00:00:00
I first met our next guest Sam mman
00:00:02
almost 20 years ago when he was working
00:00:04
on a local mobile app called looped we
00:00:07
were both backed by seoa capital and in
00:00:09
fact we were both in the first class of
00:00:11
seoa Scouts he did investment in a
00:00:14
little unknown fintech company called
00:00:16
stripe I did Uber and in that tiny
00:00:18
experiment did Uber I've never heard
00:00:20
that before yeah I think so it's
00:00:22
possible starting already you should
00:00:24
write a book Jacob
00:00:28
maybe let your winners
00:00:30
[Music]
00:00:31
Rainman
00:00:35
David and in said we open source it to
00:00:37
the fans and they've just gone crazy
00:00:40
[Music]
00:00:43
with that tiny experimental fund that
00:00:46
Sam and I were part of a Scouts is
00:00:48
sequoia's highest multiple returning
00:00:50
fund couple of low digigit Millions
00:00:52
turned into over 200 million I'm told
00:00:54
and then he did yeah that's what I was
00:00:56
told by ruoff yeah and he did a stint at
00:00:58
Y combinator where he was president from
00:01:00
2014 to 2019 in 2016 he co-founded open
00:01:03
AI with the goal of ensuring that
00:01:06
artificial general intelligence benefits
00:01:08
all of humanity in 2019 he left YC to
00:01:11
join openai full-time as CEO things got
00:01:14
really interesting on November 30th of
00:01:16
2022 that's the day open AI launched
00:01:19
chat GPT in January 2023 Microsoft
00:01:22
invested 10 billion in November 2023
00:01:25
over a crazy 5ay span Sam was fired from
00:01:28
open AI everybody was going to go work
00:01:30
at Microsoft a bunch of hard emojis went
00:01:33
viral on x/ Twitter and people started
00:01:36
speculating that the team had reached
00:01:38
artificial general intelligence the
00:01:40
world was going to end and suddenly a
00:01:42
couple days later he was back to being
00:01:44
the CEO of open AI in February Sam was
00:01:48
reportedly looking to raise $7 trillion
00:01:49
do for an AI chip project this after was
00:01:53
reported that Sam was looking to raise a
00:01:54
billion from masi yoshian to create an
00:01:57
iPhone killer with Johnny I the
00:01:58
co-creator of the iPhone all of this
00:02:01
while chat GPT has become better and
00:02:03
better and a household name is having a
00:02:05
massive impact on how we work and how
00:02:08
work is getting done and it's reportedly
00:02:10
the fastest product to hit 100 million
00:02:12
users in history in just two months and
00:02:15
check out opening eyes insane Revenue
00:02:18
ramp up they reportedly hit two billion
00:02:19
in ARR last year welcome to the all-in
00:02:23
podcast Sam mman thank you thank you
00:02:26
guys Sak you want to Le us off here okay
00:02:28
sure I mean I I think the whole industry
00:02:30
is waiting with baited breath for the
00:02:32
release of GPT 5 I guess it's been
00:02:35
reported that it's launching sometime
00:02:36
this summer but that's a pretty big
00:02:38
window can you narrow that down I guess
00:02:41
where where are you in the release of
00:02:42
GPT 5 uh we we take our time on releases
00:02:46
of major new models and I don't think we
00:02:51
uh I think it will be great uh when we
00:02:54
do it and I think we'll be thoughtful
00:02:56
about how we do it uh like we may
00:02:58
release it in a different way than we've
00:02:59
released previous BS models um also I
00:03:01
don't even know if we'll call it gbt 5
00:03:03
um what I what I will say is you know a
00:03:06
lot of people have noticed how much
00:03:08
better gbd4 has gotten um since we've
00:03:11
released it and particularly over the
00:03:12
last few months I
00:03:14
think I think that's like a better hint
00:03:18
of what the world looks like where it's
00:03:20
not the like 1 two 3 4 5 six seven but
00:03:24
you you just you use an AI system and
00:03:27
the whole system just gets better and
00:03:28
better fairly continous ly um I think
00:03:32
that's like both a better technological
00:03:34
Direction I think that's like easier for
00:03:36
society to adapt to um
00:03:39
but but I assume that's where we'll head
00:03:41
does that mean that there's not going to
00:03:42
be long training cycles and it's
00:03:44
continuously retraining or training
00:03:47
submodels Sam and maybe you could just
00:03:49
speak to us about what might change
00:03:51
architecturally going forward with
00:03:53
respect to large
00:03:54
models well I mean one one one thing
00:03:58
that you could imagine is this just that
00:04:00
you keep training right a model uh that
00:04:02
that would seem like a reasonable thing
00:04:04
to
00:04:05
me do you think you talked about
00:04:07
releasing it differently this time are
00:04:09
you thinking maybe releasing it to the
00:04:11
paid users first or you know a slower
00:04:14
roll out to get the red teams tight
00:04:17
since now there's so much at stake you
00:04:19
have so many customers actually paying
00:04:20
and you've got everybody watching
00:04:22
everything you do you know is it is it
00:04:26
moreful now yeah still only available to
00:04:29
the paid user users but one of the
00:04:30
things that we really want to do is
00:04:32
figure out how to make more advanced
00:04:34
technology available to free users too I
00:04:36
think that's a super important part of
00:04:39
our mission uh and and this idea that we
00:04:42
build AI tools and make them super
00:04:44
widely
00:04:45
available free or you know not that
00:04:47
expensive whatever it is so that people
00:04:49
can use them to go kind of invent the
00:04:51
future rather than the magic AGI and the
00:04:53
sky inventing the future and showing it
00:04:55
down upon us uh that seems like a much
00:04:58
better path it seems like more spiring
00:05:00
path I also think it's where things are
00:05:01
actually heading so it makes me sad that
00:05:04
we have not figured out how to make gp4
00:05:06
level technology available to free users
00:05:08
it's something we really want to do it's
00:05:10
just very expensive I take it's very
00:05:12
expensive yeah chamal your thoughts I
00:05:15
think maybe the the two big vectors Sam
00:05:17
that people always talk about is that
00:05:20
underlying cost and sort of the latency
00:05:23
that's kind of rate limited a killer app
00:05:26
and
00:05:27
then I think the second is sort of the
00:05:30
long-term ability for people to build in
00:05:33
an open source World versus a closed
00:05:34
Source world and I think the crazy thing
00:05:37
about this space is that the open source
00:05:40
Community is rabid so one example that I
00:05:43
think is incredible is you know we had
00:05:44
these guys do a pretty crazy demo for
00:05:47
Devon remember like even like five or
00:05:50
six weeks ago that looked incredible and
00:05:52
then some kid just published it under an
00:05:54
open MIT license like open Devon and
00:05:58
it's incredibly
00:06:00
good and almost as good as that other
00:06:02
thing that was closed source so maybe we
00:06:05
can just start with that which is tell
00:06:07
me about the business decision to keep
00:06:09
these models close source and where do
00:06:12
you see things going in the next couple
00:06:14
years so on the first part of your
00:06:16
question um speed and cost those are
00:06:19
hugely important to us and I don't want
00:06:23
to like give a timeline on when we can
00:06:26
bring them down a lot cuz research is
00:06:27
hard but I am confident we'll be able to
00:06:30
um we want to like cut the latency super
00:06:32
dramatically we want to cut the cost
00:06:34
really really dramatically um and I
00:06:37
believe that will happen we're still so
00:06:39
early in the development of the science
00:06:42
and understanding how this works plus we
00:06:43
have all the engineering Tailwinds so
00:06:47
I I don't know like when we get to
00:06:50
intelligence too cheap to meter and so
00:06:52
fast that it feels instantaneous to us
00:06:53
and everything else but I do believe we
00:06:56
can get there for you know a pretty high
00:07:00
level of intelligence and um I it's
00:07:04
important to us it's clearly important
00:07:06
to users and it'll unlock a lot of stuff
00:07:08
on the sort of Open Source close Source
00:07:10
thing I think there's great roles for
00:07:12
both I think
00:07:14
um you know we've open sourced some
00:07:16
stuff we'll open source more stuff in
00:07:17
the future but really like our mission
00:07:20
is to build towards AGI and to figure
00:07:22
out how to broadly distribute its
00:07:24
benefits we have a strategy for that
00:07:26
seems to be resonating with a lot of
00:07:28
people it obviously isn't for everyone
00:07:30
and there's like a big ecosystem and
00:07:32
there will also be open source models
00:07:34
and people who build that way um one
00:07:36
area that I'm particularly interested
00:07:38
personally in open source for is I want
00:07:41
an open source model that is as good as
00:07:43
it can be that runs on my phone and that
00:07:47
I think is going to you know the world
00:07:49
doesn't quite have the technology for
00:07:51
for a good version of that yet but that
00:07:52
seems like a really important thing to
00:07:54
go do at some point will you do will you
00:07:56
do that will you release I don't know if
00:07:57
we will or someone will but someone
00:07:59
llama 3 llama 3 running on a phone well
00:08:03
I guess maybe there's like a seven
00:08:05
billion version yeah yeah uh I don't
00:08:08
know if that's if that will fit on a
00:08:09
phone or not but that should be fitable
00:08:13
on a phone but I don't I I'm not I'm not
00:08:14
sure if that one is like I haven't
00:08:16
played with it I don't know if it's like
00:08:17
good enough to kind of do the thing I'm
00:08:18
thinking about here so when when llama 3
00:08:21
got released I think the big takeaway
00:08:23
for a lot of people was oh wow they've
00:08:24
like caught up to GPT 4 I don't think
00:08:26
it's equal in all Dimensions but it's
00:08:29
like pretty pretty close or pretty in
00:08:31
the ballpark I guess the question is you
00:08:33
know you guys released four a while ago
00:08:36
you're working on five or you know more
00:08:38
upgrades to four I mean I think to
00:08:41
Chamas point about Devin how do you stay
00:08:43
ahead of Open Source I mean it's just
00:08:45
that's like a very hard thing to do in
00:08:47
general right I mean how do you think
00:08:49
about
00:08:50
that what we're trying to do is not
00:08:54
make the sort of
00:08:57
smartest set of Weights that we can can
00:08:59
but what we're trying to make is like
00:09:01
this useful intelligence layer for
00:09:04
people to use and a model is part of
00:09:06
that I think we will stay pretty far
00:09:09
ahead of I hope we'll stay pretty far
00:09:11
ahead of the rest of the world on that
00:09:13
um but there's a lot of other work
00:09:16
around the whole system that's not just
00:09:19
that you know the the model Waits and
00:09:22
we'll have to build up enduring value
00:09:24
the oldfashioned way like any other
00:09:26
business does will have to figure out a
00:09:28
great product and reasons to stick with
00:09:29
it and uh you know delivered at a great
00:09:31
price when you founded the organization
00:09:34
you the stated goal or part of what you
00:09:37
discussed was hey this is too important
00:09:39
for any one company to own it so
00:09:41
therefore it needs to be open then there
00:09:43
was the switch hey it's too dangerous
00:09:45
for anybody to be able to see it and we
00:09:48
need to lock this down because you you
00:09:49
had some fear about that I think is that
00:09:52
accurate because the cynical side is
00:09:54
like well this is a capitalistic move
00:09:56
and then the I think
00:09:59
you know I'm I'm curious what the
00:10:01
decision was here in terms of going from
00:10:03
open we the world needs to see this it's
00:10:06
really important to closed only we can
00:10:08
see it well how did you come to that
00:10:11
that conclusion what were the disc part
00:10:12
of the reason that we released chat GPT
00:10:14
was we want the world to see this and
00:10:15
we've been trying to tell people that AI
00:10:17
is really important and if you go back
00:10:19
to like uh October of 2022 not that many
00:10:22
people thought AI was going to be that
00:10:23
important or that it was really
00:10:24
happening um and a huge part of what we
00:10:28
we try to do
00:10:30
is put the technology in the hands of
00:10:33
people uh now again there's different
00:10:34
ways to do that and I think there really
00:10:35
is an important role to just say like
00:10:37
here's the weights have at it but the
00:10:40
fact that we have so many people using a
00:10:42
free version of chat GPT that we don't
00:10:44
you know we don't run ads on we don't
00:10:46
try to like make money on we just put
00:10:47
out there because we want people to have
00:10:48
these tools um I think it's done a lot
00:10:51
to provide a lot of value and you know
00:10:54
teach people how to fish but also to get
00:10:56
the world um really thoughtful about
00:10:59
what's happening here now we still don't
00:11:01
have all the answers and uh we're
00:11:03
fumbling our way through this like
00:11:04
everybody else and I assume we'll change
00:11:06
strategy many more times as we learn new
00:11:08
things you know when we started open AI
00:11:11
we had really no idea about how things
00:11:13
were going to go um that we'd make a
00:11:15
language model that we'd ever make a
00:11:17
product we started off just I remember
00:11:20
very clearly that first day where we're
00:11:21
like well now we're all here that was
00:11:23
you know it was difficult to get this
00:11:24
set up but what happens now maybe we
00:11:26
should write some papers maybe we should
00:11:28
stand around a whiteboard and we've just
00:11:30
been trying to like put one foot in
00:11:32
front of the other and figure out what's
00:11:33
next and what's next and what's next
00:11:36
and I think we'll keep doing that can I
00:11:39
just replay something and just make sure
00:11:40
I heard it right I think what you were
00:11:42
saying on the open source close Source
00:11:45
thing is if I heard it right all these
00:11:48
models independent of the business
00:11:50
decision you make are going to become
00:11:52
asymptotically accurate towards some
00:11:55
amount of accuracy like not all but like
00:11:57
let's just say there's four or five that
00:11:58
are
00:12:00
well capitalized enough you guys meta
00:12:03
Google Microsoft whomever right so let's
00:12:05
just say four or five maybe one
00:12:08
startup and on the open web and then
00:12:12
quickly the accuracy or the value of
00:12:15
these models will probably shift to
00:12:16
these proprietary sources of training
00:12:18
data that you could get that others
00:12:20
can't or others can get that you can't
00:12:22
is that how you see this thing evolving
00:12:25
where the open web gets everybody to a
00:12:27
certain threshold and then it's just an
00:12:29
arms race for data beyond that doesn't
00:12:33
so I definitely don't think it'll be an
00:12:34
arms race for data because when the
00:12:36
models get smart enough at some point it
00:12:37
shouldn't be about more data at least
00:12:39
not for training it may matter data to
00:12:41
make it useful um look the the one thing
00:12:45
that I have learned most throughout all
00:12:48
this is that uh it's hard to make
00:12:50
confidence statements a couple of years
00:12:51
in the future about where this is all
00:12:53
going to go and so I don't want to try
00:12:54
now I I will say that I I expect lots of
00:12:58
very capable models in the world and you
00:13:01
know like it feels to me like we just
00:13:04
like stumbled on a new fact of nature or
00:13:07
science or whatever you want to call it
00:13:08
which is like we can create you can like
00:13:13
I mean I don't believe this literally
00:13:15
but it's like a spiritual point you know
00:13:18
intelligence is just this emergent
00:13:19
property of matter and that's like a
00:13:21
that's like a rule of physics or
00:13:22
something um so people are going to
00:13:24
figure that out but there will be all
00:13:26
these different ways to design the
00:13:27
systems people will make different Cho
00:13:29
choices figure out new ideas and I'm
00:13:32
sure like you
00:13:36
know like any other industry I would
00:13:39
expect there to be multiple approaches
00:13:41
and different people like different ones
00:13:43
you know some people like iPhones some
00:13:44
people like an Android phone I think
00:13:46
there'll be some effect like that let's
00:13:48
go back to that first section of just
00:13:50
the the cost and the
00:13:53
speed all of you guys are sort of a
00:13:55
little bit rate limited on literally
00:13:58
nvidia's throughput right and I think
00:14:00
that you and most everybody else have
00:14:02
sort of effectively announced how much
00:14:04
capacity you can get just because it's
00:14:06
as much as they can spin out what needs
00:14:08
to happen at the substrate so that you
00:14:10
can actually compute cheaper compute
00:14:14
faster get access to more energy how are
00:14:16
you helping to frame out the industry
00:14:19
solving those problems well we we'll
00:14:22
make huge algorithmic gains for sure and
00:14:24
I don't want to Discount that I you know
00:14:26
I very interested in chips and energy
00:14:28
but if we can make our if we can make a
00:14:30
same quality model twice as efficient
00:14:32
that's like we had twice as much compute
00:14:34
right and I think there's a gigantic
00:14:37
amount of work to be done there uh and I
00:14:42
hope we'll start really seeing those
00:14:44
results um other than that the whole
00:14:47
supply chain is like very complicated
00:14:48
you know there's there's logic Fab
00:14:51
capacity there's how much hbm the world
00:14:53
can make there's how quickly you can
00:14:55
like get permits and pour the concrete
00:14:56
make the data centers and then have
00:14:57
people in there wiring them all up
00:14:59
there's finding the energy which is a
00:15:00
huge bottleneck but
00:15:03
uh I think when there's this much value
00:15:06
to people uh the world will do its thing
00:15:08
we'll try to help it happen faster um
00:15:11
and there's probably like I don't know
00:15:14
how to give it a number but there's like
00:15:15
some percentage chance where there is as
00:15:18
you were saying like a huge substrate
00:15:20
breakthrough and we have like a
00:15:22
massively more efficient way to do
00:15:23
Computing but I don't I don't like Bank
00:15:26
on that or spend too much time thinking
00:15:27
about it what about the device side and
00:15:30
sort of you mentioned sort of the models
00:15:33
that can fit on a phone so obviously
00:15:35
whether that's an llm or some slm or
00:15:37
something I'm sure you're thinking about
00:15:39
that but then does the device itself
00:15:41
change I mean is it does it need to be
00:15:42
as expensive as an iPhone
00:15:47
uh I'm super interested in this uh I I
00:15:50
love like great new form factors of
00:15:52
computing and it feels like with every
00:15:55
major technological Advance a new thing
00:15:57
becomes possible
00:16:00
uh phones are unbelievably good so I
00:16:03
think the threshold is like very high
00:16:04
here like what like I think I think like
00:16:07
I personally think an iPhone is like the
00:16:10
greatest piece of technology Humanity
00:16:13
has ever made it's really a wonderful
00:16:14
product what comes after it like I don't
00:16:17
know I mean I was gonna that was what I
00:16:18
was saying it's so good that to get
00:16:20
Beyond it I think the bar is like quite
00:16:23
High well you've been you've been
00:16:24
working with Johnny IV on on something
00:16:26
right we've been discussing ideas but uh
00:16:29
I don't like if I knew is it that that
00:16:32
it has to be more complicated or
00:16:33
actually just much much cheaper and
00:16:35
simpler well every Mo almost everyone's
00:16:37
willing to pay for a phone anyway so if
00:16:39
you could like make a way cheaper device
00:16:41
I think the barrier to carry a second
00:16:43
thing or use a second thing is pretty
00:16:46
high so I don't think C given that we're
00:16:48
all willing to pay for phones or most of
00:16:50
us are I don't think cheaper is the
00:16:53
answer different is the answer then
00:16:56
would there be like a specialized chip
00:16:58
that would run the phone that was really
00:17:00
good at powering a you know a phone size
00:17:02
AI model probably but the phone
00:17:04
manufacturers are going to do that for
00:17:06
sure that doesn't that doesn't
00:17:07
necessitate a new device I think you'd
00:17:09
have to like find some really different
00:17:11
interaction Paradigm that the technology
00:17:14
enables uh and if I knew what it was I
00:17:17
would be excited to be working on it
00:17:19
right now but well you have you have
00:17:21
voice working right now in the app in
00:17:23
fact I set my action button on my phone
00:17:24
to go directly to chat gpt's voice app
00:17:27
and I use it with my kids and they love
00:17:29
it talking to it's got latency issues
00:17:31
but it's really we'll get we'll we'll
00:17:32
get that we'll get that better and I
00:17:34
think voice is a hint to whatever the
00:17:37
next thing is like if you can get voice
00:17:40
interaction to be really good it
00:17:42
feels I think that feels like a
00:17:44
different way to use a computer but
00:17:46
again with that by the way like what why
00:17:49
is it not responsive and you know it
00:17:52
feels like a CB you know like over over
00:17:55
it's really annoying to use you know uh
00:17:58
in that way but it's also brilliant when
00:18:00
it gives you the right answer we are
00:18:01
working on that uh it's it's so clunky
00:18:04
right now it's slow it's like kind of
00:18:07
doesn't feel very smooth or authentic or
00:18:09
organic like we'll get all that to be
00:18:12
much
00:18:13
better what about computer vision I mean
00:18:16
they have glasses or maybe you could
00:18:18
wear a pendant I mean you take the
00:18:19
combination
00:18:21
of visual or video data combine it with
00:18:24
voice and now ai knows everything that's
00:18:27
happening around you super powerful to
00:18:29
be able to like the multimodality of
00:18:32
saying like hey cha gbt what am I
00:18:35
looking at or like what kind of plant is
00:18:37
this I can't quite tell
00:18:39
um that's obvious that that's like a
00:18:41
that's another I think like hint but
00:18:44
whether people want to wear glasses or
00:18:45
like hold up something when they want
00:18:47
that like I there's a bunch of just like
00:18:50
like the the sort of like societal
00:18:53
interpersonal issues here are all very
00:18:56
complicated about wearing a computer on
00:18:57
your face um we we saw that with Google
00:19:00
class people got punched in the face in
00:19:02
the mission started a lot of I forgot
00:19:04
about that I forgot about that so so I I
00:19:06
think it's
00:19:07
like what are the apps that could be
00:19:10
unlocked if AI was sort of ubiquitous on
00:19:13
people's
00:19:14
phones do you have a sense of that or
00:19:17
what would you want to see built
00:19:21
uh I I think what I want is just this
00:19:24
always on like super low friction
00:19:29
thing where I
00:19:31
can either by voice or by text or
00:19:33
ideally like some other it just kind of
00:19:36
knows what I want have this like
00:19:38
constant thing helping me throughout my
00:19:39
day that's got like as much context on
00:19:41
as possible it's like the world's
00:19:43
greatest assistant um and it's just this
00:19:45
like thing working to make me better and
00:19:48
better uh there's there there's like a I
00:19:51
know when you hear people like talk
00:19:52
about the AI future they're imag they
00:19:54
imagine there's sort of
00:19:56
two different approaches and don't sound
00:19:59
that different but I think they're like
00:20:00
very different for how we'll design the
00:20:01
system in practice there's the I want an
00:20:04
extension of myself um I want like a
00:20:09
ghost or an alter ego or this thing that
00:20:11
really like is me is acting on my behalf
00:20:14
is um responding to emails not even
00:20:17
telling me about it is is is sort of
00:20:20
like it it becomes more me and is me and
00:20:24
then there's this other thing which is
00:20:25
like I want a great senior employee
00:20:29
it may get to know me very well I may
00:20:31
delegate it you know you can like have
00:20:32
access to my email and I'll tell you the
00:20:34
constraints but but I think of it as
00:20:36
this like separate entity and I
00:20:40
personally like the separate entity
00:20:42
approach better and think that's where
00:20:44
we're going to head um and so in that
00:20:48
sense the thing is not you but it's it's
00:20:51
like a always available always great
00:20:54
super capable assistant executive agent
00:20:57
in a way like gets out there working on
00:20:59
your behalf and understands what you
00:21:01
want and anticipates what you want is
00:21:04
what I'm reading into what you're saying
00:21:05
I think there'd be agent like Behavior
00:21:08
but there's like a difference
00:21:10
between a senior employee and an agent
00:21:13
yeah and like I want it you know I think
00:21:16
of like my I think like of it like one
00:21:19
of the things that I like about a senior
00:21:22
employee is
00:21:26
they'll they'll push back on me they
00:21:28
will sometimes not do something I ask or
00:21:31
they sometimes will say like I can do
00:21:32
that thing if you want but if I do it
00:21:34
here's what I think would happen and
00:21:35
then this and then that and are you
00:21:37
really
00:21:38
sure I definitely want that kind of vibe
00:21:41
which not not just like this thing that
00:21:42
I asking it blindly does it can reason
00:21:46
yeah yeah and push back it can reason it
00:21:47
has like the kind of relationship with
00:21:49
me that I would expect out of a really
00:21:51
competent person that I worked with
00:21:54
which is different from like a sycophant
00:21:56
yeah the thing in that world where if
00:21:58
you had this like Jarvis like thing that
00:22:01
can reason what do you think it does to
00:22:06
products that you use today where the
00:22:08
interface is very valuable so for
00:22:10
example if you look at an instacart or
00:22:12
if you look at an Uber or if you look at
00:22:14
a door Dash these are not services that
00:22:16
are meant to be pipes that are just
00:22:19
providing a set of apis to a Smart Set
00:22:22
of agents that ubiquitously work on
00:22:24
behalf of 8 billion
00:22:25
people what do you think has to change
00:22:28
in how we think about how apps need to
00:22:29
work of how this entire infrastructure
00:22:31
of experiences need to work in a world
00:22:33
where you're agentically interfacing to
00:22:35
the world you I'm actually very
00:22:37
interested in designing a world that is
00:22:40
equally usable by humans and by AIS so I
00:22:44
I
00:22:46
I I I like the interpretability of that
00:22:49
I like the smoothness of the handoffs I
00:22:51
like the ability that we can provide
00:22:52
feedback or whatever so you know door
00:22:55
Dash could just expose some a pi to my
00:22:59
future AI assistant and they could go
00:23:02
put the order and whatever or I could
00:23:04
say like I could be holding my phone and
00:23:06
I could say okay AI assistant like you
00:23:08
put in this order on door Dash please
00:23:11
and I could like watch the app open and
00:23:12
see the thing clicking around and I
00:23:14
could say hey no not this or like um
00:23:17
there there's something about designing
00:23:19
a world that is
00:23:23
usable equally well by humans and AIS
00:23:25
that I think is a interesting concept
00:23:28
excited humanoid robots than sort of
00:23:31
robots of like very other shapes the
00:23:33
world is very much designed for humans
00:23:34
and I think we should absolutely keep it
00:23:36
that way and a shared interface is nice
00:23:39
so you see voice chat that modality kind
00:23:42
of gets rid of apps you just ask it for
00:23:44
sushi it knows Sushi you like before it
00:23:46
knows what you don't like and does its
00:23:48
best shot at doing it I it's hard for me
00:23:50
to imagine that we just go to a world
00:23:53
totally where you say like hey Chachi BT
00:23:55
order me sushi and it says okay do you
00:23:57
want it from this restaurant what kind
00:23:59
what time whatever I think user I think
00:24:02
visual user interfaces are super good
00:24:05
for a lot of things um and it's hard of
00:24:07
me to imagine like a world
00:24:10
where youd never look at a screen and
00:24:15
just use voice mode only but I I can't
00:24:18
imagine that for a lot of things yeah I
00:24:20
mean Apple tried with Siri like you
00:24:22
supposedly you can order an Uber
00:24:23
automatically with Siri I don't think
00:24:25
anybody's ever done it because it's why
00:24:27
would you take the risk of not the
00:24:29
quality to your point the quality is not
00:24:31
good but when the quality is good enough
00:24:33
you're you'll actually prefer it just
00:24:35
because it's just lighter weight you
00:24:36
don't have to take your phone out you
00:24:37
don't have to search for your app and
00:24:40
press it and oh it automatically logged
00:24:42
you out oh hold on log back in oh TFA
00:24:44
it's a whole pain in the ass you know
00:24:46
it's like setting a timer with Siri I do
00:24:48
every time because it works really well
00:24:50
and it's great and I don't need more
00:24:52
information but ordering an Uber like I
00:24:55
want to see the prices for a few
00:24:57
different options I want to see how far
00:24:58
away it is I want to see like maybe even
00:25:01
where they are on the map because I
00:25:02
might walk somewhere I get a lot more
00:25:04
information by I think in less time by
00:25:07
looking at that order the Uber screen
00:25:09
than I would if I had to do that all
00:25:10
through the audio Channel I like your
00:25:12
idea of watching it happen that's kind
00:25:14
of cool I think there will just be like
00:25:16
yeah
00:25:18
different there are different interfaces
00:25:20
we use for different tasks and I think
00:25:21
that'll keep going of all the developers
00:25:23
that are building apps and experiences
00:25:26
on open AI are there few that stand out
00:25:29
for you where you're like okay this is
00:25:30
directionally going in a super
00:25:32
interesting area even if it's like a toy
00:25:34
app but are there things that you guys
00:25:36
point to and say this is really
00:25:39
important um I met with a new company
00:25:44
this morning or and bar even a company
00:25:45
it's like two people that are going to
00:25:46
work on a summer project trying to
00:25:48
actually finally make the AI tutor like
00:25:52
and I've always been interested in this
00:25:53
space a lot of people have done great
00:25:54
stuff on our platform but if if someone
00:25:57
can deliver
00:25:58
like the way that you actually
00:26:03
like H they used a phrase I love which
00:26:05
is this is going to be like a monor
00:26:07
level reinvention for how people how
00:26:08
people learn things yeah um but if you
00:26:11
can like find this new way to like let
00:26:12
people explore and learn and new ways on
00:26:14
their own I'm personally super excited
00:26:17
about that um a lot of the coding
00:26:20
related stuff you mentioned Devon
00:26:21
earlier I think that's like a super cool
00:26:23
vision of the future the thing that I am
00:26:26
Health Healthcare I I believe
00:26:28
should be pretty transformed by this but
00:26:32
the thing I'm personally most excited
00:26:34
about is the sort of doing faster and
00:26:38
better scientific discovery gp4 clearly
00:26:40
not there in a big way although maybe it
00:26:42
accelerates things a little bit by
00:26:44
making scientists more
00:26:45
productive but Alpha 43 yeah that's like
00:26:50
but Sam that will be a Triumph those are
00:26:54
not like these these models are train
00:26:58
Tred and built differently than the
00:27:01
language models I mean to some obviously
00:27:04
there's a lot that's similar but there's
00:27:06
a lot um there's kind of a groundup
00:27:08
architecture to a lot of these models
00:27:09
that are being applied to these specific
00:27:12
problem sets these specific applications
00:27:14
like chemistry interaction modeling for
00:27:17
example does you you'll need some of
00:27:20
that for sure but the the thing that I
00:27:21
think we're missing across the board for
00:27:24
many of these things we've been talking
00:27:25
about is models that can do reason
00:27:28
and once you have reasoning you can
00:27:30
connect it to chemistry simulators or
00:27:32
guess yeah that's the important question
00:27:34
I wanted to kind of talk about today was
00:27:37
this idea
00:27:38
of networks of models people talk a lot
00:27:41
about agents as if there's kind of this
00:27:44
linear set of call functions that happen
00:27:47
but one of the things that
00:27:49
arises in biology is networks of systems
00:27:53
that have cross interactions that the
00:27:56
aggregation of the system the
00:27:58
aggregation of the network produces an
00:28:00
output rather than one thing calling
00:28:02
another that thing calling another do we
00:28:04
see like an emergence in this
00:28:06
architecture of either specialized
00:28:07
models or network models that work
00:28:10
together to address bigger problem sets
00:28:12
use reasoning there's computational
00:28:14
models that do things like chemistry or
00:28:16
arithmetic and there's other models that
00:28:17
do rather than one model to rule them
00:28:20
all that's purely
00:28:23
generalized I don't know um
00:28:29
I don't know how much
00:28:31
reasoning is going to turn out to be a
00:28:33
super generalizable thing I suspect it
00:28:36
will but that's more just like an
00:28:38
intuition and a hope and it would be
00:28:40
nice if it worked out that way I I don't
00:28:42
know if that's
00:28:45
like but let's walk through the the
00:28:47
protein modeling
00:28:49
example there's a bunch of training data
00:28:52
images of proteins and then sequence
00:28:55
data and they build a model predictive
00:28:57
model and they have a set of processes
00:28:59
and steps for doing that do you envision
00:29:01
that there's this artificial general
00:29:04
intelligence or this great reasoning
00:29:05
model that then figures out how to build
00:29:07
that submodel that figures out how to
00:29:08
solve that problem by acquiring the
00:29:10
necessary data and then resolving there
00:29:13
so many ways where that could go like
00:29:15
maybe it is it trains a literal model
00:29:17
for it or maybe it just like knows the
00:29:20
one big model what it can like go pick
00:29:22
what other training data it needs and
00:29:24
ask a question and then update on that
00:29:27
um I guess the real question is are all
00:29:29
these startups going to die because so
00:29:31
many startups are working in that
00:29:32
modality which is go get special data
00:29:34
and then train a new model on that
00:29:36
special data from the ground up and then
00:29:38
it only does that one sort of thing and
00:29:40
it works really well at that one thing
00:29:42
and it works better than anything else
00:29:43
at that one thing you know there there's
00:29:44
like a version of this I think you can
00:29:49
like already see when you were when you
00:29:51
were talking about like biology and
00:29:53
these complicated networks of systems
00:29:54
the reason I was smiling I I got super
00:29:56
sick recently and I'm mostly better now
00:30:00
but it was just like body like got beat
00:30:02
up like one system at a time F like you
00:30:04
can really tell like okay it's this
00:30:05
cating thing and uh and that reminded me
00:30:10
of you like talking about the like
00:30:11
biology is just these like you have no
00:30:12
idea how much these systems interact
00:30:14
with each other until things start going
00:30:15
wrong and that was sort of like
00:30:17
interesting to see but I was using
00:30:21
um I was like using chat GPT uh to try
00:30:24
to like figure out like what was
00:30:26
happening whatever and and and would say
00:30:28
well I'm you know unsure of this one
00:30:29
thing and then I just like posted a
00:30:31
paper on it without even reading the
00:30:33
paper um like in the context and it says
00:30:36
oh that was the thing I was unsure of
00:30:37
like now I think this instead so there's
00:30:40
like a that was like a small version of
00:30:41
what you're talking about where you can
00:30:44
like can say this I don't I don't know
00:30:46
this thing you can put more information
00:30:47
you don't retrain the model you're just
00:30:48
adding it to the context here and now
00:30:51
you're getting a so these models that
00:30:53
are predicting protein structure like
00:30:55
let's say right this is the whole basis
00:30:57
and now now other molecules at Alpha
00:31:01
3 can they can yeah I mean is it
00:31:05
basically a world where the best
00:31:07
generalized model goes in and gets that
00:31:09
training data and then figures out on
00:31:11
its own and maybe you could maybe you
00:31:13
could use an example for us can you tell
00:31:15
us about Sora your video model that
00:31:17
generates amazing moving images moving
00:31:20
video and and what's different about the
00:31:22
architecture there whatever you're
00:31:24
willing to share on how how that is
00:31:27
different
00:31:28
yeah so my on the general thing first
00:31:34
my you clearly will need
00:31:38
specialized simulators connectors pieces
00:31:42
of data whatever but my
00:31:45
intuition and again I don't have this
00:31:47
like backed up with science my intuition
00:31:49
would be if we can figure out the core
00:31:51
of generalized reasoning connecting that
00:31:53
to new problem domains in the same way
00:31:56
that humans are generalized reason
00:31:59
ERS would I think be be doable it's like
00:32:02
a fast unlock faster unlock than I think
00:32:05
I I think so
00:32:10
um but yeah you Sora like does not start
00:32:13
with a language model um it's that
00:32:15
that's a model that is like customized
00:32:17
to do video uh and and so like we're
00:32:21
clearly not at that world yet right so
00:32:24
you guys so just as an example for you
00:32:26
guys to build a good
00:32:28
video model you built it from scratch
00:32:30
using I'm assuming some different
00:32:33
architecture and different data but in
00:32:36
the future the generalized reasoning
00:32:39
system the the AGI whatever system
00:32:42
theoretically could render that by
00:32:44
figuring out how to do it yeah I mean
00:32:46
one example of this is like okay you
00:32:48
know as far as I know all the best text
00:32:51
models in the world are still a lot of
00:32:52
regressive models and the best image and
00:32:54
video models are diffusion models and
00:32:56
that's like sort strange in some sense
00:32:59
yeah yeah so there's a big debate about
00:33:04
uh training data you guys have been I
00:33:06
think the most thoughtful of any company
00:33:08
you've got licensing deals now ft Etc
00:33:11
and we got to gu be gentle here because
00:33:14
you're involved in a New York Times
00:33:16
lawsuit you weren't able to settle I
00:33:17
guess an arrangement with them for
00:33:19
training data how do you think about
00:33:23
fairness in fair
00:33:24
use we've had big debates here on the
00:33:27
pod
00:33:28
obviously your actions are you know
00:33:30
speak volumes that you're trying to be
00:33:32
fair by doing licensing deals so what
00:33:35
what's your personal position on the
00:33:37
rights of artists who create beautiful
00:33:40
music lyrics books and you taking that
00:33:45
and then making a derivative product out
00:33:46
of it and and then monetizing it and and
00:33:49
what's fair here and how do we get to a
00:33:52
world where you know artists can make
00:33:55
content in the world and then decide
00:33:57
what they want other people to do with
00:33:59
it yeah and and I'm just curious your
00:34:01
personal belief because I know you to be
00:34:02
a thoughtful person on this and I know a
00:34:04
lot of other people in our industry are
00:34:07
not very thoughtful about how they think
00:34:08
about content
00:34:10
creators so I think it's very different
00:34:12
for different kinds of I mean look on
00:34:14
unfair use I think we have a a very
00:34:16
reasonable position under the current
00:34:18
law but I think AI is so different that
00:34:22
for things like art we'll need to think
00:34:23
about them in different ways but I would
00:34:25
say if you go read a bunch of math on
00:34:29
the internet and learn how to do math
00:34:34
that I think seems unobjectionable to
00:34:37
most people and then there's like you
00:34:38
know another set of people who might
00:34:40
have a different opinion well what if
00:34:41
you
00:34:45
like actually let me not get into that
00:34:47
just in the interest of not making this
00:34:48
answer too long so I think there's like
00:34:50
one category people are like okay
00:34:51
there's like generalized human knowledge
00:34:54
you can kind of like go if you learn
00:34:56
that like that's
00:34:58
that that's like open domain or
00:34:59
something if you kind of go learn about
00:35:01
the Pythagorean theorem um that's one
00:35:04
end of the spectrum and then I think the
00:35:06
other extreme end of the spectrum is um
00:35:10
is Art and maybe even like more than
00:35:14
more specifically I would say it's like
00:35:16
doing it's a system generating art in
00:35:19
the style or the likeness of another
00:35:21
artist um would be kind of the furthest
00:35:24
end of
00:35:25
that and then there's many many cases on
00:35:28
the Spectrum in between
00:35:31
uh I think the conversation has been
00:35:33
historically very caught up on training
00:35:36
data but it will increasingly become
00:35:38
more about what happens at inference
00:35:40
time as training data
00:35:43
becomes less valuable and the what the
00:35:48
system does accessing you know
00:35:52
information in in context in real time
00:35:55
or uh you know know taking like like
00:35:58
something like that what happens at
00:36:00
inference time will become more debated
00:36:02
and and how the what the new economic
00:36:04
model is there so if you say like
00:36:08
uh if you say like create me a song in
00:36:11
this in the style of Taylor
00:36:13
Swift even if the model were never
00:36:16
trained on any Taylor Swift songs at all
00:36:19
you can still have a problem which is it
00:36:21
may have read about Taylor Swift it may
00:36:22
know about her themes Taylor Swift means
00:36:24
something and then and then the question
00:36:26
is like that model even if it were never
00:36:28
trained on any Taylor Swift song
00:36:30
whatsoever be allowed to do that and if
00:36:33
so um how should Taylor get paid right
00:36:38
so I think there's an optin opt out in
00:36:40
that case first of all and then there's
00:36:41
an economic model um staying on the
00:36:43
music example there is something
00:36:45
interesting to look at from the
00:36:48
historical perspective here which is uh
00:36:50
sampling and how the economics around
00:36:52
that work this is not quite the same
00:36:54
thing but it's like an interesting place
00:36:55
to start looking Sam let me just
00:36:56
challenge that what's the difference in
00:36:59
the example you're giving of the model
00:37:01
learning about things like song
00:37:03
structure Tempo Melody Harmony
00:37:06
relationships all the discovering all
00:37:09
the underlying structure that makes
00:37:10
music successful and then building new
00:37:13
music using training data and what a
00:37:16
human does that listens to lots of music
00:37:19
learns about and and their brain is
00:37:21
processing and building all those same
00:37:23
sort of predictive models or those same
00:37:25
sort of uh discoveries understandings
00:37:28
what's the difference here and why why
00:37:30
are you making the case that perhaps
00:37:33
artists should be uniquely paid this is
00:37:36
not a sampling situation you're not the
00:37:37
AI is not outputting and it's not
00:37:39
storing in the model the actual original
00:37:41
song it's learning structure right I
00:37:44
wasn't trying to make that that point
00:37:46
because I agree like in the same way
00:37:47
that humans are inspired by other humans
00:37:50
I was saying if you if you say generate
00:37:51
me a song in the style of Taylor Swift I
00:37:54
see right okay where the prompt
00:37:56
leverages some artist I I think
00:37:59
personally that's a different case would
00:38:01
you be comfortable asking or would you
00:38:03
be comfortable letting the model train
00:38:05
itself well a music model being trained
00:38:08
on the whole Corpus of music that humans
00:38:09
have created without royalties being
00:38:12
paid to the artists that um that music
00:38:15
is being fed in and then you're not
00:38:17
allowed to ask you know artists specific
00:38:19
prompts you could just say hey pay me a
00:38:20
play me a a really cool pop song that's
00:38:23
fairly modern about heartbreak you know
00:38:25
with a female voice you know we have
00:38:27
currently made the decision not to do
00:38:30
music and partly because exactly these
00:38:32
questions of where you draw the lines
00:38:34
and you know what like
00:38:37
even I was meeting with several
00:38:39
musicians I really admire recently I was
00:38:41
just trying to like talk about some of
00:38:42
these edge cases but even the world in
00:38:46
which if
00:38:48
we went and let's say we paid 10,000
00:38:53
musicians to create a bunch of music
00:38:55
just to make a great training set where
00:38:57
the music music model could learn
00:38:58
everything about strong s structure um
00:39:03
and what makes a good catchy beat and
00:39:06
everything else um and only trained on
00:39:09
that let's say we could still make a
00:39:10
great music model which maybe maybe we
00:39:11
could um you know I was kind of like
00:39:14
posing that as a thought experiment to
00:39:15
musicians and they're like well I can't
00:39:17
object to that on any principal basis at
00:39:18
that point um and yet there's still
00:39:21
something I don't like about it now
00:39:22
that's not a reason not to do it um
00:39:25
necessarily but it is
00:39:28
did you see that ad that Apple put out
00:39:30
maybe it was yesterday or something of
00:39:32
like squishing all of human creativity
00:39:34
down into one really thin iPad what was
00:39:35
your take on
00:39:37
it uh people got really emotional about
00:39:40
it yeah it's stronger reaction than you
00:39:42
would think yeah there's something
00:39:47
about I'm obviously hugely positive on
00:39:50
AI but there is something that I think
00:39:53
is beautiful about human creativity and
00:39:55
human artistic expression and and you
00:39:57
know for an AI that just does better
00:39:59
science like great bring that on but an
00:40:01
AI that is going to do this like deeply
00:40:03
beautiful human creative expression I
00:40:05
think we should like figure out it's
00:40:08
going to happen it's going to be a tool
00:40:10
that will lead us to Greater creative
00:40:12
Heights but I think we should figure out
00:40:13
how to do it in a way that like
00:40:15
preserves the spirit of what we all care
00:40:17
about here and I I think your actions
00:40:20
speak loudly we were trying to do Star
00:40:25
Wars characters in dolly
00:40:27
and if you ask for Darth Vader it says
00:40:30
hey we can't do that so you've I guess
00:40:31
red teed or whatever you call it
00:40:34
internally yeah you're not allowing
00:40:36
people to use other people's IP so
00:40:38
you've taken that decision now if you
00:40:40
asked it to make a Jedi Bulldog or a
00:40:42
Sith Lord Bulldog which I did it made my
00:40:45
Bulldogs a Sith bulldogs so there's an
00:40:47
interesting question about your spectrum
00:40:49
right yeah you know we put out this
00:40:50
thing yesterday called the spec um where
00:40:53
we're trying to say here are here's
00:40:56
here's how our model is supposed to
00:40:58
behave and it's very hard it's a long
00:41:01
document it's very hard to like specify
00:41:03
exactly in each case where the limit
00:41:05
should be and I view this as like a
00:41:07
discussion that's going to need a lot
00:41:08
more input um but but these sorts of
00:41:12
questions
00:41:14
about okay maybe it shouldn't generate
00:41:16
Darth Vader but the idea of a Sith Lord
00:41:18
or a Sith style thing or Jedi at this
00:41:20
point is like part of the culture like
00:41:22
like these are these are all hard
00:41:25
decisions yeah and and I think you're
00:41:27
right the music industry is going to
00:41:29
consider this opportunity to make tlor
00:41:31
Swift songs their opportunity it's part
00:41:33
of the four-part fair use test is you
00:41:36
know these who gets to capitalize on new
00:41:39
Innovations for existing art and and
00:41:41
Disney has an argument that hey you know
00:41:44
if if you're GNA make Sora versions of a
00:41:47
aoka or whatever Obi-Wan Kenobi that's
00:41:49
Disney's opportunity and that's a great
00:41:51
partnership for
00:41:53
you you know to pursue so we're I think
00:41:56
this section I would label as AI in the
00:41:58
law so let me ask maybe a higher level
00:42:01
question what does it mean when people
00:42:04
say regulate AI totally Sam what does it
00:42:07
what does that even mean and comment on
00:42:09
California's new proposed regulations as
00:42:12
well a few if you're up for it uh I'm
00:42:15
concerned I mean there's so many
00:42:17
proposed regulations but most of the
00:42:18
ones I've seen on the California state
00:42:20
things I'm concerned about I also have a
00:42:22
general fear of the states all doing
00:42:24
this them themselves um when people say
00:42:27
regulate AI I don't think they mean one
00:42:30
thing I think there's like some people
00:42:31
are like B the whole thing some people
00:42:33
like don't allow it to be open source
00:42:35
required to be open source um the thing
00:42:37
that I am personally most interested in
00:42:40
is I think there will
00:42:43
come look I may be wrong about this I
00:42:45
will acknowledge that this is a
00:42:46
forward-looking statement and those are
00:42:48
always dangerous to make but I think
00:42:49
there will come a time in the not super
00:42:52
distant future like you know we're not
00:42:53
talking like decades and decades from
00:42:55
now where AI say the frontier AI systems
00:42:58
are capable of causing
00:43:02
significant Global harm and for those
00:43:06
kinds of systems in the same way we have
00:43:08
like Global oversight of nuclear weapons
00:43:11
or synthetic bio or things that can
00:43:13
really like have a very negative impact
00:43:15
Way Beyond the realm of one country uh I
00:43:18
would like to see some sort of
00:43:20
international agency that is looking at
00:43:22
the most powerful systems and ensuring
00:43:24
like reasonable safety testing you know
00:43:26
these things things are not going to
00:43:28
escape and recursively self-improve or
00:43:30
whatever the
00:43:32
criticism of this is that you're you
00:43:35
have the resources to Cozy up to Lobby
00:43:39
to be involved and you've been very
00:43:40
involved with politicians and then
00:43:42
startups which are also passionate about
00:43:44
and invest in um are not going to have
00:43:46
the ability to Resource uh and deal with
00:43:49
this and that this regulatory capture as
00:43:51
per our friend you know Bill Gurley did
00:43:53
a great talk last year about it so maybe
00:43:55
you could address that headon do do you
00:43:57
if the line were we're only going to
00:43:59
look at models that are trained on
00:44:01
computers that cost more than 10 billion
00:44:03
or more than 100 billion or whatever
00:44:04
dollars I'd be fine with that there'd be
00:44:07
some line that'd be fine and uh I don't
00:44:10
think that puts any regulatory burden on
00:44:11
startups so if you have like the the
00:44:14
nuclear raw material to make a nuclear
00:44:16
bomb like there's a small subset set of
00:44:18
people who have that therefore you use
00:44:19
the analogy of like a a nuclear
00:44:21
inspectors kind of situation yeah I
00:44:23
think that that's interesting sax you
00:44:26
have a question well go ahead you had to
00:44:28
follow can I say one more thing about
00:44:30
that of course I'd be super nervous
00:44:33
about regulatory overreach here I think
00:44:34
we can get this wrong by doing way too
00:44:36
much I or even a little too much I think
00:44:38
we can get this wrong by doing not
00:44:39
enough
00:44:41
but but I do think part of and
00:44:45
I and now I mean you know we have seen
00:44:49
regulatory overstepping or capture just
00:44:52
get super bad in other areas um and you
00:44:56
know like also maybe nothing will happen
00:44:58
but but I think it is part of our duty
00:45:00
and our mission to like talk about what
00:45:03
we believe is likely to happen and what
00:45:06
it takes to get that right the challenge
00:45:08
Sam is that we have statute that is
00:45:11
meant to protect people protects Society
00:45:14
at large what we're creating however a
00:45:17
statute that gives the government rights
00:45:20
to go in and audit code to audit
00:45:24
business um trade Secrets uh we've never
00:45:29
seen that to this degree before
00:45:31
basically the California legislation
00:45:33
that's proposed and some of the federal
00:45:34
legislation that's been proposed
00:45:36
basically requires the fed the
00:45:39
government to audit a model to audit
00:45:41
software to audit and review the
00:45:43
parameters and the weightings of the
00:45:44
model and then you need their check mark
00:45:47
in order to deploy it for commercial or
00:45:51
public use and for me it just feels like
00:45:55
we're trying to
00:45:57
reigning the the the government agencies
00:46:00
for fear and and because folks have a
00:46:03
hard time understanding this and are
00:46:04
scared about the implications of it they
00:46:06
want to control it and because they want
00:46:08
and the only way to control it is to say
00:46:10
give me a right to audit before you can
00:46:11
release it asess these people cluel I
00:46:14
mean the way that the the stuff is
00:46:15
written you read it you're like G to
00:46:16
pull your hair out because as you know
00:46:18
better than anyone in 12 months none of
00:46:20
this stuff's going to make sense anyway
00:46:21
totally right look the reason I have
00:46:23
pushed for an agency based approach for
00:46:26
for for kind of like the big picture
00:46:28
stuff and not a like write it in laws I
00:46:30
don't in 12 months it will all be
00:46:32
written wrong and I don't think even if
00:46:35
these people were like True World
00:46:38
experts I don't think they could get it
00:46:39
right looking at 12 or 24 months um and
00:46:43
I don't these policies which is like
00:46:45
we're going to look at you know we're
00:46:47
going to audit all of your source code
00:46:48
and like look at all of your weights one
00:46:50
by one like yeah I think there's a lot
00:46:53
of crazy proposals out there um by the
00:46:55
way especially if the models are always
00:46:57
retrained all the time if they become
00:46:58
more Dynamic again this is why I think
00:47:00
it's yeah but but like when before an
00:47:03
airplane gets certified there's like a
00:47:05
set of safety tests we put the airplane
00:47:07
through it um and totally it's different
00:47:10
than reading all of your code that's
00:47:12
reviewing the output of the model not
00:47:14
reviewing the insides of the model and
00:47:17
and so what I was goingon to say is I
00:47:18
that is the kind of thing that I think
00:47:21
as safety testing makes sense how are we
00:47:24
going to get that to happen Sam and I'm
00:47:26
not just speaking for open AI I speak
00:47:28
for the industry for for Humanity
00:47:29
because I am concerned that we draw
00:47:32
ourselves into almost like a dark ages
00:47:34
type of era by restricting the growth of
00:47:37
these incredible technologies that can
00:47:39
prosper human that that Humanity can
00:47:41
prosper from so significantly how do we
00:47:43
change the the sentiment and get that to
00:47:45
happen because this is all moving so
00:47:46
quickly at the government levels and
00:47:48
folks seem to be getting it wrong and
00:47:50
I'm I'm just to build on that Sam the
00:47:53
architectural decision for example that
00:47:55
llama took is pretty interesting in that
00:47:58
it's like we're going to let llama grow
00:48:01
and be as unfettered as possible and we
00:48:02
have this other kind of thing that we
00:48:04
call Llama guard that's meant to be
00:48:06
these protective guard rails is that how
00:48:08
you see the problem being solved
00:48:10
correctly or do you see at the current
00:48:12
at the current strength of models
00:48:15
definitely some things are going to go
00:48:16
wrong and I don't want to like make
00:48:18
light of those or not take those
00:48:20
seriously but I'm not like I don't have
00:48:22
any like catastrophic risk worries with
00:48:25
a gp4 level model um and I think there's
00:48:29
many safe ways to choose to deploy this
00:48:34
uh may maybe we'd find more common
00:48:37
ground if we said that uh and I like you
00:48:40
know the specific example of models that
00:48:44
are capable that are technically capable
00:48:46
not even if they're not going to be used
00:48:48
this way of recursive
00:48:51
self-improvement um or
00:48:54
of you know autonomously
00:48:57
designing and deploying a bioweapon or
00:49:00
something like that or a new model that
00:49:03
was the recursive self-improvement Point
00:49:05
um you know we should have safety
00:49:08
testing on the outputs at an
00:49:10
international level for models that you
00:49:12
know have a reasonable chance of of
00:49:14
posing a threat there uh I don't think
00:49:17
like GPT 4 of course does
00:49:22
not POS any sort of well I want to say
00:49:26
any sort cuz we don't yeah I don't think
00:49:29
the gp4 poses a material threat on those
00:49:31
kinds of things and I think there's many
00:49:33
safe ways to release a model like this
00:49:35
um but you know when like significant
00:49:39
loss of human life is a serious
00:49:43
possibility like airplanes
00:49:45
or any number of other examples where I
00:49:48
think we're happy to have some sort of
00:49:49
testing framework like I don't think
00:49:51
about an airplane when I get on it I
00:49:52
just assume it's going to be safe right
00:49:55
right there's a lot of hand ringing
00:49:56
right now Sam about
00:49:58
jobs and you had a lot of I think you
00:50:00
did like some sort of a test when you
00:50:02
were at YC about Ubi and you've been
00:50:04
result in that come out very soon I just
00:50:06
it was a fiveyear study that wrapped up
00:50:09
um or started five years ago well there
00:50:12
was like a beta study first and then it
00:50:13
was like a long one that ran but uhk can
00:50:16
you explain yeah why did you start why'
00:50:17
you start it maybe just explain Ubi and
00:50:19
why you started it um so we started
00:50:22
thinking about this in 2016 uh kind of
00:50:25
about the same time started taking AI
00:50:27
really seriously and the theory was that
00:50:31
the magnitude of the change that may
00:50:34
come to society
00:50:37
and jobs in the economy and and sort of
00:50:40
in some deeper sense than that like what
00:50:41
the social contract looks like
00:50:45
um meant that we should have many
00:50:47
studies to study many ideas about new
00:50:50
new ways to
00:50:52
arrange that um I also think that you
00:50:56
know I'm not like a super fan of how the
00:50:58
government has handled most policies
00:51:01
designed to help poor people and I kind
00:51:03
of believe that if you could just give
00:51:06
people money they would make good
00:51:08
decisions and the market would do its
00:51:09
thing and you know I'm very much in
00:51:12
favor of lifting up the floor and
00:51:15
reducing eliminating poverty um but I'm
00:51:18
interested in better ways to do that
00:51:19
than what we have tried for the existing
00:51:22
social safety net and and kind of the
00:51:24
way things have been handled and I think
00:51:26
people money is not going to go solve
00:51:29
all problems it's certainly not going to
00:51:30
make people happy but it
00:51:33
might it might solve some problems and
00:51:36
it might give people a better Horizon
00:51:40
with which to help themselves and I'm
00:51:42
interested in that I I think
00:51:44
that now that we see some of the ways so
00:51:48
2016 was a very long time ago uh you
00:51:50
know now that we see some of the ways
00:51:52
that AI is developing I wonder if
00:51:54
there's better things to do than the
00:51:57
[Music]
00:51:58
traditional um conceptualization of Ubi
00:52:02
uh like I
00:52:03
wonder I wonder if the future looks
00:52:05
something like more like Universal basic
00:52:07
compute than Universal basic income and
00:52:09
everybody gets like a slice of gpt7
00:52:12
compute and they can use it they can
00:52:14
resell it they can donate it to somebody
00:52:16
to use for cancer research but but what
00:52:18
you get is not dollars but this like
00:52:21
productivity slice yeah you own like
00:52:22
part of the productivity right I would
00:52:24
like to shift to the gossip part of this
00:52:27
gossip what gossip let's go back let's
00:52:30
go back to November what the flying
00:52:35
happened um you know I I if you have
00:52:39
specific questions I'm happy to maybe
00:52:41
I'll answer maybe you said you were
00:52:43
going to talk about it at some point so
00:52:44
here's the point what the hell happened
00:52:47
you were fired you came back it was a
00:52:49
palace Intrigue did somebody stab you in
00:52:52
the back did you find AGI what's going
00:52:54
on tell us this is a safe face
00:52:57
Sam um I was fired I
00:53:02
was I talked about coming back I kind of
00:53:05
was a little bit unsure at the moment
00:53:07
about what I wanted to do because I was
00:53:08
very upset um and I realized that I
00:53:14
really loved open Ai and the people and
00:53:17
that I would come back and I kind of I
00:53:20
knew it was going to be hard it was even
00:53:22
harder than I thought but I I kind of
00:53:25
was like all right fine um I agreed to
00:53:27
come back um the board like took a while
00:53:30
to figure things out and then uh you
00:53:33
know we were kind of like trying to keep
00:53:35
the team together and keep doing things
00:53:38
for our customers and uh you know sort
00:53:40
of started making other plans then the
00:53:41
board decided to hire a different
00:53:43
interim CEO um and then
00:53:46
everybody there many people oh my gosh
00:53:49
what was what was that guy's name he was
00:53:50
there for like a scaramucci right like
00:53:53
uh em great and I I have nothing but
00:53:56
good
00:53:58
scari um and then where were you when
00:54:02
they um when you found the news that
00:54:04
you'd been fired like taking I was in a
00:54:07
hotel room in Vegas for F1 weekend I
00:54:10
think text and they're like fire pick up
00:54:13
said I think that's happened to you
00:54:14
before J I'm trying to think if I ever
00:54:17
got fired I don't think I've gotten
00:54:18
fired um yeah I got no it's just a weird
00:54:21
thing like it's a text from who actually
00:54:23
no I got a text the night before and
00:54:24
then I got in a phone call with the
00:54:27
uh and then that was that and then I
00:54:28
kind of like I mean then everything went
00:54:31
crazy I was like uh it was
00:54:35
like I mean I have my phone was like
00:54:37
unusable it was just a Non-Stop
00:54:39
vibrating thing of like text messages
00:54:41
call basically you got fired by tweet
00:54:44
that happened a few times during the
00:54:45
Trump Administration a few uh C TW
00:54:49
before tweeting was nice of them um and
00:54:52
then like you know I kind of did like a
00:54:54
few hours of just this like absolute
00:54:56
State um in the hotel room trying to
00:55:01
like I was just confused beyond belief
00:55:03
trying to figure out what to do and uh
00:55:05
so weird and then
00:55:07
like flew home it maybe like got on a
00:55:12
plane like I don't know 3 p.m. or
00:55:13
something like that um still just like
00:55:16
you know crazy non-stop phone blowing up
00:55:19
uh met up with some people in person by
00:55:21
that evening I was like okay you know
00:55:24
I'll just like go do AGI research and
00:55:27
was feeling pretty happy about the
00:55:29
future and yeah you have options and
00:55:32
then and then the next morning uh had
00:55:35
this call with a couple of board members
00:55:36
about coming back and that led to a few
00:55:40
more days of craziness and then
00:55:44
uh and then it kind of I think it got
00:55:48
resolved well it was like a lot of
00:55:50
insanity in between what percent what
00:55:53
percent of it was because of these
00:55:54
nonprofit board members
00:55:57
um well we only have a nonprofit board
00:55:59
so it was all the nonprofit board
00:56:01
members uh there the board had gotten
00:56:03
down to six people um
00:56:07
they and then they removed Greg from the
00:56:10
board and then fired me um so but it was
00:56:14
like you know but I mean like was there
00:56:16
a culture clash between the people on
00:56:17
the board who had only nonprofit
00:56:19
experience versus the people who had
00:56:21
startup experience and maybe you can
00:56:23
share a little bit about if you're
00:56:24
willing to the motivation behind the
00:56:26
action anything you
00:56:28
can I think there's always been culture
00:56:31
clashes
00:56:34
at look
00:56:36
obviously not all of those board members
00:56:39
are my favorite people in the world but
00:56:41
I
00:56:43
have serious respect
00:56:46
for the gravity with which they treat
00:56:50
AGI and the importance of getting AI
00:56:53
safety right and even if I
00:56:56
stringently disagree with their
00:56:59
decision- making and actions which I do
00:57:02
um I have never once doubted
00:57:05
their integrity or commitment to um the
00:57:11
sort of shared mission of safe and
00:57:12
beneficial
00:57:13
AGI um you know do I think they like
00:57:17
made good decisions in the process of
00:57:19
that or kind of know how to balance all
00:57:21
of things opening I has to get right no
00:57:24
but but I think that like the intent the
00:57:27
intent of the magnitude of yeah
00:57:32
AGI and getting that right I actually
00:57:36
let me ask you about that so the mission
00:57:38
of open AI is explicitly to create AGI
00:57:41
which I think is really
00:57:43
interesting a lot of people would say
00:57:46
that if we create AGI that would be like
00:57:49
an unintended consequence of something
00:57:51
gone horribly wrong and they're very
00:57:53
afraid of that outcome but open a makes
00:57:56
that the actual Mission yeah does that
00:58:00
create like more fear about what you're
00:58:02
doing I mean I understand it can create
00:58:03
motivation too but how do you reconcile
00:58:06
that I guess why is I think a lot of I
00:58:08
think a lot of the
00:58:10
well I mean first I'll say I'll answer
00:58:12
the first question and the second one I
00:58:14
think it does create a great deal of
00:58:15
fear uh I think a lot of the world is
00:58:19
understandably very afraid of AGI or
00:58:21
very afraid of even current Ai and and
00:58:23
very excited about it and even more
00:58:26
afraid and even more excited about where
00:58:28
it's going um and
00:58:31
we we wrestle with that but like I think
00:58:35
it is unavoidable that this is going to
00:58:37
happen I also think it's going to be
00:58:39
tremendously beneficial but we do have
00:58:41
to navigate how to get there in a
00:58:42
reasonable way and like a lot of stuff
00:58:45
is going to change and change is you
00:58:47
know pretty pretty uncomfortable for
00:58:49
people so there's a lot of
00:58:52
pieces that we got to get right and ask
00:58:56
can I ask a different question you you
00:58:58
have
00:59:00
created I
00:59:01
mean it's the hottest company and you
00:59:04
are literally at the center of the
00:59:06
center of the
00:59:07
center but then it's so unique in the
00:59:12
sense that all of this value you issued
00:59:15
economically can you just like walk us
00:59:17
through like yeah I wish I had taken I
00:59:19
wish I had taken Equity so I never had
00:59:21
to answer this question if I could go
00:59:23
back in why don't they give you a grand
00:59:24
now or just give you a big option Grant
00:59:27
like you deserve yeah give you five
00:59:29
points what was the decision back then
00:59:31
like why was that so important the
00:59:33
decision back then the re the original
00:59:34
reason was just like the structure of
00:59:36
our nonprofit it was uh like there was
00:59:39
something about yeah okay this is like
00:59:43
nice from a motivations perspective but
00:59:45
mostly it was that our board needed to
00:59:47
be a majority of disinterested
00:59:49
directors and I was like that's fine I
00:59:51
don't need Equity right now I kind
00:59:54
of but like but in this weird way now
00:59:57
that you're running a company yeah it it
00:59:59
creates these weird questions of like
01:00:01
well what's your real motivation versus
01:00:03
to that's that it is so deeply un I one
01:00:07
thing I have noticed it is is so deeply
01:00:10
unimaginable to people to say I don't
01:00:12
really need more
01:00:13
money like and I well people think I
01:00:16
think I think people think it's a little
01:00:18
bit of an ulterior motive I think yeah
01:00:20
yeah yeah no so it assumes what else is
01:00:22
he doing on the side to make money
01:00:24
something if I were just trying to say
01:00:26
like I'm going to try to make a trillion
01:00:27
dollarss with open AI I think everybody
01:00:29
would have an easier time and it would
01:00:30
save me it would save a lot of
01:00:32
conspiracy theories totally this is
01:00:34
totally the back Channel you are a great
01:00:37
dealmaker I I've watched your whole
01:00:39
career I mean you're just great at it
01:00:42
you got all these connections you're
01:00:43
really good at raising money uh you're
01:00:47
fantastic at it and you got this Johnny
01:00:49
I thing going you're inhumane you're
01:00:51
investing in company she got the orb
01:00:54
raising $7 trillion to build
01:00:56
Fabs all this stuff all of that put
01:01:00
together J loves fake news I'm kind of
01:01:03
being a little fous here you know
01:01:04
obviously it's not you're not raising 7
01:01:05
trillion doll but maybe that's the
01:01:06
market cap of something putting all that
01:01:08
aside the T was you're doing all these
01:01:12
deals they don't trust you because
01:01:14
what's your motivation you you your end
01:01:16
running and and what opportunities
01:01:18
belong inside of open AI what opportuni
01:01:21
should be Sam's and this group of
01:01:23
nonprofit people didn't trust you is
01:01:25
that what happened so the things like
01:01:27
you know device companies or if we were
01:01:29
doing some chip Fab Company it's like
01:01:31
those are not Sam project those would be
01:01:33
like opening ey would get that Equity um
01:01:36
they would okay that's not the Public's
01:01:38
perception well that's not like kind of
01:01:40
the people like you who have to like
01:01:42
commentate on the stuff all day
01:01:43
perception which is fair because we
01:01:44
haven't announced the stuff because it's
01:01:45
not done I don't think most people in
01:01:47
the world like are thinking about this
01:01:49
but I I I agree it spins up a lot of
01:01:53
conspiracies conspiracy theories in like
01:01:56
Tech commentators yeah and if I could go
01:01:59
back yeah I would just say like let me
01:02:01
take equity and make that super clear
01:02:03
and then every be like all right like
01:02:05
I'd still be doing it because I really
01:02:06
care about AGI and think this is like
01:02:08
the most interesting work in the world
01:02:10
but it would at least type check to
01:02:11
everybody what's the chip project that
01:02:14
the S trillion do and where' the Seven
01:02:16
trillion number come from makes no sense
01:02:17
I don't know where that came from
01:02:18
actually I genuinely don't uh I think I
01:02:21
think the world needs a lot more AI
01:02:23
infrastructure a lot more than it's
01:02:26
currently planning to build and with a
01:02:28
different cost structure um the exact
01:02:31
way for us to play there is we're still
01:02:35
trying to figure that out got it what's
01:02:36
your preferred model of organizing open
01:02:38
AI is
01:02:40
it sort of like the move fast break
01:02:43
things highly distributed small teams or
01:02:46
is it more of this organized effort
01:02:48
where you need to plan because you want
01:02:49
to prevent some of these edge cases um
01:02:53
oh I have to go in a minute uh it's not
01:02:55
because
01:02:58
it's not to prevent the edge cases that
01:03:00
we need to be more organized but it is
01:03:01
that these systems are so complicated
01:03:04
and concentrating bets are so important
01:03:07
like
01:03:09
one you know at the time before it was
01:03:11
like obvious to do this you have like
01:03:13
deep mind or whatever has all these
01:03:15
different teams doing all these
01:03:16
different things and they're spreading
01:03:17
their bets out and you had Open the Eyes
01:03:19
say we're going to like basically put
01:03:20
the whole company and work together to
01:03:22
make gp4 and that was like unimaginable
01:03:25
for how to run an AI research lab but it
01:03:28
is I think what works at the minimum
01:03:30
it's what works for us so not because
01:03:32
we're trying to prevent edge cases but
01:03:34
because we want to concentrate resources
01:03:35
and do these like big hard complicated
01:03:38
things we do have a lot of coordination
01:03:40
on what we work on all right Sam I know
01:03:42
you got to go you've been great on the
01:03:44
hour come back any time great talking to
01:03:46
you guys yeah fun thanks for for being
01:03:49
so open about it we've been talking
01:03:50
about it for like a year plus I'm really
01:03:53
happy it finally happened yeah it's
01:03:54
awesome I really app come back on after
01:03:56
our next like major launch and I'll be
01:03:57
able to talk more directly about some
01:04:00
you got the zoom link same Zoom link
01:04:01
every week just same time same Zoom link
01:04:03
drop time just drop in just put on your
01:04:06
calendar come back to the game come back
01:04:08
to the game in a while I you know I
01:04:11
would love to play poker it has been
01:04:13
forever that would be a lot of fun send
01:04:16
where Cham when you and I were heads up
01:04:18
and you you had remind me you and I were
01:04:21
heads up and you went all in I had a set
01:04:25
but there was a straight and a flush on
01:04:27
the board and I'm in the tank trying to
01:04:29
figure out if I want to lose this back
01:04:31
when we playing small Stakes it might
01:04:32
have been like 5K pot or something and
01:04:35
then chath can't stay out of the pot and
01:04:37
he starts taunting the two of us you
01:04:39
should call you shouldn't call he's
01:04:41
bluffing and I'm like I'm going I'm
01:04:44
trying to figure out if I make the call
01:04:45
here I make the call and uh it was like
01:04:48
uh you had a really good hand and but I
01:04:50
just happened to have a s I think you
01:04:51
had like top pair top kicker or
01:04:52
something but you you made a great move
01:04:54
because the boy was so almost like a
01:04:56
bottom set Sam has a great style of
01:04:58
playing which I would call RAM and jam
01:05:00
totally you got to get I don't really
01:05:02
know if you I don't I don't know if you
01:05:03
could say about anybody
01:05:05
else I don't I don't I'm not gonna you
01:05:07
haven't seen Jam play in the last 18
01:05:09
months it's a lot
01:05:10
different much more so much fun now find
01:05:14
find hard to have you played bomb pots I
01:05:18
don't know what that is okay you'll love
01:05:19
it we'll see you is nuts
01:05:22
it's do two boards and congrats
01:05:25
everything honestly than thanks for
01:05:27
coming on and love to have you back when
01:05:29
the next after the big launch sounds
01:05:31
please do cool bye gentlemen some
01:05:34
breaking news here all those projects He
01:05:37
said are part of open AI That's
01:05:39
something people didn't know before this
01:05:40
and a lot of confusion
01:05:42
there chamat what was your major
01:05:44
takeaway from our hour with Sam I think
01:05:47
that these guys are going to be one of
01:05:50
the four major companies okay that
01:05:52
matter in this whole space I think that
01:05:54
that's clear
01:05:56
I think what's still unclear is where is
01:05:58
the economics going to be he said
01:06:00
something very discreet but I thought
01:06:01
was important which is I think he
01:06:04
basically my interpretation is these
01:06:06
models will roughly all be the same but
01:06:08
there's going to be a lot of scaffolding
01:06:11
around these models that actually allow
01:06:14
you to build these apps so in many ways
01:06:15
that is like the open source movement so
01:06:17
even if the model itself is never open
01:06:20
source it doesn't much matter because
01:06:23
you have to pay for the infrastructure
01:06:24
right there's a lot of open Source
01:06:25
software that runs on Amazon you still
01:06:27
pay AWS something so I think the right
01:06:30
way to think about this now
01:06:34
is the models will basically be all
01:06:36
really good and then it's all this other
01:06:38
stuff that you'll have to pay for
01:06:40
interface whoever builds all this other
01:06:43
stuff is going to be in a position to
01:06:46
build a really good business freeberg he
01:06:48
talked a lot about reasoning it seemed
01:06:50
like that he kept going to reasoning and
01:06:52
away from the language model did you not
01:06:53
did you note that and anything else that
01:06:55
you noted in our hour with yeah I mean
01:06:57
that's a longer conversation because
01:06:58
there is a lot of talk about language
01:07:00
models eventually evolving to be so
01:07:03
generalizable that they can
01:07:07
resolve pretty much like all intelligent
01:07:09
function and so the language model is
01:07:11
the foundational model that that yields
01:07:13
AGI but that's a I think there's a lot
01:07:16
of people that at different schools of
01:07:18
thought on this and how much my my other
01:07:20
takeaway I think is that the I think
01:07:24
what he also seem to indicate is there's
01:07:27
like so many like we're also enraptured
01:07:30
by llms but there's so many things other
01:07:33
than llms that are being baked and
01:07:35
rolled by him and by other groups and I
01:07:37
think we have to pay some amount of
01:07:39
attention to all those because that's
01:07:41
probably where and I think freedberg you
01:07:43
tried to go there in your question
01:07:44
that's where reasoning will really come
01:07:45
from is this mixture of experts approach
01:07:48
and so you're going to have to think
01:07:49
multi-dimensionally to reason right we
01:07:52
do that right do I cross the street or
01:07:54
not in this point in time you reason
01:07:56
based on all these multi-inputs and so
01:07:58
there's there's all these little systems
01:08:00
that go into making that decision in
01:08:01
your brain and if you if you use that as
01:08:03
a a simple example there's all this
01:08:05
stuff that has to go into
01:08:07
making some experience being able to
01:08:10
reason intelligently sax you went right
01:08:12
there with the corporate structure the
01:08:15
board and uh he he he gave us a lot more
01:08:19
information here what are your thoughts
01:08:21
on the hey you know the chip stuff and
01:08:24
the other stuff I'm working that's all
01:08:26
part of open AI people just don't
01:08:27
realize it and that moment and then you
01:08:30
know your questions to him about Equity
01:08:32
your thoughts on um I'm not sure I was
01:08:35
like the main guy who asked that
01:08:36
question jakal but um well no you did
01:08:39
talk about the the nonprofit that the
01:08:41
difference between question about the
01:08:44
clearly was some sort of culture Clash
01:08:46
on the board between the the people who
01:08:49
originated from the nonprofit world and
01:08:50
people who came from the startup world
01:08:52
we don't really know more than that but
01:08:54
there clearly was some sort of culture
01:08:55
class
01:08:56
I thought one of the a couple of the
01:08:58
other areas that he drew attention to
01:08:59
that were kind of interesting is he
01:09:00
clearly thinks there's a big opportunity
01:09:02
on mobile that goes beyond just like
01:09:05
having you know a chat gbt app on your
01:09:08
phone or maybe even having like a Siri
01:09:10
on your phone there's clearly something
01:09:12
bigger there he doesn't know exactly
01:09:14
what it is but it's going to require
01:09:16
more inputs it's that you know personal
01:09:18
assistant that's seeing everything
01:09:20
around you help really I think that's a
01:09:22
great Insight David because he was
01:09:24
talking about hey I'm looking for a
01:09:27
senior team member who can push back on
01:09:29
me and understands all context I thought
01:09:30
that was like a very interesting to
01:09:32
think about an executive assistant or an
01:09:35
assistant that's has executive function
01:09:38
as opposed to being like just an alter
01:09:40
ego for you or what he called a
01:09:42
sycophant that's kind of interesting I
01:09:44
thought that was interesting yeah yeah
01:09:45
and clearly he thinks there's a big
01:09:47
opportunity in biology and scientific
01:09:49
discovery after the break I think we
01:09:51
should talk about Alpha fold 3 it was
01:09:52
just announ let's do that and we can
01:09:54
talk about the the Apple ad and
01:09:56
I just want to also make sure people
01:09:57
understand when people come on the Pod
01:09:59
we don't show them questions they don't
01:10:00
edit the transcript nothing is out of
01:10:03
bounds if you were wondering why I
01:10:05
didn't ask or we didn't ask about the
01:10:06
Elon lawsuit he's just not going to be
01:10:08
able to comment on that so it' be no
01:10:10
comment so you know and we're not like
01:10:12
our time was limited and there's a lot
01:10:14
of questions that we could have asked
01:10:15
him that would have just been a waste of
01:10:16
time and com so I just want to make sure
01:10:19
people understand of course he's going
01:10:20
to no comment in any lawsuit and he's
01:10:22
already been asked about that 500 times
01:10:25
yes should we take a quick break before
01:10:26
the next before we come back yeah I'll
01:10:27
take a bio break and then we'll come
01:10:28
back with some news for you and some
01:10:30
more banter with your
01:10:33
favorite besties on the number one
01:10:36
podcast in the world the only podcast
01:10:38
all right welcome back everybody second
01:10:40
half of the show great guest Sam mman
01:10:41
thanks for coming on the Pod we've got a
01:10:44
bunch of news on the docket so let's get
01:10:46
started freeberg you told me I could
01:10:49
give some names of uh the guests that
01:10:51
we've booked for the all in Summit I did
01:10:54
not you did you've said each week every
01:10:56
week that I get to say I did not I
01:10:59
appreciate your interest in the all in
01:11:02
Summits lineup but we do not yet
01:11:05
have uh enough critical math uh to feel
01:11:08
like we should go out there well uh I am
01:11:11
a loose canidate so I will announce my
01:11:13
two guests and I created The Summit and
01:11:16
you took it from me so and done a great
01:11:18
job I will announce my guests I don't
01:11:20
care what your opinion is I have booked
01:11:22
two guests for the summit and it's going
01:11:24
to be out look at these two guests I
01:11:27
booked for the third time coming back to
01:11:29
the summit our guy Elon Musk will be
01:11:31
there hopefully in person if not you
01:11:33
know from 40,000 feet on a starling
01:11:35
connection wherever he is in the world
01:11:36
and for the first time our friend Mark
01:11:40
Cuban will be coming and so two great
01:11:43
guests for you to look forward to but
01:11:45
free BG's got like a thousand guests
01:11:46
coming he'll tell you when it's like 48
01:11:49
hours before the conference but yeah two
01:11:51
great speaking of billionaires who are
01:11:52
coming isn't coming too yes coming yes
01:11:55
he's booked so we have three
01:11:57
billionaires three billionaires yes okay
01:12:00
hasn't fully confirmed so don't okay
01:12:01
well we're going to say it anyway has
01:12:03
penciled in back we say penciled yeah
01:12:06
don't back out this is going to be
01:12:08
catnip for all these protest organizers
01:12:10
like if
01:12:11
youke the bear well by the way speaking
01:12:14
of updates what do you guys think of the
01:12:17
bottle for the all-in tequila oh
01:12:19
beautiful honestly honestly I will just
01:12:21
say I think you are doing a marvelous
01:12:24
job that
01:12:25
I was
01:12:27
shocked at the design shocked meaning it
01:12:30
is so unique and high quality I think
01:12:33
it's amazing it would make me drink
01:12:37
tequila you're going to You're Gonna
01:12:39
Want to going to it it is uh stunning
01:12:41
just congratulations and um yeah it was
01:12:44
just when we went through the deck at
01:12:47
the uh at the monthly meeting it was
01:12:49
like oh that's nice oh that's nice we're
01:12:51
going to do the concept bottles and then
01:12:52
that bottle came up and everybody went
01:12:55
like crazy it was like somebody hitting
01:12:56
like a Steph Curry hitting a half court
01:12:58
shot it was like oh my God it was just
01:13:01
so clear that you've made an iconic
01:13:03
bottle that if we can produce it oh Lord
01:13:07
it is going to be looks like we can the
01:13:11
make it yeah it's gonna be amazing I'm
01:13:13
excited I'm excited for it you know it's
01:13:14
like the Des so complicated that we have
01:13:16
to do a feasibility analysis on whether
01:13:18
it was actually manufacturable but it is
01:13:21
so or at least the early reports are
01:13:23
good so we're going to H hopefully we'll
01:13:25
have some a for the in time for the all
01:13:27
Summit I mean why not sounds I mean it's
01:13:30
great when we get barricaded in by all
01:13:32
these protesters we can drink the
01:13:34
tequila did you guys see did you see
01:13:36
Peter te Peter teal got barricaded by
01:13:39
these ding-dongs at Cambridge my god
01:13:41
listen people have the right to protest
01:13:42
I think it's great people are protesting
01:13:44
but surrounding people and threatening
01:13:46
them is a little bit over the top and D
01:13:49
think you're exaggerating what happened
01:13:51
well I don't know exactly what happened
01:13:52
because all we see is these videos look
01:13:54
they're not threatening anybody and I
01:13:56
don't even think they tried to barricade
01:13:57
him in they were just outside the
01:13:59
building and because they were blocking
01:14:01
the driveway his car couldn't leave but
01:14:04
he
01:14:05
wasn't physically like locked in the
01:14:08
building or something yeah that's that's
01:14:10
what the headlines say but that could be
01:14:12
fake news fake Social yeah this was not
01:14:14
on my bingo card this Pro protester
01:14:17
support by saaks was not on the bingo
01:14:19
card I got to say I contitution the
01:14:22
Constitution of the United States in the
01:14:25
Amendment provides for the right of
01:14:27
assembly which includes protest and sit
01:14:29
in as long as they're as long as they're
01:14:31
Peaceable now obviously if they go too
01:14:34
far and they vandalize or break into
01:14:36
buildings or use violence then that's
01:14:38
not Peaceable however expressing
01:14:40
sentiments with which you disagree does
01:14:43
not make it violent and there's all
01:14:46
these people out there now making the
01:14:47
argument that if you hear something from
01:14:50
a protester that you don't like and you
01:14:54
subjectively experience that as a as a
01:14:56
threat to your safety then that somehow
01:14:59
should be you know treated as valid like
01:15:02
that's basically violent well that's
01:15:04
that's not what the Constitution says
01:15:07
and these people understood well just a
01:15:09
few months ago that that was basically
01:15:11
snowf fakery that you know just because
01:15:14
somebody you know what I'm
01:15:16
saying we have the rise of the woke
01:15:18
right now where they're buying yeah the
01:15:21
woke right they're buying into this idea
01:15:22
of safetyism which is being exposed
01:15:25
ideas you don't like to protest you
01:15:26
don't like is a threat to your safety no
01:15:28
it's
01:15:30
not every we absolutely have snow fakery
01:15:33
on both sides now it's ridiculous the
01:15:35
only thing I will say that I've seen and
01:15:38
is this this this uh surrounding
01:15:41
individuals who you don't want there and
01:15:44
locking them in a circle and then moving
01:15:46
them out of like protesta that's not
01:15:48
cool yeah obviously you can't do that
01:15:49
but look I think that most of the
01:15:52
protests on most of the campuses have
01:15:53
not crossed the line they've just
01:15:55
occupied The Lawns of these campuses and
01:15:57
look I've seen some troublemakers try to
01:16:01
barge through the the encampments and
01:16:04
claim that because they can't go through
01:16:06
there that somehow they're being
01:16:08
prevented from going to class look you
01:16:10
just walk around the lawn and you can
01:16:12
get to class okay and you know some of
01:16:15
these videos are showing that these are
01:16:18
effectively right-wing provocators who
01:16:20
are engaging in leftwing tactics and I
01:16:24
don't support it either way by the way
01:16:26
some of these camps are some of the
01:16:28
funniest things you've ever seen it's
01:16:29
like there are like a one tent that's
01:16:32
dedicated to like a reading room and you
01:16:34
go in there and there's like these like
01:16:37
Center oh my God it's unbelievably
01:16:39
hilarious look there there's no question
01:16:41
that because the protests are
01:16:42
originating on the left that there's
01:16:44
some goofy views like you know you're
01:16:46
dealing with like a leftwing idea
01:16:48
complex right but and you know it's easy
01:16:52
to make fun of them doing different
01:16:53
things but the fact of the matter is
01:16:56
that most of the protests on most of
01:16:57
these campuses are even though they can
01:16:59
be annoying because they're occupying
01:17:01
part of the lawn they're not violent
01:17:04
yeah and you know the way they're being
01:17:05
cracked down on they're sending the
01:17:06
police in at 5:00 a. to crack down on
01:17:09
these encampments with batons and riot
01:17:12
gear and I find that part to be
01:17:14
completely excessive well it's also
01:17:17
dangerous because you know things can
01:17:19
escalate when you have mobs of people
01:17:21
and large groups of people so I just
01:17:22
want to make sure people understand that
01:17:24
large group of people large you have a
01:17:26
diffusion of responsibility that occurs
01:17:28
when there's large groups of people who
01:17:29
are passionate about things and and
01:17:31
people could get hurt people have gotten
01:17:33
killed at these things so just you know
01:17:35
keep it calm everybody I agree with you
01:17:37
like what's the harm of these folks
01:17:39
protesting on a lawn it's not a big deal
01:17:40
when they break into buildings of course
01:17:42
yeah that crosses the line obviously
01:17:44
yeah but I mean let them sit out there
01:17:45
and then they'll run out their food cars
01:17:48
their food card and they run out of
01:17:50
waffles did you guys see the clip I
01:17:51
think it was on the University of
01:17:53
Washington campus where one kid
01:17:56
challenged this antifa guy to a push-up
01:17:58
contest oh
01:18:00
fantastic I mean it's it is some of the
01:18:02
funniest stuff some of some content is
01:18:05
coming out that's just my favorite was
01:18:07
the woman who came out and said that the
01:18:09
Columbia students needed humanitarian
01:18:11
Aid and oh my God the overdubs on her
01:18:14
were hilarious I was like uh
01:18:16
humanitarian Aid the he was like we need
01:18:19
our door Dash right now we need we
01:18:21
Double Dash some boba and we can't get
01:18:23
it through the the police need our Boba
01:18:26
low sugar Boba with the with the popping
01:18:28
Boba bubbles wasn't getting in but you
01:18:31
know people have the right to protest
01:18:32
and uh Peaceable by the way there's a
01:18:34
word I've never heard very good sax
01:18:36
Peaceable inclined to avoid argument or
01:18:39
violent conflict very nice well it's in
01:18:41
the Constitution it's in the first
01:18:42
amendment is it really I've never I
01:18:44
haven't heard the word Peaceable before
01:18:46
I mean you and I are sympatico on this
01:18:47
like I I don't we used to have the
01:18:51
ACLU like backing up the KKK going down
01:18:55
Main Street and really fighting decision
01:18:57
yeah they were really fighting for I I'm
01:19:00
and I have to say the Overton window is
01:19:02
opened back up and I think it's great
01:19:04
all right we got some things on the
01:19:05
docket here I don't know if you guys saw
01:19:07
the Apple new iPad ad it's getting a
01:19:09
bunch of criticism they use like some
01:19:12
giant hydraulic press to crush a bunch
01:19:15
of creative tools EJ turntable trumpet
01:19:19
piano people really care about Apple's
01:19:21
ads and what they represent we talked
01:19:24
about that that uh Mother Earth little
01:19:26
vignette they created here what do you
01:19:29
think free BR did you see the ad what
01:19:30
was your reaction to it made me sad it
01:19:32
it did not make me want to buy an iPad
01:19:34
so huh did not seem like it made you sad
01:19:38
it actually elicited an emotion meaning
01:19:40
like commercials it's very rare that
01:19:42
commercials can actually do that most
01:19:43
people just Zone up yeah they took all
01:19:45
this beautiful stuff and heard it didn't
01:19:47
it didn't feel good I don't know it just
01:19:49
didn't seem like a good ad I don't know
01:19:50
why they did that I don't get it I I've
01:19:52
I don't I don't know I think I think
01:19:54
maybe what they're trying to do is the
01:19:55
the selling point of this new iPad is
01:19:57
that it's the thinnest one I mean
01:19:59
there's no innovation left so they're
01:20:00
just making the devices yeah you know
01:20:02
thinner yeah so I think the idea was
01:20:06
that they were going to take this
01:20:07
hydraulic press to represent how
01:20:09
ridiculously thin the new iPad is now I
01:20:12
don't know if the point there was to
01:20:14
smush all of that good stuff into the
01:20:17
iPad I don't know if that's what they
01:20:18
were trying to convey but yeah I I think
01:20:21
that by destroying all those creative
01:20:24
tools that apple is supposed to
01:20:26
represent it definitely seemed very
01:20:28
offbrand for them and I think people
01:20:30
were reacting to the fact that it was so
01:20:34
um different than what they would have
01:20:35
done in the past and of course everyone
01:20:37
was saying well Steve would never have
01:20:38
done this I do think it did land wrong I
01:20:41
mean I I didn't care that much but but I
01:20:45
I was kind of asking the question like
01:20:46
why are they destroying all these
01:20:48
Creator tools that they're renowned for
01:20:52
creating or for turning into the digital
01:20:55
version yeah it just didn't land I mean
01:20:58
shath how are you doing emotionally
01:21:01
after seeing me are you okay buddy yeah
01:21:06
I think this is uh you guys see that in
01:21:09
the
01:21:10
birkshire annual meeting last
01:21:12
weekend Tim Cook was in the audience and
01:21:17
Buffett was very laudatory this is an
01:21:19
incredible company but he's so clever
01:21:22
with words he's like you know this is an
01:21:24
incredible business that we will hold
01:21:28
forever most likely then it turns out
01:21:31
that he sold $20 billion worth of Apple
01:21:33
shares in the
01:21:34
cat we're gonna hold it forever which
01:21:37
which by the way sell if you guys
01:21:39
remember we we put that little chart up
01:21:40
which shows when he doesn't mention it
01:21:42
in the in the annual letter it's
01:21:44
basically like it's foreshadowing the
01:21:46
fact that he is just pounding the cell
01:21:49
and he sold $20 billion well also
01:21:52
holding it forever could mean one share
01:21:54
yeah exactly we kind of need to know
01:21:56
like how much are we talking about I
01:21:59
mean it's an incredible business that
01:22:01
has so much money with nothing to do
01:22:03
they're probably just going to buy back
01:22:04
the stock just a total waste they were
01:22:06
floating this rumor of buying rivan you
01:22:08
know after they shut down Titan project
01:22:10
the your internal project to make a car
01:22:11
it seems like a car is the only thing
01:22:13
people can think of that would move the
01:22:16
needle in terms of earnings I think the
01:22:17
problem is J like you kind of become
01:22:19
afraid of your own shadow meaning the
01:22:21
the folks that are really good at m&a
01:22:23
like you look at Benny off
01:22:25
the thing with benof m& strategy is that
01:22:27
he's been doing it for 20 years and so
01:22:31
he's cut his teeth on small Acquisitions
01:22:34
and the market learns to give him trust
01:22:37
so that when he proposes like the $27
01:22:39
billion slack acquisition he's allowed
01:22:41
to do that another guy you know narora
01:22:44
at pandw these last five years people
01:22:46
were very skeptical that he could
01:22:48
actually roll up security because it was
01:22:49
a super fragmented Market he's gotten
01:22:52
permission then there are companies like
01:22:54
Dan that buy hundreds of companies so
01:22:56
all these folks are examples of you
01:22:58
start small and you you earn the right
01:23:00
to do more Apple hasn't bought anything
01:23:02
more than 50 or hundred million do and
01:23:04
so the idea that all of a sudden they
01:23:06
come out of the blue and buy a 102
01:23:08
billion dollar company I think is just
01:23:10
totally doesn't stand logic it's just
01:23:13
not possible for them because they'll be
01:23:15
so afraid of their own shadow that's the
01:23:16
big problem it's themselves well if
01:23:19
you're running out of In-House
01:23:21
Innovation and you can't do m&a then
01:23:23
your options are kind of limited I mean
01:23:25
I do think that the fact that the big
01:23:27
news out of apple is the iPad's Getting
01:23:30
Thinner does represent kind of the end
01:23:32
of the road in terms of innovation it's
01:23:34
kind of like when they added the third
01:23:36
camera to the iPhone yeah it reminds me
01:23:39
of those um remember like when the
01:23:41
Gillette yeah they came out and then
01:23:43
they did the five was the best onion
01:23:44
thing was like we're doing five f
01:23:47
it but then Gillette actually came out
01:23:49
with the Mach 5 so yeah like the parody
01:23:51
became the reality what are they going
01:23:52
to do add two more cameras to the iPhone
01:23:54
you have five cameras on it no makes no
01:23:57
sense and then I don't know anybody
01:23:58
wants to remember the Apple Vision was
01:24:01
like gonna plus why are they body
01:24:02
shaming the the fat
01:24:05
iPads that's a fair point Fair Point
01:24:07
actually you know what it's actually
01:24:10
this didn't come out yet but it turns
01:24:11
out the iPad is on OIC it's actually
01:24:14
dropped a lot that would have been a
01:24:15
funnier ad yeah yeah exactly oh oh
01:24:20
ohic we can just Workshop that right
01:24:22
here but there was another funny one
01:24:24
which was making the iPhone smaller and
01:24:26
smaller and smaller and the iPod smaller
01:24:27
and smaller and smaller to the point it
01:24:28
was like you know like a thumb siize
01:24:30
iPhone like the Ben Stiller phone in
01:24:33
Zoolander or
01:24:36
correct yeah that was a great scene is
01:24:39
there a category that you can think of
01:24:42
that you would love an Apple product for
01:24:45
there's a product in your life that you
01:24:47
would love to have Apple's version of it
01:24:50
they they killed it I think a lot of
01:24:53
people would be very open minded to an
01:24:55
Apple car okay they they just would it's
01:24:57
it's a connected internet device
01:25:00
increasingly so yeah and they they
01:25:02
managed to flub
01:25:03
it they had a chance to buy Tesla they
01:25:06
managed to flub it yeah right there are
01:25:09
just too many examples here where these
01:25:11
guys have so much money and not enough
01:25:12
ideas that's a
01:25:14
shame it's a bummer yeah the one I
01:25:17
always wanted to see them do saxs was TV
01:25:19
the one I always wanted to see them do
01:25:21
was the TV and they were supposedly
01:25:22
working on it like the actual TV not the
01:25:24
little Apple TV box in the back and like
01:25:26
that would have been extraordinary to
01:25:28
actually have a gorgeous you know big
01:25:31
television what about a gaming console
01:25:33
they could have done that you know
01:25:34
there's just all these things that they
01:25:36
could have done it's not a lack of
01:25:39
imagination because these aren't exactly
01:25:42
incredibly World beating ideas they're
01:25:44
sitting right in front of your face it's
01:25:46
just the will to do it yeah all in one
01:25:50
TV would have been good if you think
01:25:52
back on Apple product lineup over the
01:25:55
years where they've really created value
01:25:57
is on how unique the products are they
01:26:00
almost create new categories sure there
01:26:02
may have been a quote tablet computer
01:26:04
prior to the iPad but the iPad really
01:26:06
defined the tablet computer era sure
01:26:08
there was a smartphone or two before the
01:26:10
iPhone came along but it really defined
01:26:12
the smartphone and sure there was a
01:26:14
computer before the Apple 2 and then it
01:26:15
came along and it defined the personal
01:26:17
computer in all these cases I think
01:26:19
Apple strives to define the category so
01:26:22
it's very hard to define a television if
01:26:24
you think about it or gaming console in
01:26:25
a way that you take a step up and you
01:26:28
say this is the new thing this is the
01:26:29
new platform so I don't know that's the
01:26:32
lens I would look at if I'm Apple in
01:26:33
terms of like can I redefine a car can I
01:26:36
make you know we're all trying to fit
01:26:38
them into an existing product bucket but
01:26:40
I think what they've always been so good
01:26:41
at is identifying consumer needs and
01:26:43
then creating an entirely new way of
01:26:45
addressing that need in a real step
01:26:46
change function from the like the the
01:26:49
the um iPod it was so different from any
01:26:51
MP3 player ever I think the reason why
01:26:53
the car could have been completely
01:26:55
reimagined by Apple is that they have a
01:26:57
level of credibility and trust that I
01:26:59
think probably no other company has and
01:27:02
absolutely no other tech company has and
01:27:05
we talked about this but I think this
01:27:08
was the third Steve Job story that that
01:27:11
I left out but in
01:27:13
200 I don't know was it
01:27:15
[Music]
01:27:17
one I launched a 99 cent download store
01:27:21
right I think I've told you the story in
01:27:22
winam and
01:27:26
Steve Jobs just ran total circles around
01:27:28
us but the reason he was able to is he
01:27:30
had all the credibility to go to the
01:27:31
labels and get deals done for licensing
01:27:34
music that nobody could get done before
01:27:36
I think that's an example of what
01:27:38
Apple's able to do which is to use their
01:27:39
political Capital to change the rules so
01:27:42
if the thing that we would all want is
01:27:44
safer roads and autonomous vehicles
01:27:47
there are regions in every town and city
01:27:49
that could be completely converted to
01:27:52
level five autonomous zones if I had to
01:27:55
pick one company that had the
01:27:57
credibility to go and change those rules
01:27:58
it's them because they could demonstrate
01:28:01
that there was a methodical safe
01:28:03
approach to doing something and so the
01:28:05
point is that even in these categories
01:28:08
that could be totally reimagined it's
01:28:09
not for a lack of imagination again it
01:28:11
just goes back to a complete lack of
01:28:13
Will and I understand because if they
01:28:15
had if you if you had $200 billion do of
01:28:18
capital on your balance sheet I think
01:28:21
it's probably pretty easy to get fat and
01:28:23
lazy yeah it is and and they want to
01:28:25
have everything built there people don't
01:28:27
remember but they actually built one of
01:28:29
the first digital cameras you must have
01:28:30
owned this right freeberg you're I
01:28:32
remember this yeah totally it beautiful
01:28:34
what did they call it was it the ey
01:28:36
camera or something quick take quick
01:28:39
take yeah um the thing I would like to
01:28:41
see apple build and I'm surprised they
01:28:42
didn't was a smart home system the way
01:28:46
Apple has Nest a drop cam a door lock
01:28:50
you know a AV system go equestron or
01:28:54
whatever and just have your whole home
01:28:56
automated thermostat Nest all of that
01:28:58
would be brilliant by Apple and right
01:29:00
now I'm an apple family that has our all
01:29:04
of our home automation through Google so
01:29:06
it's just kind of sucks I would I would
01:29:08
like that all to that would be pretty
01:29:10
amazing like if they did a crant or
01:29:11
Savant CU then when you just go to your
01:29:13
Apple TV all your cameras just work you
01:29:15
don't need to
01:29:17
yes that's the that I mean and everybody
01:29:19
has a home and everybody automates their
01:29:21
home so well everyone has Apple TV at
01:29:23
this point so you just make Apple TV the
01:29:26
brain for the home system right that
01:29:29
would be your Hub and you can connect
01:29:31
your phone to it then yes that would be
01:29:33
very nice yeah like can you imagine like
01:29:36
the ring cameras all that stuff being
01:29:38
integrated I don't know why they didn't
01:29:39
go after that that seems like the easy
01:29:41
layup hey you know everybody's been
01:29:43
talking freedberg about this uh Alpha
01:29:47
fold this folding
01:29:49
proteins and there's some new version
01:29:52
out from Google and uh also Google
01:29:55
reportedly we talked about this before
01:29:56
is also advancing talks to acquire
01:29:58
HubSpot so that rumor for the $30
01:30:01
billion market cap HubSpot is out there
01:30:03
as well free break your as our resident
01:30:06
science Sultan uh our resident Sultan of
01:30:09
Science and as an Google
01:30:12
alumni pick either story and let's go
01:30:14
for it yeah I mean I'm not sure there's
01:30:15
much more to add on the HubSpot
01:30:16
acquisition rumors they are still just
01:30:18
rumors and I think we covered the topic
01:30:20
a couple weeks ago but I will say that
01:30:22
Alpha fold 3 that was just announced
01:30:24
today and demonstrated by Google um is a
01:30:27
real uh I would say breathtaking moment
01:30:31
um for biology for bioengineering for
01:30:34
human health for medicine and maybe I'll
01:30:36
just take 30 seconds to kind of explain
01:30:38
it um you remember when they introduced
01:30:40
Alpha fold at Alpha fold 2 we talked
01:30:43
about DNA codes for proteins so every
01:30:46
three letters of DNA codes for an amino
01:30:49
acid so a string of DNA codes for a
01:30:52
string of amino acids and that's called
01:30:54
a gene that produces a
01:30:55
protein and that protein is basically a
01:30:59
like think about beads there's 20
01:31:00
different types of beads 20 different
01:31:01
amino acids that can be strung together
01:31:04
and what happens is that necklace that
01:31:06
bead necklace basically collapses on
01:31:08
itself and all those little beads stick
01:31:10
together with each other in some
01:31:11
complicated way that we can't
01:31:12
deterministically model and that creates
01:31:15
a three-dimensional structure which is
01:31:16
called A protein that molecule and that
01:31:19
molecule does something interesting it
01:31:20
can break apart other molecules it can
01:31:23
bind molecules it can move molecules
01:31:24
around so it's basically the Machinery
01:31:26
of chemistry of biochemistry and so
01:31:29
proteins are what is encoded in our DNA
01:31:32
and then the proteins do all the work of
01:31:34
making living organisms so Google's
01:31:36
Alpha fold project took threedimensional
01:31:38
images of proteins and the DNA sequence
01:31:41
that codes for those proteins and then
01:31:43
they built a predictive model that
01:31:45
predicted the three-dimensional
01:31:46
structure of a protein from the DNA that
01:31:48
codes for it and that was a huge
01:31:50
breakthrough years ago what they just
01:31:52
announced with Alpha fold 3 today
01:31:54
is that they're now including all small
01:31:56
molecules so all the other little
01:31:58
molecules that go into chemistry and
01:32:00
biology that drive the function of
01:32:02
everything we see around us and the way
01:32:05
that all those molecules actually bind
01:32:07
and fit together is part of the
01:32:09
predictive model why is that important
01:32:11
well let's say that you're designing a
01:32:12
new drug and it's a protein based drug
01:32:15
which biologic drugs which most drugs
01:32:16
are today you could find a biologic drug
01:32:18
that binds to a cancer cell and then
01:32:20
you'll spend 10 years going to clinical
01:32:22
trials and billions of later you find
01:32:24
out that that protein accidentally binds
01:32:27
to other stuff and hurts other stuff in
01:32:28
the body and that's an off Target effect
01:32:30
or a side effect and that drug is pulled
01:32:32
from the clinical trials and it never
01:32:33
goes to Market most drugs go through
01:32:36
that process they are actually tested in
01:32:39
in animals and then in humans and we
01:32:41
find all these side effects that arise
01:32:42
from those drugs because we don't know
01:32:45
how those drugs are going to bind or
01:32:47
interact with other things in our
01:32:48
biochemistry and we only discovered
01:32:50
after we put it in but now we can
01:32:52
actually model that with software we can
01:32:54
take that drug we can create a
01:32:56
three-dimensional representation of it
01:32:57
using the software and we can model how
01:33:00
that drug might interact with all the
01:33:01
other cells all the other proteins all
01:33:03
the other small molecules in the body to
01:33:06
find all the of Target effects that may
01:33:08
arise and decide whether or not that
01:33:10
presents a good drug candidate that is
01:33:12
one example of how this capability can
01:33:15
be used and there are many many others
01:33:17
including creating new proteins that
01:33:20
could be used to bind molecules or stick
01:33:22
molecules together or new prot that
01:33:24
could be designed to rip molecules apart
01:33:26
we can now predict the function of
01:33:29
threedimensional molecules using this
01:33:31
this capability which opens up all of
01:33:33
the software based design of chemistry
01:33:35
of biology of drugs and it really is an
01:33:38
incredible breakthrough moment the
01:33:41
interesting thing that happened though
01:33:42
is Google alphabet has a subsidiary
01:33:45
called isomorphic Labs it is a drug
01:33:47
development subsidiary of alphabet and
01:33:49
they've basically kept all the IP for
01:33:52
Alpha fold 3 in is
01:33:54
so Google is going to monetize the heck
01:33:56
out of this capability and what they
01:33:58
made available was not open source code
01:34:00
but a webbased viewer that scientists
01:34:03
for quote non-commercial purposes can
01:34:04
use to do some fundamental research in a
01:34:06
web-based viewer and make some
01:34:08
experiments and try stuff out and how
01:34:09
interactions might occur but no one can
01:34:12
use it for commercial use only Google's
01:34:14
isomorphic Labs can so number one it's
01:34:16
an incredible demonstration of what AI
01:34:18
outside of llms which we just talked
01:34:21
about with Sam today and obviously
01:34:23
talked about other models but llms being
01:34:25
kind of this consumer text predictive
01:34:27
model capability but outside of that
01:34:29
there's this capability in things like
01:34:32
chemistry with these new AI models that
01:34:34
can be trained and built to predict
01:34:36
things like three-dimensional chemical
01:34:39
interactions that is going to open up an
01:34:41
entirely New Era for you know human
01:34:43
progress and I think that's what's so
01:34:45
exciting I think the other side of this
01:34:46
is Google is hugely advantaged and they
01:34:48
just showed the world a little bit about
01:34:50
some of these jewels that they have in
01:34:51
the treasure chest and they're like look
01:34:52
at what we got we're going to all these
01:34:54
drugs and they've got Partnerships with
01:34:55
all thesea companies at isomorphic Labs
01:34:57
that they've talked about and it's going
01:34:59
to usher in a new era of drug
01:35:01
development design for human health so
01:35:03
all in all I'd say it's a pretty like
01:35:05
astounding day a lot of people are going
01:35:06
crazy over the capability that they just
01:35:08
demonstrated and then it begs all this
01:35:10
really interesting question around like
01:35:12
you know what's Google G to do with it
01:35:13
and how much value is going to be
01:35:14
created here so anyway I thought it was
01:35:15
a great story and I just rambled on for
01:35:17
a couple minutes but I don't know it's
01:35:19
pretty cool super interesting is is this
01:35:20
AI capable of making a science Corner
01:35:22
that David saak pay attention
01:35:25
to well it will it will predict the Cure
01:35:28
I think for the common cold and for
01:35:30
herpes so he should pay attention
01:35:33
absolutely folding cells is the app that
01:35:36
Sachs casual game Sachs just downloaders
01:35:38
playing how many uh how many chest moves
01:35:40
did you make during that segment sex
01:35:42
sorry let me just say one more thing do
01:35:43
you guys remember we talked about
01:35:44
yamanaka factors and how challenging it
01:35:46
is to basically we can reverse aging if
01:35:48
we can get the right proteins into cells
01:35:51
to tune the expression of certain gen to
01:35:53
make those cells youthful right now it's
01:35:56
a shotgun approach to trying millions of
01:35:58
compounds and combinations of compounds
01:36:00
to do them there's a lot of companies
01:36:01
actually trying to do this right now to
01:36:03
come up with a fountain of youth type
01:36:05
product we can now simulate that so with
01:36:08
this system one of the things that this
01:36:10
Alpha 3 can do is predict what molecules
01:36:13
will bind and promote certain sequences
01:36:15
of DNA which is exactly what we try and
01:36:17
do with the yanaka factor-based
01:36:19
expression systems and find ones that
01:36:21
won't trigger off Target expression so
01:36:23
meaning we can now go through the search
01:36:25
space and software of creating a
01:36:27
combination of molecules that
01:36:29
theoretically could unlock this Fountain
01:36:31
of Youth to deage all the cells in the
01:36:33
body and introduce an extraordinary kind
01:36:35
of health benefit and that's just again
01:36:37
one example of the many things that are
01:36:39
possible inredible with this sort of
01:36:40
platform I I and I'm really I gota be
01:36:42
honest I'm really just sort of skimming
01:36:43
the surface here of what this can do the
01:36:46
capabilities and the impact are going to
01:36:48
be like I don't know I know I say this
01:36:50
sort of stuff a lot but it's going to be
01:36:51
pretty for half there's um on the Block
01:36:53
post they have this incredible video
01:36:55
that they show of
01:36:57
the Corona virus that creates a common
01:37:00
cold I think the S P&M protein and not
01:37:03
only did they literally like predict it
01:37:06
accurately they also predicted how it
01:37:09
interacts with an antibody with a sugar
01:37:12
it's nuts so you could see a world where
01:37:15
like I don't know you just get a vaccine
01:37:17
for the cold and it's kind of like you
01:37:19
never have colds again amazing I mean
01:37:21
simple stuff but so powerful you can
01:37:23
filter out stuff that has offt Target
01:37:25
effect so so much of drug Discovery and
01:37:27
all the side effect stuff can start to
01:37:29
be solved for in silicone and you could
01:37:31
think about running extraordinar large
01:37:33
use a model like this run
01:37:34
extraordinarily large simulations in a
01:37:37
search space of chemistry to find stuff
01:37:39
that does things in the body that can
01:37:42
unlock you know all these benefits can
01:37:44
do all sorts of amazing things to
01:37:45
destroy cancer to destroy viruses to
01:37:48
repair cells to deage cells and this is
01:37:51
a hundred billion dollar business they
01:37:53
say I alone I feel like this is where I
01:37:57
I've said this before I think Google's
01:37:58
got this like portfolio of like
01:38:01
quiet
01:38:03
extraord what if they hit and the fact
01:38:05
and I think the fact that they didn't
01:38:06
open source everything in this says a
01:38:08
lot about their intentions yeah yeah
01:38:10
open source when you're behind closed
01:38:12
Source lock it up when you're ahead but
01:38:14
show yamanaka actually interestingly
01:38:16
yamanaka is the Japanese whiskey that
01:38:19
saak serves on his plane as well it's
01:38:21
delicious I love that Hokkaido yak
01:38:24
Jason I feel like if you didn't find
01:38:26
your way to Silicon Valley you could be
01:38:27
like a Vegas Lounge comedy guy
01:38:29
absolutely for sure yeah I was actually
01:38:32
yeah somebody said I should do like
01:38:33
those 1950s those 1950s talk shows where
01:38:35
the guys would do like the the stage
01:38:37
show somebody told me I should do um
01:38:40
like spald and gray Eric bosan style
01:38:42
stuff I don't know if you guys remember
01:38:44
like the uh the monologue is from the
01:38:45
80s in New York I was like oh that's
01:38:47
interesting maybe all right everybody
01:38:48
thanks for tuning in to the world's
01:38:50
number one podcast can you believe we
01:38:53
did it jamath uh the number one podcast
01:38:56
in the world and the all-in summit the
01:38:59
Ted killer if you are going to Ted
01:39:02
congratulations for genu flecting if you
01:39:04
want to talk about real issues come to
01:39:06
the all-in summit and if you are
01:39:08
protesting at the all-in summit let us
01:39:11
know uh what mock meet you would like to
01:39:13
have freeberg is setting up mock meat
01:39:15
stations for all of our protesters and
01:39:18
what milk you would like yeah all vegan
01:39:20
if you if your o milk nut milk
01:39:23
come five different kinds of zanth gum
01:39:26
you can from all of the nut milks you
01:39:28
could want and then they'll be
01:39:30
mindful can we have some soil likein
01:39:32
please yes on the south lawn we'll have
01:39:34
the goat yoga going on so just please
01:39:37
note that the goat yoga will be going on
01:39:38
for all of you it's very thoughtful for
01:39:41
you to make sure that our protesters are
01:39:43
going to be well well fed well taken
01:39:46
care of yes we're actually freeberg is
01:39:48
working on the protester gift bags the
01:39:50
protester gift bags they're made
01:39:54
Yak folding proteins so you're good
01:39:57
folding proteins I think I saw them open
01:39:59
for the Smashing Pumpkins in
01:40:02
2003 on fire on fire enough I'll be here
01:40:05
for three more nights love you boys
01:40:07
byebye love you besties is this the all
01:40:09
Potter open mic night what's going on
01:40:11
it's basically I'm just
01:40:15
bored let your winners
01:40:17
ride Rainman David
01:40:23
said we open source it to the fans and
01:40:25
they've just gone crazy with it love
01:40:28
queen
01:40:32
[Music]
01:40:35
of Besties
01:40:38
are my dog taking
01:40:42
driveway oh man myit will meet me we
01:40:46
should all just get a room and just have
01:40:48
one big huge orgy cuz they're all this
01:40:49
useless it's like this like sexual
01:40:51
tension that they just need to release
01:40:52
somehow
01:40:53
[Music]
01:40:58
your we need to get
01:41:04
[Music]
01:41:09
merch all right that's episode 178 and
01:41:12
now the plugs the all-in Summit is
01:41:15
taking place in Los Angeles on September
01:41:17
8th through the 10th you can apply for a
01:41:19
ticket at summit. Allin podcast.co
01:41:23
scholarships will be coming soon if you
01:41:25
want to see the four of us interview Sam
01:41:27
Alman you can actually see the video of
01:41:30
this podcast on YouTube
01:41:33
youtube.com Allin or just search all-in
01:41:35
podcast and hit the alert Bell and
01:41:38
you'll get updates when we post we're
01:41:39
doing a Q&A episode live when the
01:41:43
YouTube channel hits 500,000 and we're
01:41:45
going to do a party in Vegas my
01:41:48
understanding when we hit a million
01:41:49
subscribers so look for that as well you
01:41:51
can follow us on x x
01:41:53
/the all inpod Tik Tok is allore inor
01:41:58
talk Instagram the all inpod and on
01:42:01
LinkedIn just search for the all-in
01:42:03
podcast you can follow chth x.com chth
01:42:06
and you can sign up for a substack at
01:42:08
cho. substack docomo freeberg can be
01:42:11
followed at x.com freeberg and ohal is
01:42:14
hiring click on the careers page at ohal
01:42:16
genetics.com and you can follow saxs at
01:42:19
x.com davit saxs saxs recently spoke at
01:42:22
the American moment conference and
01:42:24
people are going crazy for it it's
01:42:26
pinned to his tweet on his X profile I'm
01:42:28
Jason kakanis I amx.com Jason and if you
01:42:32
want to see pictures of my Bulldogs and
01:42:33
the food I'm eating go to instagram.com
01:42:36
Jason inth firstname Club you can listen
01:42:39
to my other podcast this week in
01:42:40
startups just search for it on YouTube
01:42:42
or your favorite podcast player we are
01:42:44
hiring a researcher apply to be a
01:42:47
research you're doing primary research
01:42:48
and working with me and producer Nick
01:42:50
working in data and Science and being
01:42:52
able to do great great research Finance
01:42:54
Etc Allin podcast.co resesarch it's a
01:42:57
full-time job working with us the
01:42:59
besties we'll see you all next time when
01:43:01
the Allin podcast
