00:00:00
[Applause]
00:00:01
first I really appreciate just like Tim
00:00:03
has pointed out a lot of you have come
00:00:05
from very long distances so I'd like to
00:00:07
get a a sample how many people came from
00:00:10
outside the US
00:00:12
today wow nearly half the audience how
00:00:15
many are from the Stanford Community
00:00:17
this year we decided to also invite
00:00:20
students faculty and staff and so on so
00:00:22
about three or four no actually quite a
00:00:25
dozen we also extended the same there's
00:00:27
a dangerous question how many people are
00:00:29
from Berkeley
00:00:30
see that's why they're at the
00:00:33
back the thing is you're here at the
00:00:36
invitation of the dean suj kingl who's a
00:00:39
classmate of mine and as you probably
00:00:41
know she her PhD is from
00:00:44
[Laughter]
00:00:47
Stanford did you bring the axe though
00:00:49
did you follow the
00:00:51
instructions
00:00:52
no okay they have the axe this year okay
00:00:55
so with that out of the way I've I've
00:00:58
cleared the minds I'm going to talk
00:01:00
about something that is actually a very
00:01:02
hot topic but it's hot but it's actually
00:01:05
people don't agree on what the heck is
00:01:07
an agent we've used the term agent for a
00:01:09
very long time but more importantly in
00:01:12
the nature of this conference it's about
00:01:14
industrial AI so hopefully by the time
00:01:16
I'm done you'll carry away at least one
00:01:18
thing which is going to be that it turns
00:01:22
out generative AI is actually more
00:01:26
important more significant to the
00:01:28
industrial physical world
00:01:30
in my view than it is to the digital
00:01:32
world and some of you may know I have
00:01:34
some creds from the digital world I I
00:01:36
help launch Gmail right so when I say
00:01:38
this is more significant over there it's
00:01:40
probably true first I want to start out
00:01:42
this is something that we all know right
00:01:45
but I want to put it on a on a slide
00:01:46
America today we Face a
00:01:49
de-industrialization crisis this is
00:01:51
actually the economic theory that we've
00:01:53
pursued for the last 30 years right and
00:01:55
then we find one day that actually no we
00:01:57
actually still need to make things we
00:01:58
can't really just Outsource it to to
00:02:01
everyone and suffer some geopolitical
00:02:02
risks but one thing that some of you may
00:02:05
be very intimately familiar with but
00:02:08
others are not is that we have an
00:02:10
experiment in progress right and that is
00:02:13
tsmc we actually have some guests or
00:02:16
some part Partners here from tsmc so you
00:02:18
know the story as well or even better
00:02:20
than I do we have an experiment of tsmc
00:02:23
being launched new Fabs in kushu with
00:02:27
pretty much the same or even less budget
00:02:29
than the the one in Phoenix and I'm not
00:02:31
going to spend the whole talk going over
00:02:33
all the details of this but I hope I
00:02:36
don't offend anyone by pointing out that
00:02:38
the right hand side has been a failure
00:02:41
and the left hand side has been a huge
00:02:43
success and the only variable is the
00:02:46
people involved right so that that says
00:02:49
something but the main thing some some
00:02:52
have referred to it as culture but I
00:02:54
think of it specifically as what we the
00:02:56
Lost critical expertise it's quite
00:02:59
interesting at an AI conference that I
00:03:01
say what's really important is actually
00:03:03
what's in your heads what's in your
00:03:05
brains
00:03:06
still and and that's going to be true
00:03:08
for a very long time in in my view right
00:03:10
AI is not going to replace us it's going
00:03:12
to enable us so that our expertise
00:03:14
become even more important so that's the
00:03:16
message that I want to carry here you
00:03:18
may have seen the people this why saying
00:03:21
that crisis in Chinese the word is
00:03:22
Crisis and opportunity right is luck
00:03:25
right so America also has a
00:03:27
reindustrialization opportunity and in
00:03:30
that sense we're not going to be
00:03:31
bringing lowend manufacturing back it's
00:03:33
going to be Leaf frogging right so in
00:03:36
the sense there's a we talk all the time
00:03:38
about Asia and the developing World sort
00:03:40
of Lea frogging going from copper lines
00:03:44
directly to mobile and so on so America
00:03:45
actually has an opportunity to LeapFrog
00:03:47
all of the manufacturing burdens right
00:03:51
of that has not taken advantage of AI
00:03:54
but AI can help here so I want to go
00:03:57
kind of some of you may have seen me
00:03:59
done do this right and you look at the
00:04:03
screen what it is is survey that
00:04:05
essentially measures technology optimism
00:04:09
right some of you may have heard the
00:04:11
term Tech Bros here in Silicon Valley
00:04:13
usually techno Optimist right but tech
00:04:16
technology optimism is a thing right and
00:04:18
I think it's a good thing and it's no
00:04:20
surprise I think I decided to take this
00:04:23
data of this survey and I plotted versus
00:04:26
country and so you can GDB per capita
00:04:29
and it turns out first there's a strong
00:04:32
correlation right higher GDB countries
00:04:35
oh this chart is okay this is AI
00:04:38
optimism I'm so sorry okay the headlines
00:04:40
is Switched okay ignore the headline
00:04:42
listen to me and look at the chart
00:04:44
technology
00:04:46
optimism goes up with GDP per capita
00:04:49
that's no surprise because Google
00:04:51
Facebook and so on we are in the rich
00:04:53
countries we can take advantage of that
00:04:54
a lot more everybody's optimistic but
00:04:56
some are more optimistic than others the
00:04:58
surprise perhap is that there's a a
00:05:00
correlation at all and you already saw
00:05:02
the last slide when they did the same
00:05:04
thing for AI but AI is the opposite
00:05:07
that's weird
00:05:09
right the correlation is even stronger
00:05:11
65% in terms of 65% of the variation is
00:05:14
explained by the line
00:05:17
alone that's the 2022 GTP per capita and
00:05:20
so on you notice United States we're
00:05:23
less optimistic or more pessimistic than
00:05:26
China India Peru and so on and that
00:05:29
means something because if you're more
00:05:31
optimistic about something you're more
00:05:32
likely to apply it and if you're going
00:05:34
to more more likely to apply it you're
00:05:35
going to win more with it so that's
00:05:37
something important to put in mind keep
00:05:39
in mind I'm not going to go through some
00:05:41
analysis of it other than to share with
00:05:44
you this interesting so I decided to
00:05:46
take this chart and I plot the residual
00:05:48
right so think of that as optimis but an
00:05:51
extra optimism you can see some of the
00:05:53
curves here are more optimistic than
00:05:56
others like for China unreasonably
00:05:58
optimistic way up here
00:06:01
and over here like France is always
00:06:05
below the curve and you look at the the
00:06:07
residuals here Saudi Arabia is extra
00:06:11
Optimist right that is above the trend
00:06:14
line and it's not surprised that France
00:06:19
is like this on every chart in terms of
00:06:22
optimism that's okay I grew up speaking
00:06:24
French before I spoke English decided to
00:06:26
correlate that right if you look at
00:06:27
these economies that are that tend to be
00:06:30
lower GDP per capita it turns out that's
00:06:32
because they've been still making things
00:06:34
they've been doing so-called lower value
00:06:36
added stuff versus the advertising the
00:06:38
marketing that we tend to do here in the
00:06:39
U us so what do you think it'll look
00:06:42
like if we took that chart and we
00:06:44
correlated
00:06:45
against industrialized the industrial
00:06:48
content of the economy right this is as
00:06:51
a perent of their GDP do you think
00:06:54
countries that are still more industrial
00:06:57
would be more or less optimistic about
00:06:59
AI how many people think they'll be more
00:07:03
optimistic how about
00:07:06
less okay let's take a look at the
00:07:09
data huh higher the economy is still
00:07:12
making things the more optimistic they
00:07:14
are about AIS from the same data set so
00:07:17
that says something right that says that
00:07:19
Vietnam Japan China is going to be more
00:07:23
optimistic and therefore more likely to
00:07:25
adopt gen than we here in the US in fact
00:07:29
some of how many people have seen some
00:07:30
of the recent data where they survey
00:07:33
businesses like how many is your
00:07:35
business is already adopting some kind
00:07:36
of gen in your day-to-day workflow the
00:07:40
the number coming from China is 85% it's
00:07:43
amazing I I was going to say scary but
00:07:44
it's amazing right and as we may pride
00:07:47
ourselves on creating stuff but it's the
00:07:50
people who use it the people who scale
00:07:52
things right like the tsmc problem is
00:07:55
not a science problem it's a people
00:07:57
problem we cannot scale that right
00:08:00
battery manufacturing Panasonic my
00:08:02
Panasonic friends here we working with
00:08:04
with Tesla gigafactory it's a scale
00:08:06
problem you hear announcements of
00:08:08
amazing battery technologies all the
00:08:10
time none of them scales right it's a
00:08:13
scaling that matters the most to the
00:08:15
economies so let's keep this in mind I
00:08:17
decided to to just I was I like playing
00:08:19
with things so I decided to just impute
00:08:22
a model that for every percent optimism
00:08:26
above some level then that may give you
00:08:29
in this case I just say 0 2% on the GDP
00:08:33
impact and simulate that out 10 years to
00:08:36
see what happens if you can see the 2022
00:08:39
GDP here the gray where the US is Right
00:08:43
still very larger than much larger than
00:08:45
China the other data points is quite
00:08:46
interesting but we don't have time to go
00:08:48
over but just because of this extra
00:08:52
optimism right by the 2020 32 output of
00:08:56
China is going to Tower over the US
00:09:00
and along with that standards of living
00:09:01
and technology and everything else so
00:09:04
anyway I think it's something to think
00:09:06
about right this just the the the
00:09:09
optimism and the adap adoption of
00:09:11
Technology
00:09:12
alone okay but let's get more specific
00:09:16
right I I think I'm putting some
00:09:18
statements here that some of you may
00:09:20
already be familiar with quite
00:09:22
intimately but to others that you may be
00:09:24
surprising so I'm going to go over them
00:09:25
one by
00:09:27
one generic llms right maybe this is the
00:09:30
first time you hear the term generic
00:09:31
because I'm going to talk about
00:09:32
something that isn't generic but think
00:09:34
of open AI uh
00:09:37
Google and everything else as generic
00:09:40
right there they know a lot about many
00:09:42
things but they're broad but they lack
00:09:44
domain specific knowledge and that's
00:09:45
actually quite important we all know
00:09:48
about how many of the responses can be
00:09:50
hallucination and people are trying to
00:09:52
fix that in the model themselves and
00:09:54
more importantly people are trying to
00:09:55
fix it in the systems that that that
00:09:58
leverage these models
00:10:00
and something
00:10:02
that people are starting to think about
00:10:04
more particularly again in the
00:10:06
industrial
00:10:07
applications is that the very
00:10:10
probabilistic nature of these llms are
00:10:13
actually quite
00:10:14
undesirable right you can run these
00:10:16
academic uh tests in your paper and say
00:10:19
wow accuracy 90% and so on but these are
00:10:23
almost cherry-picked because in
00:10:24
industrial and in other many business
00:10:26
applications you need the answer to be
00:10:29
the the same every single time you ask
00:10:31
that same question it's a business
00:10:33
process you don't want creativity you
00:10:35
don't want the LM say every once in a
00:10:36
while I'm going to mix it up so the
00:10:39
inconsistency and the shallowness of the
00:10:42
answer I'd like to compare a fresh PhD
00:10:44
with somebody with 20 years of
00:10:45
experience fresh be is very useful when
00:10:48
you ask about dry choros silan in
00:10:50
semiconductor they can tell you all the
00:10:52
possible uses and so on then you say
00:10:54
what do I do with this problem I don't
00:10:56
know what to do I'm going to repeat
00:10:57
everything that I've learned but
00:10:59
somebody with 20 years of experience in
00:11:01
that case what you need to do is
00:11:02
increase the flow rate to 200 secm and
00:11:04
so on and so forth so think of these
00:11:06
wonderful amazing llms as fresh phds
00:11:10
right they're amazing in it of
00:11:12
themselves but they have no domain
00:11:13
specific knowledge and they don't have
00:11:15
the experience that the people in the
00:11:17
room here do have and that matters to
00:11:20
systems that we build right so domain
00:11:22
expertise is actually the key to success
00:11:26
in the physical world and this is not
00:11:29
something that I just reasoned through
00:11:30
these slides this is something that when
00:11:32
I was helping to run Global industrial
00:11:34
AI at Panasonic we actually ran into
00:11:36
problems that AI could not solve without
00:11:40
consulting with the the one or two
00:11:42
expert that Panasonic has in
00:11:43
refrigeration and with that we can
00:11:45
actually solve the predictive
00:11:46
maintenance problem the opportunity is
00:11:49
that domain expertise we always knew it
00:11:52
was valuable but it's been the reason we
00:11:54
don't use it think about this okay a lot
00:11:56
of time in technology we do something
00:11:59
and we don't realize it's actually a bug
00:12:01
it's not a feature like cyberspace is a
00:12:03
bug not a feature Cy the reason we
00:12:06
create had to coin the term cyber space
00:12:08
because our computers couldn't reach
00:12:09
into physical space so same thing
00:12:12
relying exclusively on data is actually
00:12:14
a bug it's because we couldn't use
00:12:16
knowledge because it's too difficult to
00:12:18
encode but with Gen today suddenly we
00:12:21
can speak to our machines and they
00:12:24
understand us I think that's the biggest
00:12:25
Revolution it's not that they're smart
00:12:27
it's that they know enough to listen to
00:12:30
understand for the first time we can
00:12:31
actually easily interface with the
00:12:34
analog world right natural language
00:12:37
vision and so on that's been very
00:12:39
difficult until now so that's the key
00:12:42
opportunity so with that I want to give
00:12:45
you this Vision actually the product
00:12:48
that we're building is going to happen
00:12:49
pretty soon whoa where are we okay it
00:12:53
essentially we're a lot of us work on so
00:12:57
I want to talk about this model of
00:12:59
expertise in terms of capture and apply
00:13:02
you capture the expertise and you apply
00:13:03
it sounds very obvious but you look at
00:13:05
the landscape across the landscape
00:13:07
today because of this very datadriven
00:13:11
point of view in machine learning we
00:13:13
don't think about the capture as much
00:13:14
the capture is the dirty part right we
00:13:16
think about the apply part the
00:13:18
algorithms but how do you then do the
00:13:21
same thing in the in in the knowledge
00:13:22
age how do I capture Sam Samsung experts
00:13:27
knowledge and put that that into my
00:13:29
system right I just a lot of the
00:13:32
activities today assume that these
00:13:33
documents somehow magically exist and
00:13:36
then when you look at the document they
00:13:37
tend to be manuals and so on but the
00:13:39
knowledge the experience of having gone
00:13:42
through the 20 years of work at a
00:13:44
facility in a Fab is not actually in
00:13:46
those documents so how do you so I want
00:13:48
to talk about that capture so think
00:13:50
about this as a faul diagnosis AI system
00:13:53
this is actually a real use case that
00:13:55
we're working on right but the first
00:13:58
step has to be
00:14:00
capturing the knowledge of somebody who
00:14:01
is actually about to retire and I'm not
00:14:03
saying that figuratively I'm I'm saying
00:14:05
that actually but there's somebody in
00:14:06
the audience here that's involved in
00:14:08
this project so let's start
00:14:10
talking okay go the next
00:14:17
step okay so we we can have a process
00:14:20
where I'm saying wondering etching
00:14:21
process fluctuate even a little bit this
00:14:24
is what I mean for the first time I can
00:14:25
get an expert to sit down with me with a
00:14:28
machine and just start dictating just
00:14:30
start saying extemporary
00:14:33
extemporaneously this is you know what
00:14:34
and we went through this process right
00:14:38
and then with Gen we can actually say I
00:14:40
can say okay is this what you said right
00:14:43
that's what I mean by by the easy part
00:14:45
okay let's go to the next
00:14:48
step okay yes
00:14:51
correct okay and then you can now we
00:14:54
save that and then we can now structure
00:14:56
it automatically right okay we can put
00:14:59
it in some structured form in this case
00:15:01
this yamamo file but you can imagine
00:15:03
that could be some other symbolic
00:15:05
language it could be python it could be
00:15:07
anything right you can run through this
00:15:09
process many times right and this is
00:15:11
what I mean by encoding the knowledge
00:15:13
okay and then you can go from there to
00:15:14
another transformation you can build a
00:15:17
diagnosis program okay this is a simple
00:15:20
example you can say yeah I'm just going
00:15:21
to hire some python programmer to do
00:15:23
that but you can't do that for free
00:15:26
flowing kind of and sharing of knowledge
00:15:28
and rules and so on you can do that and
00:15:30
you can simulate and see the result all
00:15:33
right so now you have one piece of
00:15:35
knowledge that is now
00:15:36
operationalized right it can essentially
00:15:38
say detect that the risk of
00:15:41
non-uniformity is very high because of
00:15:43
fluctuations in the in the plasma power
00:15:47
okay any any questions or comments about
00:15:49
this but you can repeat this process
00:15:51
right we can automate this now and that
00:15:53
I think that's the
00:15:55
revolution okay so with that I want to
00:15:58
talk about
00:15:59
the Paradigm behind this so the way you
00:16:02
can think about this is that remember
00:16:04
earlier I referred to the term generic
00:16:06
model right that's those are models that
00:16:08
know a lot but not specifically about my
00:16:11
process it's not it's dumb it's not that
00:16:14
it's DB it's not specialized enough to
00:16:16
my language to my system my processes
00:16:18
and so
00:16:19
on so what we want to have is an expert
00:16:22
model and earlier Tim spoke of semic
00:16:25
Kong and there's a booth over there
00:16:27
result of AI Alliance work
00:16:29
and then you have this expert model and
00:16:32
remember this expert model can
00:16:34
ladder you can have an industry
00:16:36
expertise and then you can have company
00:16:37
specific tool specific even process
00:16:40
specific right and that's the
00:16:41
opportunity but even that is not enough
00:16:44
because models as they go do not have
00:16:47
something that many of you may have
00:16:48
heard of planning and reasoning
00:16:50
capabilities the ability to Loop over
00:16:53
and reason through a process or a
00:16:55
problem so you'll see this emerging so
00:16:57
I'd like to share with you some a bit of
00:16:59
details about semic Kong has has been
00:17:02
spoken this is an effort under the AI
00:17:06
Alliance and hopefully throughout the
00:17:08
day today and tomorrow you'll have an
00:17:10
opportunity to talk with some of the a
00:17:11
AI Alliance Representatives here today
00:17:14
that advocates for open science and open
00:17:16
development and the alliance has a
00:17:19
particular methodology or this is or an
00:17:22
Mo that is we want to actually output
00:17:24
things not just sit there and talk about
00:17:26
nice things so semicolon is one of those
00:17:27
output right and it's built on Lama 3
00:17:30
which is an open model as well right and
00:17:33
what's the most interesting thing about
00:17:35
building something on open right the
00:17:40
slightly counterintuitive thing is
00:17:41
because it is open you hear in the
00:17:43
audience that have very secret very
00:17:46
competitive processes you can take that
00:17:49
open model and add your thank you I have
00:17:52
that and add your own knowledge without
00:17:56
having to pay a tax back to somebody
00:17:58
who's a propri model underneath so this
00:18:00
openness is very important to Innovation
00:18:03
right it means you can just if you want
00:18:04
to contribute back that's great but if
00:18:07
you need to build something proprietary
00:18:09
to tsmc or Tokyo electron or apply
00:18:11
materials you are free to do I think
00:18:13
that's what we want to promote and this
00:18:15
is for those of you that don't come from
00:18:17
the semiconductor industry This is
00:18:20
highly unusual for semiconductors right
00:18:22
if you come from software you say oh
00:18:23
that's a big deal I've been doing this
00:18:25
forever but in semiconductors even the
00:18:27
vendor and the customer have a clear
00:18:30
wall between them you may not know what
00:18:32
I'm using your equipment for right and
00:18:35
so having the ability to somehow share
00:18:38
so that we can compete on a higher plane
00:18:40
rather than at the lower levels so this
00:18:44
is an example applying actually semicon
00:18:46
being used to apply domain specific
00:18:49
agents to a semiconductor manif process
00:18:51
in this case it's an
00:18:53
reer okay and here it's using it for I
00:18:56
believe it is for process engineering
00:19:00
okay so I already talked about AI
00:19:02
Alliance take a look at the website the
00:19:05
alliance. okay today over 100 members
00:19:08
already and I'm quite proud of the fact
00:19:10
that it's not just us companies over
00:19:12
here but also a lot of the industrial
00:19:15
companies from Asia and certainly well
00:19:18
very well represented by our ja Japanese
00:19:21
members so agentic AI I mentioned
00:19:25
planning and reasoning I'll give you one
00:19:27
example this is not the the only model
00:19:29
but one example what we mean concretely
00:19:32
because you hear the term planning and
00:19:33
reasoning all the time agent in the
00:19:36
meaning that I'm using here and you will
00:19:39
see more and more people converging
00:19:41
toward this in the generic in the
00:19:43
generative AI era right we want to use
00:19:45
the word agent to
00:19:47
mean something that is goal oriented and
00:19:51
essentially it has the property of
00:19:52
planning and reasoning in order so that
00:19:54
it can iterate over a particular problem
00:19:57
statement and then it say I solved that
00:19:59
problem yet I haven't let's figure out
00:20:01
how to iterate and go there and models
00:20:03
as they are today LM by themselves don't
00:20:06
have that capability okay and to the
00:20:08
extent that you see chat GPT or others
00:20:12
that that seem to have this it is
00:20:13
actually because they have put a
00:20:15
planning and reasoning framework around
00:20:16
it right the model themselves don't do
00:20:18
that so one hierarch or one architecture
00:20:21
that we use atomatic is what called
00:20:24
hierarchical task planning that is given
00:20:26
a task break it down into subtask asks
00:20:29
right and then you ask you can your
00:20:32
system not the model itself but you ask
00:20:34
the system says given that can I solve
00:20:36
the task in one step and if the answer
00:20:38
is yes go ahead and do it if the answer
00:20:40
is no break it down further so pretty
00:20:42
straightforward in terms of that
00:20:43
planning idea within each task then
00:20:46
there's a reasoning Loop right and you
00:20:49
can use different paradigms again this
00:20:51
one is an UDA Loop right how many people
00:20:53
are familiar with the term
00:20:55
UDA great okay so I don't have to go
00:20:57
over that it's if it's good enough for
00:21:00
jet fighter pilots in in deadly
00:21:02
situations is probably good enough for a
00:21:04
lot of process engineering work but you
00:21:06
go through this process of observe
00:21:08
Orient decide and act You observe the
00:21:11
environment the resources that you have
00:21:13
available and then you Orient you decide
00:21:15
is that do I have sufficient information
00:21:19
resources to solve the problem and
00:21:21
depending on the answer you decide what
00:21:22
to do next then you take the
00:21:24
action and then now one second or one
00:21:27
day has elapsed the world has changed
00:21:29
you go through that loop again okay so
00:21:31
that's the UDA Loop so this is one very
00:21:34
significant
00:21:35
representation the the planning and
00:21:37
reasoning and I don't know if there's
00:21:40
going to be a talk with what a colleague
00:21:41
of mine Ving L is the leader in a
00:21:43
project called open SSA stand for small
00:21:46
specialist agents and so you can go to
00:21:49
you look for op SSA and there's an
00:21:51
implementation of this again it's open
00:21:53
source so you can download and use it
00:21:55
and here I just want to highlight many
00:21:58
of the the guests here that have
00:22:00
participated in this revolution
00:22:01
alongside with isomatic right and these
00:22:04
are real world use cases that have been
00:22:07
that are either in production or in in
00:22:10
development today right and
00:22:13
so I say this like it's nothing but it
00:22:18
turns out in Silicon Valley in the
00:22:20
digital world we're actually there's a
00:22:23
lot of startups that are struggling to
00:22:25
find problems for which gen is the
00:22:28
solution
00:22:29
so so I've been around long enough to
00:22:30
say this is very weird the industrial
00:22:33
companies are the first adopters of a
00:22:35
technology that's never happened before
00:22:38
right because Industrials are slow maybe
00:22:40
from Silicon Valley perspective
00:22:43
stupid I can say that because I was
00:22:45
Panasonic too but it turns out it's not
00:22:48
it's because the Industrials have been
00:22:49
working on much more difficult problems
00:22:52
when I was at Google I make I like to
00:22:54
say when I was at Google when I make a
00:22:56
mistake you click on the wrong ad but at
00:22:58
at Panasonic if I make a mistake
00:23:00
somebody dies so it's very reasonable it
00:23:03
turns out that these industries have
00:23:05
been moving more carefully and there's
00:23:08
something about genitive Ai and I think
00:23:10
it it lies in the fact that you can
00:23:12
easily suddenly you can easily capture
00:23:15
all of that domain expertise all of that
00:23:17
pent up Demand with this technology is
00:23:20
it's different this time essentially and
00:23:23
that's why the industrial world I
00:23:24
believe will leap ahead in terms of the
00:23:27
adoption of of generative AI okay so the
00:23:32
call to action here is essentially don't
00:23:34
fall behind right lead the industrial AI
00:23:37
Revolution do it with us do it with this
00:23:39
Paradigm this based llm model with the
00:23:42
expert and with agents above that and if
00:23:45
we do
00:23:46
that yeah maybe okay so I basically did
00:23:50
the same analysis said what if the US
00:23:51
was equally optimistic about AI this is
00:23:55
what the future is going to look like
00:23:57
2032 right right okay so essentially
00:24:00
that's the additional GDP growth that
00:24:03
that comes from AI adjusted Delta alone
00:24:06
okay so I think that I would like to
00:24:08
leave you with this vision for those
00:24:10
that care care about the development of
00:24:12
of the US industrial World okay so what
00:24:15
the world wants and what
00:24:17
industrialization needs is to capture
00:24:20
and apply domain expertise number one
00:24:22
very specific we can use help from AI
00:24:26
right and we we can use a lot less fear
00:24:29
and a lot more optimism and I'd like to
00:24:31
leave you with that thought thank
00:24:34
[Applause]
00:24:40
you thank you Dr W for the insightful
00:24:43
talk on AI optimism across the globe and
00:24:46
the French is pessimism specifically so
00:24:49
definitely a takeaway for me but before
00:24:52
we let you go we like to see those in
00:24:54
the audience if you have any questions
00:24:55
for Dr W
00:24:59
why are domain specific llms not
00:25:01
sufficient why do we need to combine
00:25:02
them with a gentic
00:25:05
AI okay I'm sure everyone heard the
00:25:07
question quite simply it's because llms
00:25:10
or models in general don't have planning
00:25:12
and reasoning more generally you think
00:25:15
about models As One path things right
00:25:17
there's an input there's process and
00:25:19
there output that model that that
00:25:21
Paradigm alone tells you that it cannot
00:25:23
solve problems you still need something
00:25:25
outside of that Loop to say okay given
00:25:28
that now the world has changed what do I
00:25:30
do about it there has to be a recurrent
00:25:32
Loop coming back and that Loop is the
00:25:33
planning and reasoning Loop now if you
00:25:35
build a system that has that built in
00:25:38
then that system can do the planning and
00:25:40
reasoning but llms as they are built
00:25:42
today don't have that Loop so it's hard
00:25:45
to imagine that llm is the asymptotic
00:25:47
limit of what AI is going to look like
00:25:50
and you can look at the agentic
00:25:51
workflows is basically confession that
00:25:54
you need a lot of coures to make llm do
00:25:57
something useful real life L and
00:25:59
robustly so what do you think is going
00:26:00
to be the next step after llm post llm
00:26:04
if you will post agentic llm even that's
00:26:06
a actually very sophistic question so
00:26:08
let me try to frame it in the way that I
00:26:11
understand it in other words all of this
00:26:13
coding that does the planning and
00:26:14
reasoning outside of the model that
00:26:16
seemed like a crutch right because with
00:26:18
these lolms and with these models It's
00:26:20
Magic we just feed a bunch of data into
00:26:22
it and intelligence emerges it doesn't
00:26:24
feel good to add this extra layer into
00:26:27
it let me accommodate that first and
00:26:30
then let me deconstruct it later so the
00:26:32
way I accommodate it is that in in fact
00:26:35
if you think about the architecture
00:26:37
inherently it's what I just said it
00:26:38
doesn't have recurrence right and
00:26:40
recurrence is what gives there's
00:26:42
recurrence in our own brain we actually
00:26:45
when we solve a problem we turn it over
00:26:47
our mind right we even touch the world I
00:26:50
push a car and it moves and then
00:26:51
something else happened so there's
00:26:53
recurrence outside the brain as well so
00:26:56
recurrence or that Loop is is inherently
00:26:59
necessary and there are
00:27:01
emerging it's not there um recurring
00:27:05
models have been around a long time the
00:27:08
reason we don't do it because it's too
00:27:09
expensive right and so with with more
00:27:11
compute with with recurrent becoming
00:27:13
cheap that will be built into models
00:27:16
right in fact there's a later model I
00:27:20
forget what it's called WK WQ
00:27:23
KV that that that that has a property of
00:27:27
being recurrent on the uh on on the
00:27:31
training side sorry on the it's using
00:27:34
more Transformer paralyzation on the
00:27:35
training side and recurrence on the
00:27:37
inference side and doing it in such a
00:27:39
clever way that the weights are the same
00:27:41
okay so the the general point is that
00:27:43
recurrence will come and problem solving
00:27:47
will come with it along with
00:27:49
models maybe that I can stop there but
00:27:52
I'd like to still speak for
00:27:55
code I don't think of what we're doing
00:27:57
with plan and reasoning in terms of
00:27:59
doing in high level code is a crutch I I
00:28:02
think it's inherently part of the
00:28:04
knowledge Paradigm if we can have models
00:28:07
that generate this code then why not
00:28:10
just use that at that higher abstraction
00:28:11
we don't have to go all the way down to
00:28:13
every neuron all the time so that's my
00:28:15
point of view so you showed an
00:28:17
interesting example of overcoming the
00:28:19
lack of domain specific knowledge in llm
00:28:22
is by providing an expert in this case a
00:28:24
human expert who's interacting via llm
00:28:27
one of the ch challenges and one of the
00:28:29
problems I'd be curious to hear your
00:28:30
thoughts on how to overcome is that
00:28:32
sometimes these experts will come to
00:28:34
different conclusions so two experts may
00:28:36
have disagreeing opinions or even if
00:28:39
they have the same conclusion it may be
00:28:40
wrong and how do you overcome that
00:28:43
problem when testing this requires
00:28:46
actually interacting with a physical
00:28:47
system which is expensive and has a very
00:28:49
long time scale yeah that's a really
00:28:52
good question that comes up all the time
00:28:54
right and somehow we are always
00:28:57
difficult when when we come to humans
00:28:58
well we accept machines more readily
00:29:00
what I'm getting to is machines also
00:29:03
come to different conclusions right but
00:29:05
we don't worry about us the reason is we
00:29:07
tend to think of them as much more
00:29:08
deterministic right the different
00:29:10
conclusions themselves so we no longer
00:29:13
in these systems we don't think of
00:29:14
correct versus incorrect we think of
00:29:16
better versus worse right so the
00:29:18
difference is first of all in how many
00:29:22
cases out of a 100 does that occur so
00:29:24
let's say there's 5% so the other 95% is
00:29:27
not an issue we can still deploy that in
00:29:29
those 5% there are different ways to
00:29:31
handle them in other words we can say
00:29:33
let's look at the two choices and have
00:29:35
another human or another Model come in
00:29:37
and say what do you want to do so these
00:29:39
are different options right in fact
00:29:41
that's built into the deployment today
00:29:44
we're not comfortable enough to go fully
00:29:46
autonomous there's always a human that
00:29:48
is saying he here's the recommended next
00:29:51
bet action you want to push the button
00:29:53
somewhere essentially it's resolved at
00:29:55
that
00:29:56
level okay
00:29:59
in semiconductor area the way we capture
00:30:02
the domain knowledge is through the
00:30:05
standard operation procedure which is
00:30:07
based on statistical knowledge over the
00:30:10
years so there's sop and recipe develop
00:30:15
which include the knowledge from expert
00:30:17
as well as from the machine why don't we
00:30:19
use that as a way to capture the
00:30:22
knowledge instead of going back to the
00:30:24
expert and there are so many s so
00:30:27
already ready available make you the way
00:30:30
you try to capture an knowledge
00:30:32
easier yeah that's a good question so
00:30:36
the answer is not either or the answer
00:30:38
is we have been taking advantage of all
00:30:40
those Sops already semiconductor
00:30:43
companies are very advanced in having
00:30:45
all of these things right but even that
00:30:47
is not enough there's stuff that sort of
00:30:49
escapes these things so for example I
00:30:52
had one convers so when we sit down and
00:30:54
do these interviews when we start out we
00:30:56
do this manually without these tools
00:30:58
what I always say is that I'm not
00:31:01
interested or don't talk about stuff
00:31:02
that's already been documented it's
00:31:04
quite rich but tell me about a problem
00:31:06
an incident that happened in the last 5
00:31:08
years when for some reason and I say
00:31:11
don't be embarrassed don't be too humble
00:31:14
I said when only you if you were not
00:31:16
there it would not have been solved tell
00:31:18
me about that and really interesting
00:31:20
stuff emerged so one example is one
00:31:22
gentleman from company facility in in in
00:31:26
Phoenix talked about
00:31:28
yield problems that were took a long
00:31:31
time to trace back to pressure
00:31:33
fluctuations in a
00:31:35
chamber and then they didn't know why
00:31:37
that was happening but because he's a
00:31:39
process engineer but he also talks a lot
00:31:41
to the facilities people and then he
00:31:43
remembers that there was a piece of
00:31:44
equipment that used the same gas line
00:31:46
that was installed like just a week
00:31:48
earlier right and they weren't careful
00:31:50
with it and that led to that that that
00:31:53
fluctuations would they have figured
00:31:55
that out by putting 10 20 people
00:31:57
together yeah it would may have taken a
00:31:59
week a month but because that gentleman
00:32:01
was there he's in his mind he said okay
00:32:03
take a look at that and it was solved
00:32:05
quickly and that piece of knowledge was
00:32:07
not in the standard operating procedure
00:32:10
yeah because basically now we have a
00:32:11
person who knows both facilities as well
00:32:13
as
00:32:14
process yeah good question and that's
00:32:16
why I mentioned earlier we do we we have
00:32:19
all these documents that we're ingesting
00:32:22
but I think the the grand opportunity is
00:32:24
the stuff that is not in those documents
00:32:25
already
