March 31, 2026 • 36 mins
In Episode 137, Scott Genin, Vice President of Materials Discovery at OTI Lumionics, unveils how GPU-accelerated quantum chemistry is revolutionizing material science. The discussion highlights the limitations of current quantum hardware and the role of AI in overcoming these challenges. Scott shares insights into how classical simulations can mimic quantum computers, pushing the boundaries of what's possible. He emphasizes the significance of these advancements for real-world applications, from OLEDs to new catalysts. This episode is essential for anyone interested in the future of quantum computing and material discovery. See more about the announcement here: https://arxiv.org/abs/2603.08883
Listen
Watch
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
SPEAKER_03 (00:40):
Hey Cyprien, how are you doing?
SPEAKER_04 (00:42):
Hey Patrick, I'm doing great.
Looking forward for anotherepisode of Entangle Things.
SPEAKER_03 (00:46):
Yes, I think uh we're we're we're rejoined by
Scott.
Scott's been on the show before.
Scott, do you mind reintroducingyourself to the audience?
SPEAKER_01 (00:53):
Uh yeah, thank you for having me on, Patrick and
Cyprian.
Uh I'm Scott Jenn and I'm thevice president of materials
discovery at OTI Lumionics.
SPEAKER_03 (01:02):
And you guys are, uh as I remember from the last time
we talked to you, which was veryinteresting, you're using
Quantum Inspired for your umoptical development and your
product development.
Um and there's some big thingsbrewing, from what I understand.
SPEAKER_01 (01:18):
Uh yeah.
We, you know, last year I was onto talk about the um, you know,
the theoretical foundation thatnow we've been able to implement
at scale across GPUs.
And the, you know, the speed upand the results are impressive
because it really opens up a umreally not just the scientific
(01:39):
application or you know,scientific discovery of what
quantum algorithms could do inmaterial science, but it also
demonstrates that they can bedone in kind of a reasonable
amount of time on uh classicalhardware.
SPEAKER_03 (01:52):
Cool.
So so you're using AI hardwareto like the Blackwell chips,
those kinds of rigs to reallysqueeze that quantum advantage,
but you're still using classicalcompute power.
Um when I I I I've told thestory before.
When I first started talking toCyprian about this kind of uh
technology, quantum, beforethere were even quantum uh
(02:12):
computers of any kind, there wasa company called OneQit up in
Vancouver that um was doingquantum inspired.
And so it seems like you are agood example of a company that's
taken that and run with it.
SPEAKER_01 (02:24):
Um I I think, you know, uh yeah, in in some
respects, I would say we'vetaken it to probably the
extreme.
You know, One Qubit does have abroader focus.
Like they they do they are inthe consulting goals, basically.
Yeah, they're they're consultingfirm.
Right.
And so um OTI has a singular,you know, a singular objective,
(02:47):
which is to do you know, quantumchemistry calculations as
efficiently as possible.
So really what this does thoughis it sets apart um the
algorithm, you know, thisiterative cubic couple cluster
into a different um kind of adifferent class or domain than
just some sort of likeinteresting, you know, novelty
(03:08):
to be run on a you know futurequantum computer.
Right.
Um it really now sets it intothe, you know, uh as a high
accuracy general quantumchemistry algorithm that has
much more efficient scaling andmuch better utilization of GPU
resources than other highaccuracy classical quantum
(03:31):
chemistry methods.
Now, what this what what makesthis very exciting, there's
there's kind of two aspects thatmake it exciting.
There's the scientific aspect,right?
And and then there's also theapplied aspect.
So from the you know, thescientific and theoretical
aspect, this is the firstrealization of a quantum
(03:52):
algorithm, meaning that it'sdoing quantum chemistry or
fermionic excitation operatorsand electronic structure
problems in the qubit domain ofmath, opposed to you know the
traditional fermionic domain.
This you know, it's it'simportant to also you know
(04:14):
emphasize that you know allquantum chemistry methods have
their own pros and cons, quirks,kind of oddities about them.
Yeah, and you know, that wassomething which couldn't have
which could never have beenreally explored in depth prior
(04:34):
to this.
Because fundamentally, mostresults were either like the the
system is just not converged.
So if we look at like thequantum subspace
diagonalization, you know, weyou look at the results and it's
like these results are worsethan uh CISD.
And yet it should be a you know,a true ground state FCI type
(04:56):
solution.
If it's unconverged, we can'treally do downstream uh like
electron density calculations oruh you know, we can't map it
onto there are other things thatchemists care about, like things
like natural bonding orbitals.
Um these are things that like uhyou know, actually applied
synthetic chemists care aboutmore so because it gives them an
(05:18):
intuition about the uhunderlying chemistry that is
happening.
And those things are just notpossible if you can't converge
your uh quantum chemistrycalculation to a like to a
correct or a um like to aconverged state.
And so, you know, at the scale,quantum computers have never
been able to achieve this atscale, and there has not been
(05:40):
really a quantum native methodthat has been able to do this at
scale until now.
Wow.
And so now we can actually notonly say, well, is a quantum
computer, because everyeverything, like if you look at
pretty much all the studies,they're all saying, Oh, if I had
a quantum computer that had ahundred logical error fully
(06:01):
error corrected qubits, I wouldbe able to solve this problem.
Right.
But I think it's actually fairto push back and say, well, what
will this converge to?
Because you say that it willhave this one error, but is it
actually going to converge tothe FCI solution?
And at that scale, we don'treally, we can't really
determine that.
So what this paper hasdemonstrated is that, you know,
(06:25):
historically people have usedthings like uh density matrix
renormalization group, which isa very robust, well-known uh
classical algorithm that um Ithink maybe people got carried
away in saying that it's the beall end all.
But here we objectively justshow that IQCC outperforms it.
Like it basically says I have avariational value that is lower
(06:48):
than your best attempt at DMRT.
And I can do it on fairlyavailable GPU hardware.
SPEAKER_03 (06:57):
Exactly.
Yeah, you you haven't waitedfor, oh, someday we're gonna be
able to do this.
You're doing it, you're gettingthe best of both worlds, really,
because you're you're notwaiting for, oh, well, if I had
a quantum computer, this wouldreally go.
You're you're leveraging and andcomputes exploded with AI.
Uh we we've got some Blackwellchips that we use for other
(07:19):
things, and it's justmind-blowing how much and the
backplanes and the way to youknow put them all together.
Um, and so you're takingadvantage of that.
Now, I imagine if suddenly therewere huge breakthroughs in
quantum computers and there werehundreds or thousands of logical
qubits, then you even benefitmore.
So, so uh so I think this is agreat um way to venture forth
(07:43):
and get the best of uh no matterhow it comes out.
You're basically betting bothsides of the table.
SPEAKER_01 (07:49):
Yeah.
And it's because now we can likewe can verify that a quantum
algorithm, again, working inqubit notation and domain, will
outperform some of the bestknown classical methods, right?
And that's that's fundamentallya significant milestone to like
objectively surpass them.
SPEAKER_04 (08:08):
That's um so if I'm reading correct, these these
results, the kind of if we weretrying to do some some kind of
of mapping to the world of ofquantum computers, what you
achieved is basically theequivalent of 200 qubits.
(08:30):
Like what what you would get ona quantum computer with 200
qubits, is my understandingcorrect?
SPEAKER_01 (08:36):
Yes, that is correct.
With with 200 fault tolerant,fully connected qubits, right?
That I think that's the caveatis that if you in you know in in
how we do it, you know,entangling qubit one and
entangling qubit 200 is youknow, that's instantly doable,
(08:56):
right?
From our from our uh softwarestandpoint.
SPEAKER_02 (08:59):
Yeah.
SPEAKER_01 (09:00):
But on a but on a actual physical chip, that might
be actually quite hard.
SPEAKER_04 (09:05):
Because that's that's very interesting to to to
me, right?
And we've we've touched thisseveral times on on this show
about the evolution and thebenefits of uh simulation, so to
speak, right?
But uh and and this is where Iwould like to to get a bit more
detail from you.
(09:26):
This is not really um simulatingin general a quantum computer,
right?
This is actually implementing analgorithm that runs on GPUs that
yields a result that isequivalent to what you would get
with a 200 qubit uh quantumcomputer.
Is is is is this correct?
SPEAKER_01 (09:48):
No, that that is exactly correct, right?
So the IQCC implementation on awhat makes it really scale very
well on GPU.
Um, it actually has much betterscaling on GPU than it does on
CPU.
Um, is that if you, you know,traditionally in what we
consider like if we think aboutclassical emulation of a quantum
(10:08):
computer, or trying to actuallysimulate the physical quantum
computer, you know, we haveapproaches like state vector,
where we have to store basicallythe wave function as this is a
complex state vector.
And that's where a lot of thisthe bad scaling happens.
Now there's techniques, youknow, Amazon had developed
techniques to kind of prune thisum massive state vector to so
(10:29):
that it was simulatable onclassical computers up to like
easily up to 80 qubits and alittle bit beyond that.
But fundamentally, there is alimit, right?
Especially as you want to getmore and more accurate.
What makes our implementationvery different is that it
doesn't consider really thestate vector.
It works in operator space.
And this allows us to reallyutilize integer and bitwise
(10:54):
manipulation to mimic theoutput, meaning the observable
of the quantum computer.
Right now, from yeah, I thinkfrom a you know theoretician
standpoint of a physicist, theymay think that that's you know a
little bit gimmicky.
But I would point out that youknow, my background's in
chemistry and chemicalengineering.
(11:14):
So the observable is pretty muchwhat I care about.
And this is what the quant, thisis what I'm going to see.
Right.
Right.
So it's like you're kind of ableto bypass and not I wouldn't say
necessarily bypass, but you'reable to, you know, through kind
of uh representing um you knowthe observable basically in
(11:36):
binary bit strings, you canactually kind of exploit the GPU
very efficiently to do thesecalculations and basically
reproduce these observables.
And that's what we areeffectively doing.
Yeah, that's that's one of thereasons why you know you get a
90x speed up on a blackwell overCPU.
Right.
Right?
(11:56):
Like fundamentally, if we if Iwant to be as as uh transparent
and direct as possible, IQcc hasslightly better scaling than
DMRG on CPU, but I think peoplewould really sit there, you
know, scratch their chin and belike, well, you know, 109 hours
versus 130 hours is currentlyone of the benchmarks that I
(12:18):
have for IQCC being lower thanDMRG by about one milliheart
rate.
People would sit there and belike, mm-mm mm-mm, you know,
maybe not.
But then I say, well, what aboutone hour on a GPU?
And that's fundamentally it's adifferent, you're you're in a
different question.
Two orders of magnitude, yeah.
(12:39):
Yeah.
It's it's uh it's a completelythe the implications of that is
it completely opens up what youcould do with this.
Because one of the things thatyou could do with it is you
could do geometry optimization.
And this is one of the you knowquestions that from the
chemistry, applied chemistryside, people always ask is like,
(13:00):
well, can I do geometryoptimization?
And before it'd be like, well,if each iteration takes a
hundred hours, you're gonna bewaiting a really long time
because you have to do themsequentially stepwise, right?
But if each step only takes nowone hour, that's a very
different question.
Yeah.
(13:21):
Then because the geometryoptimization of like uh you know
some OLED materials that wehave, uh, you know, they may
take 50 to 150 steps.
So you're asking really thequestion could you wait 60 hours
to have a variational geometryoptimized solution for 60 weeks?
(13:42):
Versus 60 weeks.
Exactly.
And it's it it completely opensthat up.
Oh, it's amazing.
SPEAKER_04 (13:49):
Yeah, and and just to make sure that our listeners
are are are following, right?
We were um talking here aboutIQCC and DMRG, right?
IQCC stands for iterative qubitcoupled uh uh cluster and uh DM
DMRG, right, stands for densitymatrix renormalization group.
Um that is just kind of themethods that that we are talking
(14:13):
about.
But what strikes me um asprobably as the most important
thing of this, and I know Idon't want to kind of downplay
the the practical sides of it,but for me, what's really
interesting is this might justpush out what we would consider
the boundary of quantumadvantage for for these things,
(14:37):
right?
Because I know for a fact thatin chemistry in general, right,
the the kind of consensus wasit's around 40, 50, 60 qubits,
whatever that is, right?
Now you're coming with thisresult, and you're saying, look,
sure, we're using quantuminspired and everything.
We're still running on classicalcompute, right?
(14:59):
But we can achieve a result thatis equivalent to what you would
achieve with 200 qubits.
So all of a sudden, it feels tome like it's pushing out the
boundary of what would quantumadvantage be?
Um, right?
SPEAKER_01 (15:14):
Correct.
It it also calls into questionabout the notion of like quantum
supremacy, in particular forvariational quantum eigen solver
type uh calculations as well,right?
Is there even, you know, canquantum supremacy ever be
achieved in that class of calcof uh algorithms?
And um, you know, that boundneeds to be dramatically pushed
(15:38):
out.
But even like quantum phaseestimation for quantum
chemistry, now we really haveto, you know, dramatically
consider um what is thatboundary going to be?
Because, you know, using uh youknow the uh resource estimators
that Microsoft had published,you know, using their their
(16:00):
inputs on what they think iswhat their Mariana Mayorana
quantum computer when it'sconstructed, I'm sure one day it
will be.
Um, but when it is, they saythat if everything goes well,
they will need about 200 hoursto get it within this one milli
accuracy.
And it's like, okay, well, howlong is that how much how much,
(16:23):
you know, and then it even comesfrom a practical standpoint of
like, well, how much is it gonnacost to run?
Yeah, right.
And what's one hour of a blackbelt GPU's time versus 200 hours
of quantum computers time?
SPEAKER_03 (16:35):
Time, energy, there's you know, those those
are the big trade-offs rightnow, and and AI is really
pushing.
I think we have to admit thatthree years ago this wasn't
possible with the technology.
You know, classical has hasrocketed forward and and the
goalposts are moving.
I agree.
I I don't think that's a badthing.
(16:55):
I it's probably a bad thing ifyou're trying to uh to prove
quantum supremacy, but that'sthat shouldn't really be that's
more of a marketing thing thananything else.
SPEAKER_01 (17:05):
Yeah.
I mean this this dramaticallychanges even the quantum
advantage or quantum businessutility case, right?
Especially in materials.
SPEAKER_03 (17:16):
That said, I don't I don't think there's gonna be one
of the things we've talked aboutin the past is a quantum winter,
the way AI has had severalwinters where funding's dried
up.
I I think the the Schor'salgorithm specter is is gonna
keep the money flowing to youknow not only the security side
of things, but but also the thematerial science side of things,
(17:38):
which is fine.
But I think this is giving thesetechnologies push each other.
Uh it's an oversimplification.
But when there was more than onedominant browser in the market,
we always got better browsersand better features.
And I think the fact that thequantum is challenging classical
and vice versa is a good thingfor all of us.
SPEAKER_01 (17:58):
Yeah, and it's I think it's also important, like
this, these advances areimportant for also helping
motivate why a quantum computerwould be necessary, perhaps even
in the future for quantumchemistry, because now we can
explore how is a quantumcomputer going to compute
natural bonding orbitals, how isit going to compute chemical
(18:21):
properties like electron densityfrom a native quantum
perspective, opposed to beinglike, oh, well, I have the DMRG
output, right?
You know, DMRG is fundamentallyDMRG.
Same with CCSD, these are there,there's these trade-offs on the
back on the theoretical side,which have downstream
consequences.
And chemists are very familiar,especially applied chemists, are
(18:43):
very familiar with theseconsequences.
And if we can understand thosequantum consequences of doing
them on a quantum computer,which have previously never been
able to be explored, um, it nowserves as like this is the
benchmark and standard that youknow quantum computing companies
need to kind of reach and exceedin order to have a, you know, in
(19:06):
order to really motivate it.
But you know, we might be ableto discover new interesting
physics or uh chemistry usingthis quantum-inspired uh
technique in the in the shortrun, which then would even
motivate that.
Oh, we really need a quantumcomputer that can run quantum
phase estimation perfectly.
Because it could be that, youknow, oh yeah, you know, if I
(19:28):
had another hundred qubits, youknow, 300 qubits, because we,
you know, we haven't been ableto simulate that at this point
at that scale efficiently, umthat could really then
dramatically open up a whole newrange of uh you know new
materials, innovation, anddiscovery.
SPEAKER_03 (19:46):
Uh Nvidia's Vera Rubin chip might well be able to
do that.
You know, some some of the justjust the fact that AI chips are
moving so quickly uh and so andand then custom, yeah.
I could even see if this if thisreally takes off, which it
should, um maybe we'd even getchips designed for this.
I assume that while the GPUs arereally good at it, they there
(20:09):
are optimizations at the chiplevel that could probably help
it as well.
SPEAKER_01 (20:13):
Yeah, definitely with the uh memory transfer and
sharing information betweenthem, right?
Yeah.
SPEAKER_04 (20:18):
Yeah, and I think at the end of the day, right, this
this boils down to the fact thatlike fundamentally speaking,
right, um uh DMRG is based on amodel of classical tensor
networks, right?
As uh essentially opposed to uhuh IQCC, which is qubit native,
(20:41):
right?
So you are literally describingthe problem, right, in the kind
of native language of quantum.
And I think this because if if Iam correct, um even up to like
relatively recently, DMRG wasconsidered like the more mature
kind of approach, right?
(21:03):
But what you're coming andyou're saying with your result
is that, well, it it looks likeright, IQCC can kind of surpass
clearly DMRG and get us to um uhto a point where like simulating
quantum native things is isbecoming like uh uh more
(21:25):
powerful with this with thisapproach.
SPEAKER_01 (21:28):
Yeah, and there's uh you've also hit the nail on the
head with saying that like youknow DMRG comes from this
one-dimensional model computecalculation system, right?
Um that has consequences, right,when it's expanded to 2D.
Um and so you know, IQCC doesn'thave those constraints, right?
IQCC is a lot what so one onething that we know about IQCC is
(21:55):
that it's a lot less sensitiveto the uh starting orbital.
It's much more orbitalinvariant.
You know, to be very specific,you know, and well known, you
know, in DMRG world is that uhthe order of the orbitals really
matters.
Because if they're spatially toofar apart, that increases, you
(22:17):
know, that can cause uh eitherconvergence issues or you know
requires more and more uhcomputational time.
IQCC does not have that issue.
Um it you know, it had thisissue where you had to optimize
all these uh amplitudessimultaneously at the exact same
time, which was a little bittricky.
That's what we figured out lastyear, and then implementation
(22:38):
onto GPU is what we did, I mean,within the last six months.
And so it's but the it also toyour point is that it will
produce the output that aquantum computer would produce.
And so now we can actuallyevaluate that objectively and
say, beyond just spitting out anenergy number, what else can I
(22:58):
do with this?
Which is something that, youknow, using um kind of some of
these subspace diagonalizationtechniques that other companies
are proposing and using, youcan't really do because you
haven't converged your system toits ground state.
Right.
And if you can't do that, thenbecause of the scale problem,
(23:19):
you know, they're limited tolike 77 qubits or something, you
can't go and explore it becauseyou know your picture of the
wave function of the of the ofyour Hamiltonian is already
approximated and incorrect.
Right?
That's what always allowed DMRGto have a you know somewhat of
an edge, being like, well, yeah,but I'm converged within my
(23:40):
model space, and therefore atleast my solution is stable.
That you know, you're startingfrom a better place.
Yeah, you're you're startingfrom a better place.
IQCC though, you know, DMRGoften requires like in some of
the new papers, they're like,oh, if we run a thousand
unrestricted Hartree Foxsimulations beforehand to sample
(24:00):
and do this uh sum of Slaterdeterminants or some square
slater determinant method.
And yeah, sure, it scales oh endof the six beforehand, but maybe
I'll save like four hours on thedownstream calculation.
And I IQCC just does not careabout that stuff, right?
That's what makes it a much moreuniversal or robust method, is
(24:22):
that its drawbacks are much moretied to its implementation and
its execution on classicalhardware than it is necessarily
tied to theoreticalapproximations.
SPEAKER_04 (24:32):
And what also strikes me as being very
interesting, uh, we recently uhhad uh a guest that was really
talking about now that you'restarting to have more scalable
quantum computers, you canactually investigate like the
art of the possible because youstart to see practical results,
(24:53):
right?
And I think this is exactly uhin line with what you are
saying, because you are seeingthe actual result that the
quantum computer would produce,right?
Now you can start exploringlike, okay, what else could we
do with these things?
Or what are some of the otherangles that we could take, which
(25:14):
I think is is a side ofsimulation, right?
That at least I personally umhave never kind of was aware of,
right?
Because I was in my mind,simulation was always, oh, we're
gonna simulate a quantumcomputer to understand, like
maybe how many qubits do we needfor this and that algorithms and
things like that, right?
(25:34):
I think this result kind oftakes it upside down in the
sense that we're gonna show youwhat the quantum computer of
these number of qubits for thisparticular simulation will do,
and it's gonna be the realthing, which is uh I think is
remarkable.
SPEAKER_03 (25:50):
So, so Scott, you got you now, you're also you're
not just you know a researcher,you're you're actually building
products.
Is this influencing productsthat are already in production
now or or just plans foroptimizing your product lines?
SPEAKER_01 (26:08):
Like can you even say uh I'm probably gonna have
to err on the side of cautionand say that um you know no
comment.
SPEAKER_03 (26:19):
But eventually I would assume that eventually uh
that this has real-worldimplications as opposed to you
just putting a paper out intothe ether.
SPEAKER_01 (26:27):
Correct.
I mean, one of the things that Ican at least talk about is that
this is you know integral to ouryou know forward expansion plans
into you know non-OLEDmaterials.
Um it's one of the reasons whywe did the greenhouse gas
emission catalyst system isbecause we wanted to demonstrate
that uh, you know, we're notjust all about simulating OLED
(26:50):
materials, we have many moreareas of expertise that we can
actually uh simulate andexecute.
And so in those kind of areas,yes, I can say that this is an
active area of you know tryingto predict which uh what's uh
gonna help us get the bestproduct market uh market product
fit.
Um, but unfortunately in kind ofOLED materials, I have to say no
(27:12):
comment at this point.
SPEAKER_03 (27:13):
Oh, no problem at all.
Yeah, but it's still excitingbecause you know, a lot of times
this kind of development is justin uh like, oh, we built this
and now let's see if someoneuses it, but you're definitely
gonna use it, I'm assuming.
And so we're excited to seewhere you know what comes out
and uh uh down the road.
Hopefully, you know, when youcome back on, you can tell us,
well, remember I talked aboutthis, this is what we did with
(27:36):
it, uh, even if it's next year.
Um anything else that peopleshould know about this?
Is there is there any other uhramifications?
I'm sure you know you're alwaysworking on something
interesting.
Anything else you want to getout?
Because we're coming up on 30minutes.
We've got time, uh, but uh Iwant to make sure we get
everything, every dropout.
SPEAKER_04 (27:56):
Before Scott Scott answers to that, I just want to
kind of pile up one more thinguh uh for his answer.
One of the things that thatreally kind of draw my
attention, if you can just couldcomment a little bit on that,
um, in the paper, was theparallelization strategy.
I I think that's kind of uh aunique approach, right?
(28:17):
And uh it's it's reallyfascinating to see how you folks
kind of innovated in in in thatspace, also obviously combining
it with running on on GPUs.
Um I I think that's absolutelyremarkable.
SPEAKER_01 (28:32):
No, thank you.
It's it's a you know the thewhole parallelization and the
you know strategy behind it isreal what really makes it work.
But I I want to also justreinforce you know the
importance of good softwarearchitecture, right?
Um you know it was programmed inC for a reason.
(28:54):
Um and one of the things thatI'm gonna do.
SPEAKER_04 (28:56):
So I guess it wasn't vibe coded.
SPEAKER_01 (29:00):
No, no.
SPEAKER_04 (29:01):
I I mean the the I I'm just mean to my AI uh
counterparts here.
SPEAKER_01 (29:11):
Yeah, I I mean the the vibe coding is I don't want
to say that it's like inherentlythe the AI like coding is is bad
or not.
Because like, you know, thereare many instances where I'm
like, oh, I need a script thatwill pull data from the output
of this software and manipulateinto the cruft work process that
(29:34):
you'd give to a juniordeveloper.
I don't even know if I'd give itto a junior developer.
I think I'd just look it up onStack Overflow at this point.
Yeah, copy a function.
Yeah, and and one of the thingsI'll point out that it is very,
well, I wouldn't say it's verygood at.
I would just say that it itspeeds up is the ability to
install this very esotericscientific software.
Um, like, you know, we onlyrecently have been actually able
(29:58):
to ever run DMRG because for thepast, you know, basically like
three years, nobody has beenable to actually install block
two at our at our company.
Like it had been the mostfrustrating experience.
And then, you know, using cloudcode, um, it definitely went
around in loops, and I'm noteven entirely sure what
libraries it behind the sceneshad and modified, but it
(30:21):
eventually got it to run.
Um, but there was a lot of likechanges in mathematical
libraries behind the scenes thatit was making, and it was just
like it's just because it canbasically run a thousand
commands very quickly, right?
Yeah, and I'm sure it was justone of those commands that was
needed, but it ran a thousandand it it got it to work.
(30:43):
So, you know, that's what I whatI I see I'm pretty very good at.
But in terms of likearchitecting a code from scratch
or architecting a code so thatit's incredibly efficient, no,
that that was the theparalyzation strategy, you know,
had been worked on for basicallylike almost three years.
It is incredibly complex andnuanced.
(31:04):
And um, we actually don't letthe AI actually touch that stuff
because it's like you it isserious, even though it is
incredibly well architected, youknow, commenting out one line
would probably cause the entirething to implode instantly and
start working.
SPEAKER_04 (31:19):
And I would assume there are also patterns that
because are not common in thepublic world, right?
The AI does not really have whatto learn from.
Um, it's probably there is agood part of it that's unique to
your approach.
AI.
So those are kind of thescenarios where I think at least
currently, AI models andeverything, they're just
(31:42):
breaking.
Uh they can't cope with thingsthat they haven't seen a
thousand times in a thousanddifferent places.
SPEAKER_03 (31:48):
Right.
Right.
That's the key.
I mean, the I mean, I don't wantto diversify in AI, but but it's
it's it's kind of enabled whatyou were doing.
Because if it weren't for the ifit weren't for AI, NVIDIA
wouldn't be building the chipsthat you're you're leveraging.
I I think the the right way tothink about it is is if you're
innovating intentionally, AI isnot your friend.
(32:08):
If you just need to rebuildsomething that's that's work a
day, then yeah, AI is useful.
But we still need people toguide, to point, to guide, to
judge and to blame in the AIworld.
So uh, you know, the like Isaid, the cruft work, the the
the the stu the junk stuff,that's definitely a good space
for that.
SPEAKER_01 (32:28):
Yeah, um I I would just comment that the uh the
paralyzation strategy wasdeveloped well before we had
GPUs that were designed for AI.
It was uh developed on gamingGPUs.
So I would actually attribute umcomputer games or the graphics
cards that you would have inyour desktop to be.
(32:48):
So it's Doom and Quake.
It's Doom and Quake.
It is, you know, IQCC is, youknow, it's like the Doom square
root algorithm, right?
It is kind of very similar tothat.
SPEAKER_03 (33:02):
Very cool.
Very cool.
Anything else?
So we're we're over half anhour.
We still got a little bit moretime.
Is there anything else youwanted to uh get out?
This is very in important news.
It's it's um it's great to seesomething that's been developed
that's already being used inproducts um and and or or going
forward into products thatthat'll affect everybody's
(33:24):
lives.
Anything else you want uh ouraudience to know before we we
cut we wrap it up?
SPEAKER_01 (33:29):
Well, I'd like to say, you know, this thank you
for having me.
This is uh, you know, I'm I'mvery excited to share this uh
monumental milestone.
Yeah um however I'd like to saythat this is actually just the
beginning in my eyes for this.
It's not, you know, the thejourney is not nearly completed.
Um and so you know, I'm sureeven this year we are gonna have
(33:52):
multiple other um innovationscoming out uh and new and
interesting applications, youknow, whether those are as you
know, whether they translate aswell to um, you know, direct
tangible output.
I, you know, obviously we haveto explore these things um and
kind of ponder the you know thebroader implications of this.
(34:14):
But you know, I just, you know,I know it it does sound like a
little bit of, in some respects,it may sound like a bit of doom
and gloom for um somearchitectures or certain
algorithms related to quantumcomputing, but I think this is
more so like, you know, I thinkit serves as a concrete
definable benchmark for thosehardware companies.
SPEAKER_02 (34:36):
Oh yeah.
SPEAKER_01 (34:37):
And yet it also says that, you know, some of them
need to uh start upping theirgame.
SPEAKER_03 (34:42):
Well, you're moving frontiers, and that's a good
thing.
SPEAKER_04 (34:44):
Yep.
SPEAKER_03 (34:45):
Yeah.
But no, we enjoy talking to you.
And uh we're we're we'redefinitely please please let us
know when you're ready to uhshare some more.
And uh we'd love having you onthe show.
Uh appreciate your your yourperspective and uh and sharing
this invention with us.
SPEAKER_01 (34:59):
Oh, thank you for having me on.
SPEAKER_03 (35:01):
Thanks everybody.
SPEAKER_01 (35:02):
Very exciting.
Thank you.
Thank you.
SPEAKER_00 (35:04):
See you soon.
Bye.
Cybercrime is one of the biggestthreats to businesses of all
sizes and industries.
With almost half a million opencyber positions, the problem is
compounded by the lack ofavailable talent in the
marketplace.
At Pulsar Security, our eliteteam of highly credentialed
experts collaborate with you toassess your current defenses and
(35:27):
develop solutions tailored toyour specific needs.
With services ranging fromcybersecurity education to
advanced penetration testing andred teaming, you can start
reducing your risks today.
Visit pulsarsecurity.com andlet's secure your digital future
together.
Hey Cyprien, how are you doing?
SPEAKER_04 (00:42):
Hey Patrick, I'm doing great.
Looking forward for anotherepisode of Entangle Things.
SPEAKER_03 (00:46):
Yes, I think uh we're we're we're rejoined by
Scott.
Scott's been on the show before.
Scott, do you mind reintroducingyourself to the audience?
SPEAKER_01 (00:53):
Uh yeah, thank you for having me on, Patrick and
Cyprian.
Uh I'm Scott Jenn and I'm thevice president of materials
discovery at OTI Lumionics.
SPEAKER_03 (01:02):
And you guys are, uh as I remember from the last time
we talked to you, which was veryinteresting, you're using
Quantum Inspired for your umoptical development and your
product development.
Um and there's some big thingsbrewing, from what I understand.
SPEAKER_01 (01:18):
Uh yeah.
We, you know, last year I was onto talk about the um, you know,
the theoretical foundation thatnow we've been able to implement
at scale across GPUs.
And the, you know, the speed upand the results are impressive
because it really opens up a umreally not just the scientific
(01:39):
application or you know,scientific discovery of what
quantum algorithms could do inmaterial science, but it also
demonstrates that they can bedone in kind of a reasonable
amount of time on uh classicalhardware.
SPEAKER_03 (01:52):
Cool.
So so you're using AI hardwareto like the Blackwell chips,
those kinds of rigs to reallysqueeze that quantum advantage,
but you're still using classicalcompute power.
Um when I I I I've told thestory before.
When I first started talking toCyprian about this kind of uh
technology, quantum, beforethere were even quantum uh
(02:12):
computers of any kind, there wasa company called OneQit up in
Vancouver that um was doingquantum inspired.
And so it seems like you are agood example of a company that's
taken that and run with it.
SPEAKER_01 (02:24):
Um I I think, you know, uh yeah, in in some
respects, I would say we'vetaken it to probably the
extreme.
You know, One Qubit does have abroader focus.
Like they they do they are inthe consulting goals, basically.
Yeah, they're they're consultingfirm.
Right.
And so um OTI has a singular,you know, a singular objective,
(02:47):
which is to do you know, quantumchemistry calculations as
efficiently as possible.
So really what this does thoughis it sets apart um the
algorithm, you know, thisiterative cubic couple cluster
into a different um kind of adifferent class or domain than
just some sort of likeinteresting, you know, novelty
(03:08):
to be run on a you know futurequantum computer.
Right.
Um it really now sets it intothe, you know, uh as a high
accuracy general quantumchemistry algorithm that has
much more efficient scaling andmuch better utilization of GPU
resources than other highaccuracy classical quantum
(03:31):
chemistry methods.
Now, what this what what makesthis very exciting, there's
there's kind of two aspects thatmake it exciting.
There's the scientific aspect,right?
And and then there's also theapplied aspect.
So from the you know, thescientific and theoretical
aspect, this is the firstrealization of a quantum
(03:52):
algorithm, meaning that it'sdoing quantum chemistry or
fermionic excitation operatorsand electronic structure
problems in the qubit domain ofmath, opposed to you know the
traditional fermionic domain.
This you know, it's it'simportant to also you know
(04:14):
emphasize that you know allquantum chemistry methods have
their own pros and cons, quirks,kind of oddities about them.
Yeah, and you know, that wassomething which couldn't have
which could never have beenreally explored in depth prior
(04:34):
to this.
Because fundamentally, mostresults were either like the the
system is just not converged.
So if we look at like thequantum subspace
diagonalization, you know, weyou look at the results and it's
like these results are worsethan uh CISD.
And yet it should be a you know,a true ground state FCI type
(04:56):
solution.
If it's unconverged, we can'treally do downstream uh like
electron density calculations oruh you know, we can't map it
onto there are other things thatchemists care about, like things
like natural bonding orbitals.
Um these are things that like uhyou know, actually applied
synthetic chemists care aboutmore so because it gives them an
(05:18):
intuition about the uhunderlying chemistry that is
happening.
And those things are just notpossible if you can't converge
your uh quantum chemistrycalculation to a like to a
correct or a um like to aconverged state.
And so, you know, at the scale,quantum computers have never
been able to achieve this atscale, and there has not been
(05:40):
really a quantum native methodthat has been able to do this at
scale until now.
Wow.
And so now we can actually notonly say, well, is a quantum
computer, because everyeverything, like if you look at
pretty much all the studies,they're all saying, Oh, if I had
a quantum computer that had ahundred logical error fully
(06:01):
error corrected qubits, I wouldbe able to solve this problem.
Right.
But I think it's actually fairto push back and say, well, what
will this converge to?
Because you say that it willhave this one error, but is it
actually going to converge tothe FCI solution?
And at that scale, we don'treally, we can't really
determine that.
So what this paper hasdemonstrated is that, you know,
(06:25):
historically people have usedthings like uh density matrix
renormalization group, which isa very robust, well-known uh
classical algorithm that um Ithink maybe people got carried
away in saying that it's the beall end all.
But here we objectively justshow that IQCC outperforms it.
Like it basically says I have avariational value that is lower
(06:48):
than your best attempt at DMRT.
And I can do it on fairlyavailable GPU hardware.
SPEAKER_03 (06:57):
Exactly.
Yeah, you you haven't waitedfor, oh, someday we're gonna be
able to do this.
You're doing it, you're gettingthe best of both worlds, really,
because you're you're notwaiting for, oh, well, if I had
a quantum computer, this wouldreally go.
You're you're leveraging and andcomputes exploded with AI.
Uh we we've got some Blackwellchips that we use for other
(07:19):
things, and it's justmind-blowing how much and the
backplanes and the way to youknow put them all together.
Um, and so you're takingadvantage of that.
Now, I imagine if suddenly therewere huge breakthroughs in
quantum computers and there werehundreds or thousands of logical
qubits, then you even benefitmore.
So, so uh so I think this is agreat um way to venture forth
(07:43):
and get the best of uh no matterhow it comes out.
You're basically betting bothsides of the table.
SPEAKER_01 (07:49):
Yeah.
And it's because now we can likewe can verify that a quantum
algorithm, again, working inqubit notation and domain, will
outperform some of the bestknown classical methods, right?
And that's that's fundamentallya significant milestone to like
objectively surpass them.
SPEAKER_04 (08:08):
That's um so if I'm reading correct, these these
results, the kind of if we weretrying to do some some kind of
of mapping to the world of ofquantum computers, what you
achieved is basically theequivalent of 200 qubits.
(08:30):
Like what what you would get ona quantum computer with 200
qubits, is my understandingcorrect?
SPEAKER_01 (08:36):
Yes, that is correct.
With with 200 fault tolerant,fully connected qubits, right?
That I think that's the caveatis that if you in you know in in
how we do it, you know,entangling qubit one and
entangling qubit 200 is youknow, that's instantly doable,
(08:56):
right?
From our from our uh softwarestandpoint.
SPEAKER_02 (08:59):
Yeah.
SPEAKER_01 (09:00):
But on a but on a actual physical chip, that might
be actually quite hard.
SPEAKER_04 (09:05):
Because that's that's very interesting to to to
me, right?
And we've we've touched thisseveral times on on this show
about the evolution and thebenefits of uh simulation, so to
speak, right?
But uh and and this is where Iwould like to to get a bit more
detail from you.
(09:26):
This is not really um simulatingin general a quantum computer,
right?
This is actually implementing analgorithm that runs on GPUs that
yields a result that isequivalent to what you would get
with a 200 qubit uh quantumcomputer.
Is is is is this correct?
SPEAKER_01 (09:48):
No, that that is exactly correct, right?
So the IQCC implementation on awhat makes it really scale very
well on GPU.
Um, it actually has much betterscaling on GPU than it does on
CPU.
Um, is that if you, you know,traditionally in what we
consider like if we think aboutclassical emulation of a quantum
(10:08):
computer, or trying to actuallysimulate the physical quantum
computer, you know, we haveapproaches like state vector,
where we have to store basicallythe wave function as this is a
complex state vector.
And that's where a lot of thisthe bad scaling happens.
Now there's techniques, youknow, Amazon had developed
techniques to kind of prune thisum massive state vector to so
(10:29):
that it was simulatable onclassical computers up to like
easily up to 80 qubits and alittle bit beyond that.
But fundamentally, there is alimit, right?
Especially as you want to getmore and more accurate.
What makes our implementationvery different is that it
doesn't consider really thestate vector.
It works in operator space.
And this allows us to reallyutilize integer and bitwise
(10:54):
manipulation to mimic theoutput, meaning the observable
of the quantum computer.
Right now, from yeah, I thinkfrom a you know theoretician
standpoint of a physicist, theymay think that that's you know a
little bit gimmicky.
But I would point out that youknow, my background's in
chemistry and chemicalengineering.
(11:14):
So the observable is pretty muchwhat I care about.
And this is what the quant, thisis what I'm going to see.
Right.
Right.
So it's like you're kind of ableto bypass and not I wouldn't say
necessarily bypass, but you'reable to, you know, through kind
of uh representing um you knowthe observable basically in
(11:36):
binary bit strings, you canactually kind of exploit the GPU
very efficiently to do thesecalculations and basically
reproduce these observables.
And that's what we areeffectively doing.
Yeah, that's that's one of thereasons why you know you get a
90x speed up on a blackwell overCPU.
Right.
Right?
(11:56):
Like fundamentally, if we if Iwant to be as as uh transparent
and direct as possible, IQcc hasslightly better scaling than
DMRG on CPU, but I think peoplewould really sit there, you
know, scratch their chin and belike, well, you know, 109 hours
versus 130 hours is currentlyone of the benchmarks that I
(12:18):
have for IQCC being lower thanDMRG by about one milliheart
rate.
People would sit there and belike, mm-mm mm-mm, you know,
maybe not.
But then I say, well, what aboutone hour on a GPU?
And that's fundamentally it's adifferent, you're you're in a
different question.
Two orders of magnitude, yeah.
(12:39):
Yeah.
It's it's uh it's a completelythe the implications of that is
it completely opens up what youcould do with this.
Because one of the things thatyou could do with it is you
could do geometry optimization.
And this is one of the you knowquestions that from the
chemistry, applied chemistryside, people always ask is like,
(13:00):
well, can I do geometryoptimization?
And before it'd be like, well,if each iteration takes a
hundred hours, you're gonna bewaiting a really long time
because you have to do themsequentially stepwise, right?
But if each step only takes nowone hour, that's a very
different question.
Yeah.
(13:21):
Then because the geometryoptimization of like uh you know
some OLED materials that wehave, uh, you know, they may
take 50 to 150 steps.
So you're asking really thequestion could you wait 60 hours
to have a variational geometryoptimized solution for 60 weeks?
(13:42):
Versus 60 weeks.
Exactly.
And it's it it completely opensthat up.
Oh, it's amazing.
SPEAKER_04 (13:49):
Yeah, and and just to make sure that our listeners
are are are following, right?
We were um talking here aboutIQCC and DMRG, right?
IQCC stands for iterative qubitcoupled uh uh cluster and uh DM
DMRG, right, stands for densitymatrix renormalization group.
Um that is just kind of themethods that that we are talking
(14:13):
about.
But what strikes me um asprobably as the most important
thing of this, and I know Idon't want to kind of downplay
the the practical sides of it,but for me, what's really
interesting is this might justpush out what we would consider
the boundary of quantumadvantage for for these things,
(14:37):
right?
Because I know for a fact thatin chemistry in general, right,
the the kind of consensus wasit's around 40, 50, 60 qubits,
whatever that is, right?
Now you're coming with thisresult, and you're saying, look,
sure, we're using quantuminspired and everything.
We're still running on classicalcompute, right?
(14:59):
But we can achieve a result thatis equivalent to what you would
achieve with 200 qubits.
So all of a sudden, it feels tome like it's pushing out the
boundary of what would quantumadvantage be?
Um, right?
SPEAKER_01 (15:14):
Correct.
It it also calls into questionabout the notion of like quantum
supremacy, in particular forvariational quantum eigen solver
type uh calculations as well,right?
Is there even, you know, canquantum supremacy ever be
achieved in that class of calcof uh algorithms?
And um, you know, that boundneeds to be dramatically pushed
(15:38):
out.
But even like quantum phaseestimation for quantum
chemistry, now we really haveto, you know, dramatically
consider um what is thatboundary going to be?
Because, you know, using uh youknow the uh resource estimators
that Microsoft had published,you know, using their their
(16:00):
inputs on what they think iswhat their Mariana Mayorana
quantum computer when it'sconstructed, I'm sure one day it
will be.
Um, but when it is, they saythat if everything goes well,
they will need about 200 hoursto get it within this one milli
accuracy.
And it's like, okay, well, howlong is that how much how much,
(16:23):
you know, and then it even comesfrom a practical standpoint of
like, well, how much is it gonnacost to run?
Yeah, right.
And what's one hour of a blackbelt GPU's time versus 200 hours
of quantum computers time?
SPEAKER_03 (16:35):
Time, energy, there's you know, those those
are the big trade-offs rightnow, and and AI is really
pushing.
I think we have to admit thatthree years ago this wasn't
possible with the technology.
You know, classical has hasrocketed forward and and the
goalposts are moving.
I agree.
I I don't think that's a badthing.
(16:55):
I it's probably a bad thing ifyou're trying to uh to prove
quantum supremacy, but that'sthat shouldn't really be that's
more of a marketing thing thananything else.
SPEAKER_01 (17:05):
Yeah.
I mean this this dramaticallychanges even the quantum
advantage or quantum businessutility case, right?
Especially in materials.
SPEAKER_03 (17:16):
That said, I don't I don't think there's gonna be one
of the things we've talked aboutin the past is a quantum winter,
the way AI has had severalwinters where funding's dried
up.
I I think the the Schor'salgorithm specter is is gonna
keep the money flowing to youknow not only the security side
of things, but but also the thematerial science side of things,
(17:38):
which is fine.
But I think this is giving thesetechnologies push each other.
Uh it's an oversimplification.
But when there was more than onedominant browser in the market,
we always got better browsersand better features.
And I think the fact that thequantum is challenging classical
and vice versa is a good thingfor all of us.
SPEAKER_01 (17:58):
Yeah, and it's I think it's also important, like
this, these advances areimportant for also helping
motivate why a quantum computerwould be necessary, perhaps even
in the future for quantumchemistry, because now we can
explore how is a quantumcomputer going to compute
natural bonding orbitals, how isit going to compute chemical
(18:21):
properties like electron densityfrom a native quantum
perspective, opposed to beinglike, oh, well, I have the DMRG
output, right?
You know, DMRG is fundamentallyDMRG.
Same with CCSD, these are there,there's these trade-offs on the
back on the theoretical side,which have downstream
consequences.
And chemists are very familiar,especially applied chemists, are
(18:43):
very familiar with theseconsequences.
And if we can understand thosequantum consequences of doing
them on a quantum computer,which have previously never been
able to be explored, um, it nowserves as like this is the
benchmark and standard that youknow quantum computing companies
need to kind of reach and exceedin order to have a, you know, in
(19:06):
order to really motivate it.
But you know, we might be ableto discover new interesting
physics or uh chemistry usingthis quantum-inspired uh
technique in the in the shortrun, which then would even
motivate that.
Oh, we really need a quantumcomputer that can run quantum
phase estimation perfectly.
Because it could be that, youknow, oh yeah, you know, if I
(19:28):
had another hundred qubits, youknow, 300 qubits, because we,
you know, we haven't been ableto simulate that at this point
at that scale efficiently, umthat could really then
dramatically open up a whole newrange of uh you know new
materials, innovation, anddiscovery.
SPEAKER_03 (19:46):
Uh Nvidia's Vera Rubin chip might well be able to
do that.
You know, some some of the justjust the fact that AI chips are
moving so quickly uh and so andand then custom, yeah.
I could even see if this if thisreally takes off, which it
should, um maybe we'd even getchips designed for this.
I assume that while the GPUs arereally good at it, they there
(20:09):
are optimizations at the chiplevel that could probably help
it as well.
SPEAKER_01 (20:13):
Yeah, definitely with the uh memory transfer and
sharing information betweenthem, right?
Yeah.
SPEAKER_04 (20:18):
Yeah, and I think at the end of the day, right, this
this boils down to the fact thatlike fundamentally speaking,
right, um uh DMRG is based on amodel of classical tensor
networks, right?
As uh essentially opposed to uhuh IQCC, which is qubit native,
(20:41):
right?
So you are literally describingthe problem, right, in the kind
of native language of quantum.
And I think this because if if Iam correct, um even up to like
relatively recently, DMRG wasconsidered like the more mature
kind of approach, right?
(21:03):
But what you're coming andyou're saying with your result
is that, well, it it looks likeright, IQCC can kind of surpass
clearly DMRG and get us to um uhto a point where like simulating
quantum native things is isbecoming like uh uh more
(21:25):
powerful with this with thisapproach.
SPEAKER_01 (21:28):
Yeah, and there's uh you've also hit the nail on the
head with saying that like youknow DMRG comes from this
one-dimensional model computecalculation system, right?
Um that has consequences, right,when it's expanded to 2D.
Um and so you know, IQCC doesn'thave those constraints, right?
IQCC is a lot what so one onething that we know about IQCC is
(21:55):
that it's a lot less sensitiveto the uh starting orbital.
It's much more orbitalinvariant.
You know, to be very specific,you know, and well known, you
know, in DMRG world is that uhthe order of the orbitals really
matters.
Because if they're spatially toofar apart, that increases, you
(22:17):
know, that can cause uh eitherconvergence issues or you know
requires more and more uhcomputational time.
IQCC does not have that issue.
Um it you know, it had thisissue where you had to optimize
all these uh amplitudessimultaneously at the exact same
time, which was a little bittricky.
That's what we figured out lastyear, and then implementation
(22:38):
onto GPU is what we did, I mean,within the last six months.
And so it's but the it also toyour point is that it will
produce the output that aquantum computer would produce.
And so now we can actuallyevaluate that objectively and
say, beyond just spitting out anenergy number, what else can I
(22:58):
do with this?
Which is something that, youknow, using um kind of some of
these subspace diagonalizationtechniques that other companies
are proposing and using, youcan't really do because you
haven't converged your system toits ground state.
Right.
And if you can't do that, thenbecause of the scale problem,
(23:19):
you know, they're limited tolike 77 qubits or something, you
can't go and explore it becauseyou know your picture of the
wave function of the of the ofyour Hamiltonian is already
approximated and incorrect.
Right?
That's what always allowed DMRGto have a you know somewhat of
an edge, being like, well, yeah,but I'm converged within my
(23:40):
model space, and therefore atleast my solution is stable.
That you know, you're startingfrom a better place.
Yeah, you're you're startingfrom a better place.
IQCC though, you know, DMRGoften requires like in some of
the new papers, they're like,oh, if we run a thousand
unrestricted Hartree Foxsimulations beforehand to sample
(24:00):
and do this uh sum of Slaterdeterminants or some square
slater determinant method.
And yeah, sure, it scales oh endof the six beforehand, but maybe
I'll save like four hours on thedownstream calculation.
And I IQCC just does not careabout that stuff, right?
That's what makes it a much moreuniversal or robust method, is
(24:22):
that its drawbacks are much moretied to its implementation and
its execution on classicalhardware than it is necessarily
tied to theoreticalapproximations.
SPEAKER_04 (24:32):
And what also strikes me as being very
interesting, uh, we recently uhhad uh a guest that was really
talking about now that you'restarting to have more scalable
quantum computers, you canactually investigate like the
art of the possible because youstart to see practical results,
(24:53):
right?
And I think this is exactly uhin line with what you are
saying, because you are seeingthe actual result that the
quantum computer would produce,right?
Now you can start exploringlike, okay, what else could we
do with these things?
Or what are some of the otherangles that we could take, which
(25:14):
I think is is a side ofsimulation, right?
That at least I personally umhave never kind of was aware of,
right?
Because I was in my mind,simulation was always, oh, we're
gonna simulate a quantumcomputer to understand, like
maybe how many qubits do we needfor this and that algorithms and
things like that, right?
(25:34):
I think this result kind oftakes it upside down in the
sense that we're gonna show youwhat the quantum computer of
these number of qubits for thisparticular simulation will do,
and it's gonna be the realthing, which is uh I think is
remarkable.
SPEAKER_03 (25:50):
So, so Scott, you got you now, you're also you're
not just you know a researcher,you're you're actually building
products.
Is this influencing productsthat are already in production
now or or just plans foroptimizing your product lines?
SPEAKER_01 (26:08):
Like can you even say uh I'm probably gonna have
to err on the side of cautionand say that um you know no
comment.
SPEAKER_03 (26:19):
But eventually I would assume that eventually uh
that this has real-worldimplications as opposed to you
just putting a paper out intothe ether.
SPEAKER_01 (26:27):
Correct.
I mean, one of the things that Ican at least talk about is that
this is you know integral to ouryou know forward expansion plans
into you know non-OLEDmaterials.
Um it's one of the reasons whywe did the greenhouse gas
emission catalyst system isbecause we wanted to demonstrate
that uh, you know, we're notjust all about simulating OLED
(26:50):
materials, we have many moreareas of expertise that we can
actually uh simulate andexecute.
And so in those kind of areas,yes, I can say that this is an
active area of you know tryingto predict which uh what's uh
gonna help us get the bestproduct market uh market product
fit.
Um, but unfortunately in kind ofOLED materials, I have to say no
(27:12):
comment at this point.
SPEAKER_03 (27:13):
Oh, no problem at all.
Yeah, but it's still excitingbecause you know, a lot of times
this kind of development is justin uh like, oh, we built this
and now let's see if someoneuses it, but you're definitely
gonna use it, I'm assuming.
And so we're excited to seewhere you know what comes out
and uh uh down the road.
Hopefully, you know, when youcome back on, you can tell us,
well, remember I talked aboutthis, this is what we did with
(27:36):
it, uh, even if it's next year.
Um anything else that peopleshould know about this?
Is there is there any other uhramifications?
I'm sure you know you're alwaysworking on something
interesting.
Anything else you want to getout?
Because we're coming up on 30minutes.
We've got time, uh, but uh Iwant to make sure we get
everything, every dropout.
SPEAKER_04 (27:56):
Before Scott Scott answers to that, I just want to
kind of pile up one more thinguh uh for his answer.
One of the things that thatreally kind of draw my
attention, if you can just couldcomment a little bit on that,
um, in the paper, was theparallelization strategy.
I I think that's kind of uh aunique approach, right?
(28:17):
And uh it's it's reallyfascinating to see how you folks
kind of innovated in in in thatspace, also obviously combining
it with running on on GPUs.
Um I I think that's absolutelyremarkable.
SPEAKER_01 (28:32):
No, thank you.
It's it's a you know the thewhole parallelization and the
you know strategy behind it isreal what really makes it work.
But I I want to also justreinforce you know the
importance of good softwarearchitecture, right?
Um you know it was programmed inC for a reason.
(28:54):
Um and one of the things thatI'm gonna do.
SPEAKER_04 (28:56):
So I guess it wasn't vibe coded.
SPEAKER_01 (29:00):
No, no.
SPEAKER_04 (29:01):
I I mean the the I I'm just mean to my AI uh
counterparts here.
SPEAKER_01 (29:11):
Yeah, I I mean the the vibe coding is I don't want
to say that it's like inherentlythe the AI like coding is is bad
or not.
Because like, you know, thereare many instances where I'm
like, oh, I need a script thatwill pull data from the output
of this software and manipulateinto the cruft work process that
(29:34):
you'd give to a juniordeveloper.
I don't even know if I'd give itto a junior developer.
I think I'd just look it up onStack Overflow at this point.
Yeah, copy a function.
Yeah, and and one of the thingsI'll point out that it is very,
well, I wouldn't say it's verygood at.
I would just say that it itspeeds up is the ability to
install this very esotericscientific software.
Um, like, you know, we onlyrecently have been actually able
(29:58):
to ever run DMRG because for thepast, you know, basically like
three years, nobody has beenable to actually install block
two at our at our company.
Like it had been the mostfrustrating experience.
And then, you know, using cloudcode, um, it definitely went
around in loops, and I'm noteven entirely sure what
libraries it behind the sceneshad and modified, but it
(30:21):
eventually got it to run.
Um, but there was a lot of likechanges in mathematical
libraries behind the scenes thatit was making, and it was just
like it's just because it canbasically run a thousand
commands very quickly, right?
Yeah, and I'm sure it was justone of those commands that was
needed, but it ran a thousandand it it got it to work.
(30:43):
So, you know, that's what I whatI I see I'm pretty very good at.
But in terms of likearchitecting a code from scratch
or architecting a code so thatit's incredibly efficient, no,
that that was the theparalyzation strategy, you know,
had been worked on for basicallylike almost three years.
It is incredibly complex andnuanced.
(31:04):
And um, we actually don't letthe AI actually touch that stuff
because it's like you it isserious, even though it is
incredibly well architected, youknow, commenting out one line
would probably cause the entirething to implode instantly and
start working.
SPEAKER_04 (31:19):
And I would assume there are also patterns that
because are not common in thepublic world, right?
The AI does not really have whatto learn from.
Um, it's probably there is agood part of it that's unique to
your approach.
AI.
So those are kind of thescenarios where I think at least
currently, AI models andeverything, they're just
(31:42):
breaking.
Uh they can't cope with thingsthat they haven't seen a
thousand times in a thousanddifferent places.
SPEAKER_03 (31:48):
Right.
Right.
That's the key.
I mean, the I mean, I don't wantto diversify in AI, but but it's
it's it's kind of enabled whatyou were doing.
Because if it weren't for the ifit weren't for AI, NVIDIA
wouldn't be building the chipsthat you're you're leveraging.
I I think the the right way tothink about it is is if you're
innovating intentionally, AI isnot your friend.
(32:08):
If you just need to rebuildsomething that's that's work a
day, then yeah, AI is useful.
But we still need people toguide, to point, to guide, to
judge and to blame in the AIworld.
So uh, you know, the like Isaid, the cruft work, the the
the the stu the junk stuff,that's definitely a good space
for that.
SPEAKER_01 (32:28):
Yeah, um I I would just comment that the uh the
paralyzation strategy wasdeveloped well before we had
GPUs that were designed for AI.
It was uh developed on gamingGPUs.
So I would actually attribute umcomputer games or the graphics
cards that you would have inyour desktop to be.
(32:48):
So it's Doom and Quake.
It's Doom and Quake.
It is, you know, IQCC is, youknow, it's like the Doom square
root algorithm, right?
It is kind of very similar tothat.
SPEAKER_03 (33:02):
Very cool.
Very cool.
Anything else?
So we're we're over half anhour.
We still got a little bit moretime.
Is there anything else youwanted to uh get out?
This is very in important news.
It's it's um it's great to seesomething that's been developed
that's already being used inproducts um and and or or going
forward into products thatthat'll affect everybody's
(33:24):
lives.
Anything else you want uh ouraudience to know before we we
cut we wrap it up?
SPEAKER_01 (33:29):
Well, I'd like to say, you know, this thank you
for having me.
This is uh, you know, I'm I'mvery excited to share this uh
monumental milestone.
Yeah um however I'd like to saythat this is actually just the
beginning in my eyes for this.
It's not, you know, the thejourney is not nearly completed.
Um and so you know, I'm sureeven this year we are gonna have
(33:52):
multiple other um innovationscoming out uh and new and
interesting applications, youknow, whether those are as you
know, whether they translate aswell to um, you know, direct
tangible output.
I, you know, obviously we haveto explore these things um and
kind of ponder the you know thebroader implications of this.
(34:14):
But you know, I just, you know,I know it it does sound like a
little bit of, in some respects,it may sound like a bit of doom
and gloom for um somearchitectures or certain
algorithms related to quantumcomputing, but I think this is
more so like, you know, I thinkit serves as a concrete
definable benchmark for thosehardware companies.
SPEAKER_02 (34:36):
Oh yeah.
SPEAKER_01 (34:37):
And yet it also says that, you know, some of them
need to uh start upping theirgame.
SPEAKER_03 (34:42):
Well, you're moving frontiers, and that's a good
thing.
SPEAKER_04 (34:44):
Yep.
SPEAKER_03 (34:45):
Yeah.
But no, we enjoy talking to you.
And uh we're we're we'redefinitely please please let us
know when you're ready to uhshare some more.
And uh we'd love having you onthe show.
Uh appreciate your your yourperspective and uh and sharing
this invention with us.
SPEAKER_01 (34:59):
Oh, thank you for having me on.
SPEAKER_03 (35:01):
Thanks everybody.
SPEAKER_01 (35:02):
Very exciting.
Thank you.
Thank you.
SPEAKER_00 (35:04):
See you soon.
Bye.
Cybercrime is one of the biggestthreats to businesses of all
sizes and industries.
With almost half a million opencyber positions, the problem is
compounded by the lack ofavailable talent in the
marketplace.
At Pulsar Security, our eliteteam of highly credentialed
experts collaborate with you toassess your current defenses and
(35:27):
develop solutions tailored toyour specific needs.
With services ranging fromcybersecurity education to
advanced penetration testing andred teaming, you can start
reducing your risks today.
Visit pulsarsecurity.com andlet's secure your digital future
together.
Popular Podcasts
Betrayal Weekly
Betrayal Weekly is back for a new season. Every Thursday, Betrayal Weekly shares first-hand accounts of broken trust, shocking deceptions, and the trail of destruction they leave behind. Hosted by Andrea Gunning, this weekly ongoing series digs into real-life stories of betrayal and the aftermath. From stories of double lives to dark discoveries, these are cautionary tales and accounts of resilience against all odds. From the producers of the critically acclaimed Betrayal series, Betrayal Weekly drops new episodes every Thursday. If you would like to share your story, you can reach out to the Betrayal Team by emailing them at betrayalpod@gmail.com and follow us on Instagram at @betrayalpod and @glasspodcasts. Please join our Substack for additional exclusive content, curated book recommendations, and community discussions. Sign up FREE by clicking this link Beyond Betrayal Substack. Join our community dedicated to truth, resilience, and healing. Your voice matters! Be a part of our Betrayal journey on Substack.
Crime Junkie
Does hearing about a true crime case always leave you scouring the internet for the truth behind the story? Dive into your next mystery with Crime Junkie. Every Monday, join your host Ashley Flowers as she unravels all the details of infamous and underreported true crime cases with her best friend Brit Prawat. From cold cases to missing persons and heroes in our community who seek justice, Crime Junkie is your destination for theories and stories you won’t hear anywhere else. Whether you're a seasoned true crime enthusiast or new to the genre, you'll find yourself on the edge of your seat awaiting a new episode every Monday. If you can never get enough true crime... Congratulations, you’ve found your people. Follow to join a community of Crime Junkies! Crime Junkie is presented by Audiochuck Media Company.
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.