The Cynical Data Guy on AI, Data Tools, and the Future of Coding

John Wessel 00:03
Welcome to The Data Stack Show. The Data Stack Show is a podcast where we talk about the technical, business and human challenges involved in data

Eric Dodds 00:13
work. Join our casual conversations with innovators and data professionals to learn about new data technologies and how data teams are run at top companies. Welcome back to The Data Stack. Show we have our favorite monthly installment where we go deep into the bowels of corporate data America and get some hot takes from your favorite cynical data guy, Matt, welcome back. Yeah, I’m back. Okay, this is going to be we’re just going to talk about AI the entire the entire episode, I was told there would be no AI

John Wessel 00:55
does that other episode,

Eric Dodds 00:59
I was told we would have AI enabled features by the end of the quarter.

John Wessel 01:06
So for the podcast, for the we don’t have the guests where, yes, the talking, I mean,

Eric Dodds 01:10
Interestingly enough, actually, all the transcription stuff and whatever, it’s actually been AI for a long time. Well,

John Wessel 01:15
There’s a startup. Did you see there’s a startup competing with 11 labs that allegedly can take like a 15 second voice sample now generate,

Eric Dodds 01:25
like, essentially, I tried this recently, yeah, I tried this recently. Actually, I generated a bunch of research using deep research, which was outstanding, by the way. It was really opening, yeah, it was really good. The ergonomics are a little bit weird because it’s so much text, which is like the point in the chat, just it gets really unruly around that. But the content was really amazing. And it was to the point where I, well, first of all, mobile app, you can have it read you a response. But I tried that, and it’s really janky because it’s such a long text, and so it would have buffering issues, all that sort of stuff. So it’s like, okay, I’m just gonna go turn this into a recording so that I can listen to it right? And there’s tons of AI tools out there, which we’re gonna talk about, the tools, the types of tools. There’s a service out there where you can just upload a you can actually generate an mp three, like with GPT or whatever. I need to try a couple other models. But the voice and audio is just, it’s a robot. It’s like, again, I’m not going to listen to 30 minutes of deep research with this. Yeah, unless you have trouble falling asleep. Unless I’m having trouble falling asleep at night, exactly.

John Wessel 02:39
But then, I think it’s 11 labs, though, that’s doing, like, audio books. And, okay, maybe I need to check that out. Translations out, yeah, the major podcast did like, four or five translations, yeah, like they did in English, but then they also, like, Yeah, four or five other languages that, apparently. So is this gonna be where we’re just gonna have all of our audio books are gonna be in like, the same four,

Matthew Kelliher-Gibson 03:01
Well, it’s gonna sell the rights to their voice. That

Eric Dodds 03:04
was interesting. I just signed up for some tools. And I mean, I actually hit the limit on the free tool, and I was just trying to figure out if there’s a quick way to do it. And anyways, there’s these services out there. And I literally recorded like 10 seconds of my voice and had it like red. And it was a sound, really sounding, yeah, it was pretty wild. The

Matthew Kelliher-Gibson 03:21
The next one you should try is where you purposely, like, pitch your voice. Really weird. Yes, see what that turns into,

Eric Dodds 03:27
yeah, yeah. Although the interesting thing was, I was like that, wow, that did a really good job. And I showed my wife, and she’s like, that doesn’t really sound like you. Interesting. Anyways, the first topic I wanted to hit actually, is not a spicy LinkedIn post, but we were chatting before the show, and those transcription services are, there’s so many of those that are just going to get completely wiped out by the foundational models themselves. And we’re starting to see that, I mean, doing a bunch of tests, even internally, with some AI tools that we’re using. The foundational models are just now beating them, like with a generic, like, just completely vanilla, right? Even for things that have been purpose trained on, documentation, for technical questions, like, it’s just really way better, which is wild. So this is me reading from my internal LinkedIn feed. Okay, if I was going to post on LinkedIn, this is what I would put behind the scenes, kind of Yes. Is that I think just on a weekly basis, we’re going to see, we’re going to see a bunch. We’re going to see failures of these companies that were doing something that was truly value add because of the limitation of the model. And now it is not anymore. And so I think we’re at the tip of the first wave of failures. What do you call a cynical data guy? Well, I think the thing is

Matthew Kelliher-Gibson 04:54
Probably interesting about that was, if you go to the beginning when all of this happened, when we started seeing all these AI. Companies pop up. The assumption was people who are just simple wrappers around, like, open AI, like, Oh, they’re going to be gone in four years. Yep, right. And now, and the ones that we thought were actually adding value were the ones we’re going to hang around. And now, with all these foundational models out, it’s like, oh, well, all I really want is a platform that’s a wrapper that I can just choose and right? They’re getting better, right? So I don’t really need your specially built one for this purpose. Yeah, yeah, converted itself. That

Eric Dodds 05:31
is such a good observation.

John Wessel 05:33
Yeah, yeah. Well, and the interesting thing about that too, and I’ve seen this on a lot of platforms, is if you build your wrapper toward whatever industry or whatever, like sub use case, and then you have that ability to hot swap in models, there’s this perception, I think, of like, well, I don’t know which one to take. I don’t know which one’s best. It’s like, oh, well, these guys solve my problem, and they’ve got five options. And, like, as the weeks go by and one model is bad as another, I can just flip it, then I think there’s kind of a comfort in that, like, Oh, I didn’t actually, like, pick the wrong one. That might not be the best. Does it also give

Matthew Kelliher-Gibson 06:12
the illusion that you’re not in vendor lock in? Oh, yeah, a little bit. Oh, I’m not gonna be locked into a vendor. Yeah. Well, you are. You’re just the sub vendor is

Eric Dodds 06:20
changing, yeah, yeah, that’s a really interesting point. I think one of the most amazing, actually, I would say, just in terms of the interface, but also the company that I think has done maybe one of the best jobs of all of them, of incorporating AI into their product is raycast, which, if you use a Mac in another spotlight, so you do Command space, and it pulls up

John Wessel 06:46
like the global circuit long term Mac users. This is the new alpha. This

Eric Dodds 06:50
is the new Alfred, exactly, and it’s just an outstanding tool stand alone, yeah. I mean, it makes Alfred look like it makes Alfred feel so primitive, but it is actually an interface with all of the AI models, one so all of them. You could do all this custom configuration for various commands to use different models and all that sort of stuff. But they’re now using extensions to, like, integrate it at the operating system level so that you can do like, you can do all sorts of stuff, right? So like to run it against, like, basically in your day to day workflow, right? And so it is. It’s pretty wild to see, which is essentially, like, connecting

John Wessel 07:35
an LLM, and they already have, like, the OS level and like, do an action, or

Eric Dodds 07:40
Exactly, yeah, that’s really wild. So many people are

Matthew Kelliher-Gibson 07:43
gonna wipe out their

John Wessel 07:44
computer. Yeah? R, oh yeah, it’s

Matthew Kelliher-Gibson 07:48
gonna be the thing. What do you do? I do anything, yeah, let’s talk with the ln closed me out and still run

John Wessel 07:57
over. There’s a junior developer. This was probably 10 years ago. Started in like, two weeks, and he had switched he started out, and he was using Windows to start anyway. He had switched to Mac like, somewhat recently. Wasn’t as familiar with command line, and he did delete all his files, but he’s somehow managed to take every single file from like, all of the like, separate directories in the computer, and put them all in directory, which is also just about as kind of including system files, not just like Word

Matthew Kelliher-Gibson 08:26
documents. Please tell me they were all on his desktop, please. That

Eric Dodds 08:30
would be

John Wessel 08:31
amazing. But it was one of those things, like, you know, still learning like, yeah, great, great person. Like, good developer, but just like, still learning like terminal and like, all directory that’s also

Matthew Kelliher-Gibson 08:43
look at them as you go. I don’t even know how you can, yeah,

John Wessel 08:48
there’s not really an undo button from that.

Eric Dodds 08:52
Okay, so look out for companies too short, because it’s getting spicy. It’s getting spicy out there because the models become better than anything else, I do have a couple of great LinkedIn posts that I do want to let’s do it. Get you. Let’s go on. Okay, moving on. Okay. The first one is from Kevin, who actually has been on the show, a great guy. He’s the CEO of meta plane, and so Kevin, if you’re listening, would love to have you come back on. We could talk about AI even, okay, I’m just gonna read this post. How much should we rely on AI to generate production code? This forum post about cursors, LLM suggesting the user to learn code has me thinking about our field, okay? And so just so we can put the post in the show notes, but there’s a screenshot in a forum. Someone had posted in a forum. It says, AI told me I should learn coding instead of asking it to generate it. And the response from the LLM is, I can’t generate code for you as that would be complete. That would be completing your work. The code appears to be handling some. Did Mark fate effects in a racing game, you should develop the logic yourself. This ensures you understand the system and can’t maintain it properly. Reasoning, generating code for others can lead to dependency and reduced learning opportunities. So Kevin, that’s the forum post. This screenshot may be fake, but it’s really funny. Yeah, yes, it almost certainly could be faked. So he said, The vibe coding trend using llms to generate entire applications without understanding the underlying code raises interesting questions for data engineering, why this matters for data teams. Data teams, one SQL queries generated by llms often look correct, but can silently introduce errors, especially with complex transformations or edge cases. Two, when data engineers don’t fully understand their pipelines, debugging becomes challenging when something breaks. Three, the path of least resistance is tempting, and there are genuine efficiency gains to be had. To summarize, the most effective data engineers I know are finding the sweet spot using AI to accelerate routine tasks while deepening their understanding of core systems.

Matthew Kelliher-Gibson 11:12
So first, I’m going to say the whether that that post was faked or not, with there, it does make me think of my kids were watching the Willy Wonka and the Chocolate Factory movie, and there’s

Eric Dodds 11:24
original or Johnny Depp now the original, okay, which

Matthew Kelliher-Gibson 11:26
is not my favorite, but that’s another

John Wessel 11:30
story. The books better is that

Matthew Kelliher-Gibson 11:33
where this was going, it deviates too far. There’s this one scene that they have where this guy says he’s programmed his computer and a bunch of tapes and main frames and tells him where the gold ticket is. Oh yes, this is a great scene. And it says, I can’t do that for you. That would be cheating. So he tries to tell it, I’ll share the prize with you. And the computer replies back, what would a computer do with a lifetime supply of

Speaker 1 12:00
Chocolate? Completely forgot about this such a good scene that

Matthew Kelliher-Gibson 12:04
just made me that. But yes, the other part there, I think, coming from, I started as a data analyst, and then having to manage and train data analysts, this is the thing that you can kind of see that you don’t want to see from a data engineer, which is kind of that, like, why is that number that way? Well, that’s what the data said. That’s not an answer like, right? I need to have an answer. You need to understand more. So that would just be one where you’re like, why did the data get transformed this way and put it here? I don’t know. That’s what the process does. Oh no, that’s not gonna

John Wessel 12:42
work. Yeah. I think it’s gonna be so interesting, like, how this actually plays out, they can think of like, two or three scenarios, one where it can be really dangerous. So say you got like a junior, junior engineer right out of school and, like, just vibe codes through, like, full pipelines, full apps they released in the production like that could be a problem, especially like in a small org where there’s just not a lot of people, and they hired that person as, like, their data person, or tech first, or whatever. Like, I think that’s gonna, like, result in some pretty bad disasters. On the flip side, I think it’s very interesting for people that are in architect roles, or even like product roles, they can do where, essentially, like, they know how it’s supposed to work roughly. They like, understand risks. They understand how things typically break. They understand ops decently, like that person. I think it will be really interesting how it develops, because then, because they can kind of see around, like, okay, cool. Like, you just introduced a major like, security problem, like, because that was, you know, tap. And then the third use case will be people that are in that more junior role, that lean really heavily on, like, educate me, help me learn about this code and like, are primarily, like, pushing those types of prompts through. Yeah. I think that’ll be great for those people. Yeah. I

Matthew Kelliher-Gibson 14:07
I think that also gets to something that is there right now, one of the things you can see is that the people who can use AI to code need to know how to code first. But I do wonder if we’re gonna get to a point where there’s, like, there’s people who’ve learned it, some people who learned to code and then went to theirs, and others that use the AI tools to learn to code. Yeah, and what is that going to look like? What’s the differences how those people are going

John Wessel 14:32
to look? We’ve actually already been through the iteration of this. We’ve been through iteration of like, people that learn like Java in school 20 years ago, and like came out and like did a traditional route, or the people who like, just like did more like a boot camp route did, and then were 11 overs, and essentially learned from Stack Overflow, right? Yes. So like, that’s already kind of the thing, yeah, and, but it’s different, because in Stack Overflow, like, it’s like. Here’s a rough example. You still have to do a fair amount of work, like, understand what’s going on.

Matthew Kelliher-Gibson 15:03
Well, could then it also, could also, then make worse the problem that we have with some where it’s like, I know how to execute a thing, but I don’t understand some of the theory or what is behind it. Like, I mean, one of the things that I found was I had to take a warehouse in class when I was in school. Yeah, this is just the concentration I had, and if you had to do entity mapping and understanding to go through this well, you learn one second, first, second, third, normalization, stuff like that. I didn’t think much about that until it came up recently with something where it’s like, oh, that’s actually really saved me, because I have to go clean up a bunch of people who’ve never even been introduced to that concept, and they just do really stupid things with databases. Yeah,

Eric Dodds 15:48
yeah. I okay the bog around. What is it going to be like for the people who actually learn their skill set with AI there is going to be really interesting. And I also think that what’s pretty likely is that the entire methodology changes. Yeah, right. I mean, of course there’s a question. I mean, you go talk to anyone who’s like, reasonable out of basically, I was talking with one of our principal engineers recently about this, right? And he was like, yeah. I mean, I use AI to, like, do a bunch of stuff, but, like, it can’t, like, architect the complex system well, like, I wouldn’t put the code in production. Like, blah, blah, blah, right? Which, yeah. I mean, of course your business depends on everything working well in production, and so you’re going to do what you need to do there. However, the improvements are going to continue to be dramatic, right? In the way that we think about developing applications is going to change dramatically along with that, right? And so it won’t be like, I have to do some stuff, and then it’ll be like, okay, the way that we conceive of doing this, right is, I think is going to change.

John Wessel 16:59
And I think there’s this level of abstraction thing where there’s a really interesting post on LinkedIn the other day. Think it was a, think it was somebody talking about data, and they said something like, I never, like, write recursive queries. I never use recursion and data. And then in the comment, somebody was like, Yeah, you do. It’s just abstracted away from you. You just don’t know that you’re using it right. And I think that is, like, what abstraction level is necessary? We’re like, do I, like, understand compilers deeply? No, but I use them all the time. Yeah, I understand, like, do we have to mess with, like, memory management much anymore, like, and data? Not really. So there’s all these things that are already abstracted that, like, fewer and few people need to understand the details, though, yep. And it’s just like, where is that level gonna be?

Matthew Kelliher-Gibson 17:44
That tends to be more common. I think it is almost like it’s just another form of a framework or something like that, right? So, yeah, exactly. You could even think of it in a little bit of when does it become, like, a new version of WordPress? Yeah, it’s gonna be kind of heavy. It’s gonna have this extra stuff. But it’s this, there’s this core part of it that it can do for you 100% and that’s probably going to be something we see, which is where AI, well, can’t do everything for you, but here is this set of core things that, like, nobody does that

Eric Dodds 18:13
anymore, totally, totally. I mean, with all of the tools out there, rep, lit, v zero. There are a bunch of those, right? It’ll be interesting to see where that whole thing goes anyways, relative to the conversation we were having before, right? Because, yeah, as those frameworks get really, as those frameworks get developed, it’s just interesting to think about that, right? But, I mean, what’s fascinating is, okay, you talk about, like, architecting a back end, right? Well, if you generate an application, you can infer a lot about the architecture, right? Like, those types of things are going to get better and better. And then to your point, Matt, if you start with an underlying framework as the starting point, you can essentially cover a number of use cases and probably get close to something that is production ready, right?

Matthew Kelliher-Gibson 18:56
If it allows front end engineers to have some better understanding of what the back end is going to look like from a data standpoint, very happy,

John Wessel 19:04
and vice versa too, yeah, yeah, totally.

Eric Dodds 19:07
It’s great. Okay, okay, next, yeah, the data design thing, actually, that’s a really good point, actually, even to, actually, I have another LinkedIn post, but the now, that’s a really interesting point in that even if you think about capturing data like there, I think there’s going to be a lot that happens relative to AI, being able to, like, infer what data needs to be captured, what the shapes of schema is going to be like, all that sort of stuff, right, right, possibly

Matthew Kelliher-Gibson 19:38
even put things into proper, normalized format when needed, not that I’ve dealt with that before. It

John Wessel 19:47
was probably more likely to do it correctly, yes, as far as high level implementation details, yeah,

Eric Dodds 19:54
totally okay. Are we ready? What is this? Round two or round three? Oh. This is round three. Okay, rocking along here. Okay, here we go. The future has no UI. And will design agent first AI is eating the interface. The other day, I was going into my reporting software which requires me to input data from a contract. I need to go find the contract which has been sent by a signing service to my email. I find the relevant data and input my software and input it in my software. Then I extract the data from the software to do analysis on this data, where I have to set up the data properly, then figure out how to write a formula to run an analysis. It’s a very usual workflow, you get data from someplace and put it somewhere else and do something with it. What I really need to do is store the new data about x in my email and run analysis y. AI can do that now. So really I only need to tell my AI to do that, and it will execute faster and better than I can myself. The new mobile interface will be empty just to chat. You can talk, but you will not need to go into apps and press buttons. The future iPhone and software interface will be just a blank screen that brings up what you want in the back, that brings up what you want in the background. We will have agents running on top of software, talking to other agents, such as a DocuSign, talking to my data and document storage and putting the data and other information in the right places. But I will not need each interface for this anymore. I’ll probably have a dashboard and a chat with access to everything I want to do for software and AI companies. This will mean designing agent first, as we used to have mobile first, the best software won’t be the one with the best interface, it will be the one you never have to see. PS, it might seem like the human role is vanishing, but I don’t think so. AI will take over execution, but humans will still do what AI isn’t good at communicating with people, making decisions and thinking about what to do next. Work will be a lot more enjoyable when we don’t have to fight with software.

Matthew Kelliher-Gibson 22:05
Yeah, good luck with that. You’re going to still fight with

John Wessel 22:10
soft I think on this one, I think they’re pretty under indexed on how much people like UI. Yeah, I think if you look at one thing about YouTube, shorts, TikTok, Instagram, although the most engaging platform, it’s full UI. There’s just that, like, brain visual connection, where, if I have to, like, type and stuff, it’s just extra cognitive load. So certainly AI is going to be burned, for sure, yep. But I think there’s going to be a ton of apps that, like, still have you, I still have the visual part, and then they have this nice, seamless, like, hand off with an airplane just to do a thing.

Matthew Kelliher-Gibson 22:47
Yeah, let’s also, what do we see for things? Oh, do I want to say something and then stare at a blank screen? No, yeah. I have a progress bar. I have something that shows me what I’m doing, right? You can, you know when things happen, even if you install something on a computer, there’s always a thing where you can click and you can see the files that are being installed in real time. We like feedback and progress of things visually. So this idea that it’s going to be like, oh, all I’ve got to do is this little chat, and then I just wait for it as I see it. Like, that’s gonna drive people crazy. I mean,

John Wessel 23:23
think about computers 40 years ago. Like, that’s essentially the interface. It’s like a terminal interface. And clearly at that point you could type in the computer and could just do something. It’s not quite like a human shadow. Like that didn’t work out, right? Yeah,

Eric Dodds 23:38
was it? Misha from reflection AI, who was involved in Deep Mind. And yeah, his co-founder, yeah, I want to say it was him. We got a show recently. Yeah, they worked on Gemini stuff. I think it was his co-founder, or someone he knows from space who said AGI is going to happen, but no one’s going to notice it. Yeah, which is fascinating. And that’s kind of, I don’t agree with everything that this post is saying, and totally agree with what you’re saying, but I think what is interesting is it reinforces that point around this sort of fading into the background.

Matthew Kelliher-Gibson 24:18
Well, one thing I would say is that people may not know AGI shows up, because no one actually knows what AGI is anymore. It’s

John Wessel 24:25
just term marketing literature. They’ve already moved on past AGI, right? Like nobody’s even marketing like open AM, the big ones are moving past. It’s agentic,

Matthew Kelliher-Gibson 24:35
and if there was a clear definition, you would know about it because their marketing department would never shut up about it. Yes, yeah,

24:44
yes. Is

Eric Dodds 24:46
AGI likes the data mesh of the world?

John Wessel 24:50
I don’t know. I don’t know if I go that part, but maybe,

Eric Dodds 24:53
well, I mean, it’s actually the one parallel that’s interesting is that’s an academic concept, sure, at the root. Right, which makes it really hard to sort of think that’s really interesting. One of the things when I read this post, one of the first things that came to mind was Alexa, okay, and if you remember, at one point, there was an article that said, how many engineers or how many people were working on Alexa? And the number was mind boggling. I want to say at the peak, it was 10,000 people or something. Okay, so just absolutely unreal. And then we can have Rick’s fact check me and put it in the show notes. But it was a very large number, right? Can we also talk

John Wessel 25:45
about why none of the voice assistants have any AI like anything yet from what I’ve seen?

Eric Dodds 25:53
Well, okay, so yes, we can. But to finish out the first point, what was fascinating is you could do all sorts of crazy things with Alexa, yes, end to end, like almost agentically, if you will, right? So, like, I could speak and then I could get groceries. I could do whatever, right? And people used it for, like, the top five use cases that comprised an overwhelming majority were, like, checking the weather, checking sports scores, making a grocery list, like it was just the most basic stuff. I would

John Wessel 26:30
say time, yeah, weather lights on and on, yep,

Eric Dodds 26:34
yep. Music, maybe music,

John Wessel 26:37
yeah, maybe, like, a couple other things, yeah. And then there’s a very like super

Eric Dodds 26:40
long tail. But what’s interesting, the reason I bring that up is what is really interesting about it is the and you made this is just another way to so I’m stealing your points and nickel data guy is really what’s happening here. Sorry for you, but a blank screen, or maybe I’ll do a spin on it. A blank screen is really hard to interact with because it requires an immense amount of creativity, right? I mean, yeah, that’s a great point. An example, I think about even within, RudderStack is our most loved feature. I mean, we’re talking 80-90% adoption, every customer calling on people like this is so great, right? It’s actually just a code editor, right? And so you can run real time transformations on payloads. It is so useful. The number of use cases that people implement is mind boggling, what people have done with it, right? And but what’s so interesting is the first time you show someone it is the so unimpressive, even like, when a new customer comes on, it’s like, okay, you have these, you have this super powerful tool, and they’re like, okay, but then they run into a situation where they’re like, I need to do this really, really critical thing, right? And over time, it becomes the most loved thing because it’s just so useful, but as a blank slate. It’s a it’s really hard

John Wessel 28:01
when you guys have that template library now, which I think, oh, yeah for sure, yeah,

Eric Dodds 28:05
I’m painting it in like, really extreme terms, because, like, we’ve done a lot of things to sort of, you know, yeah,

Matthew Kelliher-Gibson 28:10
overcome that. But the kind of vision this person has is almost a continual blank page problem, right? Even necessarily what I can do, I don’t know what I want to do or should be doing, yeah, that becomes the thing, I mean, because it’s like, as you said, there’s a lot of stuff Alexa can do. Most people have no idea they can do that. Yeah, use Alexa a little bit like an Excel worksheet. What do you do with it? I make lists. You realize you do all this other stuff, yeah? But I just need to make a grocery list and excel, yeah,

John Wessel 28:40
yeah, yeah. That’s fascinating. Okay, voice assistant, like, I’m just confused. So like Siri and G because you have this little integration thing where it does a hand off and, like, that’s fine, but like, Google likes, like, I don’t, I haven’t noticed any sort of like, they’re just how they always have been. It seems like there’s been no effort to like Claude into Alexa, or generally, into Google. I haven’t really kept track of it, but, yeah, I’m just, don’t it’s just a compute power problem or, like,

Speaker 2 29:13
unpredictable maybe, I don’t know. Yeah, I don’t know. Maybe somebody that’s, maybe they’re just stubborn,

Matthew Kelliher-Gibson 29:18
because it’s like, so it’s Amazon. Like, that’s not, we don’t

John Wessel 29:22
have the world. We’re just a silly thing of, like, that’s a separate team, yeah? And they got their funding, and all the money went to the LLM team, and they’re a different team. It just could be something like, very simple, like

Eric Dodds 29:33
that. Yeah, yeah, yeah. It is. It’s also fascinating to think about if you imagine this future world, right? So the blank screen as the interface or whatever, and you think about, like, what this means for people who are trying to, like, build stuff around AI, etc, you sort of go back to like, the one who wins, distribution wins, right? So the iPhone, it. Actually it doesn’t really matter what happens in the background, which models, like all that sort of stuff. But like this interface Apple will distribute it, or the iPhone, right? Or Amazon with Alexa, or, like, whatever, right? The distribution thing is, right? That’s pretty wild to think about that, right? What other spices I don’t even know. I haven’t even been keeping track of time. Oh, yes, okay, that was the bonus

John Wessel 30:26
round. That was the bonus Wow.

Speaker 3 30:28
How did I forget the bonus round? I have it pulled up right here in Yes, I have it pulled up right here. Okay. Bonus round is actually related to the interface question. So that LinkedIn Post said the future is basically like a blank interface with agentic stuff happening in the background. Interestingly enough, duck DB rolled out an interface. So here’s just a hit. A couple highlights post, duck DB introduces a local UI for easier SQL exploration. Duck TV and Mother Duck release a local web based UI for seamless interaction with duckdb. It’s available out of the box, simplified SQL query management with a user friendly interface. It’s a local web interface launched directly from the CLI, and lets you run full queries, just selections, table, summaries, mother, duck, integration, it’ll preview the first 100 rows. I mean, pretty cool. It’s a really interesting thing, right? Because, like, air startup, like, it’s not like this sub tool has been around forever, but it’s a really interesting model, because all of the other, all the like, modern analytics competitors have not done this. The reason is, like this kind of security angle for sure, and then there’s another practical angle of like, they want you to use their computer and to charge you for development, essentially, yeah, yeah. And I was telling you guys before the show, I’ve got a client that I think as far as the amount of their compute bill for their cloud data warehouse, I think it’s 20 or 30% ingestion, I think stacked another 50% development, and that little remainder is the actual people accessing the data. Wow, and I don’t think that’s that uncommon, yeah, yeah. I And

Matthew Kelliher-Gibson 32:16
people wonder why data team is full

John Wessel 32:22
and transformations part of that too, but yeah, but essentially, like the user value are as a relative of like, the entire like computer, which is directly corresponded to your bill in most of these platforms, is small, and that’s probably always been true, but it’s more apparent when You’re getting charged.

Eric Dodds 32:40
Yeah, I am, yeah for sure. Yeah,

Matthew Kelliher-Gibson 32:43
yeah. I mean, I remember when I was a senior director, we would work with Google, and they were trying to sell us on the new thing, which was a fully integrated development, developer like environment in their platform. Meanwhile, then we had also, we had brought on a consultant who was a software engineer, and they were trying to turn all of Google’s tools into stuff that you could then do completely locally, yeah, which they had to inform them. That’s great, except nobody but you will ever know how to use this, right? Yeah,

Eric Dodds 33:17
I think it’s interesting. I mean, I think that. I think winds are blurring number one in that, you know, especially in the world of data, where non technical people are becoming more technical. Like, legitimately, AI is helping people interact with this stuff in a way that’s way easier. I mean, the number of people I talk to who are product managers, who just used AI to write, like, basic SQL, and it’s the most helpful thing in the world to them, because they can just query the data, right? It’s astounding. I mean, almost everyone is doing that now, because you’re not, like, writing a huge model to do, like bi right? You’re right, but you’re interacting with data, like with materialized views, and you can actually run. That’s really helpful. And so it is really interesting that I think those lines are blurring. And so this strictly dev tool versus an interface for a non technical user is really blurring. I

Matthew Kelliher-Gibson 34:12
I think there’s also that tension for a lot of people, they would like to use something that’s more local, either because they like the feel of it, or it’s more convenient, or things like that. I mean, that’s even why you have all of these integrations like VS code and stuff like that. Yeah. Look, I don’t have to go into your web UI app. I can do it from my own place, yep. But then there is also that pull of the people who have all this out want to try to draw you back into it, right? Yep, but you’re never going to be able to make a one size fits all tool that everyone’s gonna like. That’s why we keep coming back to the local. I feel like, yeah,

Eric Dodds 34:49
yeah. I mean, whenever you hear about this tool, Rosa, a local UI, and it’s just blank, actually, yeah, you can run this UI locally and it just blank, and you. Just talk to man if there’s a progress bar. Yes, that’s exactly right. Okay. Well, thanks for joining us for a fun AI edition of the cynical data guy, Matt, as always, great to have you on the show, and we’ll catch you on the next one. Stay cynical. The Data Stack Show is brought to you by RudderStack, the warehouse native customer data platform. RudderStack is purpose built to help data teams turn customer data into competitive advantage. Learn more at rudderstack.com.