99% Correct Is Still Failure: The Last Mile for Mission-Critical AI

0:00It's great if you're, you know, 70, 80, 90, even 99 % correct, but 99 % correct is still failure when it comes to mission critical systems. A lot of these code gem tools are over here. The mission critical systems are over here and we have to help bridge that gap.

0:15Matt Paige:Welcome to the Talking AI Podcast, where we talk AI with both experts in the field and early adopters. I'm your host, Matt Paige, and we're here to demystify AI for you so you can get some value from it. Let's talk some AI. AI can now write code faster than any human alive. And most of the time, that's more than good enough. That's the magic powering the entire Vibe coding wave, the agentic coding craziness that we're experiencing right now. But there's a category of software where most of the time just doesn't cut it. The code running a fighter jet, a power grid, an autonomous vehicle, a piece of medical hardware.

0:51Matt Paige:When the code is wrong, the consequences aren't just a bug. They're a recall, an accident, a national security incident. And that's the gap that CodeMetal just raised$125 million to close. And that's why Ryan Aitay, the former CEO of Tableau, joined as president and COO to run point on what he calls safely delivering the last mile for mission-critical industries. We're going to get into what that actually means, where CodeMetal sits in this crowded market of cursor, clock code, and a host of other tools, and what an operator who ran one of the most iconic Perseed SaaS businesses thinks about this crazy world where agents are proliferating everywhere.

1:29Matt Paige:Ryan, welcome to Talking AI. It's great to be here, Matt. Thank you for having me. Yeah, this is going to be an exciting one because this is a true gap, I think, in the market. We use AI at our company 24-7. Our engineers are fully on board with it. But the point that you're solving here is critical. And just on a personal note, I was an OG Tableau user back in the day, like a complete fanboy of Tableau. But what I loved about Tableau is it wasn't just like the beautiful visualizations and charts, but it completely democratized analytics. And suddenly anybody in business could be their own data analyst.

2:10Matt Paige:And now we're living through something very similar with AI, where it's democratizing coding, design, video, writing. I mean, you name a function or purpose in a business and it feels like it's democratizing it or anybody can build. But you've stayed your next chapter on the idea that there's a category where democratization breaks down a little bit. So walk me through what you saw and what attracted you to this mission that CodeMetal is going on. Yeah. And by the way, thank you in my old capacity for being a Tableau community member because I'm still very connected to that experience. So look, I think, you know, coming into an environment which we're in now, which is this kind of AI world that we're in, I really just saw this large opportunity to work with really a lot of smart people, a lot of, you know, MIT, Lincoln Lab engineers, like our co-founder or sorry, founder and CEO, Peter Morales.

3:08and really it was you know i had all this great experience and i'm very grateful for my time at salesforce and tableau and and the things before that but it was like how do i take all these things that i've learned over you know 19 20 years and apply them to you know a new industry but also an industry and a company specifically like code metal where i could make a bigger impact what do i mean by that well it was you know make a difference make an impact because ultimately there's so many, there's a lot of AI noise, of course, it's like, you know, your podcast is called Talking AI. So it's like, there's a lot.

3:42We're part of the hype, man. We're part of the hype, but the hype's exciting, right? We're exciting to be in this environment, but there's this concept of like, how do we make AI more trustworthy? And, you know, because if you don't know enough, it could be dangerous. You know, of course there, and we'll talk probably about this today, but you know, my opportunity, I looked at it was like, how do we make an impact? There's this opportunity to make AI more trustworthy. And ultimately, I think there are a lot of people that don't understand it. So, you know, it's critical for our nation, for our military, you know, for our government, for mission critical companies, maybe that ship automobiles or airplanes or go down the list of things that we depend on every day as, you know, citizens.

4:27And if it's not done correctly or used correctly, like that could be problematic. And there are many ways and I can explain them, but like that, that could be problematic. And I think there's a lot of noise around like, well, is it going to take my job? What about is it safe for the things that I depend on every day? Like when I go get in my car or when I get on an airplane, like this is sort of the next chapter of, is it safe? Is it trustworthy? And our intent that CodeMetal is to really deliver that last mile.

4:52Matt Paige:Quick break in the pod. Our state of AI 2026 report just dropped and it breaks down what actually is changing in AI, what's hype and what leaders need to be paying attention to this year. You can grab it right now on our show notes or at hatchworks.com. Yeah. Cause that's how I looked at the opportunity. Yeah. But you're injecting this uncertainty into the world with this new, amazing thing we have AI and generative AI, but there's a lot of legacy systems that exist, these deterministic systems. So it's like, how do you graph those two together in a sense? I feel like that's part of it. And you talk about like the last mile for mission critical industries.

5:28Matt Paige:I'm curious, like, and this probably, you know, a bleeding edge of where is the last mile? Where does it start and where does it end? Or is it almost like a holistic view of how you build and leverage AI and the coding and building process? I mean, I think the last mile is relevant if we talk about things like code gen tools, and there are a lot of them today, the cloud codes, the codexes, the cursors, and they're all great, right? We know this, we use them ourselves. but when you get into and you can even ask the various AI tools today like you know can you can you generate code let's say can you translate c++ to rust and guarantee it'll be production ready and safe and they will all the tools will tell you well almost but not quite they can't guarantee it and that's the that's the tool itself right and we know that the hardest part is verification.

6:22And, you know, will it be correct? Have we tested every use case or edge case? Do we know every requirement? Is it secure? This is kind of like what I call the last mile, because there's just a lot of stuff that people forget at the end of the day that needs to happen. It's great if you're, you know, 70, 80, 90, even 99 % correct. But 99 % correct is still failure when it comes to mission critical systems. And so how do we, you know, our goal is to, you know, I think about like a lot of these code gem tools are over here, the mission critical systems are over here and we have to help bridge that gap.

6:59It's the assurance, if you will, like, how do I, how can I be assured that when I translate or modernize code, that it will actually work in a mission critical environment? And I think that is, that is really what we're trying to solve and we are solving. And, you know, it's, it's a lot, but it's also very, you know, it's a lot to unpack and understand, but it is a big opportunity.

7:19Matt Paige:Wait, it's funny. You mentioned like, if you ask AI today, if it could do, I forget the two, you mentioned Rust and something else. I feel like most of the time instead, it'll be like, heck yeah, I can do that. No big deal. In the end, you're like, you screwed XYZ up because you do have that sycophantic nature with AI where it will reassure you and say, yes, I got this, even if it doesn't have full context, which is scary. I think when you get to some of these, uh, you know, mission critical things like you're talking about, but you mentioned it though, like your teams, they use cloud code and some of these other tools.

7:54Matt Paige:So I think it'd be helpful. Like where you're not necessarily in competition with these tools. It's, it's almost like, um, tertiary or, uh, in support of these, these other tools so where if you look at the map of everything going on in this crazy world of ai coding tools and ai tools in general like where where does code metal sit in that uh that that hemisphere yeah um yeah i think at number one i think it's early uh in our in even though it feels like we're a few years into this we are but it is very early in the sense of like understanding what's possible and also, of course, identifying some of the risks and things that can be solved.

8:38We don't really see a lot of competitors, you know, at this point in terms of, you know, focused on verification, validation, assurance, the things that I've mentioned, making it trustworthy. You know, when I think about it, let me try to like, the thing I said before was C++ to Rust, right? So like, you could ask a question like, well, you know, and literally, if you go in and you say, like, go into chat GPT and say, like, can you do this in a production environment? Will you guarantee it'll work? It will say, well, yes, but like, there's always the but there. And so I think the scenario would be like, yes, you could use any code generation tool.

9:17You know, you could, they could like run an entire repository, it could be agentic, It could generate code PR. It could refactor some things. It's pretty good. But the problem becomes almost a behavior. So it's not really just a coding problem. It's more of a behavioral problem at scale. So if the tool succeeds, as I said before, 70 % of the time on a complex refactoring, you could still fail at some level and be a scenario where I think you run into challenges, right? So they can write code, but they cannot guarantee that, let's say a million lines of C++, or in this case, let's call it C++, legacy C++, that it will behave exactly like it did before once it's transformed.

10:09And that is the thing that we're solving. And there's a whole bunch of work that goes into making sure that it's guaranteed. But that is really what we're delivering. We're delivering a guarantee or an assurance that it will work. And that's the difference. Hmm.

10:22Matt Paige:That's interesting. I wonder too, cause you, you all always hear a lot of people's answer to this kind of problem you're talking about is human in the loop. Oh, we have a human in the loop here to verify all of this, which, you know, frankly, I think people get comfortable in that human in the loop, you know, verification decreases as, as time goes on. But I think that the bigger point is though, I mean, this is a process in a system like anything else and there's natural bottlenecks. And I think that the bottleneck is shifting in many ways to that human in the loop. So it's like, as this just completely progresses and scales, how do you think about human in the loop?

11:01Matt Paige:Because it's having that as the verification step, it almost feels like you have, you know, a game of Jenga. sits and you're at the very end of it where it's about to topple over and it's like yeah i mean i think and perhaps human in the loop you know there's i like to think of this as like you know well let's let's unpack a couple things so the first thing is like you know you've got ai makes everyone's jobs easier yes it used the correct way i agree i use it all the time i'm happy to talk about that as well and there's a human and you know let's say i've got an engineer and now they're not just writing code, let's say for said application, they can be doing other things.

11:38They can be thinking a bigger picture. They can be reviewing code. You know, is a, but what about the hardware, right? Because we're not just rewriting software. We're trying to connect it to hardware and do it in a production way. So we talk about our kind of our North Star here at CodeMetal, which is like, how do we, you know, basically make things as pro, like we want to make all things programmable, software, hardware, et cetera. Think of it like a, you know, if you've ever been in like a Tesla or a full self-driving experience where like you have the ability to ship over the air updates, you know, software impacting hardware, that's great.

12:16They can do it, but not everyone can do it. And so I think making sure that it works. Now, if I say, well, okay, let's just translate some old code and we'll just have a human review it. Well, what about this process of like, We need to analyze the code. We need to understand, you know, and this is, of course, some component. We do want to do test generation. So we do something called fuzzing, concolic testing. You go down the list of various things that we include in our suite to make sure that it checks every use case. It checks the source code. Is it behaving the same as it is in this new system?

12:49Is it checking every edge case? Is it checking for variables that aren't expected? So there's a, you know, and I'm going to get quickly out of my realm of expertise because I'm not an engineer. We should put a little caveat on this. But the point is, we're going a bit deeper to make sure that you can count on it ultimately at the end of the day. And that includes, in addition to human in the loop, it's hardware in the loop. So we want to be testing hardware continuously. It's a big part of like when I walked into one of our offices at first, I'm like, wow, there's all this like hardware. It's a different world, right?

13:20And so how do we make things run efficiently? you know from a small device to a mission critical device and do it in a safe way wait you lost me at was it fuzzing was that i've never heard of that term yeah i will you know think of it like um what are the different edge cases that could be a problem right and so it tests every different way and we use this thing called formal methods which is another word for math like there's a lot of this is my learning curve which is like okay what does this mean how do i you know trying to bring it all to uh english so that i like to say how can i make sure my parents can understand what i'm working on that's my uh my intent and

14:01Matt Paige:my goal as we talk about these things yeah that's a good north star but like this is separate but how amazing is it like that you have ai as this partner to like help guide you as you learn i think just just such a cool premise in general but you mentioned uh you mentioned like you're leveraging AI? Like I'm curious, companies right now, especially newer companies are growing in a totally different way because they have AI at the beginning versus existing companies going through their own innovators dilemma of how do we apply this to our existing business? I'm curious, like how are you A, as an executive leveraging AI?

14:43Matt Paige:I think that's super interesting for our audience, but also from a business context, how is AI being leveraged, whether it's in operations, it could be anything, sales, marketing, operations, name your function of the business, anything that kind of sticks out in your mind, either daily use or within the business? Yeah. You know, it's funny because we talk about this. We're a small company, you know, we're still under a hundred employees. But we, you know, we look at the opportunity, you know so we don't have like i was in theory one of the first business people non-engineers in the company yeah we have a lot more now but we can do a lot more than we could have you know if we had started this company in like you know 2000 or 2010 like even it's a it's a different world so i think you know we use you know whether it be you know even things like an m &a pipeline or you know recruiting and screening candidates more quickly like it accelerates a lot of what we can do or hey, give me an opinion.

15:43I wrote something, maybe I want to post it on LinkedIn or wherever I want to post it. I don't have a marketer yet. So hey, I need some, can you help me just like think through or give me perspective? But again, I'm not going to just take exactly what it says. I want to make sure that it's like verified. Again, it's the same premise. It's like, can I trust what is in here and how do I actually go through it? It's like, you know, in the past, you'd ask many different people before you maybe would make a decision. It's a similar thing. it accelerates that process. I do think that, you know, and of course there's, I think we found with the code generation tools is like, you know, if I'm building a simple application or a SaaS application or whatever it may be, maybe I can get started and advanced really quickly.

16:24If I'm, where we shift from that into a world where I just say like, let's be careful, let's make it sure it's trustworthy is we're gonna convert legacy code to modern code. Let's say it's a million lines, like I said before, it has to behave the same if it's gonna, let's say, you know, impact a, you know, an autonomous vehicle, right? Whether, especially in a military environment, like good on the, you can just see how this can spiral out of control very quickly. And so I think that every business right now needs to be using AI internally to run its operations, you know, like, Hey, I have built a cashflow statement.

17:00Hey, does this look right? Or maybe it can build it for you. Like, those are all great things, but you can't just like send it out. Like, I think that's the message I would want to tell people. Absolutely, if you're not using AI, the companies that aren't will lose to the companies that are, but be careful is sort of the intent, especially at this stage.

17:19Matt Paige:Yeah, there's definitely nuance about it. And I feel like the only way to get good at understanding the nuance is just by using AI a lot. I always go back to the Shopify CEO talking about, he has given a mandate for his team to use AI reflexively, just like it's just second nature. And then you do start to get, it's the nuance, it's the taste element of like, okay, I can use AI here or I know to push here. Like, I don't know about you, but I've been completely Claude code pilled over the past several, several months. And, you know, I was a very early open AI chat GBT, uh, adopter. And I went through this period of using both and now I've kind of shifted over.

18:02Matt Paige:I still use both, but I went through this exercise with our quarterly planning. I had it hooked up to our CRM and all the other data and had memory for everything we do, and I just had it build the full report for me. But to your point earlier, it wasn't just like, okay, there's the report go. Like I went through it and I found some interesting insights that I hadn't even noticed or thought of, which was just this really cool experience. in like another one we have it hooked up with again our crm and slack and whatnot and it's going through there and identifying okay what what prospects should we be following back up with that we haven't talked within a while or something's happened and it's just giving us this regular report with a you know a score attached to it and we now just have it like running and doing its thing so it's just once you see the possibilities it's like oh my god there's so much i can do with this.

18:58Yeah, I think that we're in the inning of perhaps, you know, or the stage of seeing all the possibilities, right? And then I think, you know, like we've identified code as one area where it is real and relevant and very, very useful. You know, I still think like, I'm still trying to work whether it's, you know, Claude or Picker. I'm like, hey, schedule a meeting. It's not quite my assistant yet. It's not quite perfect. It's getting better every day. You know, schedule a meeting, do this. It's like, well, I can't access your calendar or well, I can't do this or well, I scheduled it on the wrong day.

19:27So we're getting there, but I think it's education. It's just like what we're talking about here. Like, why am I doing this? It's because there's an opportunity to educate and make sure that, you know, not only impact the business, you know, or like, you know, we work a lot with various forms of our government, you know, it's pretty mission critical that things are done right. Right. It's pretty mission critical that, you know, we're making sure that if there's a drone in the air, there's a tank or whatever it is that's being operated in a safe way that is dependable. And that is a big opportunity.

19:59But not everyone understands this. So you can't just assume like, you know, if it sounds so good to be true, I always feel it probably is, right? That's a saying from the past. So just check your work. Yep.

20:09Matt Paige:It's maybe a theme, you know? What? 100%. And it definitely applies. I'd be curious though, like what you can tell us, like what are some cool use cases where you're starting to apply your approach, your methodology, your tooling with specific customers? Obviously, I'm sure some of the government ones you can't get this deep on, but are there any interesting ones that are worth noting? Yeah, I mean, maybe the one I mentioned before is a customer. It's one I can't mention the name yet soon, but the one where they basically had come to us and said, you know, we have over a million lines of legacy C++ code that they needed to basically translate and modernize, but using the same system, the same hardware that they've used.

20:59So it's kind of like, it's sort of like, I want to rewire a city, but I don't want the power to go out.

21:05Matt Paige:Yeah. It's okay, well, how do I do that without, I mean, it's impossible. So this is like an impossible task that all of a sudden we're like, well, we can do this in like a couple of weeks. It's not that hard. And we can guarantee it'll work, right? So a million lines of C++, you know, translated into Rust with zero issues in a very short timeframe is basically impossible. And now it is not impossible. I think maybe another example would be, this is the one I have to be even more vague with. I apologize. We can do this later when we can talk more about it. But, you know, think of like a Department of Defense or Department of War.

21:38They have bought software and, you know, various types of software that's been like, call it stove piped or disconnected. Some of it they can use, some of it they can't. Millions and millions of dollars has been spent on this. But at the end of the day, right, they want to make sure like, how do they modernize it? you know they can rewrite it perhaps but without like spending a bunch more money to revalidate that it's actually working so there's you know let's go through a process of like where i need to translate existing software into a modern maybe more maintainable scenario uh and i want to use that so i can do some simulation for war fighters right um pick your things right the code needs to be translated, needs to be provably correct.

22:28And then it needs to, we need to know that whatever goes in comes out and there's no need to revalidate it. So it's a, you know, again, I'm sorry I'm being vague, but like I, there's a lot of information that's sort of sitting in different places. We're bringing that all together and providing a simulation environment so they can run real world scenarios. And that, that is very powerful and really only something that we can do at this point.

22:49Matt Paige:Now that's, that's really interesting. Cause like you've, you've been in the business world for some time now like how many very long i'm very old yeah yeah how many transformation initiatives get proposed done and they just die on the vine because they become they're too complex they take too long by the time you get ready to do something things have changed like that's that's a very very real problem and you mentioned like tesla earlier there is this element of like modernization that potentially is possible now because like we talked about earlier the innovators dilemma a core part of that is just like there's so much built up infrastructure and nuance to things and ai can actually help accelerate i mean that's kind of what y 'all are doing in a sense i feel like in in a lot of ways yeah and i was trying to think of another one like you know i think robotics you know again and some of the use cases i give are more like perhaps military focused or defense focused but that's just a big big area of demand for us at the moment, but, you know, um, if you think about maybe a mid non-traditional type of environment where there's, uh, you know, let's say, uh, some kind of a tank or something that is autonomous, whatever it may be, right.

24:04Or drone or something like that. How do we, you know, how do they get more time to think about, you know, what is the environment? Like you see the environment, just like in a, in a self-driving car, you see the environment, but how do we get more time to react um and and how do i do that in a way that is safe and accurate like very important to be accurate in that environment so you know i want to compute let's say the the scene or sort of the visual faster so that that that the object can have a safer ride at higher speed gives the the, you know, the machine more time to react, that can only be done, you know, if they trust the output, right?