Data infra is printing money. Is it sustainable?

0:00Mehdi: Hello everybody, welcome to another podcast of Explain Analyze. And today I am with one of my teammates, another one, Jacob, first one on the pod, this time.

0:11Jacob: Hi, Metty. Good to be here.

0:13Mehdi: How are you doing?

0:15Jacob: I'm good man, I'm good. Bright and early here on the Seattle side. Wha how are you how are things going out there in Europe?

0:20Mehdi: Yeah, it's super rainy, end of the day. so so yeah, I'm I'm looking forward to hear a bit your thoughts on the past news because there's been a lot of things and yeah, you br you brought interesting links. do you wanna start? Actually, first I need to share my screen.

0:33Jacob: Totally. Sure. Yeah, share the screen. Let's see what we got here. Into the out of the frying pan, into the fire.

0:48Mehdi: Okay. Do you want to talk about money being raised?

0:51Jacob: let's talk about money, my favorite topic. Let's go. Yeah, you know, I think the thing that we've been seeing is that you know, companies that are building solid infrastructure are continuing to win, right? there's been a lot of

0:54Mehdi: Yeah.

1:11Jacob: Challenging market dynamics, you know, specifically for SaaS companies. And I think that we are seeing a very strong bifurcation between SaaS and infrastructure right now, indicated by, you know, this super base series F that we're seeing. Congratulations to them. It's an amazing achievement. and you know, I think also Databricks just announced Series So series as Yeah. Yeah, exactly. Series AA, yeah, yeah.

1:15Mehdi: Yeah. Exac I wonder what they're gonna do if they are past the alphabets. Like it's it's getting there. Yeah, exactly.

1:40Jacob: Anyways, so data breaks.

1:44Mehdi: Yeah, so to to to data infrastructure, why do you think there is more meats into those company to survive? Because there's still like quite a lot of competition if we take, for example, Postgres. so Super Space is you know I mean it's more than I would say probably Paul would say than just Postgres. I agree. But still there is competition up there.

1:57Jacob: Yeah. Right, right.

2:08Mehdi: Do you see like why why specifically more money over there?

2:12Jacob: Yeah, I think just you know, Javon's paradox tells us that like as the price of things go down, we use it more, right? and I think that that is a hundred percent true with infrastructure, right? There's totally if you look at, you know, macroeconomics around servers, you know, we're starting to see prices on CPUs way up, right? GPUs were just the were just the tip of the spear, right? CPUs are up, RAM is way up, solid state drives way up.

2:20Mehdi: Mm-hmm.

2:41Jacob: Right. and I think that is all just because the one thing got way cheaper, and that's writing code and making software and shipping software. Right. and you need infrastructure to do that, right? Whether it's something like Superbase, which is a database, you know, and and obviously they have a lot of other, they have a bunch of other pieces, you know, in the platform, right? So don't wanna just say, well, it's only a database. Like it's there's there's a whole bunch of stuff in there that you need to build an application. And so I think, you know, we're seeing just an explosion on that, right? So there's gonna people who win. you know, I think we'll see kind of where it goes, but you know, I think we're gonna see more and more of this. What are your what are your thoughts, Medi?

3:26Mehdi: I think the ch Yeah, I think the challenge is on the margin too, right? Because they are getting a lot of new agent user, right? I think it's neon during his acquisition to Databreaks show this graph that they had tons of it, super base. I think most of their user are coming from lovable, I guess, kind of things. and so I think they're seeing a lot of user, but they're also super aggressive on fee free tier, like

3:37Jacob: Yep, yep, yep.

3:54Mehdi: There is a lot of database being created, a lot of project. What is sustainable? What is their margin at the end? Like they sure have a lot of users. But as you mentioned, I think that's good to call out. Infrastructure has been more expensive, right? And how is that sustainable? Like all those AI agents doing one project that, you know, die. And that's why, like, you can keep your active database on Superbase like seven days. They're super aggressive on the free tier after it's

4:22Jacob: Wow, I didn't know that.

4:23Mehdi: They yeah, yeah, yeah. It's been a while. It's been a while. because yeah, they cannot afford to have like a free tier that doesn't have any queries. Sorry, yeah. Don't don't get me wrong. But so if there is no queries from for for seven days, they just yeah, they just turn it off and you have to turn it on. so yeah, I I think it's difficult to print him, but I do see we're gonna see some cracks where either some layoffs or some

4:35Jacob: Yeah, yeah, yeah.

4:51Mehdi: you know, rebudget, but I feel this is money to pay, as you mentioned, the hour that is expensive and the VC are in because they see all the AI agents getting, you know, a lot of usage. But I'm just wondering how sustainable is it on the long run? Yeah.

5:08Jacob: Yeah, I you know, I think that's a that's a good that's a good question. And I don't have an answer.

5:14Mehdi: We'll we'll revisit next time. All right. next next link. I have what am I gonna take? this one. yeah, NVDI which announced new AI chip for personal computers. and I think it's interesting that there is you know, we are at Modelar like praising, you know, local analytics with DugDB and some parts offloading in the clouds. And

5:22Jacob: Ooh, okay.

5:42Mehdi: There was a lot of noise around small small models. I don't know if you heard, like to run, you know, on really cheap hardware locally. And this is basically goes into that line. But we haven't got any major offering from the labs, you know, OpenAI or cloud to go into that offer. Like they all investing into, you know, their data centers and their service side thing.

5:48Jacob: Yep. Yep. Mm-hmm, mm-hmm.

6:09Mehdi: So I'm curious what's your thoughts on like and how and how if you have been using like personally I've been using local AI just on flights, like I download the models and sometimes just to get my grasp on like how good they have been, like the open you know, open model and running locally. But yeah, well how do you see the futures now that there is more chips going on like loc?

6:17Jacob: Yeah. Mm-hmm. Mm-hmm. I mean I think like that in the future there's there's kind of like I think there's there's g we're gonna see I think a split, right? There's a certain set of tasks that small models are gonna be able to take on. maybe like triaging your inbox or triaging your Slack messages, right? you know, looking at all of your alerts, making sure you see the right things at the right time. and I think like we're gonna see that type of stuff just start getting offloaded, you know, kind of things that are in the in the mix, things that you would expect, you know, things that you would frankly expect your iPhone to be able to handle, for example.

6:47Mehdi: Yes.

7:00Jacob: which for some mysterious reason Apple has not been able to take any ground on. That is a separate, separate, separate thing. I don't think we have any links in here about WWDC, which was this week, right? So Yeah, correct, we're all sad. so I think we'll see we'll see small models there. I think you know we may also see built-for-purpose small models. I saw a really interesting article, which now I'm I'm feeling sad that I didn't share from Harvey.

7:04Mehdi: Yeah, yeah, that's cra that's crazy. No, because it's it's sad, I feel, personally.

7:29Jacob: Around fine-tuning a in NVIDIA model for legal purposes, you know, and it approached it approached current state-of-the-art foundation models. it wasn't it wasn't as good, I don't think. Obviously, you gotta do all this like training to make it work, but I think that that we will potentially see domain-specific models come out as well that are fine-tunes of these that can run on that hardware.

7:37Mehdi: Mm-hmm.

7:55Jacob: I think that would make a ton of sense, right? I could see, for example, in the future, Mother Duck offering a model that is tuned potentially on just our feature set, right? Us local model you can run that is just part of what we offer to you as the product, right? not even running on our compute, but running on your compute, right? To go within that notion of like hybrid compute.

8:05Mehdi: Yeah. Yeah. Yeah, yeah, yeah. But that's that's that's the thing I was hinting is that I feel I don't see the interest of the labs, but I see a huge interest of like manufacturers and just in general to not have I mean it's always the same story. We have a lot of like powerful computes here, and as you mentioned, actually small model you know fine tuned to towards specific tasks makes sense. but I still yeah, yeah, feels we'll we'll still

8:15Jacob: Yeah. Yeah, totally.

8:38Mehdi: A bit down the line. But it would be super cool, right? That you can have like an agent that you download locally and that's used some part of it. And you have like a license that you pay. Like you're not paying per token, for example, if it's an AI, you just pay a monthly stuff and that's your compute. And then if you need to offload, that would be that would be the best of the

8:52Jacob: Yeah, totally, totally. Yeah, I I think I think that's right. I think like the the hybrid model makes sense, right? Because you will need you're still gonna have tasks that need that deeper intelligence. And that's gonna be available on really big computers with lots of data. And I think you know, so so you'll you'll use those, I think, together, but how exactly it works, I think remains to to shake out. you know, I haven't seen anyone really have like a really clean integration story where like you can have one interface that cleanly hands off, you know. I know there's a lot of talk about model routing and things like that.

9:26Mehdi: Yeah.

9:27Jacob: I just haven't seen anything really good there yet.

9:28Mehdi: Yeah. No, indeed. cool. You have Monday DB.

9:34Jacob: yeah. This is a fun one. Have have you seen did you see this one yet?

9:36Mehdi: The H tap for Yeah, so I'm I'm curious, maybe we can do a quick reminder for the audience what's what do we define as H T A P.

9:46Jacob: yeah, sure. So you know, yeah, we can talk about that. So in the kind of the the very purest way of thinking about databases, you have OLAP databases, which is online analytical processing is what it stands for. And OLTP databases, which is online transactional processing, I believe. I hope I got those right. and one of those is is basically good at r handling you know. thousands of concurrent transactions in a second, very small, inserting into rows and typically more write-heavy than read-heavy. And then you have an analytics database, an OLAP database that is better at kind of looking at all of that data in aggregate and performing big operations on top of them that are very expensive in an LTP database. So what MondayDB here has built is what they're calling HTAP, which stands for hybrid. Transactional analytics processing, I think, which aims to be good at both of these. And and it this is a category that's been it's been pretty hot on and off. It's it's existed for a long time. A lot of company have raised a lot of money doing it, but there isn't really like a breakthrough HTAP company so far. So it's a very hard, it's a very hard challenge. you know, they have a

10:44Mehdi: Yeah yeah. He's been there for years, huh? So just to be clear. They never succeeded. Yeah. Yeah, so Monday Monday is a software management project management. So they have here it's their engineering blog just for context, and so they've call it Monday DB. But so I think it's a pretty high claim for an engineering blog saying that they solve HTAB. It is not their core business. so yeah. So what what do you what do you think about it? So they basically if we go to maybe

11:11Jacob: Yeah, project project management. Yeah, yeah, yeah. Yeah. Ha ha ha.

11:35Mehdi: Think they have schema node on live level design here? so they have DuckDB that's basically doing the part of the analytics as part, and they have MySQL main database. So that's basically the long story short. and yeah, what do you what do you we do you think it's realistic? like do you think we can

11:39Jacob: Yeah, here we Yep, yep, yep. Yeah.

12:03Mehdi: Now with like DuckDB and others would say improvement we had these past years, claim to have an an H T A P database that's both good at both.

12:15Jacob: You know, I think I think the answer is maybe. I think the the challenge is just like being good enough at both to justify not splitting your workload. And the the reality is that just replicating data is painful and annoying. And so if you can solve the replication problem, you solve some you you solve the HTAP problem kind of in its nature. I guess like my my my joke about the thing I like to say about HTAP is like data so nice you store it twice. Right? You store it once row wise, you store it once columnar, and you keep them in sync. that's basically what they're doing here, I believe. So it's not like there's you know, that that's a very easy concept to to reason about.

12:51Mehdi: Yeah, it's true. Yeah, that's a really good point that it's not like a single storage where you have one query engine that or two query engines on the same storage.

13:03Jacob: Yeah, yeah, correct. It's it's so it's it's nice that it's a unified interface. I do think we'll see more of this as DuckDB and DataFusion and other kind of powerful analytical engines mature, that companies will just roll this. I mean, we're also seeing lots of companies embedding DuckDB and and DataFusion kind of in their products natively and what they're serving to. So you know, I think the the maturity of those really, really helps. solve lots of problems that are becoming much in more insane also kind of in the agentic era, right? Which is you've got all these traces now you need to look at.

13:40Mehdi: I think it's gonna be abstracted, right? It's not I think like it's a bit unfair to call it HTAP, like as a where there is a single hardware storage, you know, and two different query engines, but you could have a layer of abstraction, which is like what they kind of do here, and switch easily I would say compute engine based on a single story or or a slightly different version of your storage.

14:05Jacob: Yeah, yeah, yeah. Yeah, I think that's right. I think that's right. we'll see we'll see where people take these things.

14:10Mehdi: Related to that, because it's kind of the same topic. DugDB in Maria DB. So that's a lot of topic on DugDB today. but so yeah, they Maria Db officially support that. And so we see like we've seen that you know Monday.com, but there is also other big company. I think it's Bidence TikTok. They have also

14:15Jacob: yeah.

14:37Mehdi: Invested into this. I'm not sure if it's this one or another. no, it's Alibaba. It's the other. Yes, exactly. Exactly. the I was I am correcting. Biden's did invest it into DuckDb into the Rust binding. I think the one authoring the Rust binding is at Biden's, but that's I'm not sure about their online

14:40Jacob: Was it no, I think Alibaba also did they did they did DuckDB inside of MySQL, right? Yeah, yeah, yeah, yeah. Yeah, yeah.

15:04Mehdi: L T P database, but definitely. So Maria, so Pasco has with PG DuckDB, right? that we invested, but there is also other initiative now around other OLTP to support inside the DAC, but you still have constraints of course on sharing the same compute. I'm curious like what do you think what do you think is their ta you know their will over there to to do that. It's like do they expect people to actually query using the DugDB engine on the same, you know, shared compute? Or will we see no this is much more a tool for you to export into various formats, much easier?

15:48Jacob: You know, that's a good question. Can you refresh my memory, Matty? Maria D B and MySQL come from the same from the same

15:54Mehdi: Yes, yes, it's a fork. I don't know exactly the story behind it. I could actually look that up. but but yeah, but it's but it

16:02Jacob: No, we don't need to look at it right now. I can't I just I they they have a shared lineage. So I'm not surprised to kind of see MySQL and Maria DB kind of both both take a stab at this. I think to your larger question, where this actually is a really this is a really good question, right? Because the the ch the challenge traditionally with doing hybrid tran hybrid analytics transactional processing is that One of them uses a bunch of cores and one of them doesn't. Right. And so you end up jamming your your machine with analytical queries and the transactions wait, and you basically can't ever tolerate that, right? So I'm not quite sure how people are using this. Maybe, you know, I know that I know that, for example, MySQL has some really powerful sharding implementations. Vitesse is the one that I think comes to mind. So maybe people are just sharding these and like running, running different shards or like clones. or replicas and with analytics and then running, you know, transactions on primary. I actually don't I actually really don't know what the production pattern looks like to deploy this.

17:06Mehdi: Yeah. No, I'm yeah, I'm I'm I'm pretty curious. but I think like if I s if I look at actually super base, what is what they're doing is that they have PG Dog DB, also they're investing kind of, but they're mostly their strategy is to offload to object storage into a proper format.

17:23Jacob: Yeah, yeah. There that's right. That's right. And I think like, you know, there there's definitely industry industry industry moving towards some sort of openness on S3, whether it's iceberg or just parquet or whatever, right? So I think that there's lots of different ways to approach it and it just depends on your architecture and you know, frankly, what works for your users.

17:45Mehdi: Yeah. And I and I think like if you want actually to dump some data into Parquet, like the DuckDB extension is really great, right? And you have various formats. So you get like one extension because that looks like simple in theory to say, yeah, you just export to Postgres table to Parquet. But actually, yeah, you need to you have dedicated extension and so on, but it's not it's not as powerful than the the DuckDB extension. Yeah.

17:55Jacob: Yeah, yeah, totally, totally. Right. Being able to Yeah. Yeah. Yeah. That's right, that's right, that's right.

18:14Mehdi: Next, what do you want? entropic. We're not talking about Bible five, but yeah. So fun fact is that we had this link in the previous spot, but we haven't got the time to talk about it. So I'm cu I'm glad you rigged it out that up. No, but we didn't talk about it, but it was there. Yeah.

18:18Jacob: yeah. I enjoyed We got it twice. man. you didn't talk about it. Okay, great. So we get talk about it now. Very cool. you know, I think what I love about this is just that it's very, very practical and hands-on. you can

18:41Mehdi: Yeah. So just for context for people listening, don't forget. How Entropic enables self service data analytics with cloud. So that's the topic. And yeah, yeah, because we're we're we are also on YouTube. That's that's the that's the plug. If you're listening to us, you can also watch us on YouTube and see what Jacob is talking about. And I have and I have the links on mother duck.com slash explain analyzes. That's the plug.

18:49Jacob: sure. Thank you, Matty. Yeah. Amazing. Yes. Yes. yes, yes. Thank you for keeping me on the rails here. Yeah, so so this post from Anthropic about, you know, self-serve analytics, I think for me, just affirms a lot of the things I've been thinking about. you know, I was really, really excited to see this just because it has such a practical path. And I think it it calls out so on the screen we're looking at this this subtitle, Data Is Not Software. And I think this is this is part of where you have to bring a different set of thinking.

19:09Mehdi: No worries. Okay.

19:35Jacob: to handling data problems than you need to for like traditional software. And like this is where

19:39Mehdi: Yeah, but that wait yeah, I because I strongly disagree with that for some stuff, because I've been claiming data is software for people that are doing data that doesn't have a software engineering background. Everything related to BI, every single tool that BI has code are have this claim data is software. So you should version, you know, do the test and CICD because they are a software asset.

19:43Jacob: Okay.

20:06Mehdi: Right. And it's true in a sense. Like you put your business rule and so on. It's not it shouldn't be BI shouldn't be like a UI where you click and you put some rules and you cannot test it, you cannot deploy in staging and so on. So it's really interesting that like I saw that. And so maybe now you can explain like what do they mean by data is not software. Because for me it was always like if I want to promote best practice to data people that doesn't have software engineering background, I actually claim that.

20:07Jacob: Mm-hmm.

20:34Mehdi: Data is software, so you need to use all the same best practice for s software engineering.

20:39Jacob: So I think I think like the the from my perspective, I think they actually describe it pretty well, kind of in the second paragraph here. And I'll just read it, which says for analytics use cases, there's often only a single correct answer using a single correct source in which there's no deterministic way of proving correctness. Right? And so in that way, in that way, it's not like like I can if I'm writing a software that's, I don't know, a calculator or a spreadsheet, it's very easy for me to say,

20:57Mehdi: Yeah, that's true.

21:06Jacob: To write a test that says, like, make sure the math is right when we use this formula or whatever, right? that's extremely deterministic. The knowing what your revenue is, for example, it seems like it should be extremely deterministic too, but there's so many soft edges from from you know the cycle of like, you know, getting an order from a customer to fulfilling the order to recognizing the revenue that It's very hard to make that a deterministic flow. and so I think that is the core thing that you have to be accept, you know, you have to you have to kind of accept this softness in in analytics and then build build a system that can be resilient to the changes that your business people will bring. And so, like I think that that's the core of this, core of this idea, and I think really cool about what. Anthropic has posted. Of course, the solution is basically like run evals 247. So like, you know, spend more tokens. But like I I don't I don't think like I don't think that's that's wrong. And in fact, I did a little bit of testing around the ideas, surfaced in this and was able to get 100% on a very complex benchmark. So I was I would felt I felt like it was really cool to see that A you could see something like this apply it quickly and see what the results look like.

22:09Mehdi: Yeah.

22:29Jacob: and it it it seems like this is a very good way to move forward.

22:35Mehdi: Yeah, I think I mean they share also a a skill there on how you must do your validation any valves. so that's that's that's pretty nice to to get started. But yeah, I think the TLDR is that it it's like I understand like where when they say data is not soware, it's it's it's harder to to test and predict.

22:42Jacob: Yep. Yeah. Totally. Hm. Yeah.

23:00Mehdi: And so for that, you know, you need to have, you know, specific goals. and the guardrails are sometimes a bit hard. I'm thinking like just out loud, like seasonable data. I had huge challenge for that. Like you have, you know, big sales, Black Friday, for example, or other and then all your models go nuts, or you have to disable them, or you have to because basically you have a such a strong

23:17Jacob: Yep. Yep. Yeah, that's right.

23:27Mehdi: seasonality thing, but it's really a peak on your show. Like would it be an event or a period of week? And and that is super super hard to to to test beforehand, right? It's like you can estimate average, but like how much are you gonna be able to handle this? Like what would be a good numbers on Black Friday for sales to not have like an alert or there is not enough sales? Like who knows? Like that's

23:40Jacob: Absolutely. Mm-hmm, mm-hmm.

23:54Mehdi: Pretty hard. If especially if you don't have a story. like for those kind of seasonality.

23:57Jacob: Yeah, correct. yeah, I mean we gets easier with history. I mean seasonality is always is always tricky, but you know, regardless, you know, these are ultimately reality is not a deterministic system, right? And good analytics models reality as closely as possible. And so how do you get there? Well, this is one way.

24:12Mehdi: Let's t let's stay on Cloud Code. I had another blog from Delba, which has actually just joined recently Cloud Code team. she's been doing YouTube content for for a while. She was at Versal, I believe. Yes. she's great. You know you know her?

24:29Jacob: Yep, yep. Don't I don't but I've seen her content.

24:32Mehdi: you you should know. Okay. so feedback loops help clouds to complete ambitious task with less babysitting. I think that's like one of the big issue we have right now. Would it be for validation or just building, is that how do we actually evaluate, you know, what we are building and have less feedback loop. And so in a sense I think it's related to the previous blog where you want certain accuracy, but you also want to have less babisit, right? It's like I give you clear goals and you test it until until it works. so yeah she gave basically a good you know kind of like framework on write down your processes. So you say you see front hand, what's what's the validate behavior, you use the MCP. And I think like for the The browser MCP if you're doing front end. If you doing data, you define the schema, random migration error against staging, check the raw counts. I feel like she's not a data person, so that's pretty like a lightweight test, right? To just do the raw counts. But I think I've no but it's like it's it's I guess I I it's just an example, of course. but coming back, I think it's related to the blog is that. this is what I've been struggling

25:38Jacob: Yeah, yeah.

25:48Mehdi: right now is like how do I put like my validation, you know, so that he can iterate. Like I feel like UI and front end is kind of like easy. It's just a browser C P. But do you have any tips on other stuff that you've been trying to do kind of this loop with less babysitting and you you give a specific goals?

26:09Jacob: Yeah, I've done a little bit. I was doing actually some query optimization the other day on a tool I've been building. And that was an interesting one where I could kind of say, hey, go look at the query history, build a baseline, and then work, you know, work on this until you have improved it by 75%. and so that was something where I was able to use these loops. It was actually in in codex, not in not in Claude. I don't know if that's heretical, but

26:36Mehdi: Okay. That's good. That's basically it doesn't matter which one you're using.

26:38Jacob: say Yeah, yeah. But but that was actually a really interesting one where you know, it it it was able to work a little bit longer and solve solve the problems. One one thing that's interesting is that Medi, I'd be curious to see to to hear kind of your feedback. When I've been using like go goals and loops, I actually find it really hard to write long running goals and loops. Like What I mean by that is the scope of the problems that I'm working on, kind of in DevRel and, you know, just kind of in general, tend to not be large enough that I'm doing like a multi-hour run. Like usually the the last few times I've given it something that I thought was like really hard, it took like 15 minutes. And I was like, okay, well, like that was that was good. Like, am I not thinking big enough? I'm curious like what your what your thought is.

27:11Mehdi: Yes. Mm-hmm. No, it is true that depending on what you do, and we do a lot of like small tasks like a blog, a video blog, and so on. I think lately there was around the dive gallery hackathons. I plug for the hackathon if you want to go to mother dog.com slash dive gallery, where I'm running an hackathon right now. there was some stuff I needed to do and and some access because we have different region, right? And so I give him like

27:35Jacob: Yeah, sure. Mm, mm-hmm.

27:56Mehdi: My initial reaction was like, okay, let me do that and test it and can you cover the cases? And I was like, no, no, let me actually put all the credentials into one single item in one password, you know, give him access, you know, in for for the shell for the time running, right? And then say, hey, you have all the access there on reads on on read only.

28:18Jacob: Mm-hmm.

28:24Mehdi: make sure all the tests and you you try it out and I think I give a read access token from another account so basically to simulate someone going there, right? And so the point here is that I I I had a good reaction to do it little step by step and me babysitting and say, yeah, you need this, let me give you that. Right. But I think that's the way that I'm trying to change the way I'm thinking is that what are all the access that my agent needs and what can I prepare before the hands for him so that he has all the hands for the sandbox to make it happen. And then you know writing the long prompt and makes it obvious because I say, yeah you can do that, you can do that. And then he figure out himself. Even if you don't they say it's if you highlight you have access to this, you have access to that, then and you give you know a specific goal something is gonna figure out. But he needs then to have access to those I think those those contexts.

29:19Jacob: Tully agree.

29:20Mehdi: yeah. yeah, so that's I think that's the next thing for our our development. It's like just how do we make good goals for you know effective loop where we we babysit less and we still have accuracy. switching on one of your blog, this is one yours, right? I built a very small agent swarm from Tristan Handy DBT.

29:40Jacob: it is. This is one. This is mm-hmm.

29:44Mehdi: Everybody has been calling it still D V T it's Five Straum now, right?

29:48Jacob: I think I think the FiveTran DBT merger closed on June 1st. So this was just before. This was so now they are technically all, I don't know, DBT Tran. Who knows what they're calling each other internally? But yes, FiveTran now. good for them. Very fun. you know, I liked I liked I liked kind of hearing this from the the point of view of of Tristan, honestly.

29:55Mehdi: okay.

30:14Jacob: This notion of building an agent swarm, a very small one, he says. just because, you know, Tristan Brings experience as someone who, you know, was there in the early days, right? He was he was an RJ Metrics guy, I believe, kind of before DBT. So he was kind of there at the very, very ground level and was, you know, doing things when they were way harder than they are today. And so I just think his perspective on this is worth sharing. You know, what does it look like as someone who probably like honestly, like w I think what we're also seeing kind of in the AI era, is that a lot of leaders are getting back into code, right?

30:51Mehdi: Yes, that's true. That's that is that it it it is insane. Yeah.

30:55Jacob: I think like Trist Tristan is one of many, right? Like I think Toby at Shopify, I think, was always probably writing code, but now he's writing so much code, I think, from what I'm hearing. you know, he built a really nice QMD QMD library, for example, for like, you know, using using memory for agents.

31:08Mehdi: Yeah, yeah, yeah. Yeah, there's also Maxim Beauchemin from the creator of Airflow, which created his own company Preset. And he's been like on a rampage with agent or orchestration and just building building. He's originally a builder. I feel it's a lot it's a sidetrack, but a lot of like founders that were originally builders, right? Not coming from really engineering backgrounds and we are you know, kind of cold or force or whatever they're starting to leadership. And now they're like, okay, I now I have an excuse to to go back and build build again. Yeah. But so what's what's the key takeaway on the on on this one specifically for you? Aside from like

31:47Jacob: Yeah. Yeah, totally. I mean, I think I I think you know, I think a little bit here is that for me, it validates this validates a lot of the things that I've been thinking about too, which is like, how do we decompose our work so that we can do it more effectively? and he you know, again, he's sharing it from very much like a data, a data person point of view. you know, what's working for him? I think this also ties into what we're talking about about goals earlier, right? Which is like we have to rethink the way like doing the your job the same way that you were doing it two years ago is not gonna work in two years from now, right? And so this is for me a lot of this is just kind of understanding how other people are thinking about breaking these things down and like doing doing work around doing work around data. In a way that is agenc first, if that makes sense.

32:54Mehdi: Yeah. Yeah, yeah. No, yeah, it's it's interesting to see people, you know, playing around and and especially as you call out. I think leaders that that keep like are in active mode of building, like even on the side, I think is worth following up. All right, another Plug from my site. okay, so this is an interesting one. An app can be home quick meal. And I think it's all about that story. So the the way I wanted to bring this one is that basically this guy launched an app in 2020 and it's only for he him and his family. Like there are four people, zero searcher. So it's been a couple of years, like you you see an updating, you see February.

33:19Jacob: Ooh. Yeah. Yep.

33:45Mehdi: I didn't change literally nothing in the app and it's still working. This is pre-Jad GPT era. So it literally built something and just like just for context. Kind of like a be real thing where you record a short moment, you know, and then you send it on a feed, but it's a personal feed with your family, right? it's really simple, and you know, if family like it. And I was thinking, I think we're gonna see much more on that. Like I was

34:03Jacob: Mm-hmm.

34:11Mehdi: I found like super interesting to see, you know, this in the pre pre build your own stuff era. but I was wondering like what have you built stuff for you and your family that you've been using and now and you're like, yeah, I don't need like a specific service, it's like custom personal software.

34:28Jacob: So Yeah, this is really interesting. I mean, I think like a couple things come to mind. So one of them is that there was a game going viral on Twitter called 82 and 0, which was a NBA roster builder where you try to get a perfect team. And I just said, I could make a better version of that. And so I did that. And now my the last the last three nights, my daughter has like, Hey, can we like play, can we play against each other? Which has been fun. so like, you know, that was something that I would never have built before and definitely could build now.

34:57Mehdi: Yeah.

35:02Jacob: and so that that's been that's been really fun, kind of just being able to explore what what's possible there built on top of Mother Duck, of course, by the way. shameless plug. but yeah, I definitely, you know, I haven't I haven't taken the dive into something that's just for like, you know, immediate family yet, but I suspect that I will quite soon. You know, it's actually I'll actually call so my daughter plays competitive softball. So this is really interesting. I don't know if this was the vibe-coded app or not. But I was sitting next to someone yesterday at practice, and she actually built an app that tracks all of the tryouts for all the teams across the US in the summer. And so I I don't, it's a new app, but like everyone uses it. It everyone in softball knows what it is, and she's the founder of it. It's just like one person. And I was just like, I was like, whoa, like this is crazy.

35:47Mehdi: Yeah, I am. I think I think it's gonna I think there is yeah there there is a lot of stuff like this where there's still a huge like Jacob we are still like we are in a bubble and you are if you're listening and working in tech but like outside that's like I'm just like l looking for new school for my kids and I look at the website and I'm like Jesus Christ this is horrible and I'm like and my wife say

36:12Jacob: I have that thought. I have that same thought.

36:14Mehdi: They probably don't have budget. And I'm like, which budget do they need? They can just go to Lovable, right? Don't tell me they don't have 30 bucks. but they don't know that. They don't know that. And maybe they need help. And I was like, I think there is like budget to do, to just do all the the website of every school. And you just go to Lovable. Yeah. Yeah. So what I did actually because I was just telling you I was house hunting.

36:18Jacob: Yeah, yeah, yeah. Yeah, yeah. Yeah. No, correct. Totally. Met what are you building? So what are you gonna build, Medi? What are you what is your personal app?

36:42Mehdi: So I did just a scrapper. I so the market here, at least in Europe, in Belgium, is that people are putting the real estate on their own agency, real estate, and then on a main website. Like there is two or three main websites for the whole Belgium and it there is that for each country. And so the way to detect the good deals is basically to scrap those agencies because they put them on their websites and to their database and then if it doesn't work, they put it on the main websites, right?

36:54Jacob: Mm-hmm. Yeah. Yep.

37:10Mehdi: so I just built the scrapper that list all of that. Tend to just realize on Reddit someone did exactly the same. And it works. So I literally called an agent before he actually shared with his own custom like his own database customers. Like it he asked me, like how the hell did you know? Like he's just a solo person agent, right? Not even like an agency. So I have a small website. How do you how did you know that like

37:33Jacob: Yeah, yeah.

37:39Mehdi: This house available. I didn't even send an email to and I was like, Yeah, just build that robot. He he found it, you know, fascinating. And it's just to show you again, there is a huge lag on what you can actually do. I think the tryouts example is like, I'm sure people are like, Wow, but they have no idea. That's like, yeah, it's not that hard to to build, right? So yeah.

37:52Jacob: Yeah. Yeah. Incredible. Great. That's amazing. love it.

38:01Mehdi: So yeah, inspiration for you to build your own software. You need to come back on the pod and say, okay, I've built, you know, some stuff for for me or my family. Yeah. all right, I think we're getting close to the end and we have Bob this one. Doombench.

38:19Jacob: this is fun. I always love these. I always love love these. So this is introducing Doombench, can your data stack run Doom from our friends at Cedar DB, which is you know a database doing doing things very fast. and so they said, can we run can we run Doom on our database? And then of course they made it so you could test on other databases too. This is just a fun one for me. I think this actually goes into like personal apps apps that are just for you, right? So yeah.

38:52Mehdi: Yeah. Yeah. I think I think maybe we we are coming back. I mean we see if like mostly with videos like AI stuff, you know, if Harry Potter was you know all jacked. I don't know if you saw that one. Like so like crazy kind of like parallel multiverse with like looking really really real like the you know now the mod the video models have been better.

39:08Jacob: Yeah.

39:20Mehdi: But that's like just pure entertainment, like video, I would say entertainment. And I'm sure we're gonna see like the like the beginning of the internet, just weird software. that is like this one to test, you know, can you run Doom on your database? I think it's fun, it's nice, it brings a bit more like the early days of the internet where you know you had just your HTML website and anyone could do anything. so so yeah.

39:24Jacob: Mm-hmm. Yeah.

39:48Mehdi: And so d have you have you tried Doug D B and Mododoc over there?

39:52Jacob: I haven't I have not tried it yet. I've been doing other things, but you know, I think

39:56Mehdi: As they say, but your own on benchmark, push their own benchmark.

39:59Jacob: Yeah, yeah yeah, exactly. you know, these like I said, these these are these are always fun just to see because there's yeah, they're linking to all the benchmarks where that show were who's the best at what. So, you know, I think they take I mean, you know, we've talked to Jordan about benchmarking, you know, at Mother Duck and he's always kinda like, Well, now you now you're in the game of like running a benchmark. Like is that the game you want to play? Let's build an awesome customer experience. That being said, I think building fun benchmarks that are whimsical like this.

40:08Mehdi: Yeah, yeah, yeah, I see. Because the way yeah, yeah. Yeah. Yes.

40:28Jacob: I'm all for. We need we need duckbench that is like this.

40:29Mehdi: Yeah. I chose my database based on if it can root run Doom or not. That's the only thing I care about.

40:34Jacob: Yeah. Right. Yeah. Can I do ray casting with my database? This is a great question.

40:41Mehdi: Exactly. Yes. All right. Let's think to close. It's a bit outside, but I think it was a beautiful one. I haven't finished yet. so if you don't you know Curt Repo, it was renamed, it was only something. They used to do long documentary on programming language, like they have the Python documentary, which is amazing. They just released one last week on C and they do a ton of like

40:53Jacob: Like you. Mm-hmm.

41:10Mehdi: you know major contributor and software people in the industry that contribute to C or C. I think they were creator of TIES Crypto. And I think it's just beautiful to hear the story on how it started and where it has been to up to now which was raising my question with when watching this content. That's why I would recommend you to watch this content is that it just raised question on where are we going now? Are we gonna see

41:15Jacob: Mm-hmm.

41:38Mehdi: A language that we men don't understand or will never that happen? Like that is just more efficient for AI and we keep English? Or do we see still like, you know, an interface that we absolutely need? I don't know, what's your take on that? I found like I don't have specific answer on that, but I found the question very interesting.

41:58Jacob: Yeah, I mean I think, you know, the the it's an interesting question, right? Because if you think about A lot of the challenges with programming languages is you're taking something English or you know any any spoken language, which is not which is not deterministic, right? The what you're saying depends on the person you're talking to, right? they're gonna their interpretation is really important. And so what we're doing, what we do with like, you know, coding languages and and compilers is make that deterministic, right? And we did that so that we humans could talk better with computers. And so I think there's a really interesting open question, which you're getting at, which is like, well, if computers are just talking to computers, how would they talk? Right? and we don't actually know that, but I sus I suspect we'll see something like that. I'm I'm reminded of did you ever play the Sims? Do you remember like the Sims? They would talk a language that was like just Simlish, and they but you couldn't understand it, but they could understand it, right? Obviously, they're they're it's Sims, so like they they didn't really understand it. But you get my point. Like, there's some language I'm sure that will exist.

42:47Mehdi: Yeah, you do done. Yeah. Yeah. Yeah.

42:59Jacob: that's a higher level abstraction than zeros and ones that agents and and new programming languages will be able to use that will evolve.

43:07Mehdi: Yeah. But it's still like what what is the interface like today? You know, TypeScript is winning a lot because there is so much trained data also on what is a good software. but there is also some take. There was another blog on, you know, once you reach a certain level of intelligence, it's easy to have this loop of making things, you know, just better, and to create those new frameworks, those new programming language. And I was just thinking, wow, how would like a new programming language in the era of AI looks like? and like even if we have like this computer talking to computer thing is like just I'm I'm just curious because today I think we we've been on a race and just you know double down on okay, that's existing, that works, okay. But what what's what's the next level? I'm I'm I'm really curious. And I feel like, yeah, going back to the roots and the history. You know, you often like I I don't know if you I I look I read a bit of like history book and it's super interesting to see like the recycle and repetition in our, you know, modern era. And so in that way I I feel like if you learn the story about C plus but how it was being built and how they they decided to build that and it's just crazy. They were mentioning like back then the editor you had to sometimes delete the whole line if you want to modify one thing.

44:12Jacob: Yep.

44:28Mehdi: Like there was no cursor for index. It's just crazy, right? And so you just think like so much things has been built around like some early challenges. And so and now we're just in a new era. So yeah. Yeah. Yeah. Thanks. But thanks again for for joining. if you made it to the end, you can also made it to mother duck.com slash explain analyze.

44:30Jacob: Yeah, yeah. Yeah. wow. Totally, totally. So cool. So cool. Well great pod, great pod, Matty. Good links.

44:55Mehdi: We are also available on Spotify, Apple Music or YouTube. But you can get all the show notes and links we've been discussing today. And thank you, Jacob. And I hope you you'll come back to to talk about you know, money or anything else that you enjoy.

45:12Jacob: Of course, we will definitely be back on. Thank you, Maddie, for hosting. great. Thanks, everyone, for joining us.

45:15Mehdi: Yeah, yeah.