Cisco’s Tiny AI Models Just Beat GPT-5.5 on Cost by 172x

Hey everyone. Welcome to Tuesday on Techstrong Gang. The week is already off to a flying start.

We probably are two days in and seven new model releases as well, and each one claims to be better than the one before. What a crazy time. We've got a lot to talk about with our gang today.

Let me quickly introduce you to our gang. I think it's kind of our usual Tuesday crew. We've got my friend Sid Nag- Hi ...

Chris Blask, Kate Scarcella. Hey, gang. Welcome.

Yo. Mike went from the Adirondacks to the culture of Quebec, and so he's touring Montreal and the rest of the Quebec province, and so he's not here this week. But we will carry on without him.

I think we'll be fine, especially with this bunch of folks here on the gang. " I've written about this concept of free as in freedom, free as in beer before. It describes two different kinds of freedom.

One is I don't pay for something. I got free beer. Open source.

A lot of people are attracted to open source because it was supposedly free, but that turned out to be a bit of a fallacy. And then there's free as in freedom. I could use this software, change this software, modify it, do what I want with it to a certain extent, even under the open source OSI licenses.

And that's what I originally meant with free as in freedom, free as in beer. I learned that phrase from my friend Ron Gula, who started Tenable Network. But Sid, it's taken on a whole new meaning now with AI, right?

We've got open-weight models, which are not really open source, but they are open to a certain extent, and they're really getting in there. We've got just a lot of different business models vying and playing by different rules. Mm-hmm.

What do you think? Yeah. First of all, I think you've hit it right on the spot with your article.

It's so timely because we're seeing, I think, the decline of the big, large foundation model. You look at OpenAI postponing their IPO. Anthropic has announced their IPO.

I'm dying to see where that lands. But I think that your article really argues that the long-term battle in AI is not going to be won by a single sort of model that's a frontier model, and that foundational models are instead going to be a commodity, right? Yeah.

And, I think really you're going to see things like more and more value add on to these models that are going to matter. Whether it's software, whether it's observability, whether it's other orchestration layers and things of that nature. And I think at the end of the day, like I said, there's not going to be a single dominant winner.

But with the commoditization of foundation models, as intelligence becomes more abundant, you're going to see increasing economic value from the add-on capabilities that will be provided on top of this foundation model. Whether the foundation model builders go and do that or somebody else does, that remains to be seen. But clearly, I think you're seeing the advent, I think we talked about it last week, advent of Chinese models that are being offered at a really, really low price.

, I forget the exact number model, to prevent an agentic attack on their environment. And they couldn't use the frontier models that are being offered by the US vendors because they had guardrails built into it. So I think a lot of things are sort of happening at the same time, and I think the net of it is that the big hoopla we saw about foundation models being sort of the big deal in this AI conversation is going to kind of be questionable.

Yeah. Thank you. Chris?

On every level, I love this show, because once a week I get to step back and just talk about the work instead of doing it. But every week, in this last year, last year and a half, this has just been this path. " We spend this time talking about models and these one or three companies that are going to rule the world and control everything and, and, and.

And, also in the green room we're talking about, I think you have a neat article coming out tomorrow, right? Mm-hmm. This centralization doesn't work.

The internet wasn't supposed to work this way anyways. Everything from DNS to PKI to Google itself. I'll say it, Google is...

I won't even blame anybody. The companies exist because there was a void to be filled. But Google is just this big data warehouse on the horizon that just wants your information, wants to give you information, because there's no other way to do it.

And now we're in the world of AI, and yes, just to repeat ourselves, incrementally different week by week in this little cadre. It's not about the models. It never was about the models.

It isn't going to be controlled by a single either company or nation state making a model. It's about a system of systems. And in the world where we can't scale centralization more, we have to go the opposite direction.

And haven't we spoken about infrastructure a lot, right? This is just going to be a part of our infrastructure at the end of the day. And just as we have energy, we're going to have these data...

It's just going to be like this data AI pipeline that certain people, well, many of us will be running. We'll be running our own models, and this is the way it should be. It gives us autonomy.

It just gives us more control. There's other things it introduces, but this is really important. It's really important to grab the power away from a centralized system because not only does it not work, but it also...

We've read "Big Brother," we've read "1984" and George Orwell and everything else, and it's real, and we've seen it. We want to keep information in our hands. Is there some cost to that?

There is. Well, I'll give you some hope. So time itself is on the table.

org has this working group working on the time problem. You thought Y2K was something, 2038, hold my beer. And it's a whole different world.

And we had our monthly meeting this morning, and as all these working groups are, a fascinating window into some great minds thinking about this over time. And my posit on this whole thing is that there is no centralized time, literally. We have atomic clocks and NTP.

" You cannot centralize time. Doesn't work. Can't ever work.

But you can localize it and build up relational time systems from the ground up all day long. Yeah. So it doesn't matter whether you want George Orwell.

George Orwell just doesn't work. Yeah. " And part of the research, I talked about different waves of technology, and one of them was the railroads.

I don't know if you know this, but before the railroads were laid out west, here in the US anyway, every city kept its own time. So Cincinnati was a little bit ahead of maybe of Louisville, which was a little bit ahead of St. Louis.

Mm-hmm. It was the railroad, and no one really cared because things didn't move fast enough where who cared if it was 1 versus 1:20 here to there because you couldn't get there in 20 minutes anyway. But once you had trains that had to run on time and use a schedule to make sure you didn't have trains running into each other on a track, we had to standardize our clocks.

This was done without an act of Congress, without the government. The railroads imposed time or standard time across the country so that the trains would not collide. Years later, the Congress caught up.

But to my book, "The Indispensability Trap," it's part of what I wrote into this article, which is once intelligence, machine intelligence, let's call it, becomes so commoditized that it becomes that commodity layer like the railroad tracks were, like electrification, like the phone lines were, like fiber optic has become. You no longer care whether it was Anthropic or OpenAI or one of the free weight models. We're at this funny stage right now where there's still a very big distinction, and one model's claiming more freedom than the other one.

One model's claiming one's cheaper than the other. And we have the governments here jockeying for geopolitical strategy and domination. But I think this goes away when intelligence becomes ubiquitous.

Yeah. And I think the centralization of time is an excellent example. There is no centralization of time, and there is no centralization of data.

The idea that a particular estate should control all data is ridiculous. Mm-hmm. And I think what really scares me, not just the model builders are controlling all the data within their estates, and we are being asked to use only those models or one or two models, but the cabal between the data estate hoggers, if I were to use the word, and the infrastructure builders, the hyperscalers, together could be extremely worrisome.

So I would rather see a world where the data estate is more commoditized and more democratized, where either enterprises, be it Boeing or JP Morgan, whoever, have control over their data, and we as individuals have control over our data that we want to use and play with and use it for work or whatever. The idea that these things should be owned by a set of limited power-hungry oligarchs, if I may say that. Sorry, I'm not trying to be socialist.

But the idea of a few vendors owning all of this is extremely worrisome. Yeah. It's fragile.

It makes for systems that fall apart. But again, you look at analogies over time. I've lived the petrol head world, so I've wrenched on internal combustion.

But in the beginning Early 1900s, there was a couple of people making blocks and so forth. And in my era, there was the small block Chevys and so forth, and you can geek out about things. So right now, we're talking about engines.

All of this stuff, talking about AI, is like talking about whether or not the hemispherical head of a compression chamber without thinking about cars, and pumps, and motorcycles, and things you would do with an engine. Right. Engines have characteristics.

I look at all these models out there, and we use them all the time. We use them all the time. And let me overstate this just a little bit.

I don't care what the model is. I'll see the model, and I'll put it in the harness, run it up, see where it vibrates, put it in a system where I can use its capabilities. And we're all obsessed.

We literally take all of our time, 99% of our time, talking about the training data of a model. Who cares? It gives you torque.

It vibrates at a certain frequency. Put it in a system that you can count on and stop thinking about the spark plugs. Chris, back to the chapter on railroads in "The Indispensability Trap," the epigraph for it is a quote from a Harvard Business School guy who wrote this about the railroads.

The people who built the railroads thought they were in the railroad business. They didn't realize they were in the transportation business. And so, the act of building the railroad was capital-intensive on par with what we're building in data centers now.

But within 20 to 30 years, three-quarters of them were bankrupt because they were built to build the railroads. They didn't realize the real business was what went over those tracks. When Rockefeller was putting oil, transporting oil via train, when a guy named Montgomery Ward started a business, a catalog mail-order business, and another guy named Sears followed suit, and they used the railroads for their delivery.

Those became the businesses. The same thing in electric. It wasn't the people who put all the electric lines in.

They made money, but it was a huge capital investment. The people who made the real money, General Electric, making all those appliances that ran because people could just plug into their wall socket, and they didn't think about it. It was ubiquitous.

That's the point of my book. We're going to see a similar thing here with AI. Yeah.

It's what you're going to build with this terrific tool, with this intelligence, not the trillions of dollars we're going to... By the way, another chapter in my book, same thing happened in fiber with the internet in 2000, late '90s. We lit 4% to 5% of the fiber we laid back then.

It took 20 years to light up the rest of it. A trillion dollars in wealth, maybe $2 trillion, was written off by the fiber guys, companies like Global Crossing, if you remember these names. I do.

Yeah. They're familiar to all of you. MCI.

That would be a lot of money when they went down, yeah. Yes. They owed me, all of us.

Level 3, these companies, right? Yeah. Who are the Level 3, Global Crossings, and WorldComs of the AI era?

Yeah. You guys, maybe. That's it.

We're waiting for the golden toilet bowls, right? Yeah. Exactly.

Anyway, it's an interesting thing. And I'll leave this segment with the last thought. Who would have thought it was the Chinese Communist Party who's standing up for frees and freedom on open AI models?

Are they the little lone guy with a sign standing in front of the tank, the American AI tank in Tiananmen Square? What's going on here? But be that as it may, that's where we are.

Well, I'll say one more thing, if I may. Go ahead. We need to be really worried here.

I think the US is going to lose the AI race the way things are going- Yeah ... with China. Maybe a topic for the next show, but this is a serious problem.

I'm hoping nobody wins it. That'd be the worst case. It might be the best case, is no one wins it, right?

Yeah. Right. That is right.

But frees in freedom, frees in beer. But an alternate to this is let's get the intelligence off these big models and run local stuff on our stuff. You're speaking Chris's language there.

Right? Yeah, absolutely. Well, Apple appears to be following that kind of model.

And who loves Apple? Who loves Apple? Who's always sticking up for Apple here?

Kate Scarcella, that's who. But now Cisco seems to be jumping that bandwagon with something called Antares. Kate, what's this one about?

Yeah. So it's one of the more interesting announcements this week, came from Cisco with Antares. It's an open-source family of compact security AI models.

I feel like you're just talking our language here, at least with Chris and I for sure. As someone, I've spent decades in cybersecurity, and what has caught my attention wasn't that Cisco released another AI model. It was that they intentionally made it small.

It's like they're listening to us on this show, I swear. Maybe they are. But the edge has been, we've talked a lot about the edge and how we've actually, I'm going to say we because I think I'm speaking for more than just myself here.

We actually believe that AI will be moving to the edge, and the industry has a habit of assuming that bigger is always better. But bigger appliances, bigger data lakes, bigger models, history, as you have just pointed out in block A, just feeds so nicely into block B to say, "You know what? " Data swamps are real, especially- Alan living in swampland there in Florida.

I've lived in Florida, Chris. We have lived in Florida, and we understand data swamps just swallow in the shallow of water. It's just so expansive, and you just drown in it.

And so I know that this is the way to help us in the future, and that this will become a part of our every day. Even from a security operations, it's a lot better to have something small than having just this huge, big data warehouse and everything else trying to protect. So we've seen this pattern, and I believe it's a good thing.

And if I could just add one more thing. Sure. Boy, asymmetric.

If we get anything from this, if we don't start to understand that we are in a time of asymmetric warfare, cybersecurity, it's all asymmetric. And we better understand really fast that bigger is not better. Yeah.

And that's a big part of the American system. Chris, are these models bringing Mythos-level capability? Well, again, models and systems.

Look, I've been walking around the Earth, Kate has been as well, with Mythos-level capability just in our heads, right? And it just comes down to spare time, right? And Mythos, and this, again, huge model thing that can do all this stuff sitting on a stand by itself is interesting.

But, in this current world, I can do that stuff now with existing models before Mythos, just using them systemically. Anybody can, right? There's nothing new.

Mm-hmm. And small models do the thing you need them to do. So you want to make a little, tiny, small code breaking model, I guess you could do that, too.

But mostly what people need is a model that does this function, this task in a workflow, in an orchestrated bunch of little AI minions writing code or running a power system or a water system. It makes me think of AOL, right? The early internet days.

Remember, we're all sitting here going, "Oh, the internet. That's right. " And then we go to the conferences- We get free disks ...

and we get free three and a half inch diskettes from AOL. Mm. And the AOL and Microsoft argument was Bill Gates, I really wish somebody would pull this clip up.

" So AOL- ... is going to have the entire internet in a box. This one company's going to give us one internet with everything in it, and no one else needs to be involved.

That's not how any of this works. Yeah. Tiny little things.

So yes, little models running all over the place, little LLM word generators, yes, in systemic processes that are appropriate for whatever the heck people are doing. And Sid brought up Huggy Face earlier. I think this actually ties in with what you were talking about earlier, Sid, as well.

Yeah. I'm not as deep a security expert as you two guys are, or three of you are. But I really think the future, this kind of ties to the conversation we just had earlier.

The future of security AI is really unlikely to be dominated by a single sort of frontier model, right? It's going to be very organization specific. It's going to be more enterprise specific with security needs that are very bespoke for each environment.

So it kind of ties into the same philosophy of where it's not going to be a large frontier model controlled by a large company that is going to mandate security controls within the framework of their model, because there's no way that they can do it all for every possible use case or scenario, right? That's sort of my opinion. Well, and it's physically impossible.

Let's just take utilities, right? What global organization will know literally every single thing about every substation on the planet, over every second, so every device that is changed and updated and so on and so forth. Mm-hmm.

How are we going to pull all that together in one spot so that you as a new utility can come and ask for your data? It can't be done. Yeah.

It's just thermodynamically impossible. But here's the thing, and I'll throw it out to all three of you. One of the biggest pushbacks we saw with Mythos to Glasswing was sure, the top 50 guys got access to this, so Microsoft can release 500 updates on Patch Tuesday, and Google can do 300.

But what about the rest of us? We're being shut out here. We're being shut out where we know the bad guys are going to get their hands on this.

Yeah. Back to what Sid said in the earlier version. The guardrails prevented them from using to prevent an attack, prevented them from using it.

" Right? It's kind of a reverse psychological argument there. Now, Cisco releases this and says, "Okay, now it's cheap.

It can run local. " But what stops the bad guys from using this? Nothing.

But they're already using it. Crowdsourcing for the bad guys has been around forever. And I just remember IBM releasing terabytes of all this data, and they were so proud of themselves.

" And it's like, yeah, but the bad guys have been doing this for a couple decades now. " But the problem is, is that if we don't share, we're in trouble, so we have to share. And so you know what?

The bad actors already have the tools. Give us the tools, and I think that's what Cisco is doing. The cost of the capability drops, the balance shifts.

But for not only for the bad actors, but also for the defenders, so good for Cisco. Thank you. Absolutely.

And I will mention, I was at Cisco Live, I know you remember, could it be a month, maybe two months ago. And this Antares is the latest release or the latest announcement. They have several other AI-based security type of functionality that they're releasing, and I actually have some of the names here, and that they got the Foundry security spec and Code Guard.

So, a revelation for me back in my Still Secure days was when we started measuring up and sizing up the market and who our competitors were, was even back then in the early 2000s- Yeah ... Cisco was the single biggest security vendor in the world. Yeah.

Primarily through the firewall because they put one with every switch, it seemed. But they have been players in security ever since, whether now it's with Splunk and this stuff. And Nexus.

They have Nexus. Sure. They have Nexus as well.

Yeah, I have an observation to make. I don't know how you guys feel about this. I think the idea that users, be it enterprise or end users like you and I, depend on vendors like Microsoft or Cisco or whoever else to give us security comfort is going to be over.

I think it's going to be more about you being selfish and protecting yourself. So I think enterprises are going to develop their own Patch Tuesdays and implement it within their environment, but they'll donate it back to a GitHub kind of an environment, and it's going to be more of a community thing, right? " I think all that's going to disappear.

Yeah. You said so much there. Yeah, so this is literally how I do things now.

If any individual computer can, I know what I want it to do with, I know the tools are available, and I make sure it studies its own environment and maintains that visibility, period. Right? So all of the black hat scanning tools and all that stuff we know people can use, that is just built into all the nodes.

They know that whatever's available or not, and all the bloody way down, and that is hyperlocal. But the sharing through GitHub, this is the Canon node mesh thing. Mesh is sharing willfully and lawfully, not networking.

And I'm sorry if we're talking too much day jobs, guys, busy. But sharing internet protocols are, did you get the packet? A mesh protocol, the way we look at it is assuming you did- Mm-hmm ...

but in our view, nine things do you do before you do anything with it, right? Just because it arrived here doesn't mean that you actually want to do anything with it. If I'm sharing something, I want to be very specific.

Maybe it's the entire world, maybe not. Anyways, our network approach in the past has left this massive, one-dimensional, shallow system. But if you do it exactly like you said.

Start locally, figure out what you need to do, use these same tools inside the computers. You are this node in this environment, make sure you have the tools to be aware. You can generate them in bloody place.

And as you learn things, yes, sharing that, meshing that out up to global standards organizations, I think is inevitable because there's value in that, too. But you don't wait for Brussels to tell you what's secure next Wednesday. That hasn't worked yet.

Yeah. So I think that is something that we're going to move to as a world, is away from Brussels telling us, or Redmond telling us, or whatever. Maybe personal responsibility here for security is a good thing, and the fact that you could run these models locally on your own device gives you that, again, freedoms and freedom.

Yeah. And at a better price. You get the beer there, too.

It's all good in the world. Anyway. I like it.

Yeah. You like it. Let's move on to our third block for today.

And this is, again, this comes out of an article, a rather long article I wrote, but also an interview with David DeSanto, who's the CEO of Anaconda, he's a friend of mine. I've known him since he was chief product officer at GitLab. And they made a big deal this week with a company called Kilocode, which coincidentally or not, happens to be co-founded by the founder of GitLab and the chairman of GitLab, Sid Sibrande.

Chris, you followed this one. What do you think we got we're dealing with here? Well, Kay, like you said, I think at the beginning of the show, in the green room, these three blocks line up really, really well.

And that's a good sign of coherence in the production process here, because again, when the model doesn't matter that much, what does matter? The workflow. When I do coding, I was walking through a good friend of the show, Fred Cohen.

I've been working with him, I work with Fred all the time. And he is the professor down in the code approach from hell. And yesterday, I think I got him across the line understanding how much you can step back.

With CLI Codex on a node on a computer and a well-lined-up AI system of your own, the Lumina system, you've met Lumina, and the cannon behind it, so you know what you're generating from. You can just have a system where the models don't matter, to my point, right? They matter.

Just stick with OpenAI's CLI Codex. They've got a number of different ChatGPT-based coding models they have. They keep changing them over time.

They have different characteristics. You see those. But like I said in the last block, it's a motor.

It vibrates at 3,700 RPM. Tend not to run it that way. Or at least if you're going to run it at 3,700 RPM, don't use it for long, whatever.

Right? Mm-hmm. It's how these things work together systemically that let me code at speeds that are just ridiculous.

I like slowing down my coding and you can get into the anti-gravity ASHRAE. We talk about the minions, an orchestrator, and you can get really fast. I like slowing down enough that it's only 10,000 times faster than it was a year ago.

And at that level, yes, again, this topic of this segment makes sense. It's not about the bloody models, it's about the systems. And Anaconda is right there in the middle of it.

Perfect timing. Absolutely. But from my point of view, and there were two articles that back up this section, one's by John Swartz, which is kind of the hard news of the acquisition, and then the second is kind of my feeling on it.

Again, I'll go back to the indispensability trap. No one wants to be commoditized. We don't see that commercials.

I want to wake up, I want to be a commodity when I'm 30 and retired. No. Everybody tries to move up stack.

As a matter of fact, that was the second alternate title for my book, Up Stack. Because everyone wants to escape. Now, Anaconda's a great example.

Anaconda's a company that its whole reason for being is it controlled the Python packages. Everyone who wanted to reuse a Python package, basically, a script, Anaconda was the source for it. So it had a tremendous following among developers and ops people or anyone who uses Python.

I think it's the most popular language in- Yeah ... use today. But it really occupied that commodity layer now.

Everybody uses Python. No one really thinks about it. It's open source kind of stuff, and how do you move up stack?

And David DeSanto came into Anaconda about a year ago now, and he has made a few key acquisitions as well as kind of changing the company to move up the stack to be more of a tool for developers, especially in an agentic AI world. Building on, he didn't throw away what Anaconda does. Anaconda still has a damn good Python.

And actually, if you look at it, if you speak to David, he'll tell you this, CUDA, the system from NVIDIA, it's a lot of Python in there. And they do the Python packages for CUDA. So it's a huge ongoing cash-generating business.

But moving up stack, how do they become a tool in the agentic AI age for software engineers? And that's what this deal really is about, right? And this is the second kind of deal in this vein that David and Anaconda have pulled off here.

I think you hit the nail on the head. It's really about developers interacting with AI agents, right? Yeah.

As coding increasingly moves away from manual tasks to a agentic kind of model, the platform controlling model routing and agent behavior and enterprise governance is going to really start mattering. So it's definitely about agentic, but it's also about scale, right? There's some numbers quoted in the article that Kilo's model-agnostic architecture supported more than 500 AI models while orchestrating nearly 10 trillion tokens.

Yeah. So think about that. That's super scale.

And that came on in like this. Sid, I noticed on LinkedIn you wrote something about a new kind of AI model, not a stack per se, but it looked like a stack almost. Yeah, talking about the AI cost stack.

It kind of fits into this. Yeah, the AI cost stack. Yep.

Because it's really not about measuring the cost of implementing AI on the traditional tokenomics model, but it's really all about cost per business outcome. And I talk about how you can sort of model that whole thing rather than the traditional ways of trying to understand what it is going to cost an enterprise to run AI, which is very sort of myopic, I think. That's really interesting, Sid.

And I actually agree with what you're saying. It's better to look at the business outcomes, and it's like a wonderful business value assessment type of tool that people always wondered. " Here, it's just a holistic way of looking at your AI model, and I like that a lot.

I mean, but clearly it seems like we... I remember when FinOps first came to cloud. People were kind of sneaking it in the back end.

Companies that you didn't think of having anything to do with financial controls were offering FinOps. And then we started seeing standalone FinOps companies and FinOps Foundation, and all of a sudden everyone was talking about, "Hey, this cloud's a lot more expensive than we thought. It was supposed to be cheap.

" Yeah. I mean, people are literally figuring out the cost of using X number S3 buckets or EC2 instances. Right.

I mean, that's not relevant, right? No. How many applications did you run?

What outcome did you derive as a consequence of moving that application, running on premises with your data center running on premises, all the server costs and all that, transposing it to the cloud and running a software as a service mode, and the outcome you got as a consequence of doing that. That's what really matters, right? Yeah.

I think we're going into the same thing now with AI, right? Yeah. With tokenomics and maybe that is the wrong...

I saw, I guess it was an advertisement on LinkedIn, but the guy, it was again, about running a model on your own hardware, and you never have to worry about tokens again. Yeah. But what he didn't say is how much it costs to set that hardware up, right?

You're going to go set up some monster, a gigawatt rack or something. But this is kind of where I think the argument is today, and again, I do think it's part of the road to having ubiquitous intelligence, machine intelligence available, right? Because when you plug in your appliance, your vacuum, or your blender, or your toaster into the wall, you really don't think about how much electricity that's using.

Yeah, the refrigerator has those yellow tickets, right? That we all, some of us peel off right away- Right ... some of us leave on, about how much a refrigerator costs to run a year.

But we don't really think about it on a per-use basis like that, do we? And I think we're going to have the same thing. I've got a gig up and a gig down for my bandwidth at home.

Do I think about how much I've used or am I really getting a gig? Because 4 or 500 is enough for me. It's that same thing.

Chris? Yeah, I think, and again, this is a very transitional period, so we look at similar ones in the past and what characteristics happen in times like this. And yeah, the beginning of this block, you started with rising up the stack, right?

And my little comment about how I code these days, I think it's hard for us to imagine right now, what's our goal? What are we really trying to achieve with the time we have, given the tools we have? Mm-hmm.

And we're trying to say from here, it would take this much time, and our goal must be right here and nothing beyond that. When I think if we allow ourselves to move far enough up the stack, right? When I code now, literally in my head, these are the kind of words I'm thinking about.

I'm just trying to express my intent, through my language systems, my AI systems, onto some substrate, which could be a code base or a website or whatever. And whether it's using Python or microcode or exactly what, I don't care, right? And I'll care if the end result doesn't work properly to the economics of it, right?

I'm looking at this. How much effort is it taking to get those results? How much am I interested in the details in between?

And we all know, we could take it down to the gallium transistor level in every single task we do. We do not. Why not?

Because we count on those failing at a rate that works well enough for us. But we got to talk about exactly the Python library I'm using and before I talk about the fact that I'm trying to build a car. It's like, don't care.

But based on what we just talked about, I don't think the economics will matter, because if indeed we're now going to be reliant on these large models, and we're going to run our own models in our own environments, and if I can run it on my laptop that are used for 10 other things, who cares about the economics of it, right? Yeah. That's, I think, where we are gravitating towards, right?

And I think people are not going to worry about AI being some super-duper thing that is going to cost us billions of dollars. It's going to be integral part of our day-to-day lives, that day-to-day compute surface, and we won't even care. We're just going to do 10 other things.

And I actually agree. I mean, it's becoming less expensive. It's more portable and easier to integrate.

So into everyday development, into the cost and capability, it just continues to fall and- But what about the $8 trillion right? And again, back to my book, that's the whole point. Someone's got to soak up that $8 trillion.

Someone's got to pay for the $8 trillion, or are we writing it off? Well, Sid, my joke at the beginning, yeah. I had switched over to the service provider line of business at the end of my Cisco tenure to watch them all.

So there was huge amounts of money that were on paper that never showed up in Cisco's bank account. Yeah. So it gets absorbed, doesn't it?

And maybe we don't need to fight about building massive data centers in our backyard anymore. Those go away. Well, that is a fallout for that, right?

I just wrote another article on yet another data center fiasco. I mean, it's one after the next. The people are revolting against these data centers, right?

But it's also creating a lottery. I read an article, I forget if it was in "The Journal" or "The Times," in Loudoun County, Pennsylvania, maybe. Not Loudoun, Virginia, but another county in Pennsylvania.

No oil, no gas. It's data center country. They're paying farmers $10, $20, $25 million for their farms to convert the land into data centers.

Farms that were worth maybe a couple of hundred thousand dollars. Yeah. Right?

And what are the farmers supposed to do? Not take the money? They take the money.

Yeah. It's creating quite a stir in local politics out there. But this is the world we're living in right now as we go through this ubiquitous intelligence kind of move that we seem to be going in.

Anyway, we're about out of time for today. I want to thank the gang. Thank you for coming today.

Thank you for watching today. As I always say, we're live every Monday to Friday at noon here on the Techstrong network. tv, YouTube channel, Techstrong TV YouTube channel, our OTT channel for iOS, Android, Roku, Apple, Amazon, anywhere screen that you watch this stuff on.

We'll be back tomorrow with another fresh episode of the gang and more of our favorite gang members. But until then, on behalf of Kate and Chris and Sid and I, have a great day, everyone.