March 13, 2025 • 24 mins
What happens when AI diagnoses patients better than doctors? Where does artificial intelligence truly stand on the hype cycle? Is all this computational power actually benefiting society? These critical questions frame our fascinating discussion recorded live at IMAPS Device Packaging Conference in Phoenix, Arizona.
Join our expert panel featuring Hemanth Jagannathan (IBM Research), Mark Kuemerle (Marvell), and Kimon Michaels (PDF Solutions) as they tackle AI's most pressing challenges and opportunities. Their collective expertise reveals surprising insights about how AI is transforming industries while raising important considerations about its implementation.
The conversation explores AI's evolution from specialized technical applications in semiconductors to today's consumer-facing generative tools. Our experts draw fascinating parallels between AI and previous technological breakthroughs like laser technology, suggesting we've only scratched the surface of potential applications. They provide compelling examples from healthcare where AI systems demonstrate superior diagnostic capabilities by processing complex datasets beyond human capacity.
While acknowledging concerns around data accuracy, power consumption, and appropriate boundaries, the panel remains optimistic about AI's future. They emphasize that today's implementations represent merely the beginning of a transformative technology whose full impact remains largely unanticipated. Yet they also agree on applications where human judgment should remain primary – including, amusingly, matchmaking.
Dive into this thought-provoking conversation to understand why organizations must either leverage AI effectively or risk being outpaced by competitors who do. Subscribe to 3D IncItes Podcast for more cutting-edge discussions on technologies shaping our future.
Discover how technology is reshaping our lives and livelihoods.
Listen on: Apple Podcasts Spotify
IMAPS Device Packaging ConferenceInterconnects for Tomorrow’s Applications
Disclaimer: This post contains affiliate links. If you make a purchase, I may receive a commission at no extra cost to you.
Become a sustaining member!
Like what you hear? Follow us on LinkedIn and Twitter
Interested in reaching a qualified audience of microelectronics industry decision-makers? Invest in host-read advertisements, and promote your company in upcoming episodes. Contact Françoise von Trapp to learn more.
Interested in becoming a sponsor of the 3D InCites Podcast? Check out our 2024 Media Kit. Learn more about the 3D InCites Community and how you can become more involved.
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Francoise von Trapp (00:00):
This episode of the 3D Insights
podcast is sponsored by IMAPS,the premier global association
for microelectronics advancedpackaging enthusiasts.
A membership in IMAPS helpsyour company grow its advanced
packaging workforce throughprofessional education and
networking, advances your brandand supports building
relationships.
Imaps helps you learn, connectand collaborate.
Learn more at IM imapsorg.
(00:22):
Hi there, I'm Francoise vonTrapp, and this is the 3D
Insights Podcast.
Hi everyone, this week we arerecording live from the IMAPS
(00:48):
Device Packaging Conference inPhoenix, arizona.
Now AI continues to be a hottopic of discussion at industry
events like this one, and thisweek several panels were
dedicated to discussing topicssuch as are we going to have an
AI winter?
Or, as part of the GlobalBusiness Council, how are we
(01:08):
going to scale AI from datacenters to consumer applications
?
So here to talk about the keytakeaways were some of the
panelists.
From each of the panels I haveKamef Jaganathan of IBM Research
, mark Pemberley of Marvell andKim and Michaels of PDF
Solutions.
Welcome to the podcast, guys.
Mark Kuemerle (01:28):
Great to be here.
Thank you very much.
Francoise von Trapp (01:30):
So we've got a lot of questions to go
through, but before we getstarted, can you each tell me a
little bit about your companyand your role there, Kim and?
Kimon Michaels (01:39):
Sure, Kimon Michaels.
I'm one of the co-founders andEVP of Products and Solutions at
PDF Solutions.
Pdf Solutions is the big dataanalytics platform for the
semiconductor and electronicsupply chain.
Hemanth Jagganathan (01:52):
Yes, I'm Hemanth Jagannathan.
I'm from IBM Research.
Our overall charter at IBMResearch is to be part of the
future of compute.
We work on all three elementsof compute, which is bits,
neurons and qubits.
So I'm currently in charge ofthe Chippler Advanced Packaging
(02:12):
Research Group at IBM.
Mark Kuemerle (02:14):
Okay, and Mark.
Hi, I'm Mark Kumerle.
I'm VP of Technology forMarvell.
I'm responsible forarchitecture for all of our
custom products and alsodefining a roadmap to make those
products possible.
Francoise von Trapp (02:27):
Excellent.
So let's dive in on the topicof AI, which is driving this
industry.
I want to start with asking youwhere would you say AI is on
the Gartner hype cycle at thispoint?
Kimon Michaels (02:42):
In our industry.
I think it's still on theuptick in the beginning.
Manufacturing semiconductors,semiconductors in general, is
not the fastest moving industryand I think we're still getting
traction.
We're not in a trough yet,which invariably will come.
Francoise von Trapp (02:56):
So for using AI in our industry, we're
still in the up cycle.
Kimon Michaels (03:00):
I think we're still in the up cycle, but how
about as a driver of ourindustry?
You know people are starting toget concerned about the draft.
If you look at the panel lastnight, worry about is it the AI
winter?
I was at an investor conferencewhere people were very
interested in the NVIDIApresentation more about the
future, not about themcontinuing to have excellent
(03:23):
results.
So there's certainly a concern.
It remains to be seen.
Hemanth Jagganathan (03:29):
I think we're still very early in the
era of AI, because, whiletraditional AI and the solutions
out there are driving theindustry today, it's still in
its infancy.
We need to talk about anecosystem of design,
architecture, having a varietyof options for an AI need, for
(03:51):
example.
So the industry is still verymuch driven to a few specific
hardware solutions and theneverything is running on that.
So I would say there's still alot of innovation and diversity
to come.
Francoise von Trapp (04:05):
I think there's a lot of improvement
that's needed in what's outthere.
Do you feel like we kind ofrushed things a little bit?
Generative AI is the big thingright now and people are using
it for making pretty pictures,writing emails, querying Google.
These are all applications thataren't going to hurt anybody if
something goes wrong, right,Other than the fact that you get
(04:27):
not accurate information.
But what do you think the bestapplication is going to be for
AI when we get to the point ofreally using it?
Mark Kuemerle (04:37):
I have a controversial opinion that I
think the consumer applicationand continuing to kind of scale
up generative AI really has alot of potential for the world.
So when we think about, I meanthere's so many places in our
industry where we can use AI.
You know, and people have beenusing reinforcement AI for a
(04:59):
long time and in fact most ofthe place and route algorithms
that we have for putting a chiptogether are all based on
essentially what is early AI,finding local minimums and
global minimums.
So to me it's not a new thing inour industry, even though
everybody is like wow, ai is abrand new thing.
Francoise von Trapp (05:18):
Okay, so there's a difference between in
our industry and in the generalpublic, right, but in the
general public.
Mark Kuemerle (05:26):
I think it's new to the consumer and gosh what an
opportunity it has to affecteverybody's daily life, more so
than kind of using it to buildmore efficient widgets, which is
something that we've been usingsoftware for for a long time.
So I'm kind of excited aboutwhat it could mean to transform
everybody's day-to-day tasksthat they do.
Francoise von Trapp (05:51):
Anybody else?
Kimon Michaels (05:52):
Yeah, and I think, much like the invention
of the laser, people didn'tcontemplate the many ways it
could be used.
I think the use of AI and itsreal impact is still unknown.
It's going to be orders ofmagnitude higher than it is
today, both in our industry,consumers, overall.
Mark Kuemerle (06:07):
It's a fantastic analogy, I think another way of.
Hemanth Jagganathan (06:12):
Also and I agree with what both panelists
said I would say also that ifyou look at AI as a form of
compute, just like previousforms of compute, what was
available to a few is madeavailable to the masses, and
then it takes its own evolutionin terms of adoption and its
(06:33):
usability and how it's used.
So AI is going through a verysimilar cycle.
Like we in our industry,semiconductors has used AI for
many years, you know, and it's,I think, something we rely on
day in, day out, and its use ingeneral public, in more
(06:54):
enterprise solutions, is gettingto be higher and higher.
So the cycle of how new formsof compute get ingested by
society as a whole.
Francoise von Trapp (07:03):
Okay, so what industries do you think
we're likely to see AI beadopted in most?
Hemanth Jagganathan (07:10):
I think this question came up in the
panel yesterday and theunanimous answer was there is no
single industry which can besingled out.
When you are talking about anyform of computation which uses
more efficient use of data toprovide higher levels of insight
and know-how, I think theimpact can be across the board.
(07:34):
So in the world of AI, forpeople who are on the fence,
they either get to use AI or getused by AI.
So you have to choose sides.
Francoise von Trapp (07:47):
But you know, I think about this a lot
and I am a skeptic sort of youknow.
At the same time, it's reallymade Google searching much more.
You get much more information,but I do worry about the
accuracy of that information andI did hear at one point that
chat GPT has gotten dumber sinceit started because it's
(08:09):
scraping Reddit, where you'renot getting accurate information
.
And one of the speakers thismorning talked about the
forecasters predicting that AIwill be smarter than intelligent
beings somewhere between 2026and 2030.
But since AI requires trainingwith accurate data, what are the
chances of this actuallyhappening, at least in the use
(08:34):
that it's being used now by thegeneral public?
Mark Kuemerle (08:36):
I think you bring up a really interesting point,
because models are gettinglarger and larger.
One of our challenges, at leastin the hardware business, is to
make hardware that can fitthese rapidly growing models.
As the models are trained on,more and more data concerned
(09:01):
that there's going to be adilution that's actually caused
by a feedback loop where the AIis being used to generate a lot
of content, which is then beingused to train the AI.
So a lot of folks have a bigconcern and probably a relevant
concern, that that could happen.
So definitely understand tobuild these just massive models.
It'll become more useful.
We are going to have to thinkas a society how we're going to
kind of vet the content that'sout there and how these models
(09:23):
are going to use it.
Right now they just go aftereverything.
Francoise von Trapp (09:26):
Right, right.
Well, I would imagine incertain industrial applications
such as medical andsemiconductor, there are ways to
ensure that we're actuallyusing accurate data.
Kimon Michaels (09:36):
Yeah, I think the one advantage in the
technical fields is it's moredata than language driven Right,
so you can vet that your basedata set is correct, I think
much more easily than an LLMmodel doing general data, for
example.
Still there are challenges ofthe interpretation of the data
(09:57):
and, of course, like manyindustries or technologies when
they're new, I don't think we'vestepped into the legal
repercussions first.
Francoise von Trapp (10:07):
Oh, that's a whole podcast all by itself.
Kimon Michaels (10:10):
For example, you know, abs brakes, completely
accepted everywhere, nowreleased in the US several years
after the rest of the world orEurope, because of concern of
the legal aspect.
If you can brake faster thanthe car behind you, who's at
fault?
And I think that'll be anatural kind of tap the brakes
on the acceleration of AI in ourindustry, of getting to some of
(10:31):
the concerns you have.
Mark Kuemerle (10:33):
One thing that we shouldn't forget, though, is
there's a lot of different usecases for AI.
Right, as I mentioned in ourfield field, in hardware design,
we've been using AI for decades.
Right, I learned it too manyyears ago when I was in college.
We even were leveraging earlyAI.
I remember back in 2000, buyingthe textbook Multiple View
Geometry, which was anintroduction to computer vision
(10:57):
back in 2000, where matrixmultiplication was used to
figure out what's going on inthe outside world, leveraging
cameras.
So there's a lot of placeswhere we don't need to be as
concerned.
Right, reinforcement learningis essentially solving a puzzle
or playing a game that has adefined outcome.
The winner of a system likethat is always very clear.
(11:18):
That's how models were trainedto beat everybody in the world
in a game of chess.
So there are lots of placeswhere maybe we don't need to be
concerned, and having morecapability is going to just help
solve some of these complexproblems.
Hemanth Jagganathan (11:35):
I think, from the impact of AI.
There are many examples in ourindustry, as Mark mentioned, but
also there are prettyphenomenal examples in the
medical field, which ultimately,I think, are getting the core
of is AI making humanity better.
(11:57):
Right humanity better, right,right, and I think there's a lot
of promise where there are manymedical conditions, where
there's sparse data or veryunclear outcomes of whether to
go down a particular line ofprocedures for a patient, and AI
is able to not only ingest allof the information which is out
(12:18):
there, which comes at astaggering volume for no human
to ingest, but then also,looking at the prognosis and the
charts of a particular patient,to give a likelihood of what
would be if one were to go there.
Francoise von Trapp (12:33):
Right right , they can determine and advance
what the success rate would befor, for instance, for a surgery
or for a treatment plan.
Hemanth Jagganathan (12:40):
It could help tailor a more custom plan
and also predict the successrate if one were to go down that
path.
Kimon Michaels (12:48):
It's a great point.
I read a scientific studyrecently that said doctors using
AI have a much higher correctdiagnostic rate than doctors
alone.
Right right, but AI alone waseven higher, which is the
contrapositive of your concern.
Worrying about the accuracy ofAI is one side of it.
Not trusting it on data-drivenareas is the other.
Francoise von Trapp (13:11):
So the doctors didn't trust it and they
were wrong.
Kimon Michaels (13:13):
Apparently, in some cases they probably
overrode it, where, if you didthe AI straight up, it had a
higher percentage of beingcorrect.
Hemanth Jagganathan (13:19):
But it comes back to the ABS question
as well, in terms of litigationand how do you know if you're
getting the right set of data.
So, for the foreseeable future,there will always be a human
override because of some ofthese concerns.
And still it reaches the pointof ABS, where there are a set
number of rules andrepercussions of trusting it.
Mark Kuemerle (13:41):
But I think, Kimon, you've got a really great
example with the medical, andthis is why I think it's
important for us to again kindof think about different use
cases for AI in differentbuckets in our minds and how
much we worry about them.
Would you necessarily beworried about a medical
diagnosis driven by an AI thatwas using all of your vitals and
(14:02):
using a database and using areinforcement learning algorithm
that is very likely to makestatistically correct
conclusions?
In that case, it's just thebenefit to us is going to be so
significantly outweighing thethreat.
What's the?
Kimon Michaels (14:17):
real value of AI to your point.
It's not doing what humans cando more efficiently.
It's taking really highdimensionality of data and
complex sets to get to theanswer humans could not get to
if you put 1,000 men in a roomfor 1,000 days.
Hemanth Jagganathan (14:32):
And I think also, if you look at over the
past few decades, the amount ofdata which an average human
consumes has been growing at avery rapid pace.
So that also plays into thatright.
As you get more and more data,you get overwhelmed with what is
there and then you make gutreaction choices.
(14:53):
So if you're wearing your AppleWatch or if you're having a
health fitness tracker, it is insome sense processing all of
the data and giving you a scoreYou've walked enough for the day
or you've had your Right.
So taking that to the nextlevel as well, I think AI will
really provide that value aswell.
(15:15):
I agree.
Francoise von Trapp (15:17):
So in your talk, kim and you were talking
about generic AI and saying thatit's not going to solve alone.
It's not going to solve theproblems of the semiconductor
industry.
So can you explain what youmeant by that?
Kimon Michaels (15:29):
Yeah, I think if you look at the advances in ML
algorithms and AI in general,it's phenomenal.
I mentioned in the talk we runa course at Carnegie Mellon and
what undergrad grad students cando these days, even with data
in our industry, is incredible.
The challenge, though insemiconductor the data is
physical in nature.
There's physical relationshipsbetween it.
(15:51):
It's not Gaussian distributed.
It varies over time.
It has a strong temporalcomponent Two different sensors
which may have the same name,you expect to have different
values.
So understanding this industryperspective of the data the
temporal nature really drives tohow you use AI.
With the data, you want tomodel its continuously trait,
(16:13):
you want to monitor its accuracy, not only over the parameters
you built for the model, butevery other parameter, because
the next excursion may not besomething encapsulated.
So it brings the challenges ofour industry too.
Francoise von Trapp (16:26):
Okay.
So one of the questions fromthe audience at last night's
panel really stuck with me.
He said that we've all seen thecharts about how much power was
consumed by AI in 2024.
And the question was and we'vetalked about some of this now, I
think but what has AI givenback to society for all the
(16:49):
power it has?
Hemanth Jagganathan (16:50):
Well, I think there are efficiencies
which are coming into thepicture, for example, the
enterprise space.
By infusing AI into a lot oftransactional processing, you're
able to stop fraud before itactually happens.
The usual credit card fraud wascaught after a few transactions
(17:15):
go through and then you stopthat particular card.
For example, by infusing AIinto enterprise applications,
you're able to stop it before ithappens.
So it saves a huge amount ofloss to the industry by infusing
AI there.
We talked about the medicalfield, where one is able to get
a better view of the prognosisand also the proposed treatment
(17:38):
plan.
So there's a lot of benefitthere.
I think, in general, ai is alsoreally helping in bringing the
world closer together.
But there's so much of data outthere, also with natural
language processing, which, ifharnessed correctly, will be
able to help connect differentlanguages together much easier.
(18:02):
It is something which we haveto still be very careful about,
because there's a veryinteresting correlation between
the human brain and, for example, the ability to learn languages
and the propensity forAlzheimer's, for example, if you
are not having the plasticityof your brain maintained.
(18:24):
So sometimes there are thesevery interesting questions by
using AI, are you not makingyour brain as plastic Right, are
we dumbing down society becausethey don't have to use their
own critical thinking skills.
So I would say AI is given a lot, but we have to always ingest
it carefully and keep monitoringit, because there's a
(18:46):
short-term and long-term benefitand a reaction.
So we'll have to just see howthat goes and be responsible as
a society.
Francoise von Trapp (18:53):
So I have an opinion about this.
Do you think that generative AIfor consumer applications could
be considered a frivolous useof AI that will negatively
impact society due to the drainon the grid?
Mark Kuemerle (19:10):
Well, I guess I started out talking about how
consumer applications are one ofmy favorites, I know right, so
I have to jump in and defendmyself to that question.
There's a couple of key points Iwant to make here.
One is that I do really feelthat consumer applications are
going to be really important forkind of empowering people to be
(19:30):
more productive and empoweringpeople to understand data that
might be beyond whatever meansthey have or whatever experience
they have under their belt.
I also think there's a greatopportunity for us, as people
who are developing artificialintelligence or machine learning
hardware, to do everything wecan to improve the efficiency of
(19:59):
these systems, and I spend thevast majority of my time trying
to figure out how to do that.
And I believe that by helpingto make them as power efficient
as possible, we can kind ofreduce that load so that a
greater number of people can useit more freely.
And the other thing I maybewant to share is that, while it
(20:20):
may seem like a power user in mymind, at least, my personal
opinion it's a lot better thanspending that power on
cryptocurrency mining.
I'm a big fan of it from abenefit to society versus energy
consumption.
Kimon Michaels (20:36):
I would agree with you there To that point
there's nothing wrong investingin entertainment and arts.
Hemanth Jagganathan (20:44):
That's across society.
Kimon Michaels (20:45):
In the beginning .
New technologies, especiallywhen they're exciting, they're
already always overused untilpeople figure out really the
correct application, et cetera.
You know, you look at music inthe late 70s, early 80s way too
much synthesizer, Right, it wasnew.
Then People went over the top.
Francoise von Trapp (21:03):
To that point.
Are there applications whereyou think that AI shouldn't be
used?
There is a dark side and nobodytalks about that, but we have
to be really careful, right.
Hemanth Jagganathan (21:15):
That's true for any technology.
But, going to the earlierpoints, we wouldn't be talking
about AI if people didn't getexcited about labeling cats,
dogs, horses 10 years back forexample so I think that there is
something to be said aboutreally having it as an
(21:35):
accessible technology whereeverybody can relate and be part
of that technology.
But name one technology whichdoesn't have a dark side this is
true, maybe that's it.
Kimon Michaels (21:47):
Maybe I'm a romantic.
I throw out matchmaking.
Do you really want it to bepurely algorithmic driven?
Oh, yes, yeah, thank you.
Francoise von Trapp (21:52):
No, let's throw that right out the door,
and that can be a little creepytoo.
I mean, we don't need that.
Yeah, matchmaking, I wouldagree.
Kimon Michaels (22:00):
And my wife probably wouldn't have picked me
if she ran it through analgorithm.
Francoise von Trapp (22:05):
Any others.
That shouldn't be applicationsthat we should not be investing
in on the AI side.
Mark Kuemerle (22:11):
I think a thing that we're just always going to
have to be watchful for as asociety is especially when we
look at LLMs and generative AI.
You can kind of train a modelto do what you want it to do
based on what you give it asinputs.
So we do need to figure out howwe can make sure that we have
accurate input so that we're notbuilding models which are maybe
(22:35):
helping people by telling themincorrect information.
Kimon Michaels (22:40):
It's a fine line between an homage and copyright
infringement when it comes toart and writing, et cetera.
Francoise von Trapp (22:45):
Well, I have run things through ChatGPT
to see if the writing comes outbetter, but you always have to
go through and edit them andmake them personal, because,
again, that's an area wherehaving human interaction can
improve what you get output fromAI versus, maybe, the medical
decisions.
There are times where you useit as a tool to assist, not to
(23:07):
take over.
Okay, well, that's all we havetime for, but I really, really
appreciate your time.
This is a great conversationand I look forward to having you
on again.
Kimon Michaels (23:16):
Enjoyed it.
Thanks for having us.
Francoise von Trapp (23:17):
Thank you very much.
Thank you Next time on the 3DInsights podcast, recorded live
at IMAPS DPC, we talk toofficials and students from
University of Arizona andArizona State University about
the importance of anindustry-academia-government
collaboration in building asolid semiconductor ecosystem in
(23:37):
the US.
There's lots more to come, sotune in next time to the 3D
Insights Podcast.
The 3D Insights Podcast is aproduction of 3D Insights LLC.
This episode of the 3D Insights
podcast is sponsored by IMAPS,the premier global association
for microelectronics advancedpackaging enthusiasts.
A membership in IMAPS helpsyour company grow its advanced
packaging workforce throughprofessional education and
networking, advances your brandand supports building
relationships.
Imaps helps you learn, connectand collaborate.
Learn more at IM imapsorg.
(00:22):
Hi there, I'm Francoise vonTrapp, and this is the 3D
Insights Podcast.
Hi everyone, this week we arerecording live from the IMAPS
(00:48):
Device Packaging Conference inPhoenix, arizona.
Now AI continues to be a hottopic of discussion at industry
events like this one, and thisweek several panels were
dedicated to discussing topicssuch as are we going to have an
AI winter?
Or, as part of the GlobalBusiness Council, how are we
(01:08):
going to scale AI from datacenters to consumer applications
?
So here to talk about the keytakeaways were some of the
panelists.
From each of the panels I haveKamef Jaganathan of IBM Research
, mark Pemberley of Marvell andKim and Michaels of PDF
Solutions.
Welcome to the podcast, guys.
Mark Kuemerle (01:28):
Great to be here.
Thank you very much.
Francoise von Trapp (01:30):
So we've got a lot of questions to go
through, but before we getstarted, can you each tell me a
little bit about your companyand your role there, Kim and?
Kimon Michaels (01:39):
Sure, Kimon Michaels.
I'm one of the co-founders andEVP of Products and Solutions at
PDF Solutions.
Pdf Solutions is the big dataanalytics platform for the
semiconductor and electronicsupply chain.
Hemanth Jagganathan (01:52):
Yes, I'm Hemanth Jagannathan.
I'm from IBM Research.
Our overall charter at IBMResearch is to be part of the
future of compute.
We work on all three elementsof compute, which is bits,
neurons and qubits.
So I'm currently in charge ofthe Chippler Advanced Packaging
(02:12):
Research Group at IBM.
Mark Kuemerle (02:14):
Okay, and Mark.
Hi, I'm Mark Kumerle.
I'm VP of Technology forMarvell.
I'm responsible forarchitecture for all of our
custom products and alsodefining a roadmap to make those
products possible.
Francoise von Trapp (02:27):
Excellent.
So let's dive in on the topicof AI, which is driving this
industry.
I want to start with asking youwhere would you say AI is on
the Gartner hype cycle at thispoint?
Kimon Michaels (02:42):
In our industry.
I think it's still on theuptick in the beginning.
Manufacturing semiconductors,semiconductors in general, is
not the fastest moving industryand I think we're still getting
traction.
We're not in a trough yet,which invariably will come.
Francoise von Trapp (02:56):
So for using AI in our industry, we're
still in the up cycle.
Kimon Michaels (03:00):
I think we're still in the up cycle, but how
about as a driver of ourindustry?
You know people are starting toget concerned about the draft.
If you look at the panel lastnight, worry about is it the AI
winter?
I was at an investor conferencewhere people were very
interested in the NVIDIApresentation more about the
future, not about themcontinuing to have excellent
(03:23):
results.
So there's certainly a concern.
It remains to be seen.
Hemanth Jagganathan (03:29):
I think we're still very early in the
era of AI, because, whiletraditional AI and the solutions
out there are driving theindustry today, it's still in
its infancy.
We need to talk about anecosystem of design,
architecture, having a varietyof options for an AI need, for
(03:51):
example.
So the industry is still verymuch driven to a few specific
hardware solutions and theneverything is running on that.
So I would say there's still alot of innovation and diversity
to come.
Francoise von Trapp (04:05):
I think there's a lot of improvement
that's needed in what's outthere.
Do you feel like we kind ofrushed things a little bit?
Generative AI is the big thingright now and people are using
it for making pretty pictures,writing emails, querying Google.
These are all applications thataren't going to hurt anybody if
something goes wrong, right,Other than the fact that you get
(04:27):
not accurate information.
But what do you think the bestapplication is going to be for
AI when we get to the point ofreally using it?
Mark Kuemerle (04:37):
I have a controversial opinion that I
think the consumer applicationand continuing to kind of scale
up generative AI really has alot of potential for the world.
So when we think about, I meanthere's so many places in our
industry where we can use AI.
You know, and people have beenusing reinforcement AI for a
(04:59):
long time and in fact most ofthe place and route algorithms
that we have for putting a chiptogether are all based on
essentially what is early AI,finding local minimums and
global minimums.
So to me it's not a new thing inour industry, even though
everybody is like wow, ai is abrand new thing.
Francoise von Trapp (05:18):
Okay, so there's a difference between in
our industry and in the generalpublic, right, but in the
general public.
Mark Kuemerle (05:26):
I think it's new to the consumer and gosh what an
opportunity it has to affecteverybody's daily life, more so
than kind of using it to buildmore efficient widgets, which is
something that we've been usingsoftware for for a long time.
So I'm kind of excited aboutwhat it could mean to transform
everybody's day-to-day tasksthat they do.
Francoise von Trapp (05:51):
Anybody else?
Kimon Michaels (05:52):
Yeah, and I think, much like the invention
of the laser, people didn'tcontemplate the many ways it
could be used.
I think the use of AI and itsreal impact is still unknown.
It's going to be orders ofmagnitude higher than it is
today, both in our industry,consumers, overall.
Mark Kuemerle (06:07):
It's a fantastic analogy, I think another way of.
Hemanth Jagganathan (06:12):
Also and I agree with what both panelists
said I would say also that ifyou look at AI as a form of
compute, just like previousforms of compute, what was
available to a few is madeavailable to the masses, and
then it takes its own evolutionin terms of adoption and its
(06:33):
usability and how it's used.
So AI is going through a verysimilar cycle.
Like we in our industry,semiconductors has used AI for
many years, you know, and it's,I think, something we rely on
day in, day out, and its use ingeneral public, in more
(06:54):
enterprise solutions, is gettingto be higher and higher.
So the cycle of how new formsof compute get ingested by
society as a whole.
Francoise von Trapp (07:03):
Okay, so what industries do you think
we're likely to see AI beadopted in most?
Hemanth Jagganathan (07:10):
I think this question came up in the
panel yesterday and theunanimous answer was there is no
single industry which can besingled out.
When you are talking about anyform of computation which uses
more efficient use of data toprovide higher levels of insight
and know-how, I think theimpact can be across the board.
(07:34):
So in the world of AI, forpeople who are on the fence,
they either get to use AI or getused by AI.
So you have to choose sides.
Francoise von Trapp (07:47):
But you know, I think about this a lot
and I am a skeptic sort of youknow.
At the same time, it's reallymade Google searching much more.
You get much more information,but I do worry about the
accuracy of that information andI did hear at one point that
chat GPT has gotten dumber sinceit started because it's
(08:09):
scraping Reddit, where you'renot getting accurate information
.
And one of the speakers thismorning talked about the
forecasters predicting that AIwill be smarter than intelligent
beings somewhere between 2026and 2030.
But since AI requires trainingwith accurate data, what are the
chances of this actuallyhappening, at least in the use
(08:34):
that it's being used now by thegeneral public?
Mark Kuemerle (08:36):
I think you bring up a really interesting point,
because models are gettinglarger and larger.
One of our challenges, at leastin the hardware business, is to
make hardware that can fitthese rapidly growing models.
As the models are trained on,more and more data concerned
(09:01):
that there's going to be adilution that's actually caused
by a feedback loop where the AIis being used to generate a lot
of content, which is then beingused to train the AI.
So a lot of folks have a bigconcern and probably a relevant
concern, that that could happen.
So definitely understand tobuild these just massive models.
It'll become more useful.
We are going to have to thinkas a society how we're going to
kind of vet the content that'sout there and how these models
(09:23):
are going to use it.
Right now they just go aftereverything.
Francoise von Trapp (09:26):
Right, right.
Well, I would imagine incertain industrial applications
such as medical andsemiconductor, there are ways to
ensure that we're actuallyusing accurate data.
Kimon Michaels (09:36):
Yeah, I think the one advantage in the
technical fields is it's moredata than language driven Right,
so you can vet that your basedata set is correct, I think
much more easily than an LLMmodel doing general data, for
example.
Still there are challenges ofthe interpretation of the data
(09:57):
and, of course, like manyindustries or technologies when
they're new, I don't think we'vestepped into the legal
repercussions first.
Francoise von Trapp (10:07):
Oh, that's a whole podcast all by itself.
Kimon Michaels (10:10):
For example, you know, abs brakes, completely
accepted everywhere, nowreleased in the US several years
after the rest of the world orEurope, because of concern of
the legal aspect.
If you can brake faster thanthe car behind you, who's at
fault?
And I think that'll be anatural kind of tap the brakes
on the acceleration of AI in ourindustry, of getting to some of
(10:31):
the concerns you have.
Mark Kuemerle (10:33):
One thing that we shouldn't forget, though, is
there's a lot of different usecases for AI.
Right, as I mentioned in ourfield field, in hardware design,
we've been using AI for decades.
Right, I learned it too manyyears ago when I was in college.
We even were leveraging earlyAI.
I remember back in 2000, buyingthe textbook Multiple View
Geometry, which was anintroduction to computer vision
(10:57):
back in 2000, where matrixmultiplication was used to
figure out what's going on inthe outside world, leveraging
cameras.
So there's a lot of placeswhere we don't need to be as
concerned.
Right, reinforcement learningis essentially solving a puzzle
or playing a game that has adefined outcome.
The winner of a system likethat is always very clear.
(11:18):
That's how models were trainedto beat everybody in the world
in a game of chess.
So there are lots of placeswhere maybe we don't need to be
concerned, and having morecapability is going to just help
solve some of these complexproblems.
Hemanth Jagganathan (11:35):
I think, from the impact of AI.
There are many examples in ourindustry, as Mark mentioned, but
also there are prettyphenomenal examples in the
medical field, which ultimately,I think, are getting the core
of is AI making humanity better.
(11:57):
Right humanity better, right,right, and I think there's a lot
of promise where there are manymedical conditions, where
there's sparse data or veryunclear outcomes of whether to
go down a particular line ofprocedures for a patient, and AI
is able to not only ingest allof the information which is out
(12:18):
there, which comes at astaggering volume for no human
to ingest, but then also,looking at the prognosis and the
charts of a particular patient,to give a likelihood of what
would be if one were to go there.
Francoise von Trapp (12:33):
Right right , they can determine and advance
what the success rate would befor, for instance, for a surgery
or for a treatment plan.
Hemanth Jagganathan (12:40):
It could help tailor a more custom plan
and also predict the successrate if one were to go down that
path.
Kimon Michaels (12:48):
It's a great point.
I read a scientific studyrecently that said doctors using
AI have a much higher correctdiagnostic rate than doctors
alone.
Right right, but AI alone waseven higher, which is the
contrapositive of your concern.
Worrying about the accuracy ofAI is one side of it.
Not trusting it on data-drivenareas is the other.
Francoise von Trapp (13:11):
So the doctors didn't trust it and they
were wrong.
Kimon Michaels (13:13):
Apparently, in some cases they probably
overrode it, where, if you didthe AI straight up, it had a
higher percentage of beingcorrect.
Hemanth Jagganathan (13:19):
But it comes back to the ABS question
as well, in terms of litigationand how do you know if you're
getting the right set of data.
So, for the foreseeable future,there will always be a human
override because of some ofthese concerns.
And still it reaches the pointof ABS, where there are a set
number of rules andrepercussions of trusting it.
Mark Kuemerle (13:41):
But I think, Kimon, you've got a really great
example with the medical, andthis is why I think it's
important for us to again kindof think about different use
cases for AI in differentbuckets in our minds and how
much we worry about them.
Would you necessarily beworried about a medical
diagnosis driven by an AI thatwas using all of your vitals and
(14:02):
using a database and using areinforcement learning algorithm
that is very likely to makestatistically correct
conclusions?
In that case, it's just thebenefit to us is going to be so
significantly outweighing thethreat.
What's the?
Kimon Michaels (14:17):
real value of AI to your point.
It's not doing what humans cando more efficiently.
It's taking really highdimensionality of data and
complex sets to get to theanswer humans could not get to
if you put 1,000 men in a roomfor 1,000 days.
Hemanth Jagganathan (14:32):
And I think also, if you look at over the
past few decades, the amount ofdata which an average human
consumes has been growing at avery rapid pace.
So that also plays into thatright.
As you get more and more data,you get overwhelmed with what is
there and then you make gutreaction choices.
(14:53):
So if you're wearing your AppleWatch or if you're having a
health fitness tracker, it is insome sense processing all of
the data and giving you a scoreYou've walked enough for the day
or you've had your Right.
So taking that to the nextlevel as well, I think AI will
really provide that value aswell.
(15:15):
I agree.
Francoise von Trapp (15:17):
So in your talk, kim and you were talking
about generic AI and saying thatit's not going to solve alone.
It's not going to solve theproblems of the semiconductor
industry.
So can you explain what youmeant by that?
Kimon Michaels (15:29):
Yeah, I think if you look at the advances in ML
algorithms and AI in general,it's phenomenal.
I mentioned in the talk we runa course at Carnegie Mellon and
what undergrad grad students cando these days, even with data
in our industry, is incredible.
The challenge, though insemiconductor the data is
physical in nature.
There's physical relationshipsbetween it.
(15:51):
It's not Gaussian distributed.
It varies over time.
It has a strong temporalcomponent Two different sensors
which may have the same name,you expect to have different
values.
So understanding this industryperspective of the data the
temporal nature really drives tohow you use AI.
With the data, you want tomodel its continuously trait,
(16:13):
you want to monitor its accuracy, not only over the parameters
you built for the model, butevery other parameter, because
the next excursion may not besomething encapsulated.
So it brings the challenges ofour industry too.
Francoise von Trapp (16:26):
Okay.
So one of the questions fromthe audience at last night's
panel really stuck with me.
He said that we've all seen thecharts about how much power was
consumed by AI in 2024.
And the question was and we'vetalked about some of this now, I
think but what has AI givenback to society for all the
(16:49):
power it has?
Hemanth Jagganathan (16:50):
Well, I think there are efficiencies
which are coming into thepicture, for example, the
enterprise space.
By infusing AI into a lot oftransactional processing, you're
able to stop fraud before itactually happens.
The usual credit card fraud wascaught after a few transactions
(17:15):
go through and then you stopthat particular card.
For example, by infusing AIinto enterprise applications,
you're able to stop it before ithappens.
So it saves a huge amount ofloss to the industry by infusing
AI there.
We talked about the medicalfield, where one is able to get
a better view of the prognosisand also the proposed treatment
(17:38):
plan.
So there's a lot of benefitthere.
I think, in general, ai is alsoreally helping in bringing the
world closer together.
But there's so much of data outthere, also with natural
language processing, which, ifharnessed correctly, will be
able to help connect differentlanguages together much easier.
(18:02):
It is something which we haveto still be very careful about,
because there's a veryinteresting correlation between
the human brain and, for example, the ability to learn languages
and the propensity forAlzheimer's, for example, if you
are not having the plasticityof your brain maintained.
(18:24):
So sometimes there are thesevery interesting questions by
using AI, are you not makingyour brain as plastic Right, are
we dumbing down society becausethey don't have to use their
own critical thinking skills.
So I would say AI is given a lot, but we have to always ingest
it carefully and keep monitoringit, because there's a
(18:46):
short-term and long-term benefitand a reaction.
So we'll have to just see howthat goes and be responsible as
a society.
Francoise von Trapp (18:53):
So I have an opinion about this.
Do you think that generative AIfor consumer applications could
be considered a frivolous useof AI that will negatively
impact society due to the drainon the grid?
Mark Kuemerle (19:10):
Well, I guess I started out talking about how
consumer applications are one ofmy favorites, I know right, so
I have to jump in and defendmyself to that question.
There's a couple of key points Iwant to make here.
One is that I do really feelthat consumer applications are
going to be really important forkind of empowering people to be
(19:30):
more productive and empoweringpeople to understand data that
might be beyond whatever meansthey have or whatever experience
they have under their belt.
I also think there's a greatopportunity for us, as people
who are developing artificialintelligence or machine learning
hardware, to do everything wecan to improve the efficiency of
(19:59):
these systems, and I spend thevast majority of my time trying
to figure out how to do that.
And I believe that by helpingto make them as power efficient
as possible, we can kind ofreduce that load so that a
greater number of people can useit more freely.
And the other thing I maybewant to share is that, while it
(20:20):
may seem like a power user in mymind, at least, my personal
opinion it's a lot better thanspending that power on
cryptocurrency mining.
I'm a big fan of it from abenefit to society versus energy
consumption.
Kimon Michaels (20:36):
I would agree with you there To that point
there's nothing wrong investingin entertainment and arts.
Hemanth Jagganathan (20:44):
That's across society.
Kimon Michaels (20:45):
In the beginning .
New technologies, especiallywhen they're exciting, they're
already always overused untilpeople figure out really the
correct application, et cetera.
You know, you look at music inthe late 70s, early 80s way too
much synthesizer, Right, it wasnew.
Then People went over the top.
Francoise von Trapp (21:03):
To that point.
Are there applications whereyou think that AI shouldn't be
used?
There is a dark side and nobodytalks about that, but we have
to be really careful, right.
Hemanth Jagganathan (21:15):
That's true for any technology.
But, going to the earlierpoints, we wouldn't be talking
about AI if people didn't getexcited about labeling cats,
dogs, horses 10 years back forexample so I think that there is
something to be said aboutreally having it as an
(21:35):
accessible technology whereeverybody can relate and be part
of that technology.
But name one technology whichdoesn't have a dark side this is
true, maybe that's it.
Kimon Michaels (21:47):
Maybe I'm a romantic.
I throw out matchmaking.
Do you really want it to bepurely algorithmic driven?
Oh, yes, yeah, thank you.
Francoise von Trapp (21:52):
No, let's throw that right out the door,
and that can be a little creepytoo.
I mean, we don't need that.
Yeah, matchmaking, I wouldagree.
Kimon Michaels (22:00):
And my wife probably wouldn't have picked me
if she ran it through analgorithm.
Francoise von Trapp (22:05):
Any others.
That shouldn't be applicationsthat we should not be investing
in on the AI side.
Mark Kuemerle (22:11):
I think a thing that we're just always going to
have to be watchful for as asociety is especially when we
look at LLMs and generative AI.
You can kind of train a modelto do what you want it to do
based on what you give it asinputs.
So we do need to figure out howwe can make sure that we have
accurate input so that we're notbuilding models which are maybe
(22:35):
helping people by telling themincorrect information.
Kimon Michaels (22:40):
It's a fine line between an homage and copyright
infringement when it comes toart and writing, et cetera.
Francoise von Trapp (22:45):
Well, I have run things through ChatGPT
to see if the writing comes outbetter, but you always have to
go through and edit them andmake them personal, because,
again, that's an area wherehaving human interaction can
improve what you get output fromAI versus, maybe, the medical
decisions.
There are times where you useit as a tool to assist, not to
(23:07):
take over.
Okay, well, that's all we havetime for, but I really, really
appreciate your time.
This is a great conversationand I look forward to having you
on again.
Kimon Michaels (23:16):
Enjoyed it.
Thanks for having us.
Francoise von Trapp (23:17):
Thank you very much.
Thank you Next time on the 3DInsights podcast, recorded live
at IMAPS DPC, we talk toofficials and students from
University of Arizona andArizona State University about
the importance of anindustry-academia-government
collaboration in building asolid semiconductor ecosystem in
(23:37):
the US.
There's lots more to come, sotune in next time to the 3D
Insights Podcast.
The 3D Insights Podcast is aproduction of 3D Insights LLC.
Popular Podcasts
Stuff You Should Know
If you've ever wanted to know about champagne, satanism, the Stonewall Uprising, chaos theory, LSD, El Nino, true crime and Rosa Parks, then look no further. Josh and Chuck have you covered.
Dateline NBC
Current and classic episodes, featuring compelling true-crime mysteries, powerful documentaries and in-depth investigations. Follow now to get the latest episodes of Dateline NBC completely free, or subscribe to Dateline Premium for ad-free listening and exclusive bonus content: DatelinePremium.com
On Purpose with Jay Shetty
I’m Jay Shetty host of On Purpose the worlds #1 Mental Health podcast and I’m so grateful you found us. I started this podcast 5 years ago to invite you into conversations and workshops that are designed to help make you happier, healthier and more healed. I believe that when you (yes you) feel seen, heard and understood you’re able to deal with relationship struggles, work challenges and life’s ups and downs with more ease and grace. I interview experts, celebrities, thought leaders and athletes so that we can grow our mindset, build better habits and uncover a side of them we’ve never seen before. New episodes every Monday and Friday. Your support means the world to me and I don’t take it for granted — click the follow button and leave a review to help us spread the love with On Purpose. I can’t wait for you to listen to your first or 500th episode!