Microsoft’s AI Partnerships and Google’s Gemini 3 Impact | TSG Ep. 971

Microsoft's on Fire. Again, you're watching Textron. Hey everybody.

Welcome back to the Textron Gang. Today we have our, some of our usual gang members. We got John Schwartz, Terry Robinson, guy Courier, who will be giving us a report from St.

Louis in the supercomputer show shortly. But first, let's get started with the Microsoft Ignite Conference that John Schwartz attended this week. And it seems like Microsoft's partnering with everybody and anybody one more time.

They're, I guess they telegraphed this one. They said they were gonna, you know, not be exclusive with open ai, so they partnered up with Anthropic and Nvidia. But, um, John is this kind of par for the course for Microsoft, because like, sometimes I feel like maybe they built software that's kind of cool, and the rest of it is kind of like, well, they're a distribution outfit for everything else.

Yeah, it seems like that, I mean, you know, actually this, this deal, this partnership, which kind of distanced themselves a little bit more from mop ai, it's like this major cloud infrastructure partnership with philanthropic and Nvidia. It, it was so important to them that they actually announced this before the Ignite Show, which is their big conference out here in San Francisco. So they thought it was much more important than actually the products they, that, the products that Microsoft announced.

So that gives you an idea of the scope and the kind of the magnitude of this announcement. And just really quickly, um, philanthropic is pledging to purchase or buy $30 billion in computing capacity from Azure Cloud platform. NVIDIA's gonna invest 10 billion in Anthropic, and Microsoft's gonna contribute up to $5 billion in Anthropic.

So basically what they're trying to do is address this, uh, enormous computational demands of advanced AI systems. So that was first what they did. Um, and then later in the day, during a three hour, seemed like it was five hour marathon keynote speech and presentation, which, uh, just went on and on and on.

They, Microsoft unveils some AI tools to weave AI capabilities into corporate workflows rather than treating technology as an optional enhancement. So they showcase a number of things bearing this IQ branding, which is short for intelligence quos, and it extends from individual workstations to enterprise data centers. So in a sense, what they're doing is they're pushing AI beyond the experimental phase in their words, and, and making a stronger case than measurable ROIs.

Now within Reach, and as you said, Mike, it was, I think in a incredible extraordinary week for tech news. They probably won the week, but I mean, just wait until next week when Google wins the week, or who, whomever. This is just, I've never seen a series of news announcements, uh, and partnerships where there're shifting alliances.

Uh, people we don't know who's in whose corner. Uh, it's just a free for all. And I, I think it kind of underscores this land rush that's been going on for ai and it's just only gonna accelerate Guy.

Do you perceive that any of these alliances are strategic? And I'm asking this question because well, pretty sure that Microsoft somewhere is probably building its own GPU chip somewhere. And it's probably gonna say, you know, well, we're gonna offer that alongside Nvidia.

And then they're also probably saying, well, we're offering anthropic alongside, uh, open AI because well, they're gonna consume cloud resources. And, you know, as far as we're concerned, everybody's all good with us. Anybody who wants to pay us money for hardware processing is great.

So, you know, from your perspective, I mean, how significant are these alliances? I don't think they're particularly significant except in the, in the general sense of this is how, you know, this is how AI is evolving, this is what we're getting moving towards. We're moving at.

I mean, if, if you may, you may remember, uh, earlier this year and last year, I, I went on a little thing about how there's no such thing as an AI application. There's an application that uses ai. Even a chat is, you know, uh, you know, the, the fact that there's an AI powering the responses to you is, is it's a chat application.

Um, so Microsoft, uh, is doing its best to control the entire stack. Um, they did, they couldn't buy, uh, DeepMind like, uh, Google did. So instead they, um, funded OpenAI that got them started.

Um, they've been using the the GPT um, models, um, as they've come out. But Microsoft most famously owns the user end of all of this, um, more thoroughly than probably any other player in the game. And I'm including Oracle and Google in that, um, you know, Google has a full stack, lots of folks have a full stack.

But the point is, um, that, uh, Microsoft had to, uh, decouple itself, and this has been developing for a long time from open AI as its sole source. Uh, and so this is, uh, one example of it. Um, it's just chump change for them.

What is it, 15 bill or something like that, uh, with philanthropic, right? That's, uh, you know, they probably have, uh, more in Satya Nadel is a bank account to, to work with if they want to. Um, so I think just in general, this is where it's going.

Uh, someone asked me yesterday actually, uh, if, uh, I thought that, uh, we were heading towards some kind of, um, you know, sort of Microsoft, similar to Microsoft, some sort of dominant, you know, platform slash operating, the AI AI source or whatever, the world Computer. Yeah, Yeah, yeah. And, and I said, I, I just, I laughed.

Um, now that, that might indicate, um, especially your reaction just now, Mike, that I'm just completely wrong about this. But from my standpoint, this is worse than the cloud MCP from MCP to API to whatever. Like, um, the, the, when I say worse than the cloud, the, the cloud birthed the universal, you know, uh, use of APIs everywhere.

The so-called API economy, which meant that you could start doing mix and match best of breed stuff. Um, AI makes it even worse when you think about how MCP and about how agents work, where in a sense, any agent can kind of go anywhere and use anything and do whatever it wants to. And, uh, it can last for 20 minutes.

The agent can last for 20 minutes for you. So the idea of somehow getting a monopoly or a dominant model or producer of a model or whatever is absurd. And I think as usual, Satya is a little bit ahead of the game here and just say, screw this, we need to make sure, like tomorrow it might be somebody else.

Tomorrow it might be, um, cohere, it might be somebody brand new. So that's the significance of it as a strategic move. It's not about anthropic, it's about ensuring that you are delivering the right AI experience to your business and consumer users.

However you're going to do that. To that point that the guy was mentioning, I think there's an IDC report that says businesses are gonna deploy. 3 billion AI agents and automate workflows by 2028.

So in a sense, what Microsoft is doing is logical. This is what they do. You know, they, they cooperate, they work with companies until they compete with them.

So they're just hedging their bets across the board. Yeah, I think the word you're looking for is promiscuous, but Mike, They're that, and that's sort of Microsoft, isn't it? Pro promiscuous Is the right word, technically speaking, Mike.

But other reasons might not be the word you wanna use. And you might be right. I do wanna mention, I do wanna say one other thing.

'cause you talked about Microsoft building A GPU, Microsoft's not gonna build a GPU, the action is gone to XP as far as that goes. And I know that also sounds silly, given that 80%, 80% of processing spending is on GPUs right now. Um, granted, but the xus, meaning TPUs, tensor processing units, NPUs neural processing units, um, just a sort of a heterogeneous chips set of pro, not a chip set, a heterogeneous set of processing units, um, is rapidly going to develop in these systems, especially, and initially in the hyperscalers like Azure, that's Microsoft, literally has a, uh, I believe it's an NPU that they've developed themselves.

Maybe it's TPU, um, I should be better up on that. And that's really where, what, what they're gonna come out with that's well known. I don't disagree, but Terry, you know, so to John's point, thousands and maybe hundreds of thousands, even millions of AI agents and wow, all this talk couple of days long and, you know, this whole thing about like maybe how we're gonna orchestrate, govern and secure these things, footnotes if we're lucky.

What do you say? Yeah, that again, you know, it's like security gets the short shrift, I believe in this discussion. I, it's not saying that they, they don't have some plans, but I mean, they certainly didn't, uh, to my knowledge share anything that was, that was concrete about how they're gonna protect all of this stuff.

I mean, there are huge governance questions. There's, that's, and, and it just, it started, it's like mind, it's mind boggling, you know? Um, and I was thinking, didn't anthropic just put out a report on, uh, uh, a hack maybe, you know, and it's Chinese hackers are using Claude.

Yes, Yes. Yeah. Using Claude.

And, and so I just, I guess I don't understand. I mean, even why not, why they're not just at least paying some serious lip service to security when they come out with an announcement like this. And, you know, Terry, I, I, I, so I'm so glad you said that because to me there's like this classic lag between all these announcements.

They just bombard us with all these things they're gonna promise and then pie in the sky ideas. You know, they're, they're probably applicable, but there's, but it is never, we never hear from the customer or from anyone with a real concrete example because they're trying to digest all this, and they're trying to figure out how do we cover our asses? And, and when you press the companies on this topic, they, they, they switch the topic.

Yeah. And I mean, and then you have to pity the customers. I they are being responsible.

That's what I mean. Yeah. The customers, yeah.

And they're really having, uh, I think issues sorting through, but they're also having to make decisions and move on this stuff. They, uh, there's a sense that they don't wanna be like left behind. And yeah, you might be right, guy.

There's not gonna, you know, Microsoft is not gonna be able to dominate. Doesn't mean that they don't want to, you know, but it also doesn't mean that customers aren't gonna feel, you know, some sort of pressure to, you know, jump on board before the security stuff is sorted out. So why, Why would I possibly wanna bring up anything that might make you think twice about building and deploying an AI agent tomorrow morning?

Come on up. Just, just build it and it'll break and then Yeah. Build it And it'll don't worry about it.

And they'll come, well, You know, whatever. Honestly, Mike, when you told us that, that, that we would start off the show with, uh, this, uh, this announcement, which of course I saw, um, about, uh, Microsoft's investment with Anthropic, my initial thought was, you know, of course you're always thinking, what are you gonna say when Mike calls on you when you're asked about it? And what I was gonna, what my first thought was to say, what the hell is even going on anymore?

It's just, You just cannot. So, so, John, you know, to, to your point or my thoughts about what, what you, you're, you're, you know, what you said a minute ago is, I, I feel like a 10th of 1% of what is being done in AI right now is what's being talked about. 9% of what's being done in AI is a bunch of people like you, I like the folks on this call and other people we know messing around in, uh, chat GPT mostly, but also in other places, just doing stuff in chat.

And when we are hearing about how companies are doing this and organizations are doing that, and now you can do this. You can have this 12 stage, uh, super agent, blah, blah, blah, that's gonna search all your email and give you recommendations as to, you know, the nearest water closet for when you get out of the train in Shanghai or whatever. Right?

That, that is, that is really, that's, that's the cool sparkly bits that we do wanna talk about the possibilities and everything. But for those who are watching this, you know, uh, podcast are listening in, they're still at, at stage. Yeah.

Zero point. Yeah. Even if you press, If you press, for the most part, Yes.

If you press the companies who are announcing like Agent Force or what have you, I'm not gonna pick on Salesforce, but any of these companies, ServiceNow, you really press them and ask them for a customer. If you reach the customer, they give you the blandest, lowest level safest application that they're using it for. And it's usually some sort of customer service or, uh, HR internally, and it's very controlled, and it's usually overseen by humans.

So they're not practicing what they're preaching. Yeah. I mean, I'm with you guys.

I'm experimenting with this stuff on a regular basis, and I'm getting annoyed because all my experiments are kind of, eventually within about a day or so, I hit some sort of brick wall and I'm like, you know, this isn't quite gonna work, and I'm hoping that somebody will go build an agent that will do the thing that I want to do. But right now, trying to navigate my way through prompt engineering and context engineering to create some sort of automated workflow that, by the way, is not very repeatable because well, uh, uh, the LLM has no memory. So it's like basically, you know, talking to my, uh, father-in-law who doesn't remember what I told them yesterday.

So I, so, so mine being two takeaways here from this story are, uh, well, three thanks to you guys, but number one is that there is a distinction to be made between the, the application and the use of AI versus the AI model and training and stuff itself. And right now it feels a whole lot like, like, it's weird. Another weird thing to say, it's like a commodity market.

Like in perplexity, for example, you just go and you pick the model you wanna use and do that Firefox as well, whatever. Um, the second one is the really excellent point that we keep hurtling off into the stratosphere without, you know, uh, uh, keeping an eye or keeping our thoughts on the incredible, uh, uh, security, uh, implications. Um, and the third, and the third is, what is it with these keynotes now, John?

I mean, like three hours. They're just, they're so long. And then you got, you got one every day.

Also, by the way, like, it's no longer a keynote. One of the great things about the super commute con conference that I'm at right now, they do one keynote and it's got two speakers, and then they're done. So the, so that's, sorry, the keynote.

So sorry, uh, guy, the keynote yesterday reminded me of the Irishman movie. I almost, you know, like, I couldn't, I couldn't watch it, I know, decade Ago. And my wife and I stopped watching that after an hour that she was like, I can't do it anymore.

I was watch it for half an hour, then I leave the next day, watch half an hour. It took me like a week to watch it. I felt if only I could have done that yesterday, I would've felt as if I didn't l my soul was crushed.

It was, it was brutal. Sorry. Wait, sorry, Microsoft.

Wait, I don't even think that these things rise to the level of the word keynote, because they're basically 20 minutes of dialogue and two hours of videos. Yeah. Right.

And then you have the, you, the, the third party come on and, and bow to the master. They, they do their presentation, you know, the audience, they lose the audience by that point. It's like a series of eight or nine guests.

I mean, apple started this Mons monstrosity, um, but at least theirs was somewhat streamlined, although theirs is pretty tired at this point. Yeah. Well, John, you need, you need post apple traumatic syndrome therapy or something.

There you go. I would just love it if somebody somewhere, maybe stands up and gives a keynote for 45 minutes, no slides, no nothing, and just straight up, here's where we are and what's going on. And this is what it's all about.

But I would say Microsoft is not the only one of these vendors that has trouble, much like Bill Clinton parsing the word is, and a lot of things that they're talking about are gonna be true in about three years, but as of today, not so much. But anyway, that's just my take on it, and we'll see how it goes. But folks, we're gonna move on to our next topic, which has something to do with this one, but it's gonna be more of a Google take.

We'll be back in a minute. You've earned it. The spotlight, the responsibility, the weight of teams, companies, and entire industries fall on your shoulders, lives depend on your decisions, your home life included that work.

You are protected physically and digitally. Nothing gets through your team without a fight. But in a globally connected world, everyone sees you, including those who mean to cause you and your organization harm.

And now home your sanctuary attackers see an opportunity. Your digital front door is wide open. And what compromises your home can breach your boardroom.

Because the devil's greatest trick isn't targeting your workplace firewall. It's convincing you that your personal life isn't at risk. Black cloak, digital executive protection, defending the new attack surface your personal life.

All right, folks, we're back and we're talking about, well, Google announced Gemini three, which just happens to coincide with the Microsoft Ignite Conference. Hmm. Little suspicion that's about, and basically, yeah, and basically this too was telegraphed.

They said that this was coming forever in a day, and now they're finally announcing that they have this thing, and there's also some interesting vibe coding tools, and they are calling it the most powerful AI model yet. And John, as you roll your eyes, um, the question, the question I have you is, it seems like, correct me if I'm wrong, but like every other, what, 45 days, somebody stands up and says they have the most powerful AI model yet? Absolutely.

So look, think about the timeline. 5. Okay, let's see.

5. And so what they, the companies do is they've been to prove their, their superiority. They, they demonstrate, or they, they come out with these independent benchmarks.

4, which I, in context, I know, I have no idea what that means on humanity's last exam, uh, the previous record was GPT five pro. So, um, you know, it's, it's, it's all part of the, the hype. It's, it's all part of the one upsmanship, I think you said it well.

And, and on the show notes, it's L-L-L-L-M, leapfrog game. Um, and again, you know, in addition to this, I'm gonna point out that, um, the, uh, Google, CEO Sundar Paka gave an interview to BBC, I think a day before, and he kind of dropped this bomb about the AI bubble and how no one's immune, including Google. So he got attention for that.

Of course, he also said Google is better positioned than anyone, so they're gonna be fine. Um, so you're, you're right, it's just one ship. They all try to undercut one another.

Um, you know, so be it. I mean, there's, there's a lot at stake. And, um, you know, ab we'll see more.

Yeah, I believe, I believe he said there may be some irrational exuberance. I was like, maybe, Yes, as we see these, these investments that are just outrageous from company to company, you know, they're just switching alliances, switching partners. It's all, it's as, as guys said earlier, it's just like head spinning is, I can't keep track of this.

All I said was, that's probably as intended. I mean, everybody's trying to, you know, get a piece of this and, and declare superiority, and it, it, it's kind of, in my estimation, my professional opinion. Stupid.

So I, I'm sorry. 9% of my commentary. So, Well, But guy to, you may have some insights on the following.

So it's the most powerful, yet, according to a quote unquote series of benchmarks that somebody created and tracked that last time I checked, I've yet to run into a benchmark that actually reflected any real world activity. So what is the definition of, you know, most powerful? What, according to whom when?

Uh, So the benchmarks, um, are more or less, um, things like, you know, you know, tokens per second and, you know, query depth and length of, or, you know, size of context, window, and like that sort of thing. Um, actually that's not a last, that one, last one's more like a claimed feature than a benchmark. Um, all of your points are correct, uh, that, um, real world, um, real world experience with a new model, um, that scores really, really well.

Um, it does reflect that score. But one thing that is not commonly understood is, um, this sort of recursive mechanism by which, uh, well, I maybe it's more commonly understood than, than, you know, you realize, like, remember the application, um, uses the model, the model itself. Um, they, they started out as sort of what you call feed forward models, where, um, there's input and then output, um, and very quickly became recursive in how they operated, so that not only could they take larger and larger, larger and larger context windows, meaning they took a lot more input.

And by the way, a lot of that input, you don't even see you put in a prompt. And then whatever application or chat you're using, uh, add, it's not just rag so-called rag, like there could be all kinds of additional context added to help you with. That's one of the way perplexity works.

Um, most of 'em work this way to, to make the output look better. So the model doesn't see any of that. It doesn't know which is yours and which is supplied by the application.

Anyway, long, long way of saying that, um, there, there is a lot more, uh, complexity in how these models work than there used to be. And they, they, as, as indicated in Google's release, um, all of these benchmark scores and everything, including the intelligence score and everything else, are ways to, um, reflect that. It's not just simply doing a single pass, it's doing what you might think of as several passes of the same thing, breaking down the prompt, organizing it a little bit, running the different parts of the prompt through different stages, um, before providing an output even to the application, let alone to you.

Um, I actually, unlike what we were talking about earlier with Microsoft announcement, where I was going, what the heck is even going on here anymore, Mike, I got a, a, a, a funny feeling of deja vu that maybe you had as well, seeing this announcement, which is like, Hey, I know what this is. Like, this is like the old chip wars. Now we're gonna have all the models coming out and saying, you know, I've hit 61 giga flops, I've hit 61 to 62, uh, uh, gigabits, giga, whatevers like, and so on and so on.

And that just feels nicely familiar. We can maybe settle in a little bit to just watching these folks going and say, no, mine is bigger. No, no, no, no, no, mine is bigger.

No, mine is taller, mine is deeper. And that, that actually is a welcome development because if the different models are just starting to, uh, and the companies that produce 'em are starting to just compare themselves by size and speed and that sort of thing, then that helps us take them as a little bit more of a commodity play. I mean, which one really is better at coding?

Is it Gemini or is it clawed, or is it whatever? Who knows? We never get to use these models directly anyway.

We use the applications that use the models, and there's a lot in that experience that's due to the application. I have heard a new phrase, and the new phrase is called token fatigue. And it comes in two forms.

One is, I created something that was so complex that I ran it up against the LLM and it takes that LLM forever to do it. So then I gotta take my next thing and move it up to the next model. But by the time I run my thing up against the next model, it's so expensive, I can't afford to do it.

So people are like, I'm kind of, you know, lost coming and going here. But I think we're gonna get to a point soon. I hope maybe that, uh, I'm gonna create some sort of prompt that will have some form of automated context engineering, and then we'll be routed to the most efficient LLM that's fit for purpose.

And I won't have to think twice about it, because it'll figure that out for me. What do you say, guy? Can we get to that?

Um, there's a lot of fatigue building up in the system. Um, I think so. I think, I think, I think you're right.

I think that, um, I think that that, you know, we've talked a little bit here about how there's all this focus on the model, but what about the data? What about the quality of the data? It's not just a, it's not, you know, there, there's a lot that can, that, I'm talking about training data now to create the model.

I'm not talking about the data added to, you know, the inference in the model. Um, and then on the other side, I was just talking about the application. Well, listen to Gentech, AI is an application that uses ai.

And so there's a lot of work to do there. And what I have, I, I was just talking to one of the neo clouds here at the Super compute show the other day, and I told him that this moment, this AI moment we're in feels a little starting to feel a little weird to me, and that I don't know how to qualify it at any better than that. But I think talking about concepts like token fatigue, um, or recursive use of AI data to further train ai, these are all things that can build friction into the system and help, help generate the inevitable backlash that occurs in every wave.

AI is, it's just gonna have faster in ai, um, out of which we'll emerge with something better. But in the meantime, yeah, I think that that's the sort of thing that, uh, contributes to that. I'll go a step further.

I think that given the amount of money invested in all this stuff and how upside down that is that soon, that context routing engine that I'm gonna put in the middle of this is gonna be pinged by various LLMs that say, oh, pick me, pick me. I'll do it cheaper. Come on, pick me, pick me.

Wow. It's like a Kubernetes based structure for, for, for, for a, oh man, Uhhuh, just think about it. And, and, and I, and I, Now I have fatigue and I, and I only bring, and I only bring this up because there are indeed open source context routing projects underway that are gonna have these kinds of capabilities.

So stay tuned. You are so right, Mike. We will see.

All right, my friends, we'll be back in a minute with our third segment Discover Textron Group, the epicenter of tech innovation. We are your go-to for reaching IT leaders and practitioners worldwide. Our secret impactful content that sparks awareness, engagement, and top quality leads with us.

You'll access editorial websites, streaming videos, virtual events, custom content analyst research, and more. Join our satisfied clients. Let's revolutionize your tech journey.

Contact us today and tell your story to the world in the most powerful way with Textron Group. Hey folks, we're back and we're in for a little bit of a treat. Guy Courier is in St.

Louis at the SC Show, formerly known as Supercomputing, and he's been there for the last couple of days. And Guy, we talked a little bit about what Nvidia was up to at that show earlier this week, but there's a whole lot of other folks there. And, you know, walk us through some of your impressions.

'cause man, there's nothing like having feet on the ground. Uh, yeah, thanks. It's, uh, this is my favorite show, um, because its Roots are scientific and, uh, research based.

Um, and it was all about high performance computing, uh, year after year after year. Um, starting, uh, had AI related to it for a very long time, but that wouldn't have been generative ai that would've been predictive or, or recommendation engines. Um, I'll get to that in a second.

Um, but now it is, uh, a very large also commercial trade show along the lines of like a Q con or whatever, where there's, you know, uh, pursuit of development and inquiry and all that. Um, in parallel with, uh, technology and announcements and all those other sort of things. It's also one of my favorite shows 'cause it only has one keynote.

Just wanna repeat that. Um, last year, um, I reported also from, from, uh, on this, uh, program from, um, super Compute. And the big story was about cooling.

Cooling was, it was sort of, you know, AI of course was a big story last year. Again, I'll get to that again here. Um, but, uh, uh, as far as the sorts of things that are really pervasive, uh, outside of AI cooling was the big story.

Um, now what we're seeing is what you might call, um, a rapid, uh, um, a rapid emergence into the market of those cooling technologies. Uh, you, you had advanced, uh, di direct liquid cooling systems showcased from, from some of the system vendors like, uh, HPE, Dell Super Micro, um, Dell and Super Micro actually have open rack conforming, uh, or compliant, um, systems on display. Um, but they are, they weren't the only ones from smaller vendors as well.

Um, I didn't check out if Lenovo had one, but the, the now we're seeing, um, real, um, and the data center market incidentally, uh, most data center providers are, are, are now enabling direct liquid cooling a lot more easily. So that's hitting ahead now that we saw emerging last year. Um, one thing that seems relatively new this year for this show, um, is, uh, the neo clouds are here.

Um, core weave is here, neas is here. There's probably a few more that I didn't see. Um, neo clouds in case requires explanation.

Um, are, um, cloud providers, they tend to be global, but they are built recently from the ground up, you know, so to speak. Get to that in a second. But they're built from the ground up to help host AI in particular.

And the reason why I say so to speak is that a lot of them started out as crypto mining farms, um, that got repurposed, uh, to, to help host, um, ai. Um, so the question is why are neo clouds here at this show where you have a lot of tinkerers and builders and that sort of thing, people building high performance, uh, systems, uh, uh, mostly for research purposes, but also now for ai. So why would, um, a provider be here?

Um, it's because these folks are also their customers. Um, which leads me to the third thing. I'm sort of, I've sort of picked up here at Super Compute, which, um, uh, what's emerging now.

So I said like cloud is, uh, sorry, cooling was sort of emerging. Obviously cooling and liquid cooling's been around for a long time, but it was sort of emerging as this major force a year ago. Um, one of the big things emerging this year is, um, integration of distinct AI systems.

Now, I mentioned that AI has been around at HPC for quite a while. Um, mostly for in, in in predictive or recommendation engines for things like, uh, AI based job configuration and scheduling. In other words, when you're running a high performance computing job to try and simulate, I don't know, um, hurricanes on Mars or whatever they're doing to help, you know, uh, cure the world's ills.

Um, when you're doing that, the faster you can run your jobs, the faster you can. So you run a job in a high performance computing environment and then maybe you need to run it again. You need to tinker with something and run it again in order to tune it and then get the kinda research you want.

The faster you can do that, the better. The less human intervention you can have, the better. So that was a natural place for AI as a sort of a controlling system.

But now what they wanna do is they want to use generative AI in various forms in order to, um, help boost simulation effectiveness, invent or create jobs or test schemes, all kinds of analytic related stuff that can work in conjunction with HPC. So that is real buzz at this show. So on the main stage, just lots of sessions about how to integrate, build and integrate distinct ai.

So there you're thinking about the neo clouds, again, that that's a possible source for that sort of thing. That leads you to another theme, which is smaller, but like one flash session here from WWT was very well attended on workflow, workflow in HPC workflow out of AI into HPC and back, which leads me to the last bit of emerging, um, uh, uh, emerging trends. Um, which is, uh, has to do with quantum, quantum computing really looks like it's on the verge of becoming real in a number of respects.

They have, um, 150 plus qubit systems, uh, going into production right around now. 150 is a pretty considerable number, but to, to really make it work, um, you need all kinds of other systems, which I won't go into here, mostly the most of 'em have to do with error correction that leads you to ultimately a 150 qubit system, giving you five, um, five effecting, you'd even call it sort of virtual, actual reliable cubits. Five cubits is actually a lot, doesn't sound like a lot, it's a lot the way quantum computing works.

What you're doing is running sometimes multiple problems that would take months on even modern hyper computing, uh, uh, uh, um, high performance computing systems to run. You can run them in a matter of seconds. Um, it's not, it, it becomes a part of an overall high performance computing system.

And so the incorporation of quantum into HPC is a hot architectural topic here right now because these systems are starting to really go into production. So guy, are people talking about making these platforms more efficient? And I'm asking the question 'cause it seems like the number of research projects that wanna be run on a supercomputer and eventually a quantum computer far exceed the amount of available capacity and, you know, for lack of a technical term, but the conversation seems to always be about, Hey, stop Bogart in that supercomputer and let me get my project on there.

Yes. Um, that is a general theme that as like you said, it's been a relatively perennial theme at this show. 'cause there always seems to be more, uh, there's always seem to be war work, uh, demanded of these systems that are available.

Um, but, um, one of the, so the, the, um, at the keynote, um, uh, there was one of these sort of vision talks was the main part of the keynote. Um, and, um, the speaker, um, it pointed out the way that the ways that AI can ultimately help with just what you're talking about. Um, the analogy made was to, so bear with me just a second, was to electric vehicles that the initial, the current model for, uh, like personal transport with vehicles is you have your car, you own your car, and rideshare services are a way to make that a little bit more efficient.

But you still wanna own your car, but when it moves ultimately to just being a service and you can use any car, then you need fewer cars. Even as the amount of energy use goes up, the actual number of units needed goes down. Um, and so by, by analogy, Mike, the idea is that, uh, these systems can be built and connected together and used a lot more efficiently because they could be shared in a lot more efficiently.

I thanks to ai, there's, there's a lot of work to do there. Um, it, one of the sources of this workflow discussion though, has to do with just what you're talking about, Mike, is it's not so much about idle time. It's the fact that instead of it just being one of the top 500 supercomputers that you can use, maybe there are ways that you can use one of the 5,000 or so supercomputers available right now around the world to do your work.

Mm-hmm. I'm not sure I can get that metaphor because when I drive, I notice that there's more Uber and Lyft cars out there than ever. And I've concluded that basically people who previously would not have gone anywhere are now getting in cars going somewhere.

And so now there's more cars on the road than less. That's the initial stage of it. It's when it's when people start giving up, um, uh, their personal cars.

And so the analogy here is when institutions start giving up their personal systems, um, not that they wouldn't build the personal systems anymore, but it's almost the distinction we were talking about earlier between the service or not, sorry, the, the, um, the, let's call it the backend and the business end or something like that, of this overall industry and market. And when you are building, uh, in this case, the instead of foundational AI models, what you're building are foundational supercomputer systems for your own use for things, but also you're allowing other people in to use them. Or in fact, you're allowing AI agents to schedule secure jobs in those.

So Like A GPU share. Yeah, it's a, it's an Uber for super compute there that nice. I just coined a new wonderful phrase for someone, right?

It's Uber for super Last question and the Next keynote you go to, Right? And, and you were in CubeCon last week, but we didn't get a chance to talk. But one of the things we were talking about in re relates to AI at CubeCon, and I'm curious if it plays out at sc, is that, you know, there's folks who are saying you should run those AI workloads over on Kubernetes clusters because, well, the IT folks know how to manage those.

And yet when I go talk to the AI community, they have a thing called slum and that's their favorite job schedule. And they're like, you know what? We don't need no stinking Kubernetes 'cause it's too hard and we don't understand it.

S SLM is here in effect. S SLM is dominant, you're right, it's dominant for workflow. Um, it's dominant for job scheduling.

It's dominant for all of this kind of, uh, sort of application, uh, orchestration in HPC and in ai. Um, but you know, mostly ai. Um, there's only one neo cloud I know of.

Um, and I mentioned the neo clouds because, you know, they are at the forefront of, of these kind of buildouts right now. There's only one neo cloud I know of that uses Kubernetes and that's neas. Um, so it's not, um, it's not, uh, particularly common.

Um, there isn't really a lot of, um, support behind that sort of thing, but you have to remember like they're using storage systems that are really, you know, um, probably showing their age like luster, um, that's really still dominant. Um, there's lots of storage, you know, related stuff around here, um, that is, uh, more advanced. Um, so I think that moment just has not come Mike.

Um, and I don't know if I'm, I'm kind of, you know, I'm a little skeptical about Kubernetes as the ring to rule them all, but maybe that's just, you know, overall hostility to anybody saying anything is the ring to rule them all. Um, Kubernetes is definitely coming for HPC and ai, though there's no question. It's just whether or not it's gonna have, uh, the, the world, um, you know, the world dominant effect that seems to have everywhere else.

Alright, well folks, you heard it here. The one thing for sure about that show that guy goes through every year, it's kind of where the cutting edge of research and AI and physics and science all comes together. So if you do get a chance to go, you should absolutely take advantage of it.

Hey folks, thanks for sharing your thoughts and insights as usual, and thank you all for watching the latest episode of The Gang. Please stay tuned for the rest of the lineup of Techstrong tv 'cause it too will be equally awesome. And we'll see you again tomorrow.