least the generation process can
be so much faster than it could have
been before
so at least I think for us it probably
metric wise it doesn't change I think
exactly how do things now in terms of
question about kind of our companies and
as they grow kind of how we think about
them
it's a
it's probably a multi-layered uh at
least
solution that we try to I think put in
place one from from uh from technology
perspective we always try to think
um
all these Technologies actually or are
the companies that we are creating do
they actually need AI generative AI or
not a lot of the time we don't want our
companies to just kind of go surf in the
kind of generative AI hype if they don't
have a surfboard they don't know
actually how to stand on the surfboard
yet so it's uh sometimes it's it's very
easy to say oh so I'm going to use this
I've seen this it's like a hammer I'm
going to use but you don't need a hammer
you actually need the right tool that's
not a hammer and that's very that's what
I think a lot of times centrif will
actually provides to our companies
partly because we have just seen so much
more than every specific company has
seen and the second is actually
uh
kind of maybe if we increase the 20 28
that Delfino is saying hiring the right
people into the companies very early on
uh one thing maybe culturally I think we
have noticed is and this is probably
this is not Flagship specific I think
this is generally uh probably true is
some of the first people you hire in ml
in digital in AI is probably going to
define the trajectory that you are going
to have in that space it's interesting
from what I've seen personally before
even Flagship it's people don't like to
hire people that are that seem to be
smarter than they are in many cases it's
interesting observation my own but if
you if you try to hire people actually
create a collaborative environment where
they know where to take the company that
helps a lot with the transformation in
the future and hopefully 28 won't be 28
but will be more
yeah I think a lot of people have been
saying that um you know the technology
is of course a challenge and a problem
but there are a lot of other peripheral
things
um that really make it a success and one
of them is uh people and and hiring like
you said
um I I also thought
um it was a yeah a couple of interesting
points there around uh how you how you
launch the businesses out of there it's
it's pretty uh intriguing uh so I just
want to ask each of you a sort of
interesting question which I think the
audience will appreciate
um is there something surprising in the
world of gender generative AI that you
expect to take place over the next let's
say couple of years
Paul you want to go ahead
yeah
um I I think the thing that we'll see is
that it is going to move a lot faster
than we predict
um I think um you know when Eric was
talking he was talking on the horizons
of five to ten years now that's not just
generative AI they're just the AIO lab
but I think it's going to go extremely
quick look at the journey we've gone
even in the pure technical level in
terms of model sizes the number of hyper
parameters
um just even in the last six to 12
months uh and every doubling of that
represents an exponential Improvement on
the quality uh and I think so I think it
is going to surprise us how quickly this
will move and it sounds like you're
excited about the speed I yes yes
excited and a little bit worried because
uh keeping up with that is not easy it's
almost by the time you are done
releasing something it's outdated so you
got to go really quick yeah that was
going to be my question is if you have a
concern around it
um great Christian uh similar question
you know something that unexpected and
maybe a concern around it
you ask me if I expect the unexpected
which which is a neat question
um I the the one that excites me
intellectually is this notion of
emergent abilities of large language
models in particular so as Eric pointed
out the we don't entirely understand why
these models do what they do and neat
stuff seems to happen when the models
get complex enough and I sort of hand
wave when I say complex because it can
be a number of parameters it can be
training data size it could be compute
that you throw at it but they develop
these surprising abilities like the
ability to do arithmetic the ability to
do inductive logic there's there's
certain tests like that at a certain
point these things emerge for reasons
that we don't really understand and we
can't enumerate what these abilities are
um certainly some of them haven't been
codified yet you see similar things
although to a lesser extent in image
generation models that develop the
ability to do background segmentation
automatically and so I think building on
pulse Point things are going to move so
much faster but what's really exciting
to me is what are these models going to
develop the ability to do without having
necessarily been trained to do it and
I'll join you with the idea that this is
very exciting it's also moderately
terrifying because already I feel the
first five years of my career could be
replaced by an llm so now I'm thinking
man this thing's going to catch up to me
pretty soon uh so that's that's a little
bit scary yeah I think this idea of some
people call it AI explainability I think
is one of the things
it's going to get touched on on a few
different panels today it's it's a
really important topic a topic people in
the past before even generative AI have
talked about the black box of deep
learning and
um it's like a convenient term to use
but it's not convenient for long I think
so I agree with that
um Delphine any anything surprising that
you expect
so I'm actually quite excited to see
what our kids and unborn kids are going
to do with AI you know we talk about
digital natives I think we're going to
have ai natives
uh and I actually you know I mean you
guys can resonate here at MIT I was
highly frustrated that MIT wasn't
teaching me things from the textbook
initially and I realized after a while
no they're teaching me how to think and
I never had to memorize anything at MIT
and now the educational world is sort of
trying to come to terms with that but I
think it's very freeing if you don't
have to remember things and instead you
can get a computer to tell you what you
need and then you add and you think on
top of it and so I can already see with
my kids you know the kinds of questions
they asked chai GPT I would have never
thought of it I mean my 11 year old ass
chat GPT should try it are there more
wheels or doors in this world which by
the way he's asked my husband and I a
lot and you know we both being MIT
Engineers we're trying to figure out the
answer but try it out see what chat GPT
is going to say which you know allow me
to have a really interesting
conversation with them around logic by
the way because it was a logical answer
but yet it didn't really use logic I see
you're getting your your children all
ready for the McKinsey interviews you
know we're trying to keep them away from
both of our jobs but it's not working
I I just want to ask on that because it
you you gave a kind of personal example
which is really nice do you see any
concerns or risks around that if you
draw the timeline out to several years
or a couple of decades and generations
I mean I'm an optimist so yes I have
lots of concerns but I think there's
plenty of smart people in this world to
put enough safeguards around technology
so
um I I would say that the concern is if
it's makes education less accessible
which I'm hoping that's not the case but
that's always my concern if there's sort
of if it's not inclusive enough
great great and Armin
um what what surprises you
I think what surprised me in the last
few months is what Christian mentioned
actually is emergence aspect of
terrorist models
in the field that I work in now it's for
example we we took just a regular llm
large language model and with very
slight tweaking it was actually able to
fold proteins like what two years ago
before Alpha fold or ES unfold was
completely impossible to do at the level
of accuracy these models are doing right
it's not that's not something you just
expect out of the boxers models to do
very well but they do and there's again
this is just we are probably scratching
the surface of what's possible with some
of the emergency aspects of it
especially assuming these models are
going to grow in size 10x 100x thousand
decks probably in the next next few
months or years uh what I'm looking
forward to is actually how people are
going to be using
or developing middle layers on top of
his llms and probably providing types of
services that we actually don't imagine
today
and I think it's going to happen very
quickly I think these things are also
going to emerge badly on top of each
other and we are actively thinking what
those are going to be but I don't I
don't think we are the most creative
people so maybe we can do some of these
models to help us they're saying but
that's what I'm looking forward to
trying to see what middle layers are
going to come up to actually help solve
problems but the the regular base models
themselves probably don't yeah I think
this the application layer the
middleware so to speak plays a really
important role and we didn't get to
touch on that much today but um uh I
just want to say uh I just want to ask
everybody to thank our amazing panelists
because we're out of time thank you so
much