Agents, Exits and Ecosystems: Cloud Shifts and Storage Surges This Week – Infrastructure Matters EP79

This is Textron tv. All right. You are listening to watching have you choose to consume it.

I'd like to know if anyone's actually running this through the new Nova, API and getting the transcripts and just reading what we say every week. Uh, you're watching, uh, infrastructure Matters with by co-host Diane and Kimberly. Diane is becoming a habit.

We have to know where in the world is Diane, are you freezing north? Are you, uh, in the nice, sunny south? Well, I'm in the, on the coast Adriatic in Albania, where it is a very sunny day.

It's gorgeous. Spring is hitting here, so it's, uh, very pleasant. So, So thanks.

You look out my window. You can see we got snow here in Colorado. No worries.

That's awesome. Well, it's Colorado. It didn't there snow up until like July?

That's, Well, Sunday we were wearing shorts, You know, uh, the, uh, you know what, we're, we're in Chicago since we're giving weather reports. And, uh, we don't bust out the shorts until like June. It's, it's, it's safe to say that summer is going, coming one day.

All right. We have a busy week. We missed last week because one of us did not communicate well.

I'm not going to name names. Um, see, Kimberly, you're on my screen. You're pointing towards like, I'm in the middle on my screen.

So I don't know which one of us did not communicate well last week, but No, it wasn't communication. It Well, when you say communication, his, his communication, the internet was really As well. Yeah, exactly.

Oh, no. So it wasn't Diane, it was me. I was, last week was anniversary.

Oh, That's about you. No, it was, wasn't Yeah, About probably under the bus. Yeah, you, you can actually, you gotta waste my free Keith.

Oh, okay. Diane. Diane's internet connections.

You have to get a, uh, we need alternative the Sky link, and maybe we'll get into the blue Origins, uh, mess that, uh, we're not gonna get into. But let's start out open AI surprise drop. I don't think I was expecting it, but, uh, uh, oh three and oh four.

Many are out, Diane. Yes. Um, so these, uh, oh three in particular is, uh, a real flagship, uh, model.

Um, and it can, it is supposedly better than a PhD. Uh, uh, that person with a PhD in their field on any given subject, um, we haven't seen yet. It's not, it's so new, it hasn't hit the leaderboard.

So we don't know, uh, how well it's doing there yet. But all the testing from the AI experts say this is one to really, it's tuned up. It is tuned up to pass a lot of benchmarks, but it actually can actually perform really well.

Um, and it's a, it's one of the, the new fashion now in generative AI is reasoning models, right? So you'll get that whole output about all the things that it looked and evaluated and checked. Um, and so it's, it's, you know, not just a transformer model.

And so the, uh, it's also been optimized for agent use. So internally, it, it can, has full, full versatility across all the multimodal capabilities, uh, and has all the hooks to do advanced agent agent ai. So this is a, a big move.

We're still waiting for, uh, GPT five, but this is a major, uh, stepping stone on that way and should keep it when we see the, the, I'm hoping we'll see that by the weekend. We'll see how they're doing on the leaderboards, but it should put open AI back on the top Geminis up there right now on most of the benchmarks. Yeah.

So ironically, I was using, or coincidentally I was using oh four right before we started doing this oh four mini. I've been obsessed with, uh, trying to create a GPT that tweets like Keith. And I'm trying to, uh, compare the, I've been using it for about six months.

I'm trying to compare the stats of, or the effectiveness of AI to capture Keith's voice by, uh, via engagement. And oh four has been, uh, a pretty interesting approach to comparing the, the data sets. Well, that you had, you were attempting to have it write your tweets for you, and you found that was definitely not as good as you doing it yourself.

It is definitely, uh, it is there, there's like a 33% decrease in engagement. Um, there's other factors. It could be, you know, uh, traffic to Twitter.

I don't have really great control or ab uh, control set, but it, the moly, the first six months of the, uh, of the dataset versus the last six months, which I've been using the, the GPT with not, not, not, uh, not anything to write home to for as far as results. And that's interesting because you're training it on your data, your tweets. Yep.

That, that kind of stuff. There certainly not enough, which has, to me, has an implication of how much training has to happen to get solid delivery of, of, of an application. So there's a debate between how much training that has to happen.

'cause I gave it 50,000 tweets, like more than 50,000 tweets. That's a lot of, that's a lot of data to, uh, synthesize and be like, Keith. But what I'm, uh, the bias is maybe not in the data, but in the model on discovery.

So as I dig deeper, I'm discovering that it pulls, I give it an original tweet, and it's supposed to improve it like an editor. Well, its improvement is to take away my voice. So it's doing the opposite of what I asked it to do.

So a lot of what my audience, uh, and some of this is me inferring and assuming, uh, what my audience that, uh, what they tell me what they like is my auten authentic voice. Yeah. And my rawness and, uh, chat.

GPT cleans some of that up. So instead of the, you know, the, the unpolished keef, which is one of the things I put on display, it polishes keef. And I don't think that's what my audience reacts to or even wants.

And it's proved out by the data. So the giving AI data, so the, there, there's this thing with the model. So the, the, there's some basic questions that the enterprise can take from this.

Are they using the right model? Uh, so this just may be a instance where GPT chat, GPT in this case, four oh, is not the right model for what I'm trying to, uh, achieve. Yeah, that's really interesting because you think about, you know, from a marketing standpoint, how important branding is and who the company is, how you respond, and how if we're gonna use, you know, just assistance to respond to client, you know, on a website or whatever.

And yeah, maybe it's just a very basic kind of thing, but I mean, the amount of uptick that we're using in terms of using that chat bot that has your brand, and I don't know if I have seen an awful lot of discussion around how that affects your overall company and the brand. I mean, I just, I think about how important that now becomes, um, as companies move forward with using these kind of, uh, gentech AI kind of stuff to, to, to bring data to their clients. Yeah.

And I think one of the big lessons from this is observability. Uh, as these models shift, as they change as they are refined, being able to make sure that the results that you are getting, you know, three months into the project are the same results you are getting six months into the project. It is, uh, uh, as I talked to folk, folks who are actually trying to implement AI as one of the biggest challenges as they use third party agents, not, you know, agents behind their four walls, that controlling, uh, for consistency is a really good challenge.

Well, especially as models change so often, uh, and it's not clear how long that that legacy models will be kept around. So what's the end of life and AI model seems like it's gonna be a lot shorter. How do you create predictable, uh, enterprise solutions that run for years?

In some cases, some apps have to run for decades if they're in healthcare or, or other critical, uh, infrastructure that it gets, has to be certified and very hard to change afterwards, right? So it's gonna be interesting. Yeah.

Yeah. Diane, I think, uh, it sounds like we've heard this story before at some point, right? Yeah.

Uh, the, some, some third party cloud-like service moves faster than what your enterprise is capable of operating that. But we're finding that to be consistent with these third party, uh, solutions that are moving at quite the clip. So I'm going to jump to a, a, a, a relevant topic.

I was at Google Cloud last week, so as you you're watching this, uh, you're, you're probably watching me and Kimberly at, uh, AI Infrastructure Day, uh, two, which is insane. It's four days of, uh, tech field day. But the other week from when you're watching this, I was at Google Cloud next, and one of the big announcements from Google Cloud next, at least the biggest for me, because I'm a distributed or hybrid cloud person, they are the first of the major hypervisors to announce, I mean, sorry, hyperscalers, to announce that they're putting their entire model in your data center so you can get their distributed cloud service in two flavors.

You can get it where you can install the control plane on your own GPUs, and Google deploys Gemini on that. Or you can get a completely air gap solution in which you get access to, uh, the, the raw Gemini flash. And Google Mon, uh, manages it, but without any connection back to the public cloud.

And so I thought it was one of the biggest announcements of the week. Uh, and the infrastructure, uh, uh, community has been a big topic, but not in the AI infrastructure, or AI community in general, has not kinda looked at it like, oh, that's interesting, but it's a, it's a big deal in my opinion. Yeah.

So they're shipping the hardware and all the software to drop on there, and either you can run it without being connected or disconnected. What's the price tag? So that was obviously one of the first questions that I asked.

They did not, they weren't prepared to share the price tag, but they did say it is a all you can eat solution. So it's not measured based on your consumption. You, you're basically renting the infrastructure at a fixed, uh, rate.

And, uh, unlike some of the other solutions, you're not charged for the number of VMs you run, the number of, of, uh, tokens you consume. It's, it's, you know, have at it. Wow.

Yeah. Arguably it's not very elastic. But on the other hand, you have something that's predictable and controllable and, and, uh, you know, what you get, this is not glamorous, but these are the things that we need to see in Gen AI to really make them successful in the enterprise long term.

So I think it's good to see. All right. So we have some EU news.

Diane, what's going, coming on out of the eu? Well, and maybe you saw some of were cube con, um, this year, Keith, I was not at Q Kon. I missed out this year.

I did not get to go to London. Yeah. So the, uh, apparently this is where the, the, there's people started to notice that the, uh, the second tier cloud providers in Europe are getting a big uptick in local inquiries and local and closing, uh, uh, deals on workloads because of the concerns about, uh, hosting data, running workloads in the United States, or even in EU regions.

But because of the Patriot Act, that that information could theoretically still be clawed back, or at least you'd have to fight the US government for years in court. Uh, and so there's a, they, uh, vendors that cloud next are reporting significant boosts in new deals as a result. So we're seeing this potentially a bow wave of European customers, um, uh, planning to migrate, uh, uh, away from the US hyperscalers.

So that will have interesting effects, uh, in the market, and it's gonna be something to watch. And I think it's gonna be part of the, the whole geopolitical rebalancing that we're gonna see with, um, you know, the new incoming administration and the new complexion of the US federal government. So that kind of goes along with, um, the commentary that we had about, um, SaaS and their earnings, and that they had septic of, um, sales in the eu, um, versus, uh, you know, that that's the last quarter because people are, you know, making, we're making decisions about this one or this one.

So the, the entire geopolitical kind of stuff probably tips some of those decisions over to SaaS, um, and being present over there. The question I have, and I, and I don't know if anybody, either of us, any of us know this, but the Patriot Act implications, if I have an AWS region in Germany or something like that, the Patriot Act is still going to apply to the data that's on that AWS facility. So that's what we say.

Now, the German government will not agree. However, um, the US government can make it very painful for that company not to, uh, to, to fail to comply. So if they try and hide under, you know, behind German law, they can say, well, um, there's always things that, that we can do and seize from you until you do hand over that data.

Uh, and it, and it has happened a couple times, so, so we know that this is, uh, this is potentially an issue. So, um, yeah, the, the, the legalities are not well documented in case law, but there are some precedents and they're not pleasant for people who are hoping to have the, the protection of their local privacy laws. So it's interesting.

Yeah, I was just, I was talking to a Equinix, um, a Equinix architect just the other day, and he was saying how Canadian companies are really concerned about data flow and the ability to keep data within the Canadian boat borders to the point that if there's a outage, one of the big selling points of Equinix is that you can just choose to fail instead of failing over to a route that takes you through like Detroit and, uh, gives your traffic subject to inspection by the US government. So, uh, this is something that we're starting to see seriously on the ground, people engineering around, uh, the geopolitical, uh, aspects of technology. And it's ultimately going to impact availability of applications for end users.

All right. Moving on to storage news. Kimberly, what's going on in the storage rule world?

A lot of, uh, there seems to be quite a bit. Well, there has been quite a bit, and we haven't really talked about it on, so some of the stuff is a little bit old news, because Dell was earlier this month. It wasn't this week and that kind of thing.

But, um, Dell did some big updates on their key technologies for storage, um, heavily focused on what is going on in the, um, AI space. So, um, it really was across all the lines. Some of it was, was strictly a controller server upgrade that gave them, you know, double the power, double the speed, double the iOS or whatever, pick U number.

Um, but some of them were significant, and one of them was object scale, which is their object storage piece of it. And there's this debate that's going on. Will high speed object be the data, you know, platform for ai?

You know, there's a big question on that. I still don't think so. I think it's gonna be file, but you know, that, that, that's my, my prediction on that.

But, but what they did is Mean IO three, uh, AB at man io is going to get disagree into argument with you. He's gonna, I rather, he's gonna politely disagree with you, Kimberly Definitely will, and so will some others. Um, and, and it's actually, so will possibly Dell with that.

What the release is an all solid state system that is, you know, very fast. Um, the benchmarks, which I haven't, we haven't completely validated or anything like that look very impressive on there. Um, they've also done a big upgrade to their power scale, which is the also known as, uh, formerly known as ison.

Um, and that also got the upgrades. But you know, in, in one of our favorite people, this, the solid dime stuff that you do a lot of work with, and it released 122 terabyte drives. So added to you, you can get up to six petabytes now.

Wow. Who would've thought it was like amazing numbers. Um, the other big one that was this week is Hammer Space, um, little company, um, been around for quite some time.

Um, they, they've really taken off now with this AI piece of it. Um, they are a, um, PNFS parallel file system, um, also a data management system. They can manage other people's storage underneath it.

So it's kind of a virtualization play out of band. Um, they've had, and then they have this capability for doing a tier zero, which speeds up their ability to, and we've got a paper coming out on that, um, speeds up their ability to feed the GPUs. So they've had some big wins and some very visible wins, um, especially in the hyperscalers.

People like Facebook, um, I believe they're also probably done in Tesla. Um, and they're sitting, what, what in those companies, they're sitting on top of somebody else's file system. And what they have done is gone in there where certain companies, they hadn't had the speed that they needed or expected to get from a file system and Hammer space came in.

And because their separation of metadata and everything else has given the customers the need that what they need. And this is, and I I, you know, I can't talk about the companies and the names of the companies because I don't think it's all public knowledge, but, you know, pick, pick a vendor, pick a leading vendor out there that's a file system vendor, and they probably have at least one installation, which they've gone in and put their capabilities on top of that in order to get the performance. So they've had some really big doubling of numbers.

Um, and so it's a hundred million dollars big, not massive. I don't even know what the, the valuation of the firm is. I'm pretty, I'm pretty sure it's high if it's not a billion.

But otherwise, I'm sure they would've said something about it. Um, but it does, one of the things that they can do, and this gets into the argument, is do we want one big data, manage data, you know, um, one big system, like a vast system or a NetApp system to manage all of my data, or do I do something that gives me some level of virtualization to pull all the data in into a global name space and manage that and put it out there and, you know, avoid all the data movement, et cetera, that you would be faced with potentially. So that's, uh, I, I think they've got legs underneath them.

Um, and, uh, so kind of watch this space is the next big one. Yeah. Hammer Space has been at it for a while, and this is a big debate with enterprise users, right?

Do they go and put their purpose, purpose needed, uh, data sets onto something like vast or a fast system, or do they, to your point, abstract that away and use something like hammer space to, I I, I don't know the performance difference between a hammer space and a vast data system from a, uh, raw speeds and feeds. It, it could be, uh, negligible or it could be a big difference, I don't know. But I think the debate is around your enterprise standards and what is the standard way you're going to deliver AI data or ai uh, intended data to your systems.

This was a big conversation at Google Cloud next, which was how is the platform team going to provide AI capabilities to developers? And I think this is all related. Yeah.

All right. What else do you have for us? The last interesting thing that popped up was this announcement with Microsoft Western Digital and a couple small firms, is this Iowa firm out, um, called Critical Materials recycling, in which they have been able to show at scale recycling of rare Earth materials out of HDDs.

Oh, That's quite timely. Hmm. At scale.

So this is not a little bit of stuff, but they're saying at scale. So this is a real, real honest to goodness thing, and this is being done, and I was like, clicking through this and looking at who this company was. They're probably overseas.

We're probably shipping our drives overseas, and they're doing there because nobody's gonna allow them to do that here. But yes, it is in Iowa. Um, so they're doing it here on, you know, in, in the United States, rare Earth, and, um, then recycling and then sending that off to the people that are gonna manufacture everything from, well, That's very germane, given that China has just cut us off from rare Earth materials.

So, yes. Yeah. So, um, it's, and, and it's interesting technology, you know, I don't understand it how it's all done, but, um, so, you know, one, they ship the drives to somebody that takes all the parts and pieces out.

Then they ship all those parts and pieces to, um, the critical materials recycling, and then they pull out everything from gold to whatever, all the other names that I can't even pronounce, um, that they're pulling out. So I now have something to do with all of these 300 gigabyte drives that I didn't know what, how to get rid of, like, what do I do for 300 gigabyte drive? I can get that in a thumb drive now in a flash thumb drive.

Maybe not as reliable, but I I can get one. Yeah. So related to storage is, uh, Amazon snuck one in on us from at least, you know, the analyst community.

This is the type of thing they would typically send us a note on, but I had to, uh, in testing a, I'm, I'm building a AI powered RSS reader because regular, who needs a regular RSS reader when it can be AI powered. Exactly. Uh, the, and while I was testing this, it, this came along my feed S3 Express one zone is now 31% cheaper, 55% less puts, I mean, cheaper to put to put data, and 85% cheaper to get data to, for those, the charge for those re uh, requests.

So overall, 60% cheaper than transfer data. Why is this important? What is S3 zones?

It is, uh, think about it as the, the, I'm sorry, S3 express one zone. Why is this important? It is the fast version of S3 at AWS reinvent.

I think it was this past year or the year before, they announced S3 Express one Zone, and it was intended for real time streaming applications. It was AWS kind of big AI storage related announcement. Get the object storage that you need, can really, your debate on, you know, whether or not object storage or file would be the asset method.

Well, AWS is trying to take away all excuses for customers not to use S3. Yeah. The problem with the S3 express zone is that it's not this highly available multi-zone environment, so that, that's why they get, they, they can get some of the speed.

That's part of the reason why they're getting some of the speed. Yeah. And if you're architect, they, uh, promise high availability within the zone, but the multi-zone, um, the multi-zone, not multi-region, but multi-zone zone redundancy of S3 was a big deal.

It, it was this thing that you didn't have to think about. And the availability for S3, regular S expert three is insane. I think maybe you get one or two less nines with, uh, S3 zones, and you can always, of course, replicate your data at, at AWS is more than happy to have you replicate your, uh, zone data from S3 express zone to regular S3.

So your, um, belt and suspender strategy. Yes, exactly. All right.

So going back to, uh, Google Cloud X really big announcements, we'll try and get through. Uh, they spent a lot of time talking about agent space. Uh, think of this, I had a really great conversation with hugging face yesterday, and it kind of clicked.

Uh, Google wants to be the hugging face of agents is probably my shortcut of thinking about that. I'm, I'm bringing that up because Google spent a lot of time at the keynote and the analyst sessions talking about agent space and the things that you can do with it. The second announcement that I want to make sure that I get into was this, uh, opportunity for developers to basically have a virtual workspace environment, uh, on S3, I mean, I'm sorry, not on S3 on Google Cloud called Firebase.

They made a really big deal of, uh, how much free capability you were give given away for it. I'm a cursor AI vs code user, so I'm say, you know, I'll give it a try. Uh, I don't know, maybe it's the demand from the, uh, attention, but it wasn't very available to me.

So maybe check it out now. Might be a, a little bit more reliable than it was when I was planning around with it. This the week of this recording.

But the big announcement, uh, the other big announcement was agent to agent. Diane, you wanna take that one for us? Yeah.

So it was a big announcement. So Google put up a, uh, very impressive, uh, logo slide, and they've got many of the big tech companies. So, you know, the SAPs, uh, and Salesforce, they had the big management consultancies, McKinsey, PWC, and Wipro were all at the slides.

I mean, they have a lot of people on board. Uh, and, and the announcement was agent to agent, and it's a protocol to allow agents to work together. AI agents is what we're talking about.

Uh, we have a big new report that, uh, we just issued the public version that looks at all the different agent based platforms because this is how it operations and business operations are gonna be running in the future, is agents are gonna be doing a lot of the basic work and humans will be pulled in the loop when the agents, uh, can't do it. Uh, but otherwise, AI will be taking action for us. But multi-agent are really critical because if, let's say an agent that determines they can't actually do it, instead of giving up, they can go find an another agent to do it and hand the work off.

An agent to agent, um, is built on existing protocols. It's designed to be highly secure across agent collaboration. Um, and it's designed for long running agents.

So you can have this operation that can run for hours, days, maybe even weeks, that are all covered by this protocol. And this allows AI to actually go off and do things with us and agents to collaborate, um, and find each other, collaborate with each other, and do it all securely. So, big news, um, lots of, uh, information out there about it.

We're all trying to assess on how capable it really is. But, uh, absent is open AI and Anthropic and a bunch of others. So we'll, we'll see how, how much lift this actually gets.

Uh, but this is gonna be one big way that a part, a major part of the industry is gonna go down a big road they're gonna go down for, for making agents work together and, and, and making orchestration happen. So, big news. So Diane, maybe you, we could talk about, you know, to bring this to life a little bit or maybe the three of us can talk about what does agent to agent kind of applications look like?

You know, what are we, what do we, when you say it could be running for hour, days, or hours or whatever to have this analysis going on, um, you know, kind of, there's some examples that may, may bring it to life. Yeah, sure. So, um, you know, an example would be, uh, you go to an agent and you say, uh, I'd I'd like you to, to, uh, schedule this, uh, plan this trip for me, uh, schedule this whole trip for me, do all the hotel bookings, the, the, the airfare bookings, um, uh, plan my itinerary, um, schedule all my meetings for me.

So, uh, you know, you look at my calendar for that week and find out all those people, um, and make sure that I've got, you know, transportation to those places. Just handle the whole thing. I want be, have a concierge.

And it might turn out that, uh, it's in another country where Uber doesn't work. And so that agent only knows how to work with American, uh, uh, car rental services. So it could go and schedule them and maybe even actually go in and, and invoke the actual, uh, that morning of that meeting.

Actually talk to the agent, uh, AI agent for that, uh, that, uh, uh, ride sharing company and actually get that car for you without you having to do anything and just send a notification saying, I've already ordered your car. But it might have to, instead of using its own Uber, um, uh, you know, a reservation agent, it'll actually use another one. It'll go and say, oh, I, I need to find another one and talk to that one.

And it may have never viewed that before. And so it's the new interoperability where we all, we all wanted our, our software components and our, our systems to talk to each other. And we create open a, uh, APIs to do that.

This is the agent version of that. So it's all those same scenarios really, but just for AI agents that actually do things for us. And this is not to be confused with the MCP protocols and service Yeah.

That we're starting see, uh, going around. And the difference between M-C-P-M-C-P is enables you to build a interface into APIs. And so the whole point of agents has been to, uh, to reduce the complexity of interoperability when there aren't APIs available.

The obvious, I think gotcha to this or concern is human in the loop. I, the, you know, I, I gave this simple example of the chat application. You would think that you would, you know, you combine the two ideas that I'm working on, a AI RS reader, RSS reader that picks up news that Keith's audience would be, uh, interested in and then automatically creates a social PO post based in Keith's voice, is the ultimate example.

But the gatcha is that AI is probably the smartest and dumbest junior admin I've ever had. Case in point, it recommended that I retweet about, uh, having a mobile, not missing my whiteboard when I'm mobile and talk about, and it assumed that I wanted to talk about remote re remote work. In reality, I was, it was a picture of me, uh, in my Airstream saying, I missed my whiteboard 'cause I'm inside my Airstream.

So the, it, it remove context agent to agent, amazing capability. But from a practical perspective, as you're thinking about managing systems and reducing or shifting the complexity of managing systems, agents still need to be monitored by humans and that agent to agent communication, we have to figure that out of how to monitor that, that in interaction by, uh, with you with some type of human validation. So it's like having two junior people out there, that's your manage, right?

That, that you're, it's saying, okay, so you're an expert here. You're an expert here. You guys get together and work on this.

Um, so, okay. And it, and it's weird because it's a, it's a junior admin with photographic memory. Yes.

It's, it's very confusing. Oh, you mean really super smart kids. Yes.

Yes. It's like a really super smart kid and you're like, you know, a kid, the, the, the, your problem is that you don't have 10 years of experience. Like, and they just don't understand.

What do you mean? What do you mean? I can, I can regurgitate all this information.

It's like I was book information dudes. All right. So as, uh, you continue to get off our alarms, the go back, uh, read the future, uh, uh, researcher coming out of Google Cloud next, uh, my peers, uh, did a really great job of summarizing that news.

It is super relevant. It'll give you more homework than, than you would like. Uh, please feel free to ask for an inquiry.

But the, the free thing you can do outside of that is share this podcast with your peers, not just your peers. Uh, spring, uh, we're outta spring break season. The kids are gonna need something to do, uh, over the summer.

Send 'em each week of the podcast every week. They'll make sure they'll get out the house. Talk to you folks next week.

Bye team. Take care, everyone.