Enterprise Platform Choices: VMware, Codex and AI Model Proof

Hey everyone. Happy Thursday and welcome to Techstrong Gang. Well, we're still riding high a little bit.

Happy to be back at the desk here. The news doesn't stop. What's interesting is, now we maybe can look at some of the news through the eyes of the election results, and you start seeing how this is starting to influence maybe some of the positions of the companies we talk about here every day.

But I don't think we've seen the full repercussions of that yet. I think we'll start seeing it more next week. This week, though, and today we have a lot to talk about.

Some good stuff, and we've got good people to talk about it with. Let me introduce you to our gang members, some of our returning regulars. Jamie Hampton.

Jamie, it's great to see you. Hi. Good to see you.

Jamie, that's not a data center out your window, though, is it? It kind of looks like it, doesn't it? Yeah.

It's that data center kind of- I live in downtown St. Petersburg, Florida, so there's a lot of apartment buildings and things, but- Right ... right behind that is the Tropicana Field, where they- Oh, yeah.

The Rays. Yeah. Downtown St.

Petersburg is happening, too. Oh, yeah. It's become a very chi-chi kind of place.

But welcome. Joining Jamie, I haven't had the pleasure of having him on with me because he's usually on Fridays, our friend, Jack Pollard. Jack, how you doing?

Very good. Nice to be here today. Thank you.

Good to see you. And then joining Jamie and Jack and I are our Mr. East Coast, Mr.

West Coast. Kind of like Gary and the Babe. John Schwartz and Mike Bizard.

Welcome, guys. How are you? I'm not going to ask who's- I think of us as the Glimmer Twins a little bit.

Okay. All right. Not what I was going for, but okay.

I was thinking like Murderers' Row, but- No, that's good. I'll take it. Yeah.

Well, you're a Giant fan. I'm sorry to say this. No.

It's painful on so many levels, but I will get past it. Anyway, guys, it's a beautiful Thursday down here in Florida. The sun's out, not a cloud in the sky for a change, though we've had floods.

But Mike, what's going on in the world? I don't know, but I think I'm going to send John a different kind of Giants jersey because those Giants are doing well, so. There you go.

Just saying. All right. But we do live in interesting times, and so if you've been following this whole Broadcom/VMware saga for the last year or so, you've noticed that there's been a lot of consternation around various IT teams that were not happy with the changes to the licensing program, and Broadcom was making a case for more value by basically trying to drive people to acquire and deploy the VMware Cloud Foundation, which brought compute storage and networking together in one offering.

And you paid a little bit more, but you got some value out of it. " Well, come this week, suddenly we got the VMware vSphere Express offering is now available or coming out, and this is aimed at smaller IT organizations. And the idea here is that you will get what feels like traditional VMware with a couple of added bonus capabilities tossed in on top of it.

" And this is coming a week ahead of KubeCon. KubeCon is, of course, the place where all the cloud-native people go, and the cloud-native people have been banging the drum for migrating off of the VMware platform for a while now. And all this consternation has given them the opportunity to drive more workloads off of those VMware platforms, especially among smaller companies.

Now, Mirantis has one-upped the ante because VMware, to shut down a lot of these migrations, suddenly took an API that people were using to drive these migrations and made it unavailable. The open source community, being the industrious sorts that they are, went out immediately and reverse engineered that API and created an open source version that Mirantis then took and hardened, and is now making available as part of an effort to run Kubernetes on top of OpenStack, and then that gives you your virtual machine isolation, but you can migrate your workloads now from VMware over to an instance of Kubernetes. I am sure that other players will do something similar, and this drama's been playing out, I think now for the better part of two years.

But, I'll toss this to Alan first. Does this surprise you in any way, or did you think that Broadcom was going to just sit back and watch the bottom end of its customer base just migrate off and not care? As counterintuitive as it may sound, yeah, no, it did surprise me because I didn't think Broadcom cared.

Because Broadcom's model has always, I've seen them swallow up CA, VMware, others. Their model has always been, "Hey, we already have customers and all the customer bases we want, the Fortune 500, the Global 2000. We're not looking necessarily to get net new customers there.

We want to go deeper into our existing customers with Broadcom," meaning the entire portfolio. And Broadcom's an enterprise play, so the fact that they're putting this out, I mean, did they have a change of heart? Do they not get as big a VMware pickup at the enterprise under the new pricing and so forth, and so they need to shore up the bottom half?

Or are they just afraid of the ankle biters coming up? Now, Mike, I am sure that the open source community did not reverse engineer that VMware piece. They probably went into a clean room and duplicated the functionality with an independent piece of software that they then released under an open source license.

And now Mirantis has hardened that and modified it, and I'm sure they've contributed back to the open source community, what they were supposed to do under the terms of the open source license for this. Sure. But all of that aside- Was that a legal licensing?

Yeah, that was a legal disclaimer. But that being said, I do think there's a huge market below the Fortune 500 for VMware. Right?

Especially in this age with agents and bare metals and all of this stuff. And if Broadcom wasn't going to serve it, someone was. Right.

And that could be what forced Broadcom's hand here as well. Yeah. There is still a huge base of people out there who are VMware administrators, and it's what they do, and they don't necessarily want to migrate off of that.

However, enterprises are scared of a couple of things. So let's dive into this a little bit. Broadcom is saying that I think they have either, I've forgotten, it's either 2,400 or 2,600 customers that have moved to VCF 9, which is the latest version.

And that is reasonable, I think, from where they have sat. But a lot of enterprises are also sitting there going, "Yeah, we may just contain VMware, and we may start running more new workloads on other platforms," because it's a matter of trust. And the trust factor comes down to this, it's like, so VMware, I think they spent, was it $60 billion for VMware?

Broadcom spent $61 billion for VMware. That's a lot of software licenses to make up that kind of money. And if you look at it, what they're stuck with is, well, more enterprises are containing maybe their investment in VMware because they don't trust the fact that Broadcom's not going to turn around and hike the licenses up again.

It's become a cost factor when you look at all of this stuff. So I don't know. Jack, you've been in this space for a while, but it seems to me enterprise organizations have long memories.

Yes, they do. Very long memories. But I think, I'm not sure I would agree with your take that this is targeting the downmarket as much as this is another effort to prevent hemorrhaging at the top.

And I think what they're doing is saying, as you said just now, a lot of organizations, I think, are saying, "Okay, we have a fixed understanding of where VMware is and what our VMware footprint is. " There's a lot of open source solutions, there's a lot of other solutions available that are in competition with VMware. So by providing a Smaller feature set, lower cost type environment, maybe they'll get departmental pickup within those organizations or, more importantly, prevent hemorrhaging at the- So I- ...

still within that block of 2,000 that they were originally targeting I've been in companies that pursued this strategy, and I've been in the strategy rooms where we discussed it. And really for that kind of decision, Jack, the decision isn't only how are we going to have an offering that competes against competitors. A big part of the decision is are we cannibalizing our own stream.

Yes. Right? So if I only have 2,600 customers, but they're paying me $100,000 a year, and now I'm going to go sell them this express thing for 10 grand a year, though I may get an extra 10 or whatever customers, am I going to lose any of those big 100K deals to it?

And there's this thing going on here. You've got to weigh that, you're not cannibalizing your own gravy train. Yeah, I agree with you, but like I said, I think the other alternative that people are looking at is, how much am I hemorrhaging from the customers I really care about?

And if it's a choice between them spending $100,000 and them spending $10,000 or $20,000, or them spending zero, I'd rather them spend $10,000 or $20,000 than zero. Sweet. I also think there's another factor at play here, too.

It's like, just because VCF rolled up storage and networking compute doesn't mean that I as an IT organization, even if I license that, are actually going to deploy that. And I say this because VMware didn't exist in some isolation before the merger. People had networking and storage platforms already in place before Broadcom came along, and they're not ripping out Cisco and whoever else they got on the storage side because there's a new license from Broadcom.

" But if I look at the results from Cisco and the storage companies, I'm not seeing them all hurting because VCF is in the field. " A couple of weeks ago, I went to VMware's conference in Las Vegas, and although the talk was predominantly about the enterprise, there was an acknowledgement that smaller IT teams are actively exploring alternatives or hybrid splits rather than absorb these massive price hikes for storage and networking they don't need. And my takeaway is that there has been, and there will always continue to be, this deep skepticism over Broadcom's aggressive licensing overhauls, and that there's some of these dubious investments they made in third-party storage and networking hardware.

These IT teams are just so wary of future price hikes, and they just don't trust them. To your point, John, Broadcom made it a point to say that they launched this effort in response to their customers' concerns, and they were listening to that, which I imagine they got an ear full at the conference. They got an ear full.

I mentioned it back when I was at that show, but it's so hard to draw a parallel between show attendance and what's happening with the company and user satisfaction. But I've got to tell you, I was struck by, it was a very top-heavy enterprise presence, and it was very lightly attended. I don't mean this as a criticism, but to me, it showed that there was an unhappiness or a pushback among some of the- A lack of enthusiasm, certainly.

Yeah, absolutely. And I think that resonated with them, and it was stark. This is at The Venetian, where typically it is packed.

This place looked like a ghost town. No, but if you listen to Broadcom and the CEO, Hank... I forget his name.

Hank Ten. Hank Ten. 2,600 customers, that's just where they want to be, Mike.

That's the Global 2,000. Mm-hmm. That's their focus.

Yeah. " Well, here's another strategy. Maybe they figure by putting this out, the companies that might compete with VMware and vFoundation will be busy fighting it out at this $10,000 level and not moving upstream to compete with their real fort, with their real Fort Knox, which is the enterprise business.

So, Alan, to that point, one of the things that Mike and John mentioned was networking. And one of the things you can get with VMware VCF is NSX, which is a micro-segmentation solution. Yeah.

And to my knowledge, there isn't much available as an integrated micro-segmentation solution within the competitors to VMware on the low end. And so that would actually be to add micro-segmentation NSX to this low-cost offering would actually be a way- Is it in the Express? No.

No, there's no segmentation NSX thing. There's a little routing capability, which I guess you could maybe, you know, segment. But that would be the way to really draw a value proposition, a differentiator, a distinction on the low end between VMware and all of the other solutions.

Right. But Cisco is out there with micro-segmentation, and is driving that, I think, into just about every element of their portfolio they can figure out how to do it with. No.

And then there are companies, Illumio and, uh- Well, yeah. But again, it's the integrated play within the VMware environment. " Well, if they didn't need it there, why do they need it here?

Well, that's the- Mike, we're out of time on this one. Yeah. But just to be clear, they are not giving the networking stuff in the cheaper version.

Yes. That was a rumor Jack started. Well, that was a suggestion I'm making.

If the VMware folks are listening, that would be a great value offering. There you go. I don't disagree, but then you've got a problem with why don't you charge it for VCF again?

So I don't know. We'll see. Don't blame me.

Blame Jack. There you go. What do we got, Mike?

All right, let's shift a gear because the world doesn't stand still, and OpenAI turned around and said they're acquiring something called Pier Computer, which gives them a code storage capability, is how they're describing it, that it's going to be integrated into Codex. And a lot of people are already saying, "I think we have that. " So it's OpenAI, which has divorced itself a little bit from Microsoft, essentially going after GitHub and now saying, "Hey, we're going to be a core developer platform, and it's not just about giving you a cute little tool.

" And it seems to me that that's where this whole thing is going. But John, is this the next logical step from what I guess is a frenemy position between- Yes. We've been talking about this, right?

We've been looking at this as inevitable, and it's just another break for Microsoft. And in a sense, OpenAI... Microsoft still has a, what, 25% stake in OpenAI?

It's pretty significant. I don't know if it's 25%. I think it might be 10.

I think it's estimated at that high. That seems high to me, but I've seen that estimate. But anyway, regardless, it is significant, and in a sense, OpenAI is going public, so that's part of the growing pains.

So as you grow larger, and OpenAI, I'm not going to say... Well, I'll use the word. They're desperately seeking different ways to monetize because these capital expenditures are onerous, and they are branching out into various areas, and they're not the only ones who are looking to compete with GitHub.

It's not just them. So in a sense, they're branching out, so there will be more friction between them and some of their investors. So this was inevitable, and this is going to be part of the course.

You were spot on. I think, if you look over at what Anthropic is up to, they are, quote-unquote, "starting," like everybody else, to reverse engineer Git so that it will scale better in the age of agentic AI. And the minute they do that, that becomes a repository.

And they have these things like Cursor's Origin, and these other new ventures are entering the developer tooling space, so it's There's something there. It's a wave, and I suspect, I'll pass it over to Alan, sorry. But as these companies are searching for ways to haul in revenue, especially ones that want to go public, they are going to go after some of these companies they were working with, or they were cooperating with, in a certain sense.

And it's just going to become this kind of circular firing squad for a while. I like that. The keyword there is circular, and that's what OpenAI and their financing and their strategy always appears to be.

Right? This is a Bob and Carol, Ted and Alice thing. I don't know if OpenAI got the divorce from Microsoft or if Microsoft sued them for a divorce because they decided to diversify and not put all their eggs in OpenAI's basket.

But be that as it may, this is what Sam does. He makes a deal with Macy's and then Gimbels, right? That might be a little old, especially for people not from New York.

But he has no problem playing both sides of the street. And he will- Oh, by the way, this is very interesting. So OpenAI is already tacking-- I'm not going to go down the data center path, but in a sense, they're already revising their plans, they're already trying to soften their stance.

And the guy they have in charge of it, Chris Larsen, is a heavy Democratic operative, and he's now having to explain the relationship with Trump and what have you. It's just- Good luck with that ... I mean, it's just these guys are twisting themselves into knots.

Oh, no, that's what they do at OpenAI. That's what they hired him for. They play every angle.

But let's discuss this on a different level. Gentlemen, you don't have to reverse engineer Git. It's open source.

Git itself is open source. The beauty of GitHub is they took an open source repo, Git, and they made it the standard and built such great stuff around it that almost two-thirds of every developer in the world had a Git account, a GitHub account, and Microsoft was so in love with it that they paid a couple billion dollars for it. Right?

But they're not the only Git on the block. Mm-hmm. There's GitLab, there's many, many versions of Git, including my friends from Minnesota who had a different repo, but eventually wound up going with Git capable and comparable as well.

It'll come to me. But anyway, feel free to go take on GitHub. But GitHub is the granddaddy, right?

Sure. The real power of GitHub is, what is it, 42 million accounts on GitHub? Something like that the last time I looked.

42 million accounts. So I think everybody is trying to extend Git as they know it today with the exception of one. " Good for them.

42, I'll say it again, 42 million accounts. Well, so that's where it really gets interesting from a what is the current era in, because from-- and I'm no longer, but I used to sling code many, many years ago, and as an engineer, once you build your development stack and you have your CI/CD pipeline, you do not want to change that because that wreaks havoc across all of your production environment, your timelines, everything, and things go haywire very, very quickly. " Except AI and AI coding tools are now changing the game.

People are revising their entire development stack to incorporate all of these AI tools. And so if your stack is now up for change and you're now looking at, well, I'm going to throw a bunch of money at either Anthropic with Claude or OpenAI with Codex or somebody else, and I'm going to incorporate this, why don't I incorporate and get a combined package at a discount deal that's integrated in and, oh, it's the same AI that's doing the code generation as it's doing the code check-in, as it's integrated in the whole pipeline. That could be a very compelling story for a lot of people.

But I think it'll take more than Git, Jack. You're talking, in essence, OpenAI or Anthropic building a duplicate, alternative stack that you could drop right in. Here's the thing, though- But the market is right for it right now.

No, I'm not saying it's not. But this would just be the first step- First step, yes ... in building that stack.

I think there's a fundamental change happening, and you can see it at conferences. The developers used to be highly attached to Microsoft, and GitHub by extension. They are not.

At this point, they are much more starting to be attached to the AI coding tools, and whatever comes along with that is where they're going to lean into. So I think that Anthropic and OpenAI are starting to play, are trying to beat Microsoft at its own game. Own the developers, and everything else will follow.

Goes back to what I said, 42 million accounts. But I'm wondering, though, I mean, we're early in the process, but I'm wondering what type of material effect, if this is going to create a significant or a slow erosion in what GitHub-- You mentioned the 42 million. You got OpenAI and Anthropic with their sights trained on you.

Well, there's been a lot of outages in the last year on GitHub, so a lot of folks- Yeah, they've had problems. Right. Perforce, by the way- Yeah, security problems as well ...

was the other company. Perforce had- Perforce, yeah. Yeah.

I think GitLab is also building a new version of Git, too, into their... Well, I mean, Git is open source. Anybody, feel free.

But again, beyond the 42 million accounts, you've got to give GitHub credit in that they've built so many bells and whistles and add-ons and reasons why you want to use GitHub beyond just a repo. Mm-hmm. Right?

It used to be, that's the repo. I get my code there. But GitHub's far from a repo today.

It's that and so much, so much more. And that'll be how rich can OpenAI or Anthropic build their stack? Mm-hmm.

Well, I'm going to pull Jamie in on this one, but the definition of what a developer is, is changing. And we have a lot of end users who are building code, and they need something that feels like a back-end system for building those applications, and they don't have any affinity to a thing called GitHub. So, as everybody becomes a developer, do you think there's room for the Anthropics and the OpenAIs to show up and say, "Hey, we created something that's for everybody"?

I think there's room. I think that's kind of what this competition is kind of shifting from. It's not like who has the best code generator, but it's like, who owns the environment?

And I think if they can build these tools that are easy to use and already wrapped up in a platform like Codex or Anthropic's, I think that there is definitely room. Like Jack just said, there are people who are getting less attached to the tools and more interested in making their workflows easier. Yeah.

It would seem to me, at least, that the lock that GitHub and Microsoft once had is not nearly as strong as it used to be. Mike, if I might posit a question to the group is when Microsoft acquired GitHub, and obviously it wasn't 42 million users back then, but it was a massive code base. It was like 27 or something.

But it was a massive code base for which they were using it to train their AI on, if I remember correctly. And I remember there were a lot of developers originally upset that you're looking at my code to train the AI on. Has it now served its purpose for Microsoft for the AI training purpose, and now everything else is just a bonus on top of that?

I think it's a feeder into the whole Microsoft development tool chain. Well, but they open sourced Visual Source or whatever the IDE is, right? Visual Studio.

Visual Studio. Thank you. But a lot of Git is open, too.

Yeah. This ain't Steve Ballmer's Microsoft. They're not against Linux and open source.

Mm-hmm. True. So it always comes back to me, and I'm just kind of projecting here and looking in the short term, even the long term.

Given the pressure on Anthropic and OpenAI to prove to investors and prove to Wall Street they have multiple revenue sources that are significant, I wonder if what's happening to GitHub is going to start happening to other companies. Like, in a sense, OpenAI and Anthropic are going to- Absolutely. So this is right out of the indispensability trap.

Right. Right? OpenAI and Anthropic ...

do not want to be just generic infrastructure. They don't want to be in the grid of the infrastructure. They have to move upstack.

Right. So what does moving upstack mean? Creating developer tools for developers, creating specific vertical applications for legal, for healthcare, pharma, that they can sell- And we're seeing it.

We're seeing these announcements in terms of legal for- Yeah, no, they have to. It's all in the book. Mike, can you hold up a copy again?

There it is. Buy "The Indispensability Trap," and you'll find all your answers. This was a paid endorsement buy.

But I just- But wait a minute ... it's inevitable. It is.

It's inevitable. But they have to, to escape commoditization. When you've got all these open weight models coming out as fast as they are, and they're open weight, not open source, but they're open weight, and they're going to compete with you.

The data center picture, we may be having to do with smaller footprints, so we may have to run more locally. To escape commoditization and competing with all two million models in Hugging Face or whatever- Mm-hmm ... and plus everyone making their own models, they've got to move upstack.

So here's my ultimate take on this. So if you look at this DevOps space, it's been a battle between the CloudBees and GitHub and- GitLab ... GitLab and Heart.

Jemalloc. And that's all going to come back and start to look like a very small arena when you start to look at Anthropic and OpenAI starting to get in that game, and this whole thing is going up several levels. And hell, I feel like I should be selling tickets to this fight.

What did your friend Brad Feld say, Alan? If you're not in the top three, get the hell out. Mm.

Yep. Well, and Mike, don't forget that there's Jira that you have to worry about next, and- So Jira, Jack, I understand Jira is taking a huge hit in users. Yep.

Now, I don't have anything in front of me. I don't have anything in front of me either, but just what are the core developer tools that people think about today? It's things like that, and is there any reason that they...

We all talk about how the AI, vibe coding and how it makes it so easy to do it. What's the level of effort for Anthropic or OpenAI to duplicate the functionality- To re-create these things ... of Jira or the storage behind GitHub or all these things?

It's a question of focus. Yeah. They can't be everything to everyone.

They want to continue making frontier models. They want to cure cancer. They want to go after AGI.

They want to go after super intelligence. They want to go after legal verticals, health verticals. They want to go after the developers.

So do I. But even at their inflated IPO prices, do they have the ability to do everything all at once? Mm-hmm.

Or will there be a consolidation? Or do they have the ability to do everything all at once efficiently, or are they going to just churn out just- You get a lot of plates spinning, and then eggs start breaking And then a lot of them dropping, yeah. Breaking.

Mike? Or you just turn around and start buying up a couple of the existing players and get a jumpstart. That's going to happen.

Well, that would give you a jumpstart, but you still need focus. Anyway, hey, we need to focus because we've run over time. Mike, what's our third story today?

All right. So we're going to geek out a little bit with a story that Jamie put together, but basically, one of the problems we have with AI agents and models is we don't really know who did what to whom, when, where, and whatever. And these AI models, as we kind of talked about on yesterday's show, tend to squabble.

So Jamie put together a story talking about Anthropic has a thing called Flashproof, which is, as I understand it anyway, coming up with a cryptographic proof that kind of shows who did what to whom. But Jamie, explain. Yeah.

So Anthropic has developed this research prototype called Flashproof, and it's trying to solve that trust problem. How do you know the model that you were told you were using is the one that actually produced your output? And so with Flashproof, the idea is that a model provider would first create a cryptographic commitment, essentially a fingerprint tied to a specific set of model weights.

And then when the model generates an output, Flashproof can produce a zero-knowledge proof showing that the committed model performed that specific computation, and without exposing the proprietary weights. So that's the key, that zero knowledge. So it gives customers or auditors or enterprises a way to verify this model identity.

And the big limitation right now is that it's still very expensive to run, especially when generating the proofs. And also it proves which model produced output, but it doesn't verify whether or not the output is accurate or correct. But the interesting idea here is giving AI responses a kind of verifiable identity.

All right. Well, Jack, does this warm your security heart on any level? I think it warms...

I think this is, for me, the way I interpret this is this feels much more like a FinOps problem than a security problem. " By definition, all of these LLM models are non-deterministic. So I can't say, "I gave you this input, validate that that's the output you gave me or that you should have given me," because every time I give you the same input, I'm not getting the same output back.

" And you'll never be able to do that because it's supposed to be non-deterministic. So the next question is, did I use the right model? And when I ask did you use the right model, for me, 99% of the people asking that question are asking because they paid for that model, not because they're validating the responses.

So Jack- Maybe I'm interpreting it wrong ... Jack, you're saying that I may have paid for a first class AI ticket, and I'm other than steerage. Is that what you're kind of getting at?

There is that possibility. But if you got to your destination and you were comfortable, do you give a crap? Well, I- The CFO might give a crap.

I might not, but the CFO might. I might care because the travel experience wasn't as good, and also the person sitting next to me paid for something, and I got the same output, and they paid less than I did. Oh, mind your business.

No, but seriously, this is a solution looking for a problem. Mm-hmm. Anthropic didn't put this out to see if you used Fable instead of Opus.

Anthropic put this out to see if you used Astra instead of Sol. Mm-hmm. They want you to use this on other people's models.

But do we really think Anthropic and OpenAI are playing the old shell game with their models? I don't think so. I think the real use of this may be at some point where you go on Hugging Face and download one of the two million models there and then try to run it locally, and you heard such good things about this great model, open weight model, and boy, does it suck.

And now you find out, well, no, I didn't download that from the repo. I downloaded the one that's been infested with malware or something. And so it's like a file integrity checker.

It's a model integrity checker, if you will. But I don't know if this is a big problem even yet today. Again, you can't boil the ocean, Anthropic.

Pick your battles. You want to go public? Show me you're making revenue.

" Well, hey, so hey, I think, was it 40? Actually, I only know there's 41 ballot measures involving data centers. Yeah.

40 of them were, it turns out to be, however you interpret the ballot measure, was anti-data center. So, oh my God, bells and whistles. Yes, Congress is now attuned to this as a great political weapon to knock over people's heads So yes.

And I also, recent about what Anthropic is trying to do, it is kind of a, I'm not going to say it's a sign of desperation for them, but in a sense, these guys are flailing about looking for ways to haul in, again, to create more revenue streams, because the math right now is not good in terms of the losses, just for the first six months of this year, according to their own prospectus. But you know what? I'm going to go back to yesterday.

I'm still traumatized by that story about the AI agents and grappling with one another and grappling with it. Mike, you do make me lose sleep sometimes, and with some of these stories. And that one actually made me think about the revolt, the rise of the AI agents doing its own thing, but anyway, I digress.

Right. Because it may not be that the model provider is trying to commit fraud, it's just that, to Jamie's point, these AI agents are going off doing whatever they think is best, and they will argue amongst themselves and cause stuff. So, right, they're doing what they think is best and kind of go rogue.

Who's responsible? And in a sense, the Congress members are going to invariably haul up the most high-profile executives or companies to wag their fingers at them and cut them off and not allow them to answer the questions. And I can see that scenario playing out to a minor scale versus all the other subpoenas and investigations they're going to do.

Now, I know AWS has been banging a similar drum about cryptographic signage and things like that of that nature. But, Jamie, the other side of this thing too is these AI agents are operating at machine speed, so I don't see any other way to understand what it is they're doing unless we have some sort of mechanism of this nature. What do you think?

I agree. The identity and verification issues that agents bring up, everyone is coming up with a lot of novel solutions right now for this identity and access kind of situation, but I don't know if this is the one just because of how inefficient it is at this moment. Yeah.

They said producing the proof of generating it required roughly 70 to 95 times as much GPU time as just running the model normally with VLLM inference. So it's going to cost me more to find out that I got ripped off than what I got ripped off for. Exactly, yes.

All right. Alan, you think this is maybe something that is better served in a consortium somewhere, and maybe we should just kind of start a project around this and let the industry kind of collectively solve this? No.

I think you put this in a file drawer and come back to it when it's an issue. Alan mentioned keyword focus. I think it's just Anthropic making noise.

All right. I think that there's more to it than just making noise, and I think that it's a problem that's bigger than one company, so I think that's a- Well, let me do a little research. Among the five of us, raise your hand if you think you were fed the wrong model than what you thought you had.

You think so, John? Little bit. When did it happen to you?

Well, it's an issue of trust with me, and my trust is a little up with these companies. You're okay. You can have it.

Yeah, so maybe it's more kind of it's a subjective thinking on my parts, but I'm probably making more of it than it merits. I'm sorry. I just don't see this as a problem here.

I think it's a matter of trust, but I do see, when I am using different model providers, that things change in terms of what they're surfacing at any given time, and sometimes- Right. I think that's a different thing though, yeah. Yeah, but that's the non-deterministic nature of it.

Well, no, it's not. It's the model provider, or maybe it's the company I work for is doing it, and I'm not paying attention. But on any given day, suddenly I've been going from high to medium to some other model, and I don't like it.

All right. So let me be your IT help guy. If you click above the chat bar where it says what model you're supposed to be on, and version- Uh-huh ...

and you look in there and it says auto, that means yes, the AI is picking what it deems the appropriate model and weight for the task you're giving it. 1 Medium or something like that, you could do that. But when I log in, it's already decided.

There's an option. No, but you could change that. I can change that.

Right. This is America. Mm-hmm.

But something Mike said just hit me, which is there's another way to look at this, which is model attribution, and this may be- Mm-hmm ... Anthropic trying to come up with a way to say that when you did something, you did or did not use an Anthropic model. , when you say, "This AI attacked me," they can come back and say, "Hey, wait a minute.

That wasn't our model. We have signatures for our model. " So Anthropic does have that, Jack, right?

Anthropic put in the algorithms where they could tell if something was generated by Anthropic. Well, that would be for output. Think about more for an AI agent director or an AI model-directed cyber attack.

How do you attribute that? The output is a little bit different and isn't getting signed in the same way. So this may be the first start of- So we can't have a clue game where we say it was- ...

attribution ... the Anthropic agent in the living room with a fire poker. Something Mike said sort of made me think about that, but this may be the start of that part of it.

So when I wind up in court, and I have my AI agent as my defense attorney that will prove through this flash proof thing that this wasn't me, and I'm innocent, and I should be allowed to go home, Your Honor, right? " Mm-hmm. What a great show this is turning into.

I love it. They have an AGI who's reached out to the lawyers because the company that created it is giving it too much freedom. There's all kinds of AI tech plots in here.

Well- It's a great show ... Alan, in the real world, there was a judge who was recently censured because he was creating rulings based on false AI laws. Yeah.

Well, that's rampant. I read something the other day. Yeah.

That's the latest example of that happening, yeah. It's becoming an issue, yeah. I was going to say, we can come back to that topic next week.

Good enough. Hey, speaking of coming back to that topic, we've got to wrap up. Jamie, Jack, great seeing you both.

The babe and the iron horse there, thank you very much for being here. Thank you for watching. Of course, we're live every day at noon on Techstrong Gang.

tv, or our Techstrong TV OTT channel. But for now, on behalf of the gang, have a great day, everyone.