Utilizing Tech - Season 7: AI Data Infrastructure Presented by Solidigm - 3x17: Why the Heck Even Use Deep Learning?

Episode Date: January 11, 2022

Although it’s a powerful tool, deep learning is perhaps over-used in modern applications. In this episode of the Utilizing AI podcast, Rich Harang joins Chris Grundemann and Stephen Foskett to discuss the various reasons people use AI, both good and bad. In a November Twitter thread, Rich posited that the following conditions were required to use AI for real: The cost of errors must be extremely low, the decision needs to be possible but expensive, there needs to be the same kind of decision frequently, there needs to be a benefit and be better than a simple rule, you have to not care how it got the answer, the base rate must be close to even, you need a steady stream of data for training, and you must match the size and cost of the model to the application. On the other hand, these same considerations can point us to problem sets that make a great match for DL, and we should focus on using the right tool for the job. Three Questions: Chris Grundemann: Are there any jobs that will be completely eliminated by AI in the next five years? Stephen Foskett: How small can ML get? Will we have ML-powered household appliances? Toys? Disposable devices? Adam Probst of ZenML: What percentage of companies will be using ML in five years? Links: Rich’s Twitter Thread: https://twitter.com/rharang/status/1465340190919217153 Sara Hooker’s Paper, “The Hardware Lottery”: https://hardwarelottery.github.io Gests and Hosts Rich Harang, Senior Technical Lead at Duo Security. Connect with Rich on Twitter at @RHarang. Chris Grundemann, Gigaom Analyst and Managing Director at Grundemann Technology Solutions. Connect with Chris on ChrisGrundemann.com on Twitter at @ChrisGrundemann. Stephen Foskett, Publisher of Gestalt IT and Organizer of Tech Field Day. Find Stephen’s writing at GestaltIT.com and on Twitter at @SFoskett. Date: 1/11/2022 Tags: @RHarang, @SFoskett, @ChrisGrundemann 

I'm Stephen Foskett. I'm Chris Grundemann. And this is the Utilizing AI podcast. Welcome to another episode of Utilizing AI, the podcast about enterprise applications for machine learning, deep learning, data science, and other artificial intelligence topics. Each episode, we've asked many different questions from the specifics of technology to implementation details to talking to people who just got us thinking. And that's what's happening this week. Effectively, one of the things that is just always in my mind is, why are we using AI or ML or deep learning or this specific technology for this specific approach?

And we don't always know exactly why that was chosen. And Chris, I mean, sometimes it doesn't fit, right? That's absolutely right. And I think this plays a little bit to the issue of how big AI really is and how many different algorithms and different models all fall under the idea of artificial intelligence. And a lot of people want to talk about very specific things when they say AI, even though it's a general term. And I think deep learning is one of those very specific models and algorithms that we're using currently. And sometimes the answer is we use the hammer we have,

and sometimes we're using the actual correct tool. And so I think that's probably what we'll dive into here and understanding that there are other tools available and when tools are most valuable. So back in November, there was a question. Eric Lincoln posted a question in a sentence, what is AI good at? And Rich Hrang responded, well, hello there. I have some opinions. And boy, did he.

Rich's thread was great. And the whole topic of the thread was, why the hell are we even using deep learning for this particular application? And yet, it wasn't just a rant. It was actually a very smart summary of a lot of the questions that we've had. So that's why I'm thrilled to introduce Rich Harang as our guest here on Utilizing AI This Week. Hi. Yeah. Thanks for having me. I'm looking forward to the conversation. So yeah, I'm Rich Harang. I am a senior technical lead at Duo, where I lead a lot of the research into sort of basic algorithmic developments that we use to improve our product. Previously,

I did data science and machine learning at Sophos as part of the Sophos AI group. And before that, I worked on the intersection of machine learning, privacy, network security at the US Army Research Laboratory. So Rich, in your thread, like I said, it really summed up a lot of the things that I've been thinking about with AI and a lot of things that Chris and I and Frederick and I have talked with our guests on this show, but it really all comes down to the whole question of when all you have is a hammer, everything looks like a nail. And right now, everybody's favorite hammer is machine learning or deep learning specifically. And it seems like everybody's trying to do everything with deep learning. Is that how it looks to you? I think that there's, I think it's overused. I think it's used more often than it should be.

But I also think that there's a certain marketing cachet that comes with being able to say, oh, we use like the latest, greatest technology to do the thing. And there's a, there's a meme going around at one point where it's got like a dog and it's like, you know, deep learning and then like it pitches to the side, it's like machine learning. And then finally, like behind everything, it's, you know, just a picture of a dog. It says, this is just an if statement. So I think like a lot of the time when people talk about, oh, we're using AI or we're using deep learning to do this particular thing, I suspect in practice, a lot of the grunt work is still being done by

simpler, more robust techniques on the backend. But it's certainly talked about way too much, I think, or out of proportion to how useful it actually is for a lot of problems. So when you say that we're kind of over-marketing or over-hyping deep learning, maybe at the expense of talking about some of the more practical and hardworking algorithms that are actually doing a lot of the work behind the scenes, what are those other types of artificial intelligence or are there even types of artificial intelligence that is out there doing the grunt work? Oh, artificial intelligence is not a term that I think is really well-defined in a lot of cases. The stuff that most often gets used, like if you wanted to pick sort of the data science technique that gets the most traction, that gets the biggest bang for your buck, it's going to be like an SQL query, right? SQL queries and Excel spreadsheets drive business. And so if you want to stretch to the point

where you're saying, oh yeah, sure, maybe that's a kind of AI, you can say, yeah, we use AI all over the place. Once you get past things like SQL statements and, or SQL queries and Excel spreadsheets, even just simple linear models carry a lot of weight. I think the next step up from that is probably like random forests. And so by the time you get through that all the way up to like these deep learning, you know, 10 million parameter models, I think you're sort of looking at a very small part of the solution space that's left,

at least that actually requires them as opposed to where they get kind of shoehorned in because it's easy and it seems, or it seems easy and it seems fun. And it seems like, you know, oh, why wouldn't we use the most powerful tool available? As opposed to like, what, what can we actually do to generate or to generate value and avoid cost? So we're essentially talking about AI washing here to some degree, right? Where, where we're using common data science tools that have been in use for a long time. Maybe we have more data or better data than we had in the past. Maybe we don't. But then we're coming out client facing and talking about deep learning and neural networks and machine learning because that's what the audience expects us to hear.

Or is there some other fundamental reason that that's kind of being used as or being tout touted as, as the, you know, the power of the machine. Yeah, I think it is to a certain extent, AI washing, right. And if you want to say, Hey, these are the, like, nobody ever wants to be like, Oh, we, you know, we solved this problem with a 20 year old technology and it does great. Right. Nobody gets excited about that. Um, when maybe they should, because you can use the 20-year-old technology. It'll cost you 10% the maintenance burden and get you 85% of the effectiveness numbers completely made up, but in that ballpark. But there's not a great press packet you can put together through that, right? Like no one's going to get excited for like, we had this problem and we solved it with linear regression, right? Whereas if you can turn around and say, oh, like our next gen super duper, you know, self evolving, you know, non deterministic quantum, pick your buzzword AI, right? This is what we're using to do the thing. And everyone's like, oh, well, it must be good. In your thread, you posed a huge number of

requirements for AI. And I think that was pretty smart because essentially what you were saying was we need to figure out what problem set is best served by AI and deep learning. And some of those things, I think, are things that people can immediately grasp. For example, the decision needs to be something that is expensive for a human to double check, but possible for them to double check, like select all the buses. There needs to be a lot of the same kind of decision. As you said, zucchini versus volleyball, that's not just something you're going to be doing all that often. And there needs to be some benefit from making the right decision. I think that these are some of the questions maybe that people have.

And so when you're looking at things that we've talked about here on the podcast, like, for example, intrusion detection, you're making a lot of the same decisions repeatedly, there's a huge benefit to getting it right. A human can go back and check those things and say yes or no. So it's a very good application of machine learning. But of course, there are lots of them

that kind of aren't great applications of machine learning, and those tend to wash out, right? I mean, what do you wanna say? I guess, first of all, let's talk about decisions. You know, what is deep learning really good for? So I think, I mean, I don't want to like undersell deep learning in general, right? Deep learning is really good at taking huge amounts of data and essentially learning extremely complicated. And usually often there's some caveats and some wiggles we can get into around it. Usually like very high performant rules and, you know, sort of classification strategies. The two questions are, what are you, what do you have to pay to get that? Because training models, you know, the care and feeding of deep learning models can be horrendously expensive. And then is it worth the effort, right? If, if you're, if you don't actually need, you know, the four nines of performance that you can squeeze out of a very highly tuned, carefully researched, super large deep learning model and it may solve the problem you need it to solve,

but you're using way bigger guns than you actually need for the problem, right? And again, you can get by with something that'll get you 85% of the way there at 10% of the cost. And so that's, for me, what I see as the biggest disconnect, really being able to diagnose them. You have, again, just the tremendous cost of care and feeding and retraining them. And then also these deep learning models tend to be large and somewhat expensive to run, especially compared to much simpler models like linear regression. And so that actually limits the places that you can deploy this, right? If you want to deploy it to a mobile device, a deep learning model often can be challenging to do. Whereas again, like a simple linear regression or a logistic regression, or even a,

you know, a smallish random forest tend to be much more compact and much faster. And I see this a lot with IoT and, you know, devices like that too, where, I mean, nowadays we've got, everything has to have, you know, person detection and all that kind of stuff. A motion detector can be implemented with a few discrete electronic components, you know, person detection, all that kind of stuff. A motion detector can be implemented with a few discrete electronic components. You know, do we really need a deep learning model to identify that that's a human when you know, the little red thing in the corner, like, triggers the thing, just whenever there's motion? There's, there's, there's a lot of use cases like that, that I think that people are maybe going

a little overboard with deep learning. And then of course you mentioned rules, which is another thing that always gets me. In many cases, an expert can come up with some pretty good rules that we can use. I mean, we've been doing this in firewalls and so on for a long, long time. Do we really need an opaque black box machine learning system to decide if an expert could have just thought about it for a month and come up with a bunch of rules that could basically be more predictable, more understandable? Of course, there are benefits to it, but these are the questions I think that we need to be asking, you know, what kind of problem set are we dealing with? Right. Right. And I think like, I guess like the thread I do come off as highly critical of,

of deep learning and, you know, sort of these black box methods. I think there is a place for them, right? There's definitely, there's definitely's definitely places where there's almost nothing else that fits the bill, right? If you're doing facial recognition, people spend a lot of time with hog features and sift features to try and get facial recognition working. It wasn't until we got these deep learning methods that we really started to get the super high levels of performance that we see today. You know, most of like the really successful cases I can think of for deep learning land in sort of that image, image-ish kind of space. You know, you can find other places where it works really, really well.

But the number of places where you could get away with something that's simpler and cheaper and easier and more interpretable is, I think, I think a lot higher than at least sort of the tech press would have you believe, right? A lot of, a lot of the time, and I think a lot of the time, like within business, people realize that, right? When people start to look at like the cost of spinning up, you know, a full deep learning stack and, and, um, very often they, they sort of look at it and they blanch a little bit and they're like, Hmm, so, um, how about those, uh, how about those SQL queries? Right. And then even, even like beyond that, right. there's a lot of times where not necessarily with deep learning, but you can use a combination of time, you know, dredging through the data using different, you know, sort of machine learning or statistical techniques to analyze the data. And at the end of it, you distill it down, you know, maybe in consultation with a subject matter expert or using your own domain knowledge, you distill it down into a couple of rules. And then

what you deploy, quote unquote, is a bunch of if statements written in code. And everybody knows exactly what's going on with all of it, right? Whenever it hits a decision, you know why it made that decision. And if you end up with a wrong decision, then you can start looking at when you maybe need special cases or exemptions or what sort of edge cases it's failing on. And all of it is eminently debuggable and eminently understandable. And in a lot of cases where you're dealing with compliance issues, right? I don't remember if I touched on that in the thread or not, but you've got this whole space

of like regulatory and compliance issues in regulated industries where if you use a deep learning model, you could be in real trouble because this comes back and says, hey, you appear to have some sort of bias against this group or you need to prove to me

that you aren't making these selections in a biased manner. And if you have a deep learning model, you're kind of out of luck at that point. There's not a lot of great interpretability stuff that goes on in that space. Whereas if you've distilled all this data down to a chain of 15 if statements, then you're like, okay, well, here's exactly what's going on. Yeah. Here's the decision tree. Yeah. You did say you have to not care

that much about how you got to the answer, only that, you know, you got an answer. Yeah, well, and this reminds me a lot of, you know, just working in IT in general, right? And so maybe to kind of zoom out of the problem space a little bit, I've definitely had lots of times where, you know, somebody came to me and said, oh, we've got a BGP issue.

And really it was their Wi-Fi on their laptop was turned off, right? Or somebody coming and saying, hey, the server's down. And really, you know, oh, we've got a BGP issue. And really it was their Wi-Fi on their laptop was turned off, right? Or somebody coming and saying, hey, the server's down. And really they've just, you know, they don't have the application on their laptop, you know, and all these excuses. And what I mean is it's a little bit of this area of the tail wagging the dog.

And so I think anytime you're approaching a problem space in the technology world, but probably outside of technology as well, it's really good to understand, you know, what are the problems we're actually trying to solve? What are the requirements? What are the constraints? And then let's go out and solve that problem versus I think what I hear Rich saying is there's a lot of cases where folks are kind of backing into this saying, oh, we need to find a way to go deploy deep learning. Let's try and find problems to attack with the solution. In a lot of cases, right, this is a new hot thing.

We need to find ways to use this versus, hey, we have these problems. Let's figure out how to solve them. And I definitely see a lot of that in a lot of different areas of IT. And this just seems like one more area where we should be very careful to approach problems with the right tools versus approaching our business with tools and trying to find problems. Yeah. And that being said, of course, we can take this entire discussion and flip it on its head and to say, how do we identify places that we should be applying deep learning? And all of these things are really great reasons to apply deep learning. So, for example, as you mentioned there, if we have a problem set that does have lots of yes, no questions, where it's kind of an even balance between yeses and nos, where it's a similar decision happening again and again, where we've got a lot of training data, where we don't really care how we defined it.

We just need to decide whether we're going to open or close, things like that. I think that those are great applications for deep learning. And those are the kind of things that we should really be looking for, right? Yeah. And I also, I think I had that whole list of different things that make a problem a good fit for deep learning. I think I want to emphasize, it's not like if you don't have all of these things, you just give up and not use deep learning. I think that the more of those attributes you have, the more likely you are we were looking at like malware, right, detecting malware with deep learning. And in that case, the, like the base rate is unbelievably skewed. We're, we're nowhere near even odds. Um, but you can still use deep learning really, really successfully

there, which, you know, Sophos has a great anti-malware product, which deep learning is a significant part of. You know, I'm only slightly biased because I spent a lot of time working on it, but I think it's like a place like that, right? There's tons of decisions. You need to look at lots of files, right? It's hugely imbalanced. And that gives you an incredible headache when it comes to saying, okay, how well is this actually working? But it may be the case that the value you get from that model is worth that headache. And I think that's, you know, if I had to boil it, that whole thread down to one thing, it, I think it would be kind of what Chris was saying. Use the tool that's right for the job and just be aware that deep learning is sort of

like a specialized, expensive precision tool. It's not really universally, I mean, it's a weird case because you can say, oh, it's universally applicable, right? It's this class of like universal function fitters and theoretical results that like, yeah, I can learn any kind of function that has these properties and blah, blah, blah. Yes. Okay, fine. In that sense, it is, it is like the universal tool, but is it really like cost effective? Are you going to like save yourself money by jumping right to deep learning or should you sort of start small and work your way up until you finally decide like yeah i absolutely have to get that that last bit of tail error squashed and i and i know that if i do that i'm gonna come out ahead in the game

right and yeah i i've i've thought about the costs of like implementing it and the costs of like retraining the model and the cost of like curating the data and getting this huge volume of data, like brought in and analyzed and labeled and so on. And at the end of all of that, I can, you know, even with all of that, it is still worth it to me to pay for all of that, to get this last, like squeeze these last little bits of precision out of the model. Then by all means, you should be using deep learning. But like Chris said, I think in a lot of cases, it's someone comes up with a cool model and they're like, okay, crap, how can we monetize this? Or they're just trying to get kind of on the hype train

and say, hey, we use AI, right? This is where an AI driven startup, right? And the question is, okay, do you need deep learning to solve this really hard problem in a way that nobody else can solve it and this is the only way to do it and now you have sort of like this edge over everyone because you've built this technical moat that's only solvable by deep learning and nobody else wants to get into it or are you using a shiny object because you've got you know research scientists on on staff who want to use shiny objects and play with the cool new toys? Yeah, I think that goes all the way back to, you know, some of the things you were saying in the very beginning, Rich, which is, you know, that this isn't just, you know, what type of artificial intelligence is best for this use case.

But that there's some really solid data science that works in a lot of cases, and you don't even need to go to artificial intelligence, right? I mean, and you mentioned SQL queries and Excel spreadsheets. And I think that to your point, right, that they are the unsung heroes of data science. And then there's a lot of things you can do with data that doesn't involve any kind of artificial intelligence whatsoever. Is that right? Yep. Yeah, 100%. And, you know, it shouldn't, it shouldn't be the first thing you reach for, it should probably be close to the last thing you reach for. It's, you know, when all else fails, and I'm still convinced that, you know, I need it, and I can extract value from it, then okay, well, let's, let's think about

jumping into the, into the deep learning pool. But yeah, again, like the most, I don't know if I can say what it is, but the most single most impactful model I've ever written, and I put model in scare quotes there, was about a half dozen if statements and then a summation you know it's and that didn't come from like burning my own personal you know chunk of the rainforest down in gpu hours that came from a lot of jupiter notebooks and and pouring through data and talking to subject matter experts and trying to like fudge together something that made sense. And then we made something that made sense and we deployed it and it worked great.

And it's amazing what you can do with a few kind of if-then statements in a row. I mean, basically it's the old expert system, right? If you can divide up the problem set and then divide it again and divide it again and divide it again in just a few questions, you can get pretty far in many cases without having to have this black box trained to basically find that answer for you. But at the same time, there are many areas where you can't do that. And I think that that's the important thing. One more thing that occurs to me though, is that some of the answers to these questions are going to change over time. So one of the challenges that you present is the sort of kind of fitness to purpose in terms of, you know, is the model too big for the problem? You know, are we using a sledgehammer to put a tack in the wall? Right.

Well, what if technology changes and suddenly we've got lighter sledgehammers, what if we've got the ability to do this on mobile devices or IoT in a way that we don't maybe now? That could change the equation quite a lot, couldn't it? Oh, for sure. And there's cool stuff coming out in different kinds of hardware accelerators for doing deep learning or deep learning-like tasks. And if any of those... So the faster those develop, the more developed and the cheaper those get, and the lower they bring the cost of deploying these things and the more ubiquitous this kind of hardware becomes,

then I think the lower the bar gets for how hard does the problem have to be, how much value do you have to extract from it to be able to use it? What it does not solve is the backend, the data. How do I get the data? How do I get the labels? How do I get this curated and cleaned and into my system for training? That is not a hardware problem. That's in a lot of ways, it's a domain expertise problem. And it's a lot of engineering problems. you know, a lighter sledgehammer will necessarily solve, but on the hardware front for deploying it, absolutely, right? The better the hardware gets, the easier it becomes to make the case for using it because, hey, I can get this really powerful model for which I can deploy everywhere, right? Now my fridge can do facial recognition if that's a thing that you think it's really good for your fridge to be able to do. But then I think there's also, I don't know how strongly I believe what I'm about to say.

It kind of like waxes and wanes depending on the day. I think there also is a case to be made that the hardware that we're using is actually in some ways giving us tunnel vision with respect to what we do in the space of what kinds of models we fit, what kind of machine learning we use. There's a paper called The Hardware Lottery. It escapes me at the moment. I deeply apologize to the authors i i'm i'm blanking

on the names of the people that that wrote this paper but that essentially is the the argument they make that hey we've got these gpus right and it's not a huge surprise that deep learning and back propagation and all this stuff started taking off at about the time GPUs became commodity accessible and the sort of the infrastructure to do these GPU based computations became widely available. Suddenly this became an easy way to build really big models and a lot of other trains of thought kind of fell by the wayside. And suddenly it was just everything is backprop driven models. And so in some ways, are we sort of victims of our own success here, right? We've got the hardware that gives us the cool models.

The cool models inspire us to build better hardware, better hardware, and we're just sort of like spinning around this very, you know, dare I say it, a local optima perhaps of back prop models that get trained through back propagation and play nicely with like sort of these parallel, simple parallel computations you can do on a GPU. So in some ways, I think if we had hardware that worked really well with different classes of models, that might sort of open stuff up to new and interesting techniques within the machine learning research space. I think that you're referring to Sarah Hooker's paper, likely. But what I was going to say is that such is the nature of the history of technology as well, in that the tail we tend to continue to go in that direction. I mean, I think that's why, for example, we're continuing to use internal combustion engines everywhere. It's just because that technology has matured to such an extent over a hundred years that even though it's not

the right technology for many, if not most problem sets in transportation, we're still using them because it's well-proven. And I think that we may end up going in that direction with this technology as well. Well, it's not just that. It's also the entire infrastructure that gets built around it, right? Because once you've got internal combustion, once you've got roads that work great with internal cars that run on internal combustion, there's built to withstand the speed range that internal combustion engines travel over. And you've got gas stations and you've got gas pipelines and all of this gets built up. And it makes it, it's almost sort of like a lock-in, a technology lock-in, right? Because to move to something like electric vehicles, right? You've got to build this whole

parallel infrastructure of like charging stations when we've got all these gas stations already. So I think the analogy does kind of hold because if we focus on these backprop, GPU driven models, then we build stuff that runs them. And then suddenly these are the devices that any model you want to deploy has to run on. And so it, Rich has not been prepared for these questions, so we're going to get his off-the-cuff answers right here and right now. We're also going to change things up again here in Season 3 by having a question from a previous guest asked of Rich as well. So, Chris, why don't you take the first of the three questions? Sure. Thanks, Stephen.

So, Rich, we've talked a lot about some of the things that maybe machine learning and deep learning and artificial intelligence in general won't do or can't do or maybe shouldn't do quite yet. I want to flip that over and ask you, do you see any jobs, any roles that are going to be completely eliminated by artificial intelligence in the next five years? Honestly, I'm going to say no. I don't, I can't think, that doesn't mean there isn't one. Maybe I just don't have the imagination to come up with it. I can't think of anything that will be completely eliminated. I can think of some things that might change or will be impacted, right? You can imagine a world where at some point, you know, facial recognition and image processing and video processing and anomaly detection get good enough that things like, you know, security positions where you've got people watching cameras all day, right? Those will be significantly impacted and you see a reduction in the number of people required to do that sort of work because

they're sort of enabled and made more efficient by some sort of like machine learning driven technology. You know, I periodically hear, maybe this is just like the weird corner of Twitter that I'm dragged into, but you periodically hear things like legal discovery and paralegals are going to years, I cannot imagine a world where machine learning or AI or whatever you want to call it gets good enough that you can completely take humans out of the picture. You can think you can come up with stories where like sometimes humans end up in more like supervisory roles where they're sort of like double checking what the machine learning does and keeping an eye on it to make sure it doesn't go off the rails. But if for nothing else other than just like purely compliance reasons, right, you got to have someone to blame if something goes wrong. I have a hard time seeing a class of jobs just being completely wiped out. Well, thanks for that, Rich. That's one of our classic questions. Another classic question I'm going to throw to you,

since we did talk about this a little bit during the episode, how small can ML get or deep learning? Will we have household appliances running this stuff? How about toys? How about disposable devices running this stuff? How about toys? How about disposable devices running deep learning? It depends on what exactly you want to put around or what sort of box you want to put around deep learning. Right now, we've got sort of this vague notion that, oh, yeah, it's like these very deep multi-layer perceptrons. Um, and so like,

that's sort of what a lot of people stick on, on deep learning. Um, what about just like a single, like a single hidden layer? Is that still deep learning or is that, you know, just a, just a perceptron? Um, so I think you can have machine learning. You do have machine learning models that are very small and are already embedded, right? Like you've got face detectors and in consumer grade cameras, getting something like, I don't know, GPT-3 onto a disposable device seems like a long shot to me. Can't imagine a world where that's coming up anytime soon. But if you're willing to sort of think outside the box, going back to the hardware thing, right? If we come up with new and interesting kinds of hardware, there's a company, I'm going to say LightIO, maybe, that's looking at like essentially using optical techniques to get really complicated features,

pseudo random features built out of an input stream using optical techniques. And that gives like an immensely rich feature space than that they slap much more straightforward machine learning models on top of. So you can imagine maybe some sort of hardware-based approach like that gets you something that kind of acts like deep learning and a much smaller footprint um that i could i could see that conceivably happening you know you can keep invoking moore's law and say oh things will just keep getting smaller and smaller and faster and faster but i think we're kind of topping out on that and we're faking it by just adding a lot more like hardware to things. The nature of deep learning models is going to make it hard to get them sort of ubiquitously deployed,

barring sort of massive improvements in hardware, which could very well come. But all of that, again, comes with a corresponding significant increase in the cost of like fabrication of those chips and research of those chips and stuff like that. So all of which is to say I'm not super, super up on saying, oh, yeah, we'll get, like, disposable stuff right around the corner um if i had to put money down i would say there's going to be like some sort of like different kind of hardware architecture that lends itself more to like different kinds of techniques and maybe it'll still have sort of like the power of deep learning models and you can squint and say yeah it's kind of deep learning ish but uh yeah like with our current sort of like hardware paradigm, if that's the right word. IoT devices, sure. Fridges, toasters, sure. If that's a thing that seems like it's exciting to somebody, right? Again, I don't know why you would want facial recognition on your toaster, but sure, maybe you can put it there. But yeah, disposable devices seem like a

long way off to me. I'm not a hardware expert by any stretch of the imagination, so take that opinion for exactly what it's worth. Well, thanks, Rich. And then finally, our final question here, as promised, we're going to use a question from a previous guest. Adam Probst is one of the co-founders of ZenML. Adam, take it away. Hello, I'm Adam from ZenML. And my question is, how much percentage of companies will be using ML in five years? In five years. So if we're going to distinguish machine learning from deep learning, and so anything like starting at linear regression, right, or didn't use some kind of machine learning somewhere in its business in five years. I suspect the percentage with sort of this loose definition of machine learning is extremely low even now, right? People are, you know, everything

from A-B testing to like things like looking at customer health and customer churn to, you know, recommender systems, right? There's the number of applications for machine learning for like sort of quickly summarizing and analyzing data and making it actionable. It's, you know, huge, huge, huge, just areas of application for all of this stuff. So yeah, I mean, making it their primary product or like integrating it directly into whatever it is they're selling, maybe a slightly different question of using it somewhere within their business operations. It's got to be close to 100%. Well, thanks very much, Rich. And again, thank you for joining us here

on the podcast today. We do look forward to your question for a future guest if you happen to have one. And to our listeners, if you want to be part of the fun, just let us know. Contact us at host at utilizing-ai.com and we'll record your question for a future guest. So Rich, thanks for joining us.

Where can we connect with you and follow your thoughts on enterprise AI, deep learning, machine learning, data science, other topics? Well, you can find me, I probably spend most of my time on Twitter. It's just at rharang, R-H-A-R-A-N-G. And fair warning, I do tend to post a lot of really terrible memes and bad jokes. So

if the thoughts about enterprise AI don't attract you, maybe the memes will. How about you, Chris? What's going on in your life? Yeah, everything is over at chrisgrundemann.com. You can also follow me or chat with me on Twitter at Chris Grundevin. And LinkedIn is another place to connect and chat about anything IT or AI related. And as for me, I'm looking forward to our next AI Field Day event, which is gonna be coming up here in Q2.

So we're gonna be having some companies present to folks like us about their products and technologies. We'll ask them questions and discuss and figure out how it works, kick the tires, all on live streaming video. You can find more information about that at techfieldday.com, or you can find us on YouTube. Just go to YouTube slash Tech Field Day to see videos of our previous AI Field Day event

and other topics as well. Well, thank you very much for listening to the Utilizing AI podcast. If you enjoyed this discussion, please do subscribe, rate, and review on just about any podcast application, and please do share this with your friends. The podcast is brought to you by gestaltit.com, your home for IT coverage from across the enterprise. For show notes and more episodes, go to utilizing-ai.com or find us on Twitter at utilizing underscore AI. Thanks for joining us and we'll see you next week.