00:00:02
We've got a special guest who's going to
00:00:04
come join us. This always happens.
00:00:06
Another bread, everybody. Oh my god.
00:00:09
Somebody told me you uh started
00:00:11
submitting code and it kind of freaked
00:00:14
everybody out that daddy was hungry. All
00:00:17
models tend to do better if you threaten
00:00:19
them. If you threaten them like with
00:00:20
physical violence. Yes. Management is
00:00:23
like the easiest thing to do with AI.
00:00:25
Absolutely. It must be a weird
00:00:26
experience to meet the bureaucracy in a
00:00:28
company that you didn't hire. But on the
00:00:30
other side of it, I would say it's
00:00:31
pretty amazing that some junior muck
00:00:33
muck can basically look at you and say,
00:00:35
"Hey, go yourself." No, but I'm serious.
00:00:37
That's a sign of a healthy culture,
00:00:39
actually. You're punching a clock, man.
00:00:41
I hear the reports. You and I have
00:00:43
talked about it. You're going to work
00:00:44
every day. Yeah. It's been, you know,
00:00:47
some of the most fun I've had in my
00:00:48
life, honestly. And uh I retired like a
00:00:52
month before CO hit in theory. Yeah. And
00:00:55
I was like, you know, this has been
00:00:56
good. I want to do something else. I
00:00:58
want to hang out in
00:00:59
cafes, read physics books. Yeah. And
00:01:02
then like a month later, I was like, uh,
00:01:05
that's not really happening. So then I
00:01:08
just started to go to the office, you
00:01:10
know, once we could go to the office.
00:01:12
And
00:01:13
um actually to be perfectly honest,
00:01:16
there was a guy uh from uh OpenAI, this
00:01:20
guy named Dan, and I I ran into him at a
00:01:23
little party and he said, you know, look
00:01:26
what are you doing? This is like the
00:01:28
greatest
00:01:30
transformative moment in computer
00:01:32
science ever. Completely like and you're
00:01:34
a computer scientist. I'm a computer
00:01:35
scientist. Forget that. Founder of
00:01:37
Google, but you're a PhD student for
00:01:39
computer science. I haven't finished my
00:01:40
PhD yet, but working on it. Keep
00:01:42
working. Yeah, we'll get there.
00:01:44
Technically on leave of absence, right?
00:01:46
And uh he he told me this and I'd
00:01:48
already started kind of going into the
00:01:49
office a little bit and I was like, you
00:01:51
know, he's right. And uh it has been uh
00:01:54
just incredible just well you guys all
00:01:57
obviously follow all the AI technology
00:02:00
but being a computer scientist it is you
00:02:02
know the most exciting thing you know of
00:02:05
my life just technologically and the
00:02:08
exponential nature of this the pace of
00:02:11
it dwarfs anything we've seen in our
00:02:13
career it's almost like every thing we
00:02:16
did over the last 30 or 40 years has led
00:02:19
up to this moment and it's all compound
00:02:22
pounding on itself. The pace maybe you
00:02:24
could speak, you know, you you had a
00:02:26
company Google that grew from, you know,
00:02:29
a 100 users and 10 employees
00:02:32
to now you have over two billion people
00:02:35
using I think six products or five
00:02:37
products have over two billion. It's
00:02:39
it's not it's not even worth counting
00:02:41
because it's the majority of the people
00:02:42
in the planet touch Google
00:02:44
products. Describe the pace. Yeah. I
00:02:47
mean the excitement of the early web
00:02:50
like I remember using Mosaic and then
00:02:52
later Netscape. Uh how many of you
00:02:55
remember Mosaic actually? Am I a weirdo?
00:02:58
And you remember there was a what a
00:03:00
what's new page? The what's new page is
00:03:01
great. Right. Like you go through two or
00:03:04
three new web pages again. Yeah. It was
00:03:05
like in this last week these were the
00:03:07
new websites. Yes. And it was like such
00:03:09
and such elementary school such and such
00:03:12
a fish tank. Yeah. And you were like
00:03:15
Michael Jordan appreciation page. Yeah.
00:03:17
Whatever it was, these were the three
00:03:18
new sites on the whole internet. So
00:03:21
obviously the web you know developed
00:03:23
very rapidly from there and that was a
00:03:25
very uh exciting and then we've had
00:03:28
smartphones and whatnot. But you know
00:03:30
this the developments in AI are just
00:03:33
astonishing I would say by comparison uh
00:03:36
just because of you know the web spread
00:03:39
but didn't technically change so much uh
00:03:42
from you know month to month year to
00:03:44
year but these AI systems actually
00:03:47
changed quite a lot quite a lot you know
00:03:49
the like if you went away somewhere for
00:03:53
a month and you came back you'd be like
00:03:55
whoa what happened somebody told me you
00:03:57
uh started submitting code and it kind
00:04:01
of freaked everybody out that daddy was
00:04:02
home.
00:04:05
Okay. Daddy need a PR. What happened?
00:04:07
The code I submitted wasn't very
00:04:09
exciting. I think I needed to like add
00:04:11
myself to get access to some things and
00:04:14
you know a minor CL here or there. Um
00:04:17
nothing nothing that's going to win any
00:04:19
awards. uh but but I you know you need
00:04:23
to do that to um to do basic things run
00:04:26
basic experiments and things like that
00:04:29
um and I've I've tried to do that and
00:04:32
touch different parts of the system so
00:04:35
that you know I so I first of all it's
00:04:38
fun and secondly I know what I'm talking
00:04:40
about um it's really feels privileged to
00:04:43
be able to kind of go back to the
00:04:46
company not have any real executive
00:04:48
responsibilities but be able to actually
00:04:51
go deep into every little pocket. Are
00:04:54
there parts of the AI stack that
00:04:57
interest you more than others right now?
00:04:58
Are there certain problems that are just
00:05:00
totally captivating you? Yeah, I started
00:05:02
uh you know like sort of um I don't know
00:05:05
a couple years ago and maybe a year ago
00:05:08
uh I was really very close with uh the
00:05:11
what we call pre-training. Yeah. Um,
00:05:13
actually most of what people think of as
00:05:15
AI training, whatever people call it,
00:05:17
pre-training for various historical
00:05:20
reasons. Uh, but that's sort of the big
00:05:23
super, you know, you throw huge amounts
00:05:26
of computers at it. Um, and,
00:05:29
uh, I I learned a lot, you know, just
00:05:31
being deeply involved in that and seeing
00:05:33
us go from model to model and so forth
00:05:35
and running little baby experiments, but
00:05:38
uh, kind of just for fun. So I could say
00:05:41
I did it. Uh and uh more recently the
00:05:44
post-training especially as the thinking
00:05:46
models have uh come around. Um and
00:05:49
that's been you know another huge step
00:05:52
up in general in AI. So uh you know we
00:05:56
don't really know what the ceiling is.
00:05:59
When you um explain what's happening
00:06:02
with prompt engineering then to deep
00:06:04
research and what's happening there to
00:06:07
like a civilian. How would you explain
00:06:09
that sort of step function? because I
00:06:11
think people are not hitting the down
00:06:13
carrot and watching deep research in
00:06:16
Gemini's mobile app and you got a mobile
00:06:18
app and it's pretty great and I by the
00:06:19
way I got the uh fold after you and I um
00:06:22
were talking about it and okay Google
00:06:24
kicks Siri's ass now like it actually
00:06:26
does what you ask it to do when you ask
00:06:28
it to open up it does stuff but the
00:06:30
number of threads the number of queries
00:06:32
the number of follow-ups that it's doing
00:06:34
in that deep research is 200 300 maybe
00:06:39
explain that jump and then what you
00:06:40
think the jump after that is. To me, the
00:06:43
exciting thing about AI, especially
00:06:45
these days, I mean, it's not like quite
00:06:47
AGI yet as people are seeking or it's
00:06:50
not superhuman intelligence, but it's
00:06:54
pretty damn smart and uh can definitely
00:06:57
surprise you. So, I so I think of the
00:07:00
superpower is when it can do things in a
00:07:03
volume that I cannot. Yes. Right. So,
00:07:06
you know, by default when you use some
00:07:08
of our AI systems, you know, it'll suck
00:07:10
down whatever top 10 search results, you
00:07:13
know, and kind of pull out what you need
00:07:16
out of them, something like that. Um,
00:07:18
but I could do that myself to be honest,
00:07:20
you know, maybe it would take me a
00:07:22
little bit more time. But if it sucks
00:07:24
down the top, you know, thousand results
00:07:27
and then does follow on searches for
00:07:28
each of those and reads them deeply,
00:07:31
like that's, you know, a week of work
00:07:33
for me. Like I can't do that. This is
00:07:35
the thing I think people have not fully
00:07:37
appreciated who are not using the deep
00:07:38
research projects before we had our F1
00:07:41
um driver on stage. I'm a neopight. I
00:07:44
don't know anything about it. I said,
00:07:45
"How many deaths occurred per decade?"
00:07:47
And I said, "I want to get to deaths per
00:07:49
mile driven it." And at first was like,
00:07:50
"That's going to be really hard." I was
00:07:51
like, "I give you permission to make
00:07:53
your best shot at it and come up with
00:07:55
your best theory. Let's do it." And it
00:07:57
was like, "Okay." And it was like
00:08:00
there's this many teams, there's this
00:08:01
many races. Which model did you use it?
00:08:04
Uh, no. I'm I use Gemini. Gemini
00:08:07
fabulous version, the fabulous one. And
00:08:09
it was like, but I treat it like I get
00:08:12
sassy with it and it kind of works for
00:08:14
me. You know, it's a weird thing. It's
00:08:16
like you drinking the wine. We don't
00:08:18
circulate too much in the AI community.
00:08:21
Uh, but the not just our models, but all
00:08:24
models tend to do better if you threaten
00:08:25
them. If you threaten them like with
00:08:27
physical violence. Yes. But like that's
00:08:30
people feel weird about that, so we
00:08:31
don't really talk about that. Yeah, I
00:08:33
was threatened them with not being
00:08:34
fabulous and it responded to that as
00:08:36
well. Yeah, that's historically you just
00:08:38
say like, "Oh, I'm going to kidnap you
00:08:40
if you
00:08:41
don't." Yeah, they actually Can I ask
00:08:43
you a more specific But hold on. But it
00:08:45
went through it and it literally came up
00:08:48
with a system where it said, "I think we
00:08:50
should include practice miles." So,
00:08:52
let's say there's 100 practice miles for
00:08:54
every mile on the track. And then it
00:08:56
literally gave me the deaths per mile
00:08:59
estimated. And then I started cross
00:09:00
referencing it and I was like oh my god
00:09:02
this is like somebody's term paper for
00:09:05
undergrad you know like whoa done in in
00:09:09
minutes. It's yeah I mean it's amazing
00:09:11
and all of us have had these experiences
00:09:13
where you suddenly decide okay I'll just
00:09:16
throw this AI I don't really expect it
00:09:18
to work and then you're like whoa that
00:09:20
actually worked. So, as you as you have
00:09:22
those moments and then you go home to
00:09:25
your just life as a dad, have you gotten
00:09:27
to the point where you're like, "What
00:09:30
will my children do? And are they
00:09:32
learning the right way? And should I
00:09:35
totally just change everything that
00:09:37
they're doing right now?" Have you had
00:09:39
any of those moments yet? Yeah. I mean,
00:09:41
I look, I I don't really know how to
00:09:44
think about it to be perfectly honest. I
00:09:45
don't have like a magical way. I mean I
00:09:47
see I have a a kid in high school and
00:09:50
middle school and you know I mean the
00:09:52
AIS are basically you know already ahead
00:09:57
you know I mean obviously there's some
00:09:59
things AIs are particularly dumb at and
00:10:00
they you know they make certain mistakes
00:10:03
a human would never make but generally
00:10:05
you know if you talk about like math or
00:10:07
calculus or whatever like they're pretty
00:10:10
damn good like they you know can win
00:10:12
like math contests and coding contests
00:10:14
things like that against you know some
00:10:16
top humans and and then I look at you
00:10:19
know okay he's whatever my son's going
00:10:22
to go on to whatever from sophomore to
00:10:24
junior and what is he going to learn and
00:10:26
then I think in my mind and I talk to
00:10:28
him about this well what is the AI going
00:10:30
to be inact
00:10:33
yeah yeah and it's like comparable right
00:10:36
obviously are there areas where you
00:10:38
would tell your son look don't or not
00:10:41
not yet I don't know if you can like
00:10:43
plan your life around this I mean I
00:10:46
didn't
00:10:47
particularly plan my life to like I
00:10:51
don't know be an entrepreneur or
00:10:52
whatever. I was just liked math and
00:10:53
computer science. I guess maybe I got
00:10:55
lucky and it worked out to be you know
00:10:57
useful in the world. I don't know. I
00:10:59
guess I I I think you know my kids
00:11:01
should do what they like. Hopefully it's
00:11:04
somewhat challenging and they can you
00:11:05
know overcome uh different kinds of
00:11:08
problems and things like that. What
00:11:09
about specifically? What about college?
00:11:11
Do you think college should is going to
00:11:13
continue to exist as it is today? I mean
00:11:15
it seems like college was already
00:11:16
undergoing this kind of uh revolution
00:11:20
even before this sort of AI challenge of
00:11:22
people are like is it worth it? Should I
00:11:24
be more vocational? What's actually
00:11:26
going to be useful? So we're already
00:11:28
kind of entering this kind of situation
00:11:31
uh where there's sort of questions asked
00:11:33
about colleges. Yeah, I think you know
00:11:36
AI obviously puts that at the forefront.
00:11:39
As a parent, I think a lot about, hey,
00:11:42
so much of education in America and the
00:11:45
middle class, upper class is all
00:11:47
about what college, how do you get them
00:11:50
there? And honestly, lately, I'm like, I
00:11:52
don't think they should go to college.
00:11:53
Like, it's just fundamentally, you know,
00:11:55
my son is a rising junior and his entire
00:11:59
focus is he wants to go to an SEC school
00:12:01
because of the culture.
00:12:04
And two years ago, I was I would have
00:12:07
panicked and I would have thought,
00:12:09
should I help him get into a school,
00:12:10
this school, that school? And now I'm
00:12:12
like, that's actually the best thing you
00:12:14
could do. Be socially well adjusted,
00:12:17
psychologically deal with different
00:12:18
kinds of failures, you know, enjoy a few
00:12:20
years of exploration. Yeah. Yeah. Yeah.
00:12:23
Sergey, can I ask you about hardware?
00:12:26
You know, years ago, Google owned Boston
00:12:28
Dynamics, maybe a little bit ahead of
00:12:30
its time, but the way these systems are
00:12:33
learning through visual information and
00:12:36
sensory information and basically
00:12:38
learning how to adjust to the
00:12:40
environment around them is triggering
00:12:42
these kind of pretty profound like
00:12:43
learning curves in hardware and there's
00:12:46
dozens of like startups now making
00:12:48
robotic systems. What do you see in
00:12:51
robotics and hardware? Is this a year or
00:12:54
are we in a moment right now where
00:12:55
things are really starting to work? I
00:12:57
mean, I think we've uh you know,
00:12:59
acquired and later sold like five or so
00:13:02
robotics companies and uh you know,
00:13:04
Boston being one of them. I guess if I
00:13:06
look back on it, we built the hardware.
00:13:08
We also had this more recently we built
00:13:10
out u everyday robotics internally and
00:13:14
then later had to transition that. You
00:13:16
know, the robots are all cool and all,
00:13:18
but the software wasn't quite there.
00:13:23
Um, that's every time we've tried to do
00:13:25
it to, you know, to make them truly
00:13:28
useful
00:13:29
and presumably one of these days that'll
00:13:32
no longer be true, right? But have you
00:13:34
seen anything lately that Yeah. Do and
00:13:36
do you believe in the humanoid form
00:13:37
factor robots or do you think that's a
00:13:39
little overkill? I'm probably the one
00:13:41
weirdo who doesn't who's not a big fan
00:13:43
of humanoids, but maybe I'm jaded
00:13:46
because we've, you know, we at least
00:13:48
acquired at least two humanoid uh
00:13:50
robotics startups and later sold them.
00:13:53
Um but but the reason is I mean the
00:13:56
reason people want to do humanoid robots
00:13:58
for the most part is because the world
00:14:00
is kind of designed around this form
00:14:02
factor and you know you can train on
00:14:04
YouTube, we can train on videos, people
00:14:06
do all the things. Um, I personally
00:14:09
don't think that's given the AI quite
00:14:12
enough credit. Like AI can learn, you
00:14:16
know, through simulation and through
00:14:18
real life pretty quickly how to handle
00:14:20
different situations. And I don't know
00:14:22
that you need exactly the same number of
00:14:24
arms and legs and wheels, which is zero
00:14:26
in the case of humans, as humans to make
00:14:28
it all work. And that so I'm I'm
00:14:31
probably
00:14:32
less bullish on that. But to be fair,
00:14:34
there are a lot of really smart people
00:14:36
who are making humanoid robots. So I
00:14:38
wouldn't discount it. What about the
00:14:40
path of being a programmer? That's where
00:14:42
we're seeing with that finite data set.
00:14:45
And listen, Google's got a 20-y old code
00:14:47
base now. So like it actually could be
00:14:48
quite impactful. What are you seeing
00:14:50
like literally in the company? You know,
00:14:53
are the 10x developer is always this
00:14:55
like ideal that you can, you know, you
00:14:57
get a couple of unicorns once in a
00:14:58
while, but are we going to see like all
00:15:00
developers like, you know, their
00:15:02
productivity hit that level 8 n 10 and
00:15:04
they're just going to or is it going to
00:15:06
be all done by computers and we're just
00:15:08
going to check it, make sure it's not
00:15:09
too weird. Um, because it could get
00:15:12
weird if you vibe code. Yeah, I'm
00:15:15
embarrassed to say this. Okay. I like
00:15:17
recently I just had a big tiff inside
00:15:20
the company because we had this list of
00:15:23
what you're allowed to use to code and
00:15:25
what you're not allowed to use to code
00:15:26
and the Gemini was on the no list. You
00:15:30
Oh, you have to be pure. You can't I
00:15:32
don't know for like a bunch of really
00:15:33
weird reasons that it would like boggled
00:15:36
my mind that you know you couldn't vibe
00:15:38
code on the Gemini code. I mean, nobody
00:15:40
would like enforce this rule, but um but
00:15:43
there was this, you know, actual
00:15:46
internal web page. For whatever reason,
00:15:47
historical reason, somebody had put this
00:15:49
and I had a big fight with them and I,
00:15:51
you know, I cleared it up after a
00:15:54
shocking period of time. You escalated
00:15:56
to your boss. Oh, I I definitely told
00:16:00
about it and I Sorry, I don't know if
00:16:02
you remember, but you got super voting
00:16:05
founders. You are the boss. You can do
00:16:07
what you want. It's your company still.
00:16:09
No, no, it was uh he was very
00:16:11
supportive. I was more like uh uh I was
00:16:14
like I talked to him. I was like I can't
00:16:16
deal with these people. You need to deal
00:16:17
with this. Like I just like I'm beside
00:16:20
myself that they're like saying it's
00:16:22
weird that there's bureaucracy like in a
00:16:23
company that you must be a weird
00:16:25
experience to meet the bureaucracy in a
00:16:26
company that you didn't hire. But but on
00:16:28
the other side of it I would say it's
00:16:30
pretty amazing that some junior mucky
00:16:32
muck can basically look at you and say
00:16:34
hey go yourself. No but I'm serious.
00:16:36
That's a sign of a healthy culture
00:16:38
actually I guess. So anyway, it did get
00:16:40
fixed and uh people are using it. So
00:16:42
they got fired.
00:16:45
That person's working in Google Siberia.
00:16:48
No, we're trying to, you know, roll out
00:16:51
every possible kind of AI and and trying
00:16:53
external ones, you know, be whatever the
00:16:55
cursors of the world, all all of those
00:16:58
uh to just see what really makes people
00:17:01
more productive. Um I mean for myself
00:17:04
definitely makes me more productive
00:17:06
because I'm not do you do you think the
00:17:08
number of foundational models like if
00:17:10
you look three years forward
00:17:12
will they start to cleave off and gets
00:17:14
highly specialized like beyond the
00:17:16
general and the reasoning maybe there's
00:17:19
a very specific model for chip design
00:17:22
there's clearly a very specific model
00:17:23
for biologic precursor design protein
00:17:26
folding like is the number of
00:17:27
foundational models in the future Sergey
00:17:30
a multiple of what they are today the
00:17:32
same something in between. That's a
00:17:35
great question. I kind
00:17:37
of if I I mean look I don't know like
00:17:40
you guys take a guess just as well as I
00:17:42
can but um if I had to
00:17:46
guess you know things have been more
00:17:49
converging
00:17:50
uh and uh this is sort of broadly true
00:17:53
across machine learning I mean you used
00:17:54
to have all kinds of different kinds of
00:17:56
models and whatever convolutional
00:17:59
networks for vision things and you know
00:18:02
you had um whatever RNN's for text and
00:18:07
speech and stuff and uh you know all
00:18:09
this has shifted to transformers
00:18:11
basically uh and uh increasingly it's
00:18:15
also just becoming one model um now we
00:18:18
do get a lot of oomph occasionally we do
00:18:20
specialized models uh and it's it's
00:18:24
definitely
00:18:25
scientifically a good way to iterate
00:18:28
when you have a particular target you
00:18:29
don't have to like do everything in
00:18:31
every language and handle whatever both
00:18:34
images and video and audio and uh in one
00:18:37
go. Um but we are generally able to
00:18:42
after we do that take those learnings
00:18:46
and basically put that capability into a
00:18:48
general model. So there's not that much
00:18:51
benefit. Um you know you could you can
00:18:54
get away with a somewhat smaller
00:18:55
specialized model a little bit faster a
00:18:58
little bit cheaper but the trends have
00:19:00
not gone that way. What do you think
00:19:02
about the open source closed source
00:19:04
thing? Has there been big philosophical
00:19:07
movements that change your perspective
00:19:09
on the value of open source? Um, we're
00:19:12
still waiting on this o, you know, open
00:19:15
AI. Oh, yeah. Open source drop. I mean,
00:19:17
we haven't seen it yet, but
00:19:19
theoretically it's coming. I mean, have
00:19:21
to give credit uh to where credit's due.
00:19:24
I mean, Deepseek released a really
00:19:26
surprisingly
00:19:28
powerful model uh when it was January or
00:19:32
so. So, that that definitely closed the
00:19:34
gap to proprietary models. We've pursued
00:19:37
both. So, we released Gemma uh which are
00:19:40
our open- source or you know open models
00:19:44
and um those perform really well.
00:19:47
They're small, dense models, so they fit
00:19:48
well on one computer. Um, and
00:19:52
uh, they're not as powerful as Gemini.
00:19:55
Uh, but I mean, the jury's out which way
00:19:58
that's going to go. Do you have a point
00:19:59
of view on what human computing
00:20:02
interaction looks like as AI progresses?
00:20:06
It used to be, thanks to you, as a
00:20:09
search box. You type in some keywords or
00:20:12
a question and you would click on links
00:20:14
on the internet and get an answer. Is
00:20:16
the future typing in a question or
00:20:18
speaking to a AirPod or thinking or
00:20:23
thinking or like what's the what's the
00:20:24
Yeah. And then the answer is just spoken
00:20:26
to you. I mean by by the way just to
00:20:27
build on this it was Friday, right?
00:20:30
Neuralink got breakthrough designation
00:20:32
for their human brain interface. I mean
00:20:34
that's a very big step in allowing the
00:20:36
FDA to clear everybody getting it
00:20:38
implanted. Yeah. Is it like if you could
00:20:40
just summarize what you think is kind of
00:20:42
the most common place human computer
00:20:45
interaction model in the next decade or
00:20:47
whatever. Is it a you know there's this
00:20:49
idea of glasses with a screen in the
00:20:51
glasses and you tried that a long time
00:20:53
ago. Yeah, I kind of messed that up.
00:20:54
I'll be honest.
00:20:56
Uh got the timing totally wrong on that.
00:20:59
Early again. Yeah. Uh right. Right. But
00:21:02
early. There are a bunch of things I
00:21:03
wish I had done differently, but
00:21:05
honestly it was just like the technology
00:21:07
wasn't ready for for Google class. Uh
00:21:10
but nowadays these things I think are
00:21:12
more sensible. I mean there's still
00:21:13
battery life issues I think that you
00:21:17
know we and others need to overcome. Uh
00:21:20
but I think that's a cool form factor. I
00:21:23
mean when you say 10 years though you
00:21:25
know a lot of people are saying hey the
00:21:26
singularity is like five years away. So
00:21:30
your ability to see through that into
00:21:34
the future.
00:21:36
I mean it's very important. But do you
00:21:37
have anybody else? Sorry. Just let me
00:21:39
ask about this. Do you There was a
00:21:40
comment that Larry made years ago that
00:21:44
humans were a stepping stone in
00:21:46
evolution. Okay. Can you comment on
00:21:49
this? Like do you do you think that this
00:21:52
AGI super intelligence or really silicon
00:21:54
intelligence exceeds human capacity and
00:21:57
humans are a stepping stone in you know
00:22:00
progression of evolution? Boy, I think
00:22:03
like sometimes us nerdy guys go and get
00:22:05
have a little too much wine. I've had
00:22:08
two glasses and um I'm ready to go. I I
00:22:11
need some more for this conversation. Um
00:22:15
human implants. Let's go. I mean, I
00:22:17
guess we're starting to get experience
00:22:19
with these AIs that can do certain
00:22:21
things, you know, much better than us.
00:22:23
Um, and they're definitely, you know,
00:22:25
with my skill of math and coding, I feel
00:22:28
like I'm better off just turning to the
00:22:32
AI now. And how do I feel about that? I
00:22:34
mean, it doesn't really bother me, you
00:22:36
know, I use it as a
00:22:37
tool. So, I feel like I've gotten used
00:22:40
to it, but you know, maybe if they get
00:22:43
even more capable in the future,
00:22:46
um, I'll look at it differently. Yeah,
00:22:48
there's a moment of insecurity, maybe. I
00:22:50
guess. So, as an aside, management is
00:22:52
like the easiest thing to do with the
00:22:54
AI. Yeah, absolutely. And I did this,
00:22:57
you know, uh, at Gemini on some of our,
00:23:00
you know, work chats, um, kind of like
00:23:02
Slack, but we have our own version. We
00:23:04
had this AI tool that actually was
00:23:06
really powerful. We unfortunately anyway
00:23:08
temporarily got rid of it. I think we're
00:23:10
going to bring it back and bring it to
00:23:11
everybody. But it it could suck down a
00:23:14
whole chat space and then answer pretty
00:23:16
complicated questions. So I was like,
00:23:18
"Okay, summarize this for me." Okay, now
00:23:21
assign something for everyone to work on
00:23:23
and uh and then I would paste it back in
00:23:26
so people didn't realize it was the AI.
00:23:28
I I admitted it pretty soon. Um and
00:23:31
there were a few giveaways here or
00:23:32
there, but it worked remarkably well.
00:23:35
And then I was like, well, who should be
00:23:37
promoted in this chat space? Uh, and I
00:23:41
actually picked out this woman, this
00:23:43
young woman engineer who like, you know,
00:23:44
I didn't even notice she wasn't very
00:23:46
vocal uh particularly in that PRs kicked
00:23:50
ass. No, no, it was like and then uh I
00:23:53
don't know something that the AI had
00:23:55
detected and I went I talked to the
00:23:56
manager actually and and he was like,
00:23:58
"Yeah, you know what? You're right. Like
00:24:01
she's been working really hard all these
00:24:02
things." Wow. I think that ended up
00:24:04
happening actually. Uh
00:24:07
so I don't know. I guess after a while
00:24:09
you just kind of take it for granted
00:24:10
that you can just do these things. I
00:24:12
don't know. It hasn't really Do you
00:24:13
think that there's a a use case
00:24:16
for like an infinite context link? Oh
00:24:20
100%. I mean all of Google's codebase
00:24:23
goes infinite but sure you should have
00:24:25
access infinite. Yeah. Stateful. Yeah.
00:24:29
and then multiple sessions so that you
00:24:30
could have like 19 of these things, 20
00:24:32
of these things running or just evolve
00:24:33
itself. Eventually, it'll evolve itself.
00:24:35
Yeah. I mean, I guess if it knows
00:24:37
everything, then you can have just one
00:24:39
in theory. You just need to somehow
00:24:41
disambiguate what you're talking about.
00:24:43
Uh but yeah, for sure there's no limit
00:24:46
to use of uh context and there, you
00:24:50
know, there are a lot of ways to make it
00:24:52
larger and larger. There's a there's a
00:24:53
rumor that internally there's a Gemini
00:24:55
build that is a quasi infinite context.
00:25:00
Is it is it a valuable thing? Like I
00:25:01
don't know. Well, you say what you want
00:25:03
to say, but I mean for any such cool new
00:25:07
idea in AI, there are probably five such
00:25:09
things internally. Um and uh you know
00:25:12
the question is how well do they work?
00:25:13
And um yeah, I mean we're definitely
00:25:15
pushing all the bounds um in terms of
00:25:18
intelligence, in terms of context, in
00:25:20
terms of um speed, you know, you name
00:25:24
it. And what about the hardware? Like
00:25:25
when you guys build stuff, do you care
00:25:28
that you have this pathway to Nvidia or
00:25:32
do you think eventually that'll get
00:25:33
abstracted and there'll be a transpiler
00:25:36
and it'll be Nvidia plus 10 other
00:25:38
options, so who cares? Let's just go as
00:25:39
fast as possible. Well, we mostly for
00:25:41
for Gemini, we mostly use our own TPUs.
00:25:45
So, um but we also do support um Nvidia
00:25:48
and we we're one of the big uh uh
00:25:52
purchasers of Nvidia chips and we have
00:25:54
them in Google Cloud available for our
00:25:57
customers uh in addition to
00:25:59
TPUs. Um at this stage it's uh for
00:26:05
better or for worse not that abstract
00:26:06
and maybe someday the AI will abstract
00:26:09
it for us but you know given just the
00:26:11
amount of computation you have to do on
00:26:12
these models you actually have to think
00:26:14
pretty carefully how to do everything
00:26:16
and exactly what kind of chip you have
00:26:18
and how the memory works and the
00:26:20
communication works and so forth are
00:26:23
actually pretty big factors and it
00:26:25
actually yeah maybe one of these days
00:26:28
the AI itself will be good enough to
00:26:30
reason through that today. It's not
00:26:31
quite good enough. I don't know if you
00:26:33
guys are having this experience with the
00:26:35
interface, but I find myself even on my
00:26:37
desktop and certainly on my mobile
00:26:39
phone, going immediately into voice chat
00:26:42
mode and telling it, "Nope, stop." Uh,
00:26:45
that wasn't my question. This is my
00:26:46
question. Nope. Uh, let's say that again
00:26:48
in shorter bullet points. Nope, I want
00:26:49
to focus on this. It's so quick now.
00:26:53
Last year was unusable. It was too slow.
00:26:55
And now it like stops. Okay. And then
00:26:58
you sell it. I would like It's what I
00:27:00
want to go to. I don't want to type. I
00:27:01
want to use voice. And then
00:27:02
concurrently, I'm watching the text as
00:27:05
it's being written on the page and I
00:27:07
have another window open and I'm doing
00:27:08
Google searches or second queries to an
00:27:11
LLM or writing a Google doc or a notion
00:27:15
page or typing something. So, it's
00:27:17
almost like that scene in um Minority
00:27:20
Report where he has the gloves or in
00:27:22
Bladeunner where he's, you know, in his
00:27:24
apartment saying, "Zoom in, zoom in."
00:27:25
Closer to the left, to the right. And
00:27:27
there's something about these language
00:27:29
models and their ability to the response
00:27:32
time which was always something you
00:27:33
focused on response time the is there
00:27:36
like a response time thing where it
00:27:37
actually is worth doing voice and where
00:27:40
it wasn't previously everything is
00:27:43
getting better and faster and so forth
00:27:45
you know smaller models are more capable
00:27:48
there are better ways to do inference on
00:27:49
them that are faster you can also stack
00:27:52
them like you know this is like Nico's
00:27:53
company 11 labs it's an exceptional TTS
00:27:56
SSD stack like there's I mean there are
00:27:59
other options. Whisper is really good at
00:28:01
certain things, but those this is where
00:28:03
I I kind of believe you're going to get
00:28:05
this like compartmentalization where
00:28:08
there'll be certain foundational models
00:28:10
for certain specific things. You stack
00:28:11
them together. You kind of deal with the
00:28:13
latency and it's like pretty good
00:28:16
because they're so good. Like Whisper
00:28:18
and 11 for those speech examples that
00:28:20
you're talking about are
00:28:22
kickass. I mean, they're exceptional.
00:28:24
Well, wait till you turn on your camera
00:28:26
and it sees your reaction to what it's
00:28:29
saying and you go and before you even
00:28:31
say that you don't want it or you put
00:28:32
your finger up, it's pauses. Oh, did you
00:28:34
want something else? Oh, I see you're
00:28:36
not happy with that result. You know,
00:28:38
it's going to get really weird. It's a
00:28:40
funny thing, but we have the, you know,
00:28:41
we have the big open shared offices, so
00:28:44
during work, I can't really use voice
00:28:46
mode too much. I usually use it on the
00:28:48
drive. The drive is incredible. Yeah. I
00:28:51
don't feel like I could. I mean, I would
00:28:53
get its output in my headphones, but if
00:28:56
I want to speak to it, then everybody's
00:28:57
listening to me. So, it's weird. I just
00:28:59
think that would be socially awkward,
00:29:01
but I should I should do that. In my car
00:29:03
ride, I do chat to the AI, but then it's
00:29:06
like audio in, audio out. But I feel
00:29:08
like I honestly maybe it's a good
00:29:10
argument for a private office. I should
00:29:12
spend more time like you guys are. You
00:29:14
could talk to your manager.
00:29:17
They might get one. I like being out in
00:29:19
the
00:29:21
I like with everybody. Uh, but I do
00:29:23
think that there's this AI use case that
00:29:25
I'm missing which I should probably
00:29:27
figure out how to try more often. If
00:29:30
people want to try your new product, is
00:29:32
there a website they can visit or
00:29:34
something or special code or go check? I
00:29:37
mean, honestly, there's a dedicated
00:29:38
Gemini app. If you're using Gemini, just
00:29:40
like you're going through the Google
00:29:42
navigation from your search, just get
00:29:43
the download the actual Gemini app. It's
00:29:45
kickass. It really is the best models. I
00:29:47
think it is. You should use 2.5 Pro. 2.5
00:29:51
Pro. Pay the It's It's a You got to pay,
00:29:53
right? Uh yeah, you got a few query, you
00:29:56
got a few prompts for free, but uh you
00:29:57
know, if you do it a bunch, you need to
00:29:59
make all these like 20 bucks a month.
00:30:02
You got a vision for like making it free
00:30:03
and throwing some ads on the side. Yeah.
00:30:05
One step down in hardware cost, the
00:30:07
whole thing will be free. Well, okay.
00:30:08
It's free today without ads on the side.
00:30:10
You just got a certain number of the top
00:30:11
model. I think we're likely are going to
00:30:13
have always now like sort of top models
00:30:15
that we can't supply infinitely to
00:30:18
everyone right off the bat. But, you
00:30:21
know, wait 3 months and then the next
00:30:23
generation. Seems to me like if I'm
00:30:24
asking all these queries, you know, just
00:30:26
having a little on the sidebar of things
00:30:28
I might be a running list that changes
00:30:30
in real time of things I might be
00:30:32
interested in. All for, you know, really
00:30:34
good AI advertising. I just um
00:30:38
uh I don't think we're going to like
00:30:39
necessarily our latest and greatest
00:30:41
models which are you know take a lot of
00:30:44
computation. I don't think we're going
00:30:45
to just be free to everybody right off
00:30:48
the bat. But as we go to the next
00:30:51
generation you know it's like every time
00:30:52
we've gone forward a generation then the
00:30:55
sort of uh the new free tier is usually
00:30:57
as good as the previous pro tier uh and
00:31:02
sometimes better. All right, give it up
00:31:03
for Sergey Brit. Thank you.
00:31:07
[Applause]
00:31:09
Okay, thanks everybody for watching that
00:31:10
amazing interview with Sergey Brent and
00:31:12
thanks Sergey for joining us in Miami.
00:31:14
If you want to come to our next event,
00:31:16
it's the All-In Summit in Los Angeles,
00:31:19
fourth year for All-In Summit. Go to
00:31:22
all-in.com/events to apply. A very
00:31:24
special thanks to our new partner, OKX,
00:31:27
the new money app. OKX was the sponsor
00:31:30
of the McLaren F1 team, which won the
00:31:32
race in Miami. Thanks to Haidider and
00:31:35
his team, an amazing partner and an
00:31:37
amazing team. We really enjoyed spending
00:31:39
time with you. And OKX launched their
00:31:41
new crypto exchange here in the US. If
00:31:43
you love Allin, go check them out. And a
00:31:45
special thanks to our friends at Circle.
00:31:48
They're the team behind USDC. Yes, your
00:31:51
favorite stable coin in the world. USDC
00:31:54
is a fully backed digital dollar
00:31:56
redeemable one for one for USD. It's
00:31:59
built for speed, safety, and scale. They
00:32:01
just announced the Circle Payments
00:32:03
Network. This is enterprisegrade
00:32:05
infrastructure that bridges the gap
00:32:07
between the digital economy and outdated
00:32:09
financial reality. Go check out USDC for
00:32:12
all your stable coin needs. And special
00:32:15
thanks to my friends including Shane
00:32:17
over at Poly Market, Google Cloud,
00:32:18
Salana, and BVNK. We couldn't have done
00:32:22
it without y'all. Thank you so
00:32:25
much. We'll let your winners ride.
00:32:29
[Music]
00:32:33
And it said we open source it to the
00:32:35
fans and they've just gone crazy with
00:32:36
it. Love you queen of quinoa.
00:32:40
[Music]
00:32:46
Besties are gone.
00:32:48
That is my dog taking notice your
00:32:50
driveways.
00:32:53
Oh man. My dasher will meet up. We
00:32:56
should all just get a room and just have
00:32:58
one big huge orgy because they're all
00:32:59
just useless. It's like this like sexual
00:33:01
tension that they just need to release
00:33:02
somehow.
00:33:06
Wet your feet.
00:33:09
We need to get Murky's I'm going all
00:33:14
[Music]
00:33:18
in. I'm going all in.
