AI Scaling Nobody Explains From Industry Leaders

Of all the technological advancements of our time, none feel as unpredictable as artificial [music] intelligence.

Artificial intelligence artificial intelligence both uh heterineous models [music] with architecture that that that's the kind of quite exciting and challenging problem we also try to solve.

Today the average drug cost is almost 2.6 billion. The performance is really

2.6 billion. The performance is really important. But when you achieve this

important. But when you achieve this performance, you need required. Then

what's the next important [music] thing is your efficiency.

AI is awesome at what it's awesome at, but it kind of sucks at everything else.

We deployed AI systems across 60 manufacturing sites globally in Sani.

The even fundamental bottleneck of running AI [music] we think uh is the energy. How some of the smaller language

energy. How some of the smaller language models might become smaller and smaller and potentially embedded.

But why are we only looking at productivity? Why not also look at

productivity? Why not also look at quality?

Thank you everyone uh for making it uh this uh Thursday [music] evening. It's a

weekday indeed. Um but you know what uh there is no shortage [music] of AI panels here in Boston or here uh around MIT. I know that. But what makes uh this

MIT. I know that. But what makes uh this panel uh exceptional [music] and really unique is the mix, right? Uh so uh in this room we've got enterprise engineering, frontier chips, mRNA

[music] strategy and also real world financial AI systems. So uh I think we're really [music] uh in for a treat.

And if scalable AI was the goal, then I think we have the Avengers here. And uh

first I will introduce ourselves. Um and

uh I guess I'll go first. My name is Re Lim and um I like to say I have three lives. So uh I was a financial reporter

lives. So uh I was a financial reporter uh at CNBC uh soul. Um so that was my first life. Second life as an MIT

first life. Second life as an MIT co-founder for an startup focusing on 3D bio imaging. And then now uh I have my

bio imaging. And then now uh I have my boss somewhere here but I I am a partner at Oneway Ventures where we back exceptional immigrant entrepreneurs. If

you have an opportunity to walk around a little bit and look out all the different directions, you can see out to the to the west to the south over the river where you've got of course MIT and and BEu. Um, of course this new so you

and BEu. Um, of course this new so you really can see pretty much all of Cambridge and beyond from right here.

And um, so I really to me this is a very special place. Um, and in fact I've been

special place. Um, and in fact I've been here a long time because um, I was a graduate student in the early 90s uh, at MIT. I did my PhD at MIT just right over

MIT. I did my PhD at MIT just right over there and um with our founder to Tom Tom Leighton uh who was one of my PhD adviserss and I joined Okami in the

summer of 1999. The company at the time was about one year old. Um and it's been pretty wild experience ever since now 26 and a half years. Um done lots of

different things. I've run

different things. I've run organizations, I've done engineering, architecture, all sorts of uh different interesting things here at Okamay. We've

been um changing throughout from CDN, static content, dynamic content security, cyber security, now cloud computing and of course um we recently announced our our um our inference cloud

with obviously a big focus now on supporting our customers um with AI.

Thanks Bobby. Uh I'm Shokaduri. I'm a

global head of strategy and also intelligence at Sanopi uh at MRNA center of excellence. uh so before joining son

of excellence. uh so before joining son of I spent uh uh more than 10 years in biotech doing a business development corporate strategy uh so during the time

I also did my MBA from MIT slow that's why I met MIT community and re I'm happy to really you know be here and discuss what we're doing at soni how we're scaling I so looking forward to it thank

you hi my name is Jun uh I am um CEO and co-founder of furious AI um thank you for hosting this wonderful

event and it's a pleasure to be here. Uh

actually I um started my semiconductor uh engineering career. Uh my first job was actually here in Boston actually. I

worked AMD uh Boston as a co-op student uh which was actually CPU design. It was

like in 2006 actually uh and um then I worked a semiconductor company uh for almost 10 years. Then I

um started this furious AI around 2017.

So it has been almost like nine years of journey uh from as a venture funded startup to uh to develop this uh uh the high performance AI chip. uh hopefully

we can uh share some some of our perspective uh during this panel session. Thank you.

session. Thank you.

Um hello, my name is Jen Shan. Thank you

for inviting me. Um I am head of AI solutions engineering applications at Fidelity Investments and also an IT fellow. Been there for about 16 years

fellow. Been there for about 16 years and uh prior to that I was chief architect of S Street Global Markets.

Uh well um you know today uh session is really about sustainability and I think sustainability uh can mean different things. It can be efficiency, it can be

things. It can be efficiency, it can be ethical deployment or sometimes business viability, right? So, um I want to learn

viability, right? So, um I want to learn from the leaders of each industries. Um

how are you looking at sustainability in terms of scalable AI? Uh and what does that actually look like in practice today?

Well, yeah, it's a great question. Um

and I really like the way the question was framed uh broadly because oftentimes when people think scalability or efficiency all the focus goes to the

infrastructure layer and that is super important and I know we're going to hear more about that because frankly if you can't scale at the infrastructure level the layers above aren't going to be able

to scale either. So super important, but you do need to think about those layers above from infrastructure to software to systems um to to processes, workflows

and and ultimately up into entire organizations. And and we've seen I

organizations. And and we've seen I think again and again studies have shown that enterprises are still struggling with the challenge of scaling AI in a

way that really generates meaningful impact to the enterprise, real P&L impact. Um, and given that when I think

impact. Um, and given that when I think about scalability and sustainability, I take a a pretty broad holistic view and I always start with matching the tool to

the problem. And that sounds almost

the problem. And that sounds almost naive, but it's important especially in the age of AI because so often the hype

around AI leads enterprises to start with AI as the goal, not value. So again

and again I hear companies saying my goal is to use AI more.

Okay maybe. But how about start with the problem? Start with what is the problem

problem? Start with what is the problem that you're trying to solve that can really improve your product, your customer experience, your employee experience. What is the problem you're

experience. What is the problem you're trying to solve? And then match the tool to the problem. AI is not always the right tool. And I and I want want to say

right tool. And I and I want want to say that again because especially with agents, you keep hearing about agents being talked about as employees, part of your workflow.

I don't like that. A AI and agents also they are tools and you need to pick the right tool for the problem and AI is not always the right tool. It often is, but it's not always the right tool. Um

sometimes I say that, you know, AI is awesome at what it's awesome at, but it kind of sucks at everything else. So use

it for what it's awesome at but not for the everything else. Let me just one more thing on that which is that also when AI is the right tool there's lots of flavors of AI. Another pitfall that I

keep hearing is equating AI with large language model. There's a whole world of

language model. There's a whole world of AI out there beyond large language models that can often times solve your problem much better than a large

language model and potentially at 1 1,000th maybe 1 millionth of the cost of what it would cost what it would take to do it with an LLM. So, pick the Oh, and

I should also add even when when LL when an LLM is the right thing, you don't always have to use a gigantic ask me anything LLM, right? You've probably

heard about smaller language models.

Instead of SLM being small language model, I like to say they are specialized language models because if you take a small model and specialize it, it can oftentimes do better than

what the large model would do at again potentially 1/ 1,000th of the cost. So

match the tool to the problem. Don't

just jump in and use AI everywhere.

Thanks Bobby. I think one thing I like to talk about former point of view, right? So re I think you ask what is

right? So re I think you ask what is sustainable. I think what is not

sustainable. I think what is not sustainable in pharma is the biggest issue is the drug development cost is going up and up. So it's almost eight times more compared 30 years ago for each drug to come into clinic and also

go to commercial. So and today the average drug cost is almost 2.6 billion for the new drug actually to come into commercial stays actually. So it takes more than 10 years. Yes we seen the technology advancement but we didn't

really seen the improvements actually in drug drugs to really coming to market and a cheaper cost actually. So I think connecting dots to what Bobby said actually yes at Senophi we do think

about actually uh fit for purpose actually like what we're building for.

So the four pillars actually we we do use uh Senophi is of course publicly we communicated environmental impact uh by

20 uh 30 actually 55% reduction in emissions and also by net zero by 2045 actually. So why a to come into picture

actually. So why a to come into picture actually that's where we are using actually what how to really draw the metrics actually to support this environmental promise that we made and second pillar is actually operational

efficiency so we are really looking into manufacturing sites for easy to solve the problems actually so we do have a real example we deployed AI system so across 60 manufacturing sites globally

in Sani and we seen almost 80% of shortfalls actually you know we prevented actually before even getting to the uh I I before getting to the our hands actually. So I think that also

hands actually. So I think that also saved money and also efficiency actually. So second thing is also about

actually. So second thing is also about ethical deployment. I will get into

ethical deployment. I will get into maybe examples later actually in our discussion. So third one is ultimately

discussion. So third one is ultimately you know we are in a business actually you know serving patients actually. So

patient impact actually. So how to measure the metrics. So that's also we're looking into. So really the four pillars of sustainability. So we're

looking into here.

I'll share more perspective from the computer side. Um so

computer side. Um so so my analogy is uh like uh it's like from like uh transitioning from the

gasoline car to let's say electric car.

We think uh electric car is much more sustainable than um than the gasoline car because um um the technology is

fundamentally saving the more uh energy.

uh I think we need uh the same kind of uh transition uh uh on this computer side as well. Um like uh the the GPU I

mean the the fundamental cost of running AI these days even fundamental bottleneck of running AI we think uh is the energy right now. So uh we need

quite um more innovation and breakthrough uh to uh make this uh AI compute more uh energy efficient. Um and

when I say the energy efficient it doesn't mean we sacrifi sacrifice performance uh like electric car is not slower than the gasoline car right even it's even

faster. So um so we like to uh think

faster. So um so we like to uh think about new uh engine to uh make this compute way more uh efficient. Uh that's

our kind of focus on this uh making this uh I think it's a very fundamental layer uh to make this AI ecosystem sustainable.

Sure. Um I would just like to say everything Bobby said was spot on. Same

with Ashoka and June as well. I think on the one hand a lot of us uh consumers that um are on the other side on the software side of compute are looking forward to the cost for going down the efficiency curve of going up. But I

think one of the things that I'm noticing is this is a real opportunity for us to look at how we're actually doing things within the enterprise and within the business and within the processes that we have to make them more

efficient using tools uh that are kind of fit for purpose. And uh one thing that I've seen get gained tremendous benefit from that is really how we're going from sort of the hype of the vibe

coding that happened in the earlier part of this year to more spectriven development and then once that started happening inspecting how our software development life cycle is kind of strung together and trying to look at

efficiencies there. Um but I agree with

efficiencies there. Um but I agree with you Bobby. I think a lot a lot of the

you Bobby. I think a lot a lot of the onus on large enterprises is really making sure that we're finding the right ROI value for the particular use case and um you know AI models are only

growing larger and uh they're being distributed right as we speak um and Akami has experience uh scaling globally right uh so from uh the lessons that

you've learned and of course everybody else in your respective industries u how can you apply those lessons to build scalable AI foundations Yeah, I I I do

think that we're sort of at the beginning of potentially a repeat of something that happened about 27 28 years ago. Um, and that ultimately led

years ago. Um, and that ultimately led to the invention of the CDN and the founding of Okami.

Backtracking to that time, some of you might remember at that time the worldwide web was just taking off and the inventor of the worldwide web, Tim

Berners Lee, um, moves from, uh, I think he was in Switzerland at CERN, but he moves here to MIT and they established

the worldwide web consortium, W3C, at MIT. And through I think accident or

MIT. And through I think accident or happen stance they ended up on the third floor of the building that hosted the lab for computer science. The third

floor happened to be the the theory floor and that's where Tom Leighton had his office and his students his in particular his graduate student Danny

Leuen and Tim Berners Lee was fretting over what was then called the worldwide weight. the worldwide web was getting

weight. the worldwide web was getting congested. And the problem, and he knew

congested. And the problem, and he knew this, the problem was that the web as it had been built so far, it was all

self-hosting and people were delivering websites that were getting very popular from a very small number of locations and that was creating bottlenecks and congestion. That's what led to the

congestion. That's what led to the worldwide weight. Fortunately, just down

worldwide weight. Fortunately, just down the hall on the third floor, you had somebody who had all the tools, all the mathematical tools to solve the problem.

That was Tom and Danny. And that's what led to the founding of Okami. I think

the same thing is is about to happen, but with inference.

As we've trained these models, now we have to use these models to do valuable things. And for all the caveats that I

things. And for all the caveats that I gave you earlier, as I said, LLMs are awesome at what they're awesome at. and

they're actually awesome at a lot of things and we are going to be seeing incredible amounts of inc incredible

systems built with LLMs. But if we serve them the inference in the way that's currently being done, which is pretty centralized, we're going to end up in

the equivalent of the worldwide weight.

And here's a challenge for you. Help me

out. I need an analog. So, back then we called it the worldwide weight. So, I

need I need a catchy term. So, here's

the best I've come up with so far, and it's it's pretty bad. Um, I've come up with large language mess, large language morass. I've come up um a

colleague said, "How about large language molasses?" Um, so help me out

language molasses?" Um, so help me out here. If you can come up with something,

here. If you can come up with something, I'll I'll give you all due credit. We'll

use it. Um, but I'll give you all all due credit. Yeah. So we need to find a

due credit. Yeah. So we need to find a way to do um deliver inference more efficiently and I think distribution is going to play a role. Um that now

obviously I'm saying that in there's something somewhat self- serving here because that's what we do. Um, but it one of the reasons why we're investing

in being an inference provider for AI is because we believe we have that capability to potentially solve the large language molasses that I think is coming.

Large language molasses. It's the best I got so far.

molasses. It's the best I got so far.

Help me out.

So yeah, I I I totally agree that the uh the the the future of um this AI inference should be more much more

distributed one uh and um it it doesn't have to be just a big centralized place

uh and um I think small and um like many different specialized models working in orchestration to build the the the the

create the value will be the the ultimate goal of the inferencing. Uh I

think uh um they'll they'll these uh like a service area will have a huge innovation and transformation down the

road. We believe uh and uh especially as

road. We believe uh and uh especially as we go to more distributed uh area that's those are also very also power

constraint as well. Power distribution

the power will be very important matters as well. Uh so uh compute solution also

as well. Uh so uh compute solution also needs to be uh quite aligned with those quite uh be compatible with those kind

of uh edge data centers. uh and um another but from our perspective another challenge is the computer solution not only efficient but also needs to be

quite programmable enough to support many uh heterogeneous models with the architecture that that that's the kind of quite exciting and challenging

problem we also try to solve actually um I actually think you know everything that June and and uh Bobby said absolutely uh agree with But I also wonder how how some of the smaller

language models might come become smaller and smaller and potentially embedded. Um, in fact, you know, earlier

embedded. Um, in fact, you know, earlier this year, there's a company called M5 Stack that builds embedded uh modules uh for $49 off of Ali uh Alibaba.com. You

can get a programmable module that has a Quen 2 billion parameter model distilled from DeepSeek. Um, and you can use that

from DeepSeek. Um, and you can use that home automation. I think over time um

home automation. I think over time um we'll certainly have more distributed inferencing clouds and distributed architectures for inferencing but I think some of these models are going to shrink and get a lot smaller as well.

Yeah. So I just add one line actually so right. So what we are doing is also fit

right. So what we are doing is also fit for purpose as I said.

So three different uh type of AI we are using now actually. So one is expert AI where we use really LLMs bigger LMS actually to really look into new targets for drug discovery and then other one is

actually you know geni. So basically you know very daily task actually for employees in the company. So the one most interesting one actually is snackable AI that we have in our app in our phone we can see uh the scenarios

what if scenarios actually what's happening in manufacturing side what's happening sales you know that really computes actually over 1 billion data points actually so I think so using the

models in right way and right purpose actually that's where I think we see the meaningful use actually so for solutions yeah snackable AI huh first time using

or hearing about it but yeah Um, pretty cool. And um, June, you know, Furiosa,

cool. And um, June, you know, Furiosa, you've mentioned um, uh, basically efficiency versus performance, right? It

can't always be, I guess, um, uh, a tradeoff, but um, Furiosa AI is tackling the biggest AI, uh, bottleneck, which is compute, right? But it's not just about

compute, right? But it's not just about chip design. It's about energy use, data

chip design. It's about energy use, data management, algorithm design, all of it, right? So, how do we actually think

right? So, how do we actually think about resource efficiency versus that high performance? Is this like a chicken

high performance? Is this like a chicken and egg problem or what's kind of the mental model here?

So as I as I said like um um like AI chip fundamentally must be very high performance machine. It's it's called

performance machine. It's it's called sometimes we are accelerator. So um um so it's one of the ultimate kind of performance engineering we we believe.

So performance is really important. But

when you achieve this uh uh performance you need required then what's the next important thing is your efficiency. So

how much power you are uh consuming and uh what's the capex for your machines all those things are quite uh important

metric actually we need to achieve. So

um so I think uh the actually performance and efficiency is not kind of trade-off relationship. Uh actually

we need to uh achieve both performance and efficiency as much as we can. But

the the the tradeoff the challenging trade-off is actually this uh efficiency and also programmability actually cuz um

um like uh the the challenging part so the AI chip design is not is quite different from designing this bitcoin mining chip because uh you know the if

you think about bitcoin mining the mining algorithm is actually fixed algorithm actually it's not it's quite stable. It doesn't change that much. So

stable. It doesn't change that much. So

you can optimize your chip to be very optimized for let's say uh the bitcoin uh mining uh algorithm. So these days all these bitcoin mining is happening

all the very specialized aging. But the

the difference between between mining and AI algorithms are AI algorithms are one of the the algorithm that evolve very fast that changes like the one of

the the fastest changing algorithm we have ever seen in terms of scale in terms of its kind of architecture uh meaning that uh AI chip uh needs

quite fundamental general architecture to support uh very changing workload.

So it means it needs some flexibility at the same time you like to achieve uh quite the efficiency as well. So the

finding the right tradeoff of this efficiency and flexibility is uh one of the quite um the challenging part of AI chip design which actually we are very

excited about actually.

Anyone want to chime in on the scalable AI u efficiency versus u performance angle? No. Um, I just want to say thank

angle? No. Um, I just want to say thank you for making it faster and cheaper.

[laughter] All right. Um, well, I guess the next

All right. Um, well, I guess the next question is to Ashoka and Jyn because you guys are in industries where trust is paramount, right? Like healthcare and

finance, right? Um, so how do you scale

finance, right? Um, so how do you scale AI while ensuring that you protect trust and privacy and still deliver on the real world impact? Like seems like

that's a pretty big mission. So I think uh we're in patient serving industry right so all the all the components you mentioned is crucial for us actually to develop a drug that successful actually

to reach the patient. So I think uh one of the philosophy that we are using in the company is raise uh basically it's a responsible AI at son of for everyone so

that goes through so much scrutiny and uh uh of risk that we're actually taking actually with the systems actually. So,

so there is a guardrails uh we put together actually in the in the you know in the process. So I think total we pushed more than 50 actually AI systems to do this process actually both our

internal processes and also working with the developers actually outside. So I

think that is really crucial actually you know to really uh especially dealing with the data from the patients and also there's a lot of uh data coming from outside historical data actually. So I

think this is uh this is something actually you know the ethical point of ethical deployment uh uh this is really crucial for us actually while building our drugs actually. So

yep um Ashok and I were talking earlier I think our industries have very similar problems similar solutions in terms of guard rails and whatnot. Um, one of the things that comes to mind specifically

in this area is um, and I've talked to lots of peers and other companies uh, as well is that we have to find the right balance between approaching go AI governance in a way that is not

technology governance because technology governance in large enterprises tends to be very bureaucratic and and and so um, at the same time we still have to let all the innovation flow. there's so much

excitement and energy down to the individual develop developer to UXD uh user experience designers using Figma make to generate uh proto interactive prototypes for example um to other

functions as well. So I think that's part of the challenge um where we you would find the sweet spot to be able to figure out a way to have the same kind of mechanisms for safety and whatnot and

technology management that you would do in typical technology governance but kind of speed up the learning loop and speed up the iteration rate so that you can still let the innovation kind of flow and bubble up.

Yeah. So I just add one more thing right. So I think the philosophy also is

right. So I think the philosophy also is like AI you know advice human decide. Uh

basically you know uh another mantra that we also use is that basically AI plus uh human uh you know uh basically is the one actually making you know human is the one making decision actually. So more responsible taking

actually. So more responsible taking accountability for the human actually.

So that is also very important actually you know when it comes to ethics. So

moving on to the next question. Uh so I want to rethink about scaling right. Um

because I think when we think about uh scaling AI, it's always about um uh you know maybe the parameters, the model model parameters and the size, right?

But maybe uh there's something else here like maybe there's fairness, long-term resilience um uh you know and uh these uh in some ways um abstract things that

we don't think about quite as often. So

um do we have any thoughts on that?

Yeah. Well, when it comes to yet measuring um there's a huge challenge here and that sort of on on one extreme, you know, all the way down sort of at the at the software infrastructure

layer, it's fairly easy to measure say, you know, cost per token or or watts per token or something like that. At the

other end, you know, at the enterprise level, it's not hard to measure something like, you know, profit and loss. We have to do that. Um it's in the

loss. We have to do that. Um it's in the middle. Like when we look at say a

middle. Like when we look at say a developer using an AI tool, has they has it improved? has what's the impact and

it improved? has what's the impact and how does that you know impact on the on the company's bottom line but the measurement yeah so the measurement is very difficult in the middle and and one

challenge that I always point out is that almost all the effort that I've seen where people are trying to put in place metrics to measure this stuff it always goes down to productivity you know and obvious and productivity is

hard to measure especially like a developer or something like that but why are we only looking at productivity why not also look at quality Um, this this bothers me a lot that everybody talks about AI as a productivity improver and

therefore they talk about how many people we can eliminate from the job force. How annoying is that, right? How

force. How annoying is that, right? How

about talking about how much it can improve the work that the people do and make the people more valuable, make the ROI on the people better because they're

able to do better work. um you never hear that and I I admit it's really hard to measure this stuff but it's also hard to measure the productivity part of it anyway. So my my plea is simply as we

anyway. So my my plea is simply as we work on metrics let's not focus just on productivity let's focus on on quality.

So I can add uh from pharma again point of view right you know I think metric that we aim for is you know patient impact for a you know compute unit that is something very complex to measure

actually but we do have a measurements actually for example right you know we're combining both the digital twins uh like basically virtual patient with AI AI systems we a we are able to

actually you know develop the drugs much faster actually so using digital twins actually we have three examples that we shared publicly uh is uh one of them is um you know when there is a drug

approved in adults how we can really interpret actually you know make a model actually that can work in a you know pediatric population actually so so that you can avoid actually large number of clinical trials you know patients in in

controls actually when they're not getting medicine right placebo that is one of the clear example actually and also through digital twins and AI systems we also seen actually uh how the

approved drug can be applied to new disease new indications actually before even getting into clinical trial actually. So I think these are the you

actually. So I think these are the you know clear examples actually you know where you can look into to really reduce the cost for drug development also improve the patient's lives actually. So

I think yes there is a lot of metrics actually but you know if you have clearly measuring actually you know what the impact looks like before AI implement or after implement you can really see a change actually. Um, sure.

Um, the the thought that comes to mind is around coding agents. That's been on my brain a lot lately to be honest with you. And, uh, I agree with what Bobby

you. And, uh, I agree with what Bobby said. You have to take into quality as

said. You have to take into quality as well. And some of the analogies that

well. And some of the analogies that I've used at work has been uh, you can take the Iron Man suit and you can give it to Tony Stark or you give it to Homer Simpson and you're going to have two different outcomes. And and and what I

different outcomes. And and and what I mean by that is senior engineers uh, coding agents are like they just give them superpowers. You can do so much.

them superpowers. You can do so much.

But a lot of the thinking and discussions that we're having uh across my peer group is how do we uh use with spectrum development how do we kind of

ensure that the prompts to the coding agents do local linting local uh code static code quality checks. Um there was a paper published I think a week ago or last week uh from Carnegie Melon and I

think they scanned about 800 or so GitHub repositories into two different buckets. uh one that did use coding

buckets. uh one that did use coding agents and one that did not. And what

they found that was that you immediately get an uptick in productivity, which is basically lines of code uh committed, but then you immediately follows with an uptick in static code quality warnings

and lines of code kind of exploding. Um

so I think the quality comment that you made, Bobby, is spot on. Uh we have to be able to use these tools very efficiently. But uh you know I I think

efficiently. But uh you know I I think uh we think about uh the kind of the AI race or war right globally um where uh people think oh China is gonna take over the AI race right the US got to move

fast and Korea saying we're going to be number three now right uh but how do we think about collaboration uh across uh actually countries and uh

also across uh research inter industry and policy uh what's uh maybe um a blue ocean or a north star uh with that regards uh for crossborder collaboration

for scalable AI to make it sustainable and fair and equitable.

I mean it's an area that I yeah not much of an expert on on policy but let me just just offer one quick thought just personally. Most of my life I feel like

personally. Most of my life I feel like has been characterized at least when it comes to the the question you're asking by increased international cooperation.

there's been more and more free trade in terms of ideas but also commerce. Um,

and the last few years of course are characterized at least certainly at the at the federal level by the opposite and and frankly to me it's bewildering. Um,

just maybe that's all I got to say. I'm

bewildered and I don't know that I can offer it much else.

Good good bewildered or uh no bad bad bad bewildered.

Okay. So brace ourselves.

Well, hopefully we can reverse it. It's

one thing that I think hybrid approach is the best approach you know you can do in a company certain things you know you can really always you know you know collaborate you know borrow the thing that you can't really you know

build inside so we need actually you know multiple collaborations uh with open AI and also formation bio for different kind of you know for technology to drug discovery so coming

back to policies right you know we also working with the FDA EMA in Europe actually you to really understand actually when the AI systems are implementing for drug discovery making sure that actually they also sees the

same way that we are seeing actually because as you can see going from Europe to US and Asia right now every every agency sees things very differently actually so making sure that there is a

harmony and standardization actually so I think really the collaborations global collaboration is really important actually so the drug appro should work same way in the in Europe and everywhere else in

the world actually So I think AI system same thing actually it's the same same drug same systems actually. So but we want to make sure that all these policy makers actually understand same thing actually. So they speak same same

actually. So they speak same same language. I

language. I I think uh as you notice a lot of uh countries are talking about sovereign AI

these days. uh they like to um kind of

these days. uh they like to um kind of um cultivate and uh their own AI

capability and um also control their technology as much as as they can but in

reality actually uh the the other than US I guess the other US and China I'll say it's not

easy to to build entire things sometimes by themselves. Uh so

by themselves. Uh so so I think the the collaboration uh global collaboration

those uh countries uh like uh who uh don't want to rely just on just u single company or a single nation uh can

potentially collaborate with each other to to to build uh their own sovereign AI concept I believe. One thought that I

had is um I personally have never witnessed or experienced so much energy from public and private in academia in open source you know open sourcing code

open sourcing modelates I love that open AAI open source GPO GPTO OSS for example um it's just I think that's a great

solution because there's so much human energy you know from all various levels um just focused on um the things I guess

what you said Bobby the things that LLMs and genai for example do very well it feels very very magical because the sophistication of the language that's output is is very high so

just at one point actually talking collaborations right so uh also it's very unconventional actually so we partnered with the formula one to really understand uh their thinking actually

about efficiency as you can imagine actually bit stop they record is actually 1.80 in 0 seconds. So, so the the reason we cro it actually with this company right to really understand what

goes through it actually to reach that really milliseconds of precision actually the time savings right so we're building that kind of a philosophy actually in the company actually through AI systems actually especially to start

our manufacturing we seen some efficiency improves actually so even you see these collaborations actually to really understand what they are doing and borrow some of the thinking actually to our industry side

so Bobby the world is not doomed seems like there's a lot of uh yes silver linings. Uh my last question and we'll

linings. Uh my last question and we'll open up the floor but is a rapid fire style so 30 seconds each. So if you could come up with a global metric to measure sustainable AI what would that

be?

Happiness per inference and I think of everyone using ro AI return on investment of AI investments.

I think if you go one layer deeper right for our industry especially the industry we have in pharma patient impact patient impact for a you know compute unit actually it's so complex you know that's a northstar we are aiming for

health per inference yeah for sure so that I think I think it's complex but I think if you keep this as a north star right you can aim for it actually can you know deploy your yeah in different particles as you

combine actually that you see the impact so so I'll say the the intelligence per energy or we can say value per energy I think how many how much value we can

create per value kind of relate to the quality as well. I think about value per energy

as well. I think about value per energy can be the metric for sustainability.

Um I would say something a little bit different. I think it's just the overall

different. I think it's just the overall STEM literacy in youth. And I loved Bobby that you mentioned that your wife leads day and AI for example because I kind of fear that before if you were

lucky uh you had access to a computer in the classroom back in the 80s if you're a Gen Xer uh most of you look like you're a lot younger than I am but um but then it was access to content in the

worldwide web. Now it's kind of access

worldwide web. Now it's kind of access to intelligence and reasoning. And I

really think especially in in economically disadvantaged uh populations uh we really need to increase literacy in STEM.

Well uh thank you uh very much for all your questions.