Stop Using a Sledgehammer to Kill a Fly: The Case for Practical, Stratified AI | Utilizing AI Episode 48

AI is inescapable, but many uses are too big, too often, and just too much. This episode of "Utilizing AI" considers how much AI we should be using and features Alan Shimmel, Guy Currier, and myself from Futurum Media. Welcome to "Utilizing AI," the podcast focused on practical applications of artificial intelligence and part of the Futurum Media Podcast Network.

Every Wednesday, we explore news and use cases of the ways in which AI is transforming enterprise IT and the industries it serves. I'm your host, Stephen Foskett, president of the Tech Field Day business unit here at Futurum Media. Before we dive into this discussion, let's meet who's on the panel today.

Hey, everyone. I'm Alan Shimmel. Stephen, thanks for inviting me.

It's my first time on this podcast. I'm looking forward to it. For those who don't know me, I am the founder of Techstrong Group, which is now part of the Futurum Media family, where I am the CSO.

" I'm very excited about that. I also host the "Techstrong Gang," so I do a fair amount of video. I'm thrilled to be here.

Hi, I'm Guy Currier, and congratulations, Alan, on your book launch. Can't wait to read it. I am a research director at the Futurum Group, and I'm also the chief analyst at Futurum Media, and I study AI and application development, AI services, and infrastructure.

Well, it is awesome having the two of you here on "Utilizing AI" this week. As you can hear, there's a bit of a theme. We're all part of the Futurum Media division, which is really a big reorganization of all of the media assets within the Futurum Group.

Check out LinkedIn, ask me questions. I'd love to talk about what we're doing there. But essentially, we're doing this.

com. We are bringing together a lot of the assets that had formerly been somewhat segmented underneath the Futurum Group. So this week, we are diving into a fresh and full of life "Utilizing AI" podcast talking about AI, and we're going to take a bit of an interesting angle here.

The point that we're going to make here is that you really don't need too much AI in your life. Well, and by that, I don't mean you don't need AI, because frankly, it's inescapable. It's 2026.

There's going to be AI everywhere. But you don't need too much AI, too big AI, too strong AI. Guy, this was your suggestion, so I'm going to let you kick this off.

Tell us, you don't really need too much AI, right? Right. We all don't need too much AI, and I think there's two or three angles to this, any one of which we can talk about.

The first one is, to your point, Stephen, AI is inescapable, but does that mean it needs to be everywhere? Do I really need to ask an AI bot where I should go shopping for shoes? Maybe it'll be helpful, but maybe not.

So the frequency with which we use it, that's one angle. Another angle is, I think the main one that I had in mind, which is AI models continue to develop at a spanking pace. We can definitely talk about why that is.

Alan, I think your book, to some degree, covers the forces that are racing us and accelerating us and this industry towards bigger and better and deeper and faster, blah, blah, blah models. That is another way of thinking of too much AI. Maybe I do want to use an AI to figure out where to shop for shoes, but do I need to use the same AI that can also be used to vibe code an app that tracks all of my plants inside and outside and how they're doing by photos?

I mean, the same model to do both, that's kind of what we're doing now. So that's really what I'm more interested in because it's from there that we are getting these forecasts of doom and disaster and so forth because these AIs are getting more and more capable. Not more and more intelligent, please, but more and more capable.

So I'm going to take a different tack to both of you. I think we could use as much AI as we could eat, right? I think, like in Jevons paradox, the more AI we have to use, the more we're going to use it.

I think the real question here is, do I need a sledgehammer to kill a fly? Because what we're developing is stratification in our AI models, right? I don't need the AI that supposedly can cure cancer or fold proteins or nuclear kind of physics to pick out, Guy, as you said, where's the best shoe store for me to get a pair of loafers this weekend, right?

I think as the AI market matures, we're going to see this stratification, and it's more than just picking the right model for the job. It is what model am I using and where does it live? The one that helps me pick my loafers for the weekend, it could live on my phone.

It doesn't need to go beyond that. I shouldn't be paying tokens for that or anything else. That's a small model, it could live on my phone.

And quite frankly, if we're talking Apple hardware, those models aren't so small that they can run on your phone anymore. When I'm doing vibe coding, yeah, maybe I want something in the server closet that's a little more beefy, open weight, though, so I can still code as much as I want and keep iterating. And then when I'm talking solving cancer or something that really demands the very top, maybe then I'm going to one of those data centers in the $8 trillion worth of data centers we're building and using that model where I need it.

And to me, that's AI nirvana, guys. That's where we want to go to. Yeah, I agree, and I would even go a little bit further there and say that it's not just stratification between big and small, between cloud and local, it's also stratification between sort of exactly what the model is and what it was made to do.

One of my most common uses of AI is using the incredible Whisper model to do transcriptions of audio into text and make that text more usable, searchable, indexable, that sort of thing. I do that every day, and that model is not only tiny and not only local, but more importantly, is special purpose. It is not going to tell me what shoes go with my shirt.

It is going to tell-- or how to code. It is only a model for transcribing audio into text, and it's brilliant at it. And increasingly, I don't know if we're going to get too far into it, but there's been a lot of interest in the industry in special purpose models that don't boil the ocean, that do specific tasks, do those tasks well, and serve as sort of LLM duct tape in conventional deterministic computing processes.

And I'm pretty excited about using that sort of thing. I think that having your own decision framework, let's say, and the more sophisticated you can get with that, the better in terms of what am I trying to accomplish? How do I determine-- It's almost like a development process of its own, where you're figuring out what model is the right model, where it should be hosted, what size it needs to be based on a business purpose or a personal purpose, I suppose.

Like Whisper, for example, Stephen, there are many transcription, there are many models that are suitable for transcription. Maybe if you're developing an app that relies on transcription, you need to start off with a general purpose model just for the expediency of it. But part of your development process isn't just coding the app, incorporating the service, it's determining which model you-- Right now we have model routers, right?

Those are supposed to be dynamic and up to date. The dimensions of this are almost daunting. But the ultimate aim is just enough model right now for your scale and what you're trying to do with the flexibility to change models, change scales, what have you, as you go forward.

That's one of the many reasons I think we say that we're just at the beginning of this, because that level of sophistication is getting built rapidly, but it can only be absorbed so fast by the humans, which, by the way, all of this is for, is for humans. So Guy, I don't disagree. Stephen, I don't disagree, but again, I want to put a little caveat in there.

I think we are going to have several different, again, stratification flavors, use models going forward. One is I use AI to cure cancer. One, Stephen, is I'm using AI for transcription.

I'm using a whole application that is AI-powered to transcribe. I think a bigger market may be in, I think we're seeing this with the Signal 1 Jeff stuff and with the very narrow focus, is the same way we build software now, right? Call it Frankenstein software, where we stitch together a lot of open source components and a lot of pieces that do a specific task.

And I'm forgetting, it's a company that was in Techstars out of Boulder. They did the email piece of your app, right? So when you signed up for an app and you had to get a verification email, all it does is it's the email gateway, so you can now email enable your app.

And it's in every app, not every app, but it's in a majority of the apps we use, right? We are going to see little AI pieces like that. You want to call them artifacts, but components that when we're vibe coding our apps, we will use these prefab AI components to make our apps.

The whole app may not be AI, it may only be that little piece of it, but it's going to make it a hell of a lot easier to build these things, right? And then that flywheel keeps spinning round and round. Stephen?

I think that that's a real important aspect because, and again, that's sort of the Unix philosophy. I'm an old school Unix admin. Build a tool that does a thing and is able to be connected to other tools that also do a thing.

And philosophically, I think that that really can help us all. And frankly, I don't see that that is antithetical to this sort of agentic concept. In fact, I think it's way better to have agents that are simply little tools.

I was excited about MCP at first because I was hoping for a world that looked a lot more like process automation and a lot less like monolithic LLMs that are trying to boil the ocean. Unfortunately, as we talked about last week, and as I talked about last week on Security Boulevard and so on, a lot of these agents that are being introduced are absolutely the worst possible solution because they are monolithic, giant LLMs that are intended to ingest all the data and do all the things. And frankly, that's not going to work, that's never going to work, and it's just frankly a bad idea.

Instead, we should- Oh, no, Stephen. Okay, Guy, go ahead. It's just, look, they have their place.

The frontier model and the most advanced frontier model has its place. If I'm thinking practically as an application manager or developer, I don't want to use a model that can do 90 million things at scale in production when it's only 50 things I need it to do. Same thing for agents.

But when I'm in an early stage of development, think of it as a sandbox-ready model or something like that. I'm messing around, my team is messing around. We don't know which those 90 are, and we don't want to spend the time to develop or to work up a model or an agent or whatever that has the likely 9,000 of them.

So I'm just going to go with the one that can do 90,000, 90 million, whatever it is, just to begin with. Initial scoping, if you like. I think there's a place for such things, Stephen.

But the problem that I think you're highlighting is that place is not everywhere. It's not everything. That is facile, if I may use this fancy posh word for stupid.

It is facile. Yeah . It's facile.

We almost seem to be getting conditioned by, I don't know, all these important people around or whatever, to find the easy button for freaking everything. The easy button is the most advanced frontier model. Let's find the best one.

That's what we're going to use. We're all smarter than that. We're all more capable than that.

We can spend the week, the one week, let's say horrors of horrors, one week of planning to get started on a project in a more efficient way. So it's not really a question of easy, Guy. Humans are always going to pick easy.

I think it's hardwired into our brains. We pick the easy way, the shortcut, if we can. The real issue becomes what I led the show off with, which is why use a sledgehammer when you need a fly swatter, right?

Humans are the greatest tool users that, well, that we've ever had on our planet anyway, and AI is a tool. It's about picking the right AI tool for the job, and that I think is really where we're going. There's always going to be a place for the frontier models, right?

Just like in a car, right? There are people who want to drive around. I live in Boca Raton.

There are people driving around in Rolls-Royce SUVs. Don't ask me why they have Rolls-Royce SUVs, but they do. Other people have BMWs and Ferraris and Audis, and people have Fords and Chevys and Hyundais and what have you.

We're going to have the same kind of thing with models. Use the right tool for the job. " And here's the thing.

It's different. I think it's fundamentally different and actually more liberating in AI than it is with cars. Because a car, I know you don't have to own them.

There are other options of using cars, but many people choose to buy a car. And so I don't really fault them for buying the biggest, baddest, most flexible, most capable car on the road, even if they don't use it. And it reminds me when I lived in Texas, and everybody had a pickup truck, a giant four-door, super long, extended whatever pickup truck.

And when we moved in with a Miata and a Saab, well, first they looked like we were aliens. But then they literally said to us, like, "How are you going to go to Home Depot? How are you going to go to the grocery store?

" And I don't really blame them. They were wrong. I don't really blame them for that attitude because frankly, I get it.

You've got to buy a car. You might as well buy one that could theoretically go on the cross-country journey and carry 500 bags of cement and all these different things. And so you might as well buy the most capable one because that's the one you're stuck with.

But none of us are stuck with an AI model. " But we don't have to make that choice. We can use 100 different ones.

We can use different-- We can make a special purpose choice for a special purpose. And again, that comes back to that sort of sysadmin mentality that I've got built into me. And so when I'm looking at this, I'm like, what's the best transcription model?

Okay. What's the best summarization model? Okay.

What's the best JSON extractor? Okay. What's the best...

And I can chain all those together, and I end up, I think, with a much better solution than just upload the audio file to ChatGPT and say go. And I think that's objectively true. I think it's a better solution, but I don't think that I'm really representative of the world.

Well, that's what we're here to talk about. Right? Three bits of practical advice, I just gave one.

Breathe, calm down. Alan, to your point, maybe most people are, but you can be considerate and just not necessarily go with the default. The second piece of practical advice, maybe you can use a crappy model.

Why don't you try that? What you think of... I would err more on the side of less capability and see what happens, something free.

These open source models, I think two of them came out last week, one from France, or three of them or something. There was a new Mistral, I think a Beam, which is from what, Reflection AI or something. They had a new one.

Use something free and cheap and see how that goes. And don't worry too much about it. Again, back to point number one, don't worry too much about it because if it fails, then you learn something, right?

And then the third bit of practical advice, I think is, essentially to your point, Steven, there are lots of approaches to this. What I take from you, Steven, is the sysadmin in you lives still today. You take that, it's like we have new tools and new capabilities, but we don't have to change how we approach problems fundamentally.

Don't feel like you're missing out or you are behind the times or whatever. Guys, Guy, your advice is missing one key piece. There's no polygamy or monogamy rules with AI.

You're not wed to one model. And it's not like buying that Rolls Royce either, where I'm plunking down a few hundred thousand dollars. I could, and I think we all will, we will have multiple models.

The same way we talk about we're going to have 10X agents per person, 100X agents per person, we're going to have 10X and 100X models per person. And we may go into our tool chest and pull out the model I want for this job, then put it back in that drawer. I may not open that drawer for five more years and pull out the next model for the right model for my next job.

That's the beauty of this whole thing. We're going to have all these different agents that do different jobs. We may have a master agent that kind of manages them.

We're going to have models that these agents will plug into via harnesses and stuff. That's the beauty of this. We're not wed to a ChatGPT or Claude or whatever comes next, or one of the Chinese open weights or any of them.

We're going to use the best tools for the job. And I would take it even further and say that the structure of the AI industry as it stands, OpenAI and Anthropic and Google and Microsoft and so on, the structure of the industry as it stands and the way that they're rolling out their products, I think is fundamentally going to drive users in that direction. You have heard many people say that these companies don't have a moat.

In fact, Anthropic's recent IPO, pre-IPO filing, they spend a lot of time talking about the fact that their users can abandon them at any time and switch to another competing model at any time. And it's because they fundamentally made a huge mistake in not engineering a system that has any kind of a moat, because it doesn't. And every time you start a new chat with ChatGPT, you are crossing a moat.

Every time you use a different tool, you're crossing that moat. And everybody is learning, I think, that there is no moat. And furthermore, and this is the thing that drives me absolutely up a wall, even if you want there to be a moat, you can't get one.

So for me, I'm trying to do business process automation, and I'm wiring it to ChatGPT. Well, guess what? 5 and then et cetera, et cetera.

They take these things off the market, which means that my tools that are hardwired into these things, they're going to break after six months. And so far, the model routers haven't been able to solve that problem. " And unfortunately, that broke, too.

And so I think what they're- Bedrock helps there ... they're teaching us, I think, not to be loyal customers. They're teaching us to be polygamists and to say, not even polygamists.

They're teaching us to just be- They're teaching us how to be swingers. I didn't want to go there, Guy. I'm glad it was you and not me.

But they're teaching us to have no loyalty whatsoever. And Anthropic is right. Sorry, dude.

I'm not married to Claude, and tomorrow I could decide bye-bye and never touch it again. And I would be fine with that, and they probably shouldn't be fine with that. But that's the market.

Yeah. If they make the best tool for a job, you'll use it. Assuming economics, everything, right?

That's the market at play. And this, not to harp on it, but this is exactly what I say in my book. As this becomes more indispensable, as more people use it, we will gravitate towards doing it the cheapest, most efficiently, easiest way possible.

That's human. And ultimately, while we talk about AI taking over the world and making us extinct and everything else, it's our own humanity that's going to let us use this the best way possible. We'll use the right tool.

Let me point out where I think you're missing a part of this, Alan. Which is, because when you think about, I guess when I was talking about we all want to do the easy thing, the easy button, find the easy button. And you, I think, wisely sort of corrected me in that.

That's human nature. It's not particular to this. Here's the thing.

How often in our lives do we take some sort of easy route that turns out to be more complicated later, as opposed to doing something a little more complicated now that makes it a lot easier for us later? This is part of, obviously, maturing and having more experience and understanding. I think maybe that's my basic point here, which is this is, and we're describing it here, an almost insanely complex, super rapidly developing, hard to really put your finger in a lot of places on what's going on type of technology and development.

And so, you can be forgiven, like you say, for trying to find simple ways to move forward. The issue and the question is that you want to ask yourself, and maybe the pause I'm asking you to take is, think a few steps beyond. It's not that you don't already, it's just that getting that sort of understanding of this technology and market as it's going might mean doing something harder now that's going to make it a little easier for you later, Steven.

MCP or the model routers or that sort of thing seemed like simplifications that actually turned out to create great complications as well. It's the time factor I'm talking about. It's the time factor.

Yeah. So I think as we get better with these model routers or harnesses, as a lot of people talk about harnesses, we're going to leave that to the AI, too. You're going to tell the AI harness, "This is the job.

" Right? Why not? They're interchangeable at some point.

There's no moat. And I would love to see more of that. Frankly, I have to say, I was deeply disappointed with OpenRouter because to me, it didn't solve the fundamental problem that I wanted.

What I wanted it to do was I wanted to be able to say, "I want the cheapest model that will quickly get this job done, and I want it to work forever. " It didn't do that. It didn't do anything like that.

And frankly, that made me abandon it after just a short time of trying it, because what I would love is the ability to just basically say Just fricking work. Will you just fricking work, and will you make it cheap? And frankly, that has been something that I think the platform companies have maybe done a better job of and maybe have a better idea about.

You look at what Microsoft and Apple and Google are doing in AI, and for the most part, you're not making choices. You're just using the thing, and then the thing does the right thing in the background in a way that works. So for example, the new iOS and macOS 27 have shortcuts built in.

That's their automation framework. I don't mean shortcuts like keyboard shortcuts. Anyway, the Shortcuts application lets you use Apple Intelligence to do things.

And when you build a shortcut that includes Apple Intelligence, it will do that thing either locally or in the cloud with whatever model, and it'll just kind of do it. Now, I will say that they actually broke that between version 25 and 27. But be that as it may, I think they didn't intend to break that.

And I think that increasingly that's maybe how people are going to use these things. Essentially, they're going to be in Google Workspace or in Zoom or in Microsoft Windows or whatever, and they're going to have AI do a thing for them, and they don't care where it's running or what it is or anything. It's just going to do it.

And I kind of like that idea. I would love to see this all be a lot easier and a lot less monolithic and a lot less centric around the big frontier model companies. But we're kind of getting to the end of the conversation here.

Guy, I want to throw it to you first. What is your advice to people who are thinking about doing too much with AI or not doing too much with AI? " There's so many choices, so many options.

I don't think that theme has changed, but the prescription from back then was, well, try something, do anything. Almost every route is potentially profitable. At least you'll learn something.

So my feeling out of this discussion where we've reached right now, it's kind of the same, but maybe keep in mind all the forces out there in the market compelling you towards the latest, hottest frontier models, the latest, hottest technology. Think about that next step, whatever it is you do, whatever direction it goes in as, at the very least, a learning experience, but also quite likely, even if it's cheaper, lower key, more experimental, to be really quite productive. And keep your head down, keep your nose to the grindstone, and if you can do any work in that position, then you're likely to be productive in AI.

Daniel Newman wrote the foreword to my book. I can't think of a bigger bullish person on AI than Daniel. And my position in the book, and I make it clear, is I am very bullish on AI.

I think it's going to be one of the greatest technologies, greatest tools we've ever invented. But we've got to remember, we're two years into this. Think about where we were with the first two years after the very first computing machines or any other major technology.

How long did it take the internet to go from DARPA to the internet, right, and commercial and all of that? We're not even at the end of the beginning. We're still a little bit past the beginning of the beginning.

And I'm excited every day to wake up and see what's new because there is so much new, and the market is going to shape this tool for us to use it the most efficient ways, the best ways, because that's what humans do. And if I had to sum it up, that's where I am. I think the good news is this is evolving so quickly, right?

We barely touched on this, Jeff, and that whole philosophy where purpose-built, single kind of question answers. There's going to be so many evolution, right? Like a Cambrian explosion, if you will, of AI in the next couple of years.

I'm excited to see it. And I will agree with you on that, Alan, amazingly enough. I will agree with you, except that I expect that this Cambrian explosion is going to result in a lot more little muskrats and mouse and bat type creatures than dinosaurs.

I think we're going to see an explosion of cool little alternative things that actually have some practical use instead of this weird pseudo-religious belief that we're going to build a mind out of- Super intelligent ... a text prediction machine. So we'll see how this all plays out and if I'm the one on the bad end of the text prediction mind on its naughty list.

But we do have to wrap up. Before we go, thank you for joining me on this podcast. Before we go, tell me where can we continue this conversation, and where have you written, what have you written, and where can we read more?

Guy? com/in/guycurrier. social and on X at Guy Currier's feed.

Not on Mastodon yet, Stephen. Oh, and also, I will be at AI Infrastructure Field Day Number 6 in just a few weeks' time, the first week of November. So look for me there.

I'm not hard to find. Just look for Shimmy. But I do write an awful lot on Techstrong across the entire portfolio.

Well, this week I'm back out in Austin for Qualys Rock On. Next week I'm at the Couchbase user event. The week after that I'm at Security Field Day at Santa Clara.

And then I think two weeks after that we're at KubeCon where we're doing our first Futurum Summit at KubeCon Monday morning. Google it. I don't have it in front of me, but you can go register for the Futurum Summit Monday morning.

Great way to kick off your KubeCon week, and we'll have a Tech Field Day at KubeCon. Techstrong will be there broadcasting live all week from the show floor as well. So excited.

Then I come home for Thanksgiving and right back out into Vegas, baby, for re:Invent. So it's going to be a busy couple of weeks. Do check out my book wherever you buy books, "Indispensability Trap," and you can watch, if you like this kind of thing, Techstrong Gang every weekday at noon Eastern Time.

I was just going to mention that because you'll actually see the three of us on Techstrong Gang, right, Guy? Yes. The three of us on Techstrong Gang in the coming weeks.

I'm usually on on Mondays. Alan, I think he's on all the time. I don't know.

Not really all the time. Not while I'm traveling. But Mike Azard is on there, and just some incredible gang members.

So do check out Techstrong Gang live streaming every day, every weekday at noon Eastern Time. Also, I will mention that it doesn't look like it, but I am currently at AI Field Day. We actually recorded this ahead, so I'm not actually there.

But be that as it may, today and tomorrow, Wednesday and Thursday, are AI Field Day. We've got some great presentations from some companies, but also from individuals in the community, including me, and maybe some curveball surprises for you. One of the things that we're going to be doing at AI Field Day, in fact, is we're going to be recording a couple of live roundtable format episodes that we'll publish in this podcast feed in the coming weeks and months as a way to just bring more voices in here.

It isn't just Futurum people that join us here. It can be almost anyone. And we really enjoy bringing in independent voices in everything we do, and so we're going to be recording a few episodes of that as part of AI Field Day.

So thank you for listening to the "Utilizing AI" podcast today. If you enjoyed this discussion, please do subscribe. You'll find us on YouTube or in your favorite podcast application, and consider giving us a rating or a review.

This podcast is brought to you by the analysts and experts from the Futurum Group, where insights meet AI. ai, the "Utilizing AI" YouTube channel, or the Techstrong TV app. Thanks for listening, and we'll catch you next week.