00:00:00
you've been an inspiration a lot of the
00:00:01
current AI revolution has to pay credit
00:00:03
to you Mustafa suan Mustafa the CEO of
00:00:07
Microsoft AI CEO of Microsoft AI
00:00:11
co-founder of Deep Mind One of the ogs
00:00:13
of the AI World spent more than a decade
00:00:16
at the Forefront of this industry before
00:00:18
we even had gotten to feel it in the
00:00:21
past couple of years now there's going
00:00:22
to be a point in the future where these
00:00:24
AI agents go out and make money by
00:00:26
themselves just take a moment to think
00:00:28
about how crazy it is that you can ask
00:00:30
that question if anyone can go out there
00:00:32
and just say hey spin me software that
00:00:34
does ABCD what happens to the entire
00:00:37
field of software as a service instead
00:00:39
of just a computer being able to say
00:00:41
things it will actually be able to do
00:00:43
things it'll take actions it'll learn to
00:00:45
use apis it'll buy things it'll write
00:00:48
emails it'll make phone calls just as a
00:00:51
entrepreneur
00:00:53
would ladies and gentlemen today I'm
00:00:56
with Mustafa sulaman who you know to be
00:00:58
honest you've been an inspiration ever
00:01:00
since deep mind and I think uh you know
00:01:03
a lot of the current AI revolution has
00:01:06
to pay credit to you so thank you so
00:01:08
much for doing everything you do and
00:01:10
thank you so much for agreeing to do
00:01:11
this very quick session I'm just going
00:01:13
to start with questions I'm just going
00:01:14
to dive in the first question I had and
00:01:16
I remember reading this one tweet about
00:01:19
you a long time ago where somebody spoke
00:01:21
about your version of the modern during
00:01:23
test so you had spoken about the modern
00:01:25
touring test where you said that there's
00:01:27
going to be a point in the future where
00:01:29
these AI agents
00:01:30
they're sort of able to uh you know go
00:01:33
out and make money by themselves right
00:01:35
without too much human intervention and
00:01:37
we've start to see some of that In
00:01:38
Pockets right there's still some human
00:01:40
intervention like we are a good example
00:01:41
we use avatars on YouTube and Instagram
00:01:44
and that's last month we did close to
00:01:46
about 80 million views right at some
00:01:48
point that's going to be an agentic
00:01:49
process where there need not be humans
00:01:51
involved my question to you is now that
00:01:52
we're seeing some of this evidence what
00:01:54
are your predictions
00:01:56
now first of all just take a moment to
00:01:59
think about how crazy it is that you can
00:02:02
ask that question just just really think
00:02:04
people have spent 50 or 60 years being
00:02:07
inspired by the touring test right and
00:02:10
so I I think many of your viewers will
00:02:12
know what the touring test is but just
00:02:14
for those who don't it is uh a long ago
00:02:18
prediction by the amazing computer
00:02:19
scientist and mathematician Alan churing
00:02:21
who basically said that intelligence is
00:02:24
when you can sort of convince somebody
00:02:26
else that the AI system is a human and
00:02:31
they can't really tell whether it's a
00:02:32
human or whether it's an AI system
00:02:34
because it speaks so fluently right that
00:02:37
was the Turing test the imitation game
00:02:39
and in the last year or so we've kind of
00:02:42
just breezed past that moment right
00:02:45
we've sort of passed the touring test
00:02:47
and nothing seems to be uh you know
00:02:50
nothing seems to have changed that's
00:02:52
really profound um so like with that
00:02:55
said I sort of you know kind of proposed
00:02:57
a different version of that where
00:03:00
um you know instead of just a computer
00:03:02
being able to say things it will
00:03:05
actually be able to do things it'll take
00:03:07
actions right it'll learn to use apis
00:03:10
it'll buy things it'll write emails
00:03:12
it'll make phone calls um just as a
00:03:15
entrepreneur would or an office worker
00:03:18
would and um we're now at the moment
00:03:21
where it hasn't quite approached that
00:03:24
but sometime in the next few years you
00:03:27
know you're right it's probably going to
00:03:28
be the case that an AI could generate
00:03:31
new videos on YouTube um try and promote
00:03:34
them on social media to optimize their
00:03:36
watch time um you know potentially try
00:03:39
and monetize them in different ways
00:03:41
maybe create merchandise for the show
00:03:44
all kinds of things and that is a you
00:03:46
know it is a pretty wild moment to be in
00:03:49
it's it's amazing capabilities that I
00:03:51
think are going to make us all much more
00:03:53
sort of creative how does that change
00:03:55
the economy I'm sure there's going to be
00:03:56
a bunch of smart people that take
00:03:57
advantage of this and a lot of other
00:03:59
people who feel like they're being out
00:04:00
competed Now by agents you had other
00:04:02
humans to compete with in the past now
00:04:03
you also have agents to compete with
00:04:05
yeah I mean you know think about it like
00:04:07
this I mean we've we all sort of adopt
00:04:11
Technologies at different times um but
00:04:14
net net we all end up using televisions
00:04:18
using cars having
00:04:20
smartphones um and you know something
00:04:23
that feels sort of profound and scary
00:04:26
and uncertain and you know suddenly
00:04:29
feels every day and mundane right it's
00:04:32
kind of incredible that you know I can
00:04:34
leave a voice note for my friend and
00:04:36
dispatch that to another part of the
00:04:38
universe it's incredible that I could
00:04:41
have an idea and in two sentences
00:04:45
visualize that idea in an image um and I
00:04:48
I just in general think sort of people
00:04:51
underrate how profound the current
00:04:53
moment is um and we get desensitized to
00:04:58
um you know things that are already
00:05:00
around us and are actually super
00:05:02
significant all the time and so to be
00:05:04
creative In This Moment is actually to
00:05:07
be open-minded and to test and iterate
00:05:11
and absorb and integrate and play with
00:05:14
everything at our disposal and then you
00:05:16
get to see the limitations of it right
00:05:18
because if you don't engage with
00:05:20
something you over idolize it as though
00:05:23
it's like you know so much more than it
00:05:25
actually is but when you really play
00:05:27
with these models you see okay they're
00:05:28
amazing and inspire iring in some ways
00:05:30
but they're also super Limited in some
00:05:32
ways and that gives you an intuition for
00:05:34
how to use them when not to use them how
00:05:37
to mitigate their
00:05:39
weaknesses um and uh you know I think
00:05:42
that's the sort of way to think about
00:05:44
adoption I think one follow-up question
00:05:46
here is a lot of our audience is
00:05:49
software engineers and a lot of the work
00:05:52
we do in our company in the past we of
00:05:55
course wrote code now we're increasingly
00:05:58
seeing us use GitHub copy a bunch of
00:06:00
other tools and starting to see that
00:06:01
we're offloading some of our work there
00:06:03
to the point where it feels like we're
00:06:05
writing English right which is awesome
00:06:09
in the first place but secondly what
00:06:10
does that do to software because if
00:06:12
anyone can go out there and just say hey
00:06:14
spin me software that does ABCD and then
00:06:17
just get the the app or the tool and
00:06:19
they're able to use it what happens to
00:06:21
the entire field of software as a
00:06:23
service yeah it's a great question and
00:06:25
and you know maybe if you look back um
00:06:28
it's kind of easier to see that this has
00:06:30
been the course of software development
00:06:33
for many decades right we've been
00:06:35
creating languages of increasing
00:06:38
abstraction all the way from binary and
00:06:40
Assembly Language up to you know now low
00:06:43
code and no code languages for a long
00:06:45
time and even when you know you write
00:06:47
python you're really calling on a bunch
00:06:49
of existing libraries there's many many
00:06:51
libraries that you're not writing from
00:06:52
scratch every time and um so just as you
00:06:56
strung together a whole series of
00:06:57
different libraries now you're sort of
00:06:59
giving a single instruction to compose
00:07:01
an entire you know subcomponent of a
00:07:03
program or maybe even the entire program
00:07:05
itself so what is that mean it means
00:07:08
that it reduces the barrier to entry to
00:07:11
getting stuff done it makes it easier
00:07:13
for anybody with fewer skills if you
00:07:16
like um to write a piece of software um
00:07:20
create an application design something
00:07:23
see it produced in the real world what
00:07:25
does that mean well that means that we
00:07:27
have more experimentation of different
00:07:29
different ideas faster and so we're
00:07:31
trying out as a species collectively all
00:07:35
the different Paths of possible
00:07:36
combinations of ideas that we could
00:07:39
combine um and that in itself is going
00:07:42
to increase the rate of discovery of new
00:07:45
things new business models new product
00:07:47
applications you know new scientific
00:07:50
ideas which in turn is collectively
00:07:53
going to drive the overall productivity
00:07:55
of our civilization unlike anything
00:07:57
before because after all invention has
00:07:59
been the thing that has driven you know
00:08:01
our species from the beginning of time
00:08:03
to reduce human suffering to find
00:08:05
shelter to make it easier to get food to
00:08:07
reduce conflict and so that trajectory
00:08:09
is about to go exponential another
00:08:11
question here is around you know
00:08:14
cognitive effort right I was talking
00:08:15
about this before the Pod went live
00:08:17
where you I was looking at a really big
00:08:18
number with a lot of commas and instead
00:08:21
of me figuring out what that number was
00:08:23
I just picked it up put it into an llm
00:08:26
and said hey what number is this right I
00:08:27
wanted the number in billions or
00:08:28
trillions or whatever ever and it just
00:08:30
gave me the answer you know that's
00:08:31
something I would have figured out in 3
00:08:33
seconds but I didn't take the effort to
00:08:35
take those 3 seconds I just dumped it in
00:08:37
sometimes with YouTube videos I just
00:08:38
pull out the entire video dump it into
00:08:41
uh you know an llm and say please give
00:08:43
me a summary of this so do you feel like
00:08:46
at some point we're starting to get
00:08:48
cognitively lazier and a lot of the
00:08:51
tasks we used to do on a Microsoft
00:08:52
Office or in Excel they kept the brain
00:08:54
sharp right do you feel we're losing the
00:08:56
bicycle for the
00:08:57
mind you know
00:08:59
throughout evolutionary history we as a
00:09:03
species acquire Knowledge and Skills
00:09:06
which enable us to adapt to whatever the
00:09:08
environment requires of us and then we
00:09:10
we lose those Knowledge and Skills as
00:09:13
you know civilization evolves like in
00:09:15
the past it would have been important
00:09:16
for us as foragers to know which berries
00:09:19
were poisonous or you know which animals
00:09:22
were likely to attack us or not right
00:09:24
and we sort of drop that information
00:09:26
over time it becomes less and less
00:09:28
valuable to you know memorize you know
00:09:31
certain pieces of information and more
00:09:33
and more valuable to focus on creativity
00:09:36
and judgment and so on so in know as I
00:09:39
was saying sort of I think it reduces
00:09:41
the barrier to entry to creativity right
00:09:43
it make because I don't have to spend
00:09:47
Years Learning to be a great programmer
00:09:49
I can now try something out quickly then
00:09:52
what the kind of evolutionary story is
00:09:54
optimizing for is you know my ability to
00:09:56
in invent a new idea A New Concept an
00:09:59
abstract layer rather than my ability to
00:10:02
be able to execute on it um so in a way
00:10:06
it's sort of exercising a different part
00:10:07
of our brain and yes atrophying a
00:10:10
certain part of our brain like for sure
00:10:12
you know now that we all have
00:10:13
smartphones we don't memorize telephone
00:10:15
numbers like we might have done back in
00:10:17
the day um or maybe we don't memorize
00:10:20
our times table as much right or maybe
00:10:22
we're a bit less good at Maps navigation
00:10:25
because we all have a you know a map on
00:10:27
our phone should do that affect
00:10:29
education for youngsters now cuz we
00:10:32
still do the time stable and we still do
00:10:33
a lot of things that at this point none
00:10:35
of us would do yeah you think education
00:10:38
like especially kwell education has to
00:10:39
change now that agent thing yeah
00:10:41
definitely I mean I think I think it's
00:10:43
already changing I mean if you compare
00:10:45
it to sort of 20 30 years ago pre
00:10:47
pre-digital it is a different
00:10:50
educational environment to know that you
00:10:51
now have access to information that your
00:10:53
fingertips and obviously people worry
00:10:54
that it means people are just going to
00:10:56
sort of copy and not be you know um you
00:10:59
know not not sort of ingest or learn or
00:11:02
memorize new information and that's
00:11:04
going to remain important um but I think
00:11:07
we just have to be kind of deliberate
00:11:09
about it it is a in a new phase every
00:11:11
tool gives us an amplification of some
00:11:14
skills and a kind of atrophying of other
00:11:17
skills and we just have to sort of
00:11:18
balance that in intention with one
00:11:20
another interesting I want I want to
00:11:22
talk about Microsoft a little bit right
00:11:24
I think you guys have a huge campus here
00:11:27
in India in Hyderabad and you know
00:11:30
something like 20,000 plus employees uh
00:11:33
what is what is Microsoft's commitment
00:11:35
to doing AI well in India like are there
00:11:37
Engineers here who are contributing to
00:11:39
copilot like what's the what's the India
00:11:41
play here yeah I I mean actually even
00:11:43
here in Mangalore I have team members
00:11:46
who are contributing to co-pilot and
00:11:49
contributing to our advertising stack
00:11:51
contributing to our search relevance
00:11:53
search quality um so you know this is
00:11:56
some of the best engineers in the world
00:11:58
India itself is one of the most popular
00:12:01
uh markets for co-pilot it's growing
00:12:03
faster than most other markets it's
00:12:05
really quite impressive so it's a it's a
00:12:07
big part of what we're going to do
00:12:09
interesting I have one last question for
00:12:11
you which is can you tell me the top
00:12:14
three like from your you know backend
00:12:17
data right what are the top three use
00:12:18
cases of co-pilot today and then also
00:12:20
paint a picture of where copilot goes
00:12:22
three years from now and what that looks
00:12:24
like what are the top three use cases
00:12:26
you know in 2027 or 2028 yeah I mean top
00:12:30
top use cases are I think replacing your
00:12:32
search engine yeah I mean many many
00:12:36
people are using it for everyday queries
00:12:39
um you know what is the GDP of this or
00:12:41
where do I find that or you know what
00:12:43
did this person do you know generic
00:12:46
queries that you might otherwise put to
00:12:47
a search engine second is I think
00:12:49
education you know it's definitely a
00:12:51
help for college and school and third is
00:12:54
I think increasingly I'm seeing people
00:12:56
use it for companionship uh an emotional
00:12:59
support helping to think through a
00:13:00
tricky problem that you're working
00:13:02
through because you know co-pilot today
00:13:04
is not judgmental it doesn't put you
00:13:07
down it asks you know it's always there
00:13:09
to talk to you in a very sort of simple
00:13:12
and calming way so I think those are the
00:13:14
kind of big use cases but tomorrow or
00:13:16
sort of in years to come um you know
00:13:19
these models are going to sort of have
00:13:21
near Perfect Memory right so they're
00:13:23
going to be very uh useful as a kind of
00:13:27
second brain for you like anything
00:13:29
anything that occurs to you any idea
00:13:30
that you have any open question that you
00:13:32
have any historic documents that you
00:13:34
have any tasks that you have it's really
00:13:36
going to be a way for you to sort of
00:13:38
augment your everyday thought process
00:13:41
and brain and that's kind of what I mean
00:13:43
by an AI companion that's always at your
00:13:45
side seeing what you see hearing what
00:13:47
you hear and living life alongside you
00:13:49
do you think it's going to be a friend
00:13:50
do you think I'd be able to give it a
00:13:51
name because I do that a lot I spar with
00:13:53
copilot a lot right which is hey I have
00:13:55
this idea can you help me flesh it out
00:13:57
and I and I know a lot of smart people
00:13:58
that do that at some point you think I'd
00:14:00
give it a name and it knows me very very
00:14:02
well and it becomes my friend I think
00:14:04
it's going to become a friend yeah
00:14:06
definitely I think it's going to you
00:14:08
know really live life alongside you it's
00:14:10
going to see what you see both in the
00:14:13
physical world and on you know the
00:14:15
digital world and the strange thing
00:14:18
about it is that you're going to be able
00:14:19
to sort of point to things like say you
00:14:21
know take a look at this what do you
00:14:23
think of that right and it'll just know
00:14:25
what this or that is because it has a
00:14:27
kind of presence living life with you
00:14:29
it'll know what your style is what your
00:14:31
tone is and how you like to talk do you
00:14:33
think I'd be able to send it to work
00:14:35
instead of me I think that you're going
00:14:36
to apply for a job with it you know your
00:14:40
your personal AI is going to get to know
00:14:42
you so well and fill in a lot of the
00:14:44
kind of gaps that you have to allow you
00:14:47
to be your best self you know it because
00:14:49
it's infinitely malleable and
00:14:52
adaptive you know and because I think
00:14:54
we're designing it with such intention
00:14:56
it should complement the areas where you
00:14:59
deliberately want it to augment you to
00:15:03
enable you to shine in the ways that you
00:15:05
know you are sort of uniquely capable of
00:15:08
doing and proud so I think of it as a
00:15:10
kind of like jigsaw puzzle connecting
00:15:12
with the areas that you choose to um you
00:15:15
know sort of amplify in your life and
00:15:17
yeah I can imagine you introducing it to
00:15:20
your friends or to your parents or
00:15:23
taking it to work with you or switching
00:15:24
jobs with you Etc um I think it's going
00:15:28
to be an important part of life so in a
00:15:30
way nobody has to really suffer with
00:15:31
loneliness like there's a huge
00:15:32
loneliness epidemic right nobody has to
00:15:35
suffer or at least has tools to help
00:15:36
them out with that period and in the
00:15:38
meantime I think it'll always have
00:15:40
utility like this image was generated
00:15:42
with co-pilot that's awesome so it'll
00:15:44
always have cool image yeah it will
00:15:46
always have utility I guess till that
00:15:48
time Point very cool thank you so much
00:15:50
Mustafa this was like a super exciting
00:15:53
session I had like these three four
00:15:54
questions for you and I know we added a
00:15:56
couple more questions because I wanted
00:15:57
to explore some paths um I learned a lot
00:16:01
and thank you so much for being here
00:16:02
this has been great it's a lot of fun
00:16:04
and thanks a lot cheers
