The Safest Bet in AI Is Hardware

0:00Mehdi: Hello and welcome

0:00Jacob: Okay, amazing.

0:01Mehdi: to another podcast episode of Explain Analyze Podcast. and this week I am with Jacob. Jacob, how's it going?

0:12Jacob: Hi Matty, how are you doing?

0:14Mehdi: I'm doing great. I mean it's still a bit summer vibe, and you know, things are a bit slow.

0:20Mehdi: But I would say on not really on news. I don't know if you feel the same way. It feels like, you know, some part of the world are on holidays and some don't.

0:31Jacob: Yeah, I think that's exactly right. I guess the agents are still working while we go on vacation. May maybe that's the

0:36Mehdi: Ooh.

0:36Jacob: destination for everyone.

0:37Mehdi: That's a that's a really good framing. so as usual

0:40Jacob: huh.

0:41Mehdi: we have roughly around ten links of news around data and ai that we just shared. I haven't seen the links of Jacob, he hasn't seen the links of mine and we just have a discussion around it. Jacob, you want to start maybe with this one?

1:02Jacob: Yeah, sure, sure, sure. all right, so meta anthropic intox for potential 10 billion compute lease deal sources say. I think you know, one of the things that's been really interesting in the kind of AI or the business of AI has been that we're seeing winners and losers kind of play out in terms of adoption of their models. But even the folks that are, you know, kind of losing in terms of what they ended up building, in this case meta.

1:35Jacob: they're still winning because now Anthropic is going to lease some of their compute. And so I

1:39Mehdi: Yeah.

1:39Jacob: think it's been very interesting to see that, like you know, the companies who made big bets on hardware maybe were not able to execute on software. And the companies that are executing like crazy on software, Anthropic and OpenAI, for example, are underindexed on hardware. And you know, in the in the past.

2:01Jacob: In the past, the hyperscalers just handled all this for us, right? and now AWS and GCP and Azure don't have enough hardware for us, right? so what do you do? And I think that's what we're seeing here. It's been really, really interesting to see it play out.

2:18Mehdi: Have

2:18Mehdi: you have you seen also Mistral is building data center in EU?

2:23Jacob: Okay, I did not know that, but it makes sense. If I if I was if I was them, I would do it too, right? I think like,

2:25Mehdi: Yeah, yeah. It's a big yeah.

2:29Jacob: you know, we're seeing we're seeing a ton of strain on power, you know, power consumption here in the US as well. you know, I would assume that in Europe that is even more top of mind. and so how do you how do you build these things? And I think you know, you you if you want to be a serious player, and I think it's important for there to be a European lab.

2:53Jacob: like Miss Troll that that is you know, take taking their approach to this and training the model they it the way they want to train it and not making us all reliant on OpenAI and anthropic is amazing. And so super, super cool to see that happening too.

3:07Mehdi: But I think what's interesting is that there was another blog long time ago that said that all this companies are invest investing billions in data center today and it's all it's still a bet if they will succeed, like Mistrial if they will succeed.

3:24Mehdi: But this story is interesting because I haven't thought about it where actually investing infrastructure can be a a safe plan B, you know? If Mistra doesn't

3:33Jacob: Yeah, absolutely.

3:34Mehdi: work on the software then yeah, we can still lease our hardware in EU for a tropic. Who knows?

3:42Jacob: Yeah, yeah, yeah. Total totally. Tot exactly. Or another European lab, right? so

3:45Mehdi: Yes, probably that.

3:48Jacob: hopefully. you know, yeah, that's it's been you know, I think you know, tons of big we we've seen lots of other big winners in the infrastructure space, right? Mother Duck lives more in the infrastructure space than the AI space. So like how do you you know, what does it look like? I I I think there's a lot of tailwinds for infrastructure comp or companies building at the infrastructure layer, right? Enabling

4:10Mehdi: Yeah.

4:10Jacob: AI. So

4:11Mehdi: Yeah.

4:11Jacob: Very interesting.

4:13Mehdi: Cool. next let's talk about this one. So there is I have actually three three links. and they all had a similar link. so there is this is the

4:32Mehdi: talks from the AI engineer Worldfare. so they just

4:37Jacob: Mm-hmm.

4:37Mehdi: published it on their YouTube channel, a series, all the I'm I think almost all the talks. And you have this one, Active Graph Agent Run runtime. you have your moat is a data model from Mike Phillip Phipps, and here is another one from a person from Neo4G, so that's it's a bit biased, of course. graph.

5:02Mehdi: database but they all have a same thing where they talk how we are moving towards graph context for agents

5:14Jacob: Mm-hmm.

5:15Mehdi: so you've been working on features around context layer for for mother duck so do you do you have any opinion or thoughts on this and you have you have you read a couple of things around this

5:27Jacob: I mean, yeah, I think so first off, the these are videos are awesome. I'm gonna have to like look at the show notes for this one because I'm gonna have to watch these. This looks amazing. I think what we're discovering i especially in the notion of

5:44Jacob: You know, empowering AI to use context, that a a graph is really important for making sure that it finds the right thing super reliably. And one thing that we've kind of turned off in our brains is the fact that like we have access to this amazing graph that Google has curated, right? but it only works for kind of the public open internet. And

6:11Jacob: Now we have n now those same principles actually can be applied to our own internal data and how do we find the right things super reliably is really interesting. Really, really interesting, right? because like when you're a human in the loop, it's okay if you get like five links and only one of them is right and four are bad.

6:27Mehdi: Yeah, and then you do you

6:29Mehdi: but you do you do kind of like a graph, you know, cross through a mind in the sense that you go through one link, you see a connection that that

6:37Jacob: Yeah, yeah, yeah. Yeah.

6:39Mehdi: that is to the other link, right? And

6:41Jacob: Mm-hmm, mm-hmm.

6:42Mehdi: then you try to, you know, make a decision based on that.

6:46Jacob: Yeah, that's exactly right. That's exactly right. yeah, I mean I I'm curious, Medi, like, you know, has that changed how you have approached, you know, what you're thinking about from a data engineering perspective? Or is this kind of just like, you know, new ideas for you at the moment?

7:03Mehdi: so to your answers I I think what's interesting is that we had graph database as a really niche database. I'm a bit skeptical on

7:20Mehdi: How valid the model is to use actually a graph database. I think the graph modeling is maybe better framing in a sense that I need to see on the application side and runtime is that do we actually need the complexity of a graph database or is just

7:44Mehdi: you know, a specific framework that is designed as a graph that enables the agent to to basically cross the information. As we see

7:52Jacob: Yeah.

7:52Mehdi: here, for example, as a state, right? And if it's a state, maybe it's just a gra you know a database in memory and and that's it. And that would be maybe end handled by the labs for you and you know and the way you will mostly interact it to fit to fit that graph.

8:13Mehdi: So all in all is that I'm really it's since the vector database, I'm really skeptical about niche database. I think the only two worth existing I three existing databases are anytime database, transactional database, and time series database. yeah, for the rest I think it's really still niche. we saw that

8:39Mehdi: s a lot of people said that they needed a vector database and most of the case, yeah, they don't. they it still needed

8:45Jacob: Yeah, sure, sure.

8:46Mehdi: for some things, but if you remember like just like two two, three years ago, you know, how they're called again, Wave and others, Chroma were all about the rage. yeah,

9:01Jacob: Pine cone, yeah, yeah.

9:02Mehdi: Pinecone.

9:04Mehdi: And I think they're losing steam because other things are being abstracted by the labs, right?

9:11Mehdi: Or they we realize that actually we don't need that much because if you remember it was all about doing embeddings, to so that

9:19Jacob: Yep, yep.

9:20Mehdi: the agent has a memory and so on, but no, it's it's almost built in. I'm not saying we don't need it, I'm just saying it's it's really a niche use case. So maybe a lab needs in, but not you as a data engineer. That's that's the thing. So I'm trying still trying to figure out what is happening here. Is there really a transition where we're gonna need to

9:41Mehdi: get our hands dirty around grift graph database and setting up in our stack or is it more like vector database a way of thinking and that will be abstracted and used by the labs and yeah you just had to know that has it's you know what's an embedding here you know today

9:59Jacob: Yeah.

10:00Mehdi: you you know what's a graph notes tomorrow

10:03Jacob: Hmm, interesting. Yeah, well we'll see, I guess. Also, I'm a little

10:08Mehdi: Yeah.

10:08Jacob: offended. You didn't have ledger databases as one of your categories for databases? Come on.

10:15Mehdi: That is that is true. I don't know, I it was just on top it

10:17Jacob: Shout out Tiger Beetle. Yeah, yeah, yeah.

10:20Mehdi: was just on top of my bat. But alright.

10:22Jacob: Ha ha ha.

10:23Mehdi: no comment no comment on that. all right. use my no AI slop skills to remove

10:31Jacob: Yeah, this is

10:32Jacob: this is interesting. So I think one of the things we're all seeing, and we can talk there there's a little bit of a through line, I think, in what in my articles, but I'll we'll we'll get to it. well let me ask you this. Medi, have you have you tried to use AI to do writing?

10:53Mehdi: Yeah, yeah, of course. Yeah. And I think I I think I fall

10:55Jacob: And w wha how's your experience how's your experience been?

10:58Mehdi: I think I fall into the trap to use too much. and I think it's okay.

11:05Mehdi: As long as you recognize this is too much, like there is too much pattern, it's not respecting your voice. So the the workflow I'm using today is I, you know, index it. The the cool thing is that for me and for you too as well, we have tons of content and data where we didn't use AI to write, right?

11:23Jacob: Yep. Yep. Yeah.

11:25Mehdi: So

11:27Mehdi: When I say I use too much I felt like I was losing my voice. It was like there was all this EI

11:31Jacob: Yeah, I agree.

11:31Mehdi: slop in there and I felt a bit ashamed that people recognize that. and I think it's okay. I

11:36Jacob: Yep, me too.

11:37Mehdi: think as as as long as you you catch this up. and so now I have a workflow that just, you know, look at all my existing blogs and I have those kind of skills like to remove the the slop skills. But the best way I still use for writing is that

11:57Mehdi: You know, do most of the draft yourself and then, you know, correct the grammar and extend the missing gap. But even though when you have like a good I would say 70% there, sometimes the thirty percent still contains a lot of slab. Well what is what

12:14Jacob: Yeah, I totally agree.

12:16Mehdi: is what is your experience? I I'm curious like to today, latest thing of about AI and writing.

12:22Jacob: Yeah.

12:23Jacob: So I'm like I think like the pendulum keeps swinging for me. What you know, the initial initially like let's say a year or a year and a half ago, did

12:32Mehdi: Mm-hmm.

12:33Jacob: a bunch of experimentation with it, you know, kind of an around the notion of like, okay, like what SEO is missing in our in our

12:40Mehdi: Yeah, I remember.

12:41Jacob: website and like what can we build? And it was just like, my god, like this is horrible.

12:48Jacob: And then I spent some time like tuning like voice skill and say, okay, like I wrote all this stuff, help me, you know, tune this on my voice. And it did okay. But then upon reflection, like a couple months later, I look back at those those pieces and I'm like, this is rough. Like you can tell this was written with AI,

13:04Mehdi: Yeah.

13:05Jacob: even though I spent a bunch of time on that. and so what I'm kind of thinking, what what I've been doing lately is actually been really using it to help me shape narrative. Like, here's the points to hit in what order, and like

13:17Jacob: pressure test, you know, the story, but then also but I do the writing. So it's almost like my and it's more

13:23Mehdi: Okay.

13:24Jacob: of an assistant than like on primary. And then the other thing I like to do and that I found the the what I would say is like the best AI assisted articles I've written have been like where I interview a subject matter expert and I take that transcript and I use that to turn it into the article. It does a really good job of summarizing and concisely explaining it without losing the author's voice.

13:41Mehdi: Yeah. Yeah. Yeah. I

13:46Mehdi: I yeah, I had a big yeah,

13:46Jacob: In a way that's better than me. Yeah.

13:49Mehdi: I had I had a project around that to basically summarize talks and convert them into blogs and I think it's i it's still worth it to do. but yeah, converting a talk to a blog is really nice because you have much more I would say

14:08Mehdi: colours in twenty-five minute talks, right? To get kind of the voice of the author. but but yeah I do I do agree that I haven't thought so you just to to to understand you is that you you ask l roughly the AI to do like a table of contents with a specific thing and then you do the writing.

14:33Jacob: Yeah, that's what that's what I found is the best way to like help me solve the blank page problem. A lot of

14:39Mehdi: Yeah.

14:39Jacob: the challenge that I run into with writing is just like staring at the blank page. And it's almost like I it's almost like switching the role of the prompter, where I'm like, okay, AI, write, write a prompt for me to to write the article. But it totally works.

14:52Mehdi: Yeah. And

14:55Jacob: I've I found the results are good so far.

14:57Mehdi: Okay. No, it's something I haven't tried. I mostly just try to put a eye at the end, right? Rather than at the beginning and then fine-tuning.

15:06Jacob: Yeah. Mm-hmm. Mm-hmm. Yeah, I think ultimately

15:10Jacob: the thing I I care about is like I want to show respect to my audience. Right. Like

15:14Mehdi: Yeah.

15:14Jacob: if it's just like me brain dumping and then AI editing it, like I don't know how interesting that is for them to hear. you know the

15:22Mehdi: I really don't know.

15:23Mehdi: I really don't know. Like it's a really good question that actually there is a blog I I have here that is obviously written by AI, but the content was good. So I would say the frame got me a bit, you know, triggered. But I still put it there this week because it's it's it's a good there. But just to close

15:41Jacob: Yeah, yeah, yeah.

15:42Mehdi: on d on this one. so did you try it? Because I looked at the skills actually yesterday.

15:47Mehdi: And I was like, yeah, okay, the the skill is is really pretty light. I mean Peter Young has a good audience, so yeah I say I love all like the stars matrix on GitHub are completely done now. Like a project

16:02Jacob: Yeah, yeah, sure.

16:03Mehdi: of that size for like basically n ninety-three lights. Wow. so

16:06Jacob: Yeah, like how many lines are on the file? Yeah, yeah. Yeah.

16:14Mehdi: yeah, did did you test it or

16:17Jacob: I haven't tested it, I just saw going viral. I've been using the humanizer skill, which is very similar to this. But the problem that

16:21Mehdi: Yeah. So very similar.

16:23Jacob: the problem that I s that I have found is that like, for example, like, all right, don't use dashes. And so then you what you'll notice is that like people will write their sentences and they'll just put periods there instead of using an M-dash. And it actually flows worse, but they're like, well, dashes are an obvious AI tell, so I can't use it. And like that, they make their own writing worse. Yeah. Yeah.

16:41Mehdi: Yeah yes. They took that from us. They will never get back.

16:46Jacob: No, the dashes are dead.

16:47Mehdi: They're gonna

16:48Mehdi: take other punctuation. And then what what's left for us human?

16:50Jacob: So like Yeah, exactly. Exactly. I

16:54Jacob: do think that that in general the now that I've been using the humanizer skill and I'll check out what Peter built here, which I thought was really interesting.

17:06Jacob: Now I now I see people who have written with AI and then have ran these skills on them and then thought that was good enough. And I find I

17:12Mehdi: Yeah.

17:13Jacob: I notice that now too. So I'm like, like

17:15Mehdi: Mm.

17:16Jacob: and I see that in my in some of the writing that I did over the last few months too. I'm like, crap, I just like, you know, I need to rewrite these. This is this is not capturing the the real art of what it takes to write here. And so I think like ultimately,

17:27Mehdi: Yeah.

17:28Jacob: ultimately what we all want is an AI editor. and that's really, really hard to do well.

17:35Mehdi: Yeah.

17:35Jacob: So good luck.

17:36Mehdi: I think there is the there is still the workflow to figure it out, right? If you

17:41Jacob: Yep.

17:42Mehdi: Just write everything yourself. But it's a I mean it's a I think it's an interesting way. We do see like things are changing right now. Like people spotting

17:51Jacob: Mm-hmm.

17:51Mehdi: up and calling out and and using less. So I think that's that's good. We are just balancing. we speak over another blog that was written by I. No, I don't know how much is it written by I but it is obvious I find. this is

18:13Mehdi: Wait, wait, wait. Let me find quickly. I think you know the the the the negation stuff like this is not X, this is this, it's it's there

18:23Jacob: Yeah.

18:23Mehdi: some somewhere. But yeah, all in all the the content is still is still pretty good. so

18:33Mehdi: why it worked on my machine still happens in twenty twenty-six. so it's pr it's it's a really long actually blog. And the first part is all about I think

18:48Mehdi: showing that what's what is important is actually runtime that differs still like you know ai has been good at at coding infrastructure as a code and so on but you still have a really you know big difference between your local machine and and remote production and that most

19:08Jacob: Mm-hmm, mm-hmm. Mm-hmm.

19:11Mehdi: error actually comes from from configuration

19:15Mehdi: Like you say, for example, when an engineer has a new variable that chains has to provide every environment rel reliably. And and so I had that and I think as you as you do complex application with you know authorization sign up, so you have all token, authorization domain. So and you want to test that locally, you have to

19:42Mehdi: do a bunch of things to mimic that, right? But it's still not close to to what's actually run in production. so yeah I found like this is actually interesting to feel that there is still that problem happening and there's still major like I would say infrastructure challenge is talking about you know pass

20:10Mehdi: So platform as infrastructure and mentioning that team we're moving to that, but you know, initially

20:18Mehdi: People were just used to use Kubernetes and their tooling. So they have that knowledge existing. And so there is a movement going there slowly. But I can see, for example, Versel doing really well at that, where you know, you just provide, I have it on the screen here, a

20:34Jacob: I love Russell.

20:35Mehdi: a YAML file. I mean it's it's all of them, right?

20:37Jacob: Yeah, yeah, yeah.

20:38Mehdi: but YAML file, GitHub Action is also, you know, platform as as a service. So just a YAML file and you get your runtime and everything.

20:48Mehdi: So yeah, I think that was just an interesting reflection on this. I don't know. Did you do you still have this issue or have you you haven't thought on it or yeah.

21:01Jacob: No, I think definitely you know, so I have this issue less, I think, because the choices that I make are in acknowledgement of this, right? So when I'm doing something that's going on the web, I'm using Next.js. Right? So that like I know that the deploy is me going into the CLI and typing Vercell prod.

21:25Mehdi: Yeah.

21:27Jacob: so I think like that also limits

21:29Jacob: some extent the scope of the problems I'm solving. But like, you know, in general, the things that I'm building work well enough this way. Our our tooling has gotten a lot better, but I think, you know, once you get more complex, you know, I'm not doing I'm not doing anything that's running Kubernetes, for example, or like not at any reasonable scale that would matter. And so I don't see this as much since I am kind of not working at that high scale of a space. But like

21:57Jacob: It make the the the conceit here makes sense to me. And I think like I'd be curious I'd be curious around like what scale people start to need to go away from something like a Vercel and into something where they're running it themselves on AWS. Cause that seems like the the tipping point. Okay, this is costing me so much to run on Vercel. so

22:24Mehdi: I don't I

22:25Mehdi: don't think there is a the sc a scale issue there. I mean there it's one there there is that, but I think it's just in the blog what I mentioned is just history. You have a setup, you have a Kubernetes cluster to deploy a website. Why would

22:39Jacob: Mm-hmm. Mm-hmm. Mm-hmm.

22:40Mehdi: you go to Versal? You have this knowledge, you have a dedicated team that manage this,

22:43Jacob: Mm-hmm, mm-hmm.

22:44Mehdi: right?

22:44Jacob: Mm-hmm.

22:45Mehdi: and so getting away from this is hard. A bit kind of like the same analogy on you know hyperscaler versus managing your own cloud in your own data center, right? why would you do this? Do you drop this? Do you drop the team and then the team needs to learn that tool, right? And it's still

23:05Jacob: Mm-hmm.

23:05Mehdi: a bit of a black box and every specific workflow. So there is that. but

23:12Mehdi: For me, I mean also the other thing that I mentioned is around dependencies and trust transient

23:17Jacob: Mm-hmm.

23:19Mehdi: you know dependency. And we've seen like tons of hack, you know, in especially you you're talking about NextGS and the NPM ecosystem, where, you know, an update is being done somewhere and boom, things are not working anymore.

23:35Mehdi: so that also that blew up, I feel over time, and that will blow up even more because it's super easy to build a package. So it's super easy to build a dependency

23:46Jacob: Yep. Yeah. Yep.

23:49Mehdi: that will just be critical at some point. I like I

23:53Jacob: Yes.

23:54Mehdi: build libraries, you build libraries, right? What

23:56Jacob: Yep. Yep.

23:58Mehdi: like how many like I have this challenge too?

24:02Mehdi: It works on my machine. Like how do you trust today libraries that you install given that there is so many vibe coded that will never be maintained again?

24:12Jacob: Yeah, that's such a good question. I mean, I think like the ecosystem took a critical dependency on friction being high to deploy a new package, for example. And now it's so easy that, you know, I've published a package to NPM that has, you know, thousands of downloads per week now.

24:31Mehdi: Yeah.

24:32Jacob: that doesn't mean it's not useful and it's not valuable. Like I think it is. That's why I built it. But also, that would have been really hard for me to do three years ago.

24:41Jacob: And you know, how do you trust that that yeah, how do you trust it? It's a really interesting question.

24:47Mehdi: So yeah, so you don't have the same. So you you're doing this side project. I guess it's a side project, right? Correct me if I'm wrong.

24:58Jacob: Yeah, sure, yeah, yeah, yeah.

24:59Mehdi: is it the MVs package? What is it? What package is it? Okay, so

25:02Jacob: Yeah, the invis package, yeah. Mm-hmm. Mm-hmm.

25:03Mehdi: so that you you have your plug. so but

25:05Jacob: Yeah, thank you.

25:08Mehdi: but so you don't have the same expectation maybe that the one is gonna use it.

25:15Mehdi: Maybe the person that's using it has crazy CI with testing or you know production

25:20Jacob: Yeah.

25:21Mehdi: stuff. And then you're like, I'm doing a new version for this side project. There is a breaking change. Meh, who cares? That's true. There's nothing wrong with that. It's

25:27Jacob: Correct, yeah. Yeah.

25:31Mehdi: just the expectation, right? Expectation from the other and expectation from the user. so

25:37Jacob: Correct, correct. Yeah.

25:39Mehdi: So yeah, then then those kind of things will still will still happen and I think still happen more. That's

25:44Jacob: Yeah, we'll see what happens.

25:45Mehdi: the the the conclusion. All right. y y Jacob, you're always bringing like dollars block.

25:53Jacob: I am bringing dollars. we

25:55Jacob: can we can we'll talk briefly about this one. I think I had intended for these these kind of businessy links to be like grouped together. So we'll just talk about this one quickly. But

26:02Mehdi: No no, it's nice. It's nice. It's nice to know.

26:05Jacob: my bad. AMD commits up to five billion to anthropic. I think this just goes into more of the hey, it's actually pretty safe to make an infrastructure bet. and the reason is because there's so much demand for this stuff.

26:22Jacob: That even if Anthropic doesn't use it, let's say that like their next model's a flop and everyone

26:25Mehdi: Yeah.

26:26Jacob: goes to Meta or something, who cares? They're still consumers for the new GPUs.

26:30Mehdi: Yes.

26:32Jacob: and actually I think one one of the things that's actually coming out of this that is an interesting call-out from this article is that the financing the financing for these data centers and so on are looking a lot more like real estate plays than they look

26:44Mehdi: Mm-hmm.

26:44Jacob: like technology infrastructure. And so just something to keep an eye on. Like, is it like the

26:51Jacob: That sometimes that can be indicative of hey, there's a bubble here and maybe there is. but also it tells us that the market has a lot of faith in these being good investments as they move kind of to more mature instruments. So very interesting to see that too.

27:06Mehdi: Okay. there is there was another one dollars? This one? no, that's the same.

27:10Jacob: that was this one. But this one th

27:12Jacob: the this is the Financial Times link. This one actually ha had the mo most of the detail about

27:17Mehdi: Okay.

27:18Jacob: about the real estate side of it. But anyways, we we can we can move on because that's just kind of the business the business side of, you know, AI, which is thriving.

27:25Mehdi: Yeah, yeah.

27:26Mehdi: okay. I'll I'll have a business link and then we stop.

27:29Jacob: Yeah.

27:32Mehdi: exactly Domo which was basically a BI tool, right?

27:38Jacob: Mm-hmm.

27:39Mehdi: I'm not sure what they rebranded this last year's has been acquired for 400 millions and

27:50Mehdi: Yeah, so it is it is feels like it's the end of BI tools or at least the end as a standalone vendor, probably.

28:01Jacob: Mm-hmm, mm-hmm.

28:03Mehdi: and I don't know, it feels

28:10Mehdi: It feels still a lot, but I'm not sure a l a lot of money, four hundred millions for for them. It's press probably like acquiring the the team and the and the customer. What's your what's your thought on it?

28:24Jacob: Yeah, I mean I think, you know, they they had a mature a mature product, not a lot of growth in the business, and you know, not a clear pathway to more growth. And I think their their valuation reflected that in the in the public market. not to mention, you know,

28:39Mehdi: yeah yeah, they were public, right?

28:41Jacob: yeah, they were public. They still are public. They're still public. but what's what's left in the public company is just some cash

28:45Mehdi: Yeah, sorry. Well yeah, of course, yeah. Ooh.

28:49Jacob: and net operating losses. Yeah, yeah, yeah, yeah. it was a rough ride.

28:51Mehdi: Okay. Covid Covid Covid

28:54Mehdi: was not good for them.

28:55Jacob: Yeah.

28:57Jacob: yeah, it was good. It was good until the lockup ended, I think, is what that looks like.

29:01Mehdi: Yeah.

29:03Jacob: but y you know, I think they they built something that was very much indicative of kind of the previous generation of tooling and but like not as good as Tableau. and you know, Tableau Tableau kind of getting acquired by Salesforce was I think the last big BI acquisition. Actually that's not true. was Looker was Looker just after that?

29:26Jacob: So like Looker

29:26Mehdi: Yeah.

29:26Jacob: and Tableau were the big winners. You know, both those companies got grabbed by larger companies

29:30Mehdi: Yeah.

29:30Jacob: who recognized the platform play. And

29:33Mehdi: Yeah, but there's still

29:34Mehdi: I mean Google is differently, but Tableau is still Tableau, right? So there is a bigger play. I think here

29:40Jacob: Yeah, yeah.

29:40Mehdi: the story might be a bit different. Like when I said to sell substitutionally all assets, it felt like you know, we're just gonna divide you know, what's on pieces and and put it somewhere. No? Okay.

29:53Jacob: I don't think that's true. I actually I think

29:55Jacob: that's a little bit too pessimistic. I think like really what this means is is because pro there's some this is some weird tax stuff, Medi, that's kind of arcane, which is basically

30:05Mehdi: Okay.

30:06Jacob: the company has some in in theory valuable net operating losses that another company can use, but they're not useful to progress software.

30:14Mehdi: Okay.

30:15Jacob: So because progress can't use them, they're not acquiring it. and so they just bought everything else.

30:21Jacob: I think it it's kind of just a so so it's a weird it's weird from that standpoint. I expect they're gonna keep running it. I don't think this is a you know, anything more than just like, okay, this is a stable software business that can be run by a private private company and is better suited

30:34Mehdi: Okay.

30:34Jacob: that way. so

30:37Mehdi: Okay.

30:37Jacob: I yeah, who knows who knows where it will go in the future? you know, I wouldn't be too worried. But yeah, it's interesting

30:43Mehdi: Yeah.

30:44Jacob: to see that, you know, ultimately something that was valued at what, 20x the

30:49Jacob: 20x the price it just got purchased

30:50Mehdi: Yeah.

30:51Jacob: for, you know.

30:53Mehdi: But yeah, it

30:54Mehdi: like they also like it's like DBT acquisition by Factor and we I think it was never disclosed, right? But DBT also had a a crazy valuation.

31:03Jacob: Yeah, yeah, DBT, I mean,

31:06Jacob: you also saw Dagster getting absorbed by Prefect a few days ago, you know.

31:08Mehdi: Yeah, indeed. Yeah, yeah. That was

31:12Mehdi: that was not on my bingo list, by the way. like in the first first

31:16Jacob: Me neither. Me neither, I would say.

31:17Mehdi: it was like okay, prefect can acquire Dexter. Like that's that that would be even the first question. all right.

31:26Jacob: Yep. Yeah.

31:29Mehdi: last business. Joe. Joe was is talking about

31:32Jacob: Yeah, I love this.

31:34Jacob: This is this is this is an older article, but I always love resurfacing

31:38Mehdi: yeah.

31:38Jacob: this because so this is where data engineering is heading in 2026, from Joe Reese. I always love his takes on this stuff. I think like he's very well connected into where where we're going. and you know I think this this is definitely informed.

32:01Jacob: How I'm thinking about what we need to build on the mother duck devrel side and how I'm thinking about a lot of this stuff. you know, obviously we are extremely AI pilled.

32:10Mehdi: Yeah.

32:11Jacob: and you know, it's I I I I look at the content I've produced over the last six months and the quality bar and the volume is like, man, like probably four or five times higher than what I had done the previous six

32:23Mehdi: Yeah. Yeah.

32:24Jacob: months. Like it's it is wild. It is wild how how how this is moving.

32:33Mehdi: Yeah, it's just talking about data modeling crisis in semantics as a second trend.

32:38Jacob: Yep, yep.

32:38Mehdi: Orchestration get consolidated or abandoned. Is there so at least that we have

32:41Jacob: Well he had one he's nailed one. he nailed yeah. that's

32:48Mehdi: proof with the acquisition from from Prefect of Daxters and it was talking he's talking about airflow?

32:57Mehdi: Dexter at twelve percent in small company versus okay. So is it is actually breaks into enterprise orchestration as a category get absorbed into platform. It's funny because it talks

33:10Jacob: I

33:10Mehdi: about those two, but not in the right scenario.

33:13Jacob: Yeah, that's pretty funny. Yeah, yeah, yeah.

33:15Mehdi: Yeah, the Lake House versus warehouse war hands in the draw. What do you think about this one?

33:22Jacob: I think it's interesting. like, you know, in the old days, like when I built my first data warehouse, we built it on SQL Server. On physical hardware, we ran into data center. And we had this need for physical hardware for performance reasons. and we were limited, like when you needed more space, it was non-trivial to get more space to it. Yeah, we had a

33:48Jacob: like a big sand backing it so we could add more space to it. It wasn't like impossible. But like it was painful. And I think like, you know, lake houses kind of came out of that world of like, well, you have a lot of data, a data warehouse needs all of this disk space, when you use it, you know, kind of as a database. So what does it look like in the future? You know, I think a lot the the trade off on a lot of that was latency and like how long it take you to warm the cache and things like that. Right.

34:18Jacob: You know, without giving too much away, if I look at like a mother duck's architecture, it's not that different from a from a lake house and it feels like a warehouse. And so I think like, you know, we ul ultimately I think they're they're they're converging, but ul but the trade-off I think will ultimately be in cases where you can keep data cold, a lake house makes sense. There's a certain set of data that you need to keep hot.

34:45Mehdi: Yeah.

34:46Jacob: and hot hot might mean on

34:48Jacob: An SSD, hot might mean in RAM, you know, it might mean something else. There's a bunch of different ways to think about what that cat how that cache is maintained. But like, you know, what we're seeing is for example, Mother Duck works as an awesome, or even Duck TV by itself works as an awesome cache on top of Parquet. Okay, well, without a tech like that, you know, your lake house is cold and slow. And now if you can get better, you know, we'll we'll see how it goes. yeah.

35:13Mehdi: But but also like

35:15Mehdi: i I think just in general both category has been getting features from the others. So Lake Ho like yeah. Because

35:21Jacob: Yeah, that's right. They they're stealing the best cow yeah, exactly.

35:23Mehdi: Data Lake initially was DataLake versus warehouse in Lake House because they have table format and now they have a easy transaction like a data warehouse. And

35:31Jacob: Yeah, yeah. Yep.

35:33Mehdi: then you know, all ML stuff and training were happening in Data Lake because of the scaling needed.

35:43Mehdi: And now this is built in in a lot of like warehouse. you can train a model

35:47Jacob: Mm-hmm. Mm-hmm.

35:48Mehdi: directly with just SQL.

35:51Mehdi: So yeah, there is tons of features that here and there have been inspired, I think at the end of the day. And so Databricks has been building their warehouse feature, right? Because then they're like, okay, how do you access the lake? you put the clients, you put Spark, but that's pretty heavy. warehouse is easy, it's just SQL. But now

36:12Jacob: Yeah. Yeah.

36:13Mehdi: Databricks has put their warehouse with SQL. So yeah, it's just

36:18Jacob: Yeah, that's right. That's right.

36:20Mehdi: I I don't think it's a draw, it's mostly like a merge of features and it started to be really hard to say you know what's good and bad but what's what is funny is that this this war is purely marketing like if you look at most of the mature data company and tech company they all had both

36:44Mehdi: They all had a data leak and every house. So it is really funny that, you know, online people speak about I mean online people I would say Snoffbreak and Databricks speak about lake house

36:57Jacob: Sure. Yeah.

36:58Mehdi: and and and warehouse versus in reality

37:03Mehdi: No, it's a bit different, right? Because of economics, because of what, you know, how Databricks is moving and so on. So I do see people my emerging, you know, migrating everything on Databricks or Snowflake, more on the first one. But before, like, yeah, all the tech, big tech company, I at least I know in Berlin, they have all had both. And every platform

37:23Jacob: Yeah, yeah.

37:24Mehdi: I worked had both because there was two different separate use cases. So so yeah.

37:30Jacob: Mm-hmm. All right, number five. This is a fun one. Leadership becomes the bottleneck everyone talks about. this is really, really interesting. I think I think what I'm observing is that teams that are moving really quickly are empowered by leadership to kind of move fast and break things to borrow from something that's now 15 years old.

37:59Mehdi: Yeah.

38:00Jacob: on the analytics side. And I think that teams are that are not empowered to do that are moving really slow. And that manifests, I think, is people complaining about leadership. But you know, I I I think that we are in the kind of Wild West era of things. And as such, leadership is very uneven in terms of what they're thinking about how to apply these new technologies to their business. I don't know. I don't know what your what your take is here, Medi, but that's kind of my charitable take.

38:25Mehdi: Yeah.

38:26Mehdi: No, I f I I think I think you're right. The other side of

38:32Mehdi: This is that with AI coming in, I mean the lack of leadership direction has always been there because it's not easy to make money from a data team. That's just the reality. Like to justify

38:43Jacob: That's right, that's right. Yeah, yeah, yeah, yeah, yeah.

38:45Mehdi: the cost of a data infrastructure in a data team. And

38:48Jacob: That's right.

38:49Mehdi: now you have AI on top of that, which kind of like change everything. Like now your business people, you know, can solve their own question, but they're they don't have the, you know

39:02Mehdi: governance and they're not defining revenues as you do in your column. I was, you know, I had I did a talk at this big retailer Decathlon which is one of the biggest

39:14Jacob: Mm mm.

39:15Mehdi: sports retailer in Europe and and yeah they were mentioning they have team outside that do stuff and they have like contracts on data set they maintain.

39:28Mehdi: But it's I said like

39:28Jacob: Mm-hmm.

39:29Mehdi: there's still team doing Wild Wests, and it's always like the discussion after where, Okay, you compute that number, but we serve that number.

39:39Mehdi: You know, why did you change that? Right? Why wh wh which business rule did you add? That makes more sense to use that. So and the problem with that is that before it was happening, like kind of like Shido IT on the team side doing their business stuff, but now it's accelerating because of AI. So and and that that data engineer I kind of like

40:03Mehdi: don't have power to do anything there, right? It needs to be kind of a top-down decision to say, yeah, we shouldn't trust those numbers. And people should be empowered to think together. But yeah, it's it's a really hard problem to solve in my opinion.

40:19Jacob: Yeah.

40:20Jacob: I think one thing that I saw, I saw really so Ryan Dolly is writing a book about AI enabled BI. And

40:28Mehdi: Mm-hmm.

40:29Jacob: one of the things he's doing is writing a glossary of like defining terms. And one of the things he said is like he's looking for a good term on like metrics, like how to define the word metric. And what he's what his insight was was like, a metric is actually a social social contract. How it manifests to data engineers is like, to us, it's just like, this is just like a piece of code.

40:47Mehdi: Yeah.

40:47Jacob: But like

40:48Jacob: it's actually a social contract. It's an agreement between leaders on what something means. And I think that's really interesting insight and kind of goes into all of this, which is like a lot of we're we're trying to it's sometimes it's a square peg and a round hole where it's like we're trying to force a very specific programmatic definition on something that is squishy. And how do we how do we actually model that? Anyways, it's

41:07Mehdi: Yeah.

41:08Jacob: very interesting.

41:10Mehdi: Yeah, was cool blog. Nice to bring that up. I think it's nice to bring old blog to see the situation right now. We are you know mid twenty twenty-six. That was a fun one. All right, next I have tweet which is w interesting. Do you see do you saw that one from TL T Radore? Okay,

41:28Jacob: have no, this is new. This is a new one.

41:30Mehdi: so I'm pretty sure his record is our fault. So people were completing a lied about how to see the

41:40Mehdi: the how much like code base has been bycoded and you know his quality it's just by looking at his records you're not talking machine if you don't have at least 20 copy of this function in your repo and so it's a

41:54Jacob: Mm-hmm. Mm-hmm.

41:55Mehdi: really stupid function actually and so why is codex so upset with his record and so TL draw which is kind of infinite it's an infinite canvas if you don't know

42:08Mehdi: they wrote a blog on saying I think it's our fault. I'm sorry for that. they're open source and they show

42:13Jacob: Ha ha.

42:14Mehdi: they show basically in their code base, you know, in the past where where it's that and why they put that. you know, it's not sure it's about, you know, the because of them, but it's still interesting that you know, apparently it could it's there is high chance and it makes you think also.

42:37Mehdi: about how you know models are being trained on, you know, sorry for the word, shitty codes, or you know, not really good. I'm not saying TLDRA has shitty code, but like this function obviously is not, or it's maybe something really specific that doesn't need to be a standard, obviously.

42:56Jacob: Mm-hmm, mm-hmm.

42:57Mehdi: and so interesting to see that like whatever you're doing today, you know, you and me, we could potentially

43:03Jacob: Yeah.

43:04Mehdi: influence.

43:06Mehdi: Thousands of other projects and I haven't thought about that. So

43:08Jacob: that's so funny. I hope so. That's great.

43:12Mehdi: so yeah, I've I'm feeling just a bit more scared now when I'm pushing card to GitHub, like you know, to see like okay, you know, you were talking about side project

43:20Jacob: that's so funny.

43:22Mehdi: and people installing your library and having dependency on that. That's like one thing,

43:26Jacob: Yeah. Mm-hmm. Mm-hmm.

43:27Mehdi: but just getting, you know, influence in new code base because of like, you know.

43:35Mehdi: wrong stuff you've been coding wrong or I don't know or wrong architecture. that's

43:39Jacob: Mm-hmm. Mm-hmm.

43:40Mehdi: interesting. Another take related was that it's gonna be really hard. People have this hypothesis that it's gonna be really hard for new framework to po pop out. Like you know, everybody is on React and X.js for the web, I mean mostly, right? And before we had a new JavaScript framework every three months, right?

44:05Mehdi: And I think now it's changing. We maybe have, but nobody's taking care attention because all the models know really well React, right? and

44:14Jacob: Yep, that's right.

44:15Mehdi: so how do you do innovation on the web where you have such a a weight of existing code and inspiration, right?

44:26Jacob: Yep.

44:27Mehdi: so yeah, to d have you thought about this angle too?

44:31Jacob: I definitely have. I mean, I think one of the big challenges I ran into with like Python is and data frame libraries is it wants to write panda syntax everywhere. Right. So if you try to use polars, it will just like it'll use like I think like

44:49Jacob: Polars is like group I with an underscore, and Pandas is like group I with no underscore. And it will just use

44:54Mehdi: Yeah. Yeah, yeah.

44:55Jacob: or something like that. You know, so there's a bunch of examples like that where they're like totally valid design choices, but like it it confuses them. and I think that's something to consider, right? Like incumbent languages have a huge advantage from being in the training data set. SQL is actually extremely

45:13Mehdi: Yeah. Yeah,

45:14Mehdi: yeah.

45:14Jacob: overrepresented probably in the training set because there's so many different databases and they all use this common language that does have different syntax, you know, different

45:22Mehdi: Yeah.

45:22Jacob: functions. Or I guess the core syntax is the same, but the functions are different. And I think like this is this actually makes it really powerful and and means that LMs are insanely good at writing it, which is a great tailwind for us, but is also like if you're trying to build something better than what already exists, you also have to build the training, the training piece of it. And that's hard.

45:44Jacob: I think like how do you get how do you get all of your documentation into the data as quickly as possible? Right. Into the training data, I mean. definitely gonna be hard to do that. So we'll see where that goes.

45:55Mehdi: Alright, let's wrap up. Did you add another I think we cover all your links? Yes. All right.

46:01Jacob: yeah. Yep. Now here we go. Let's do it.

46:07Mehdi: Okay, I have I I have a one to close. I think I mean it's just it's just blog from Claude's from Anthropic on loop engineering. and I think it just got me thinking that, you know, today we we get some tasks on people. I think we we actually already talked about that in the previous part around how you know you use goal and so on. But I think it's

46:37Mehdi: I say that, you know, how clothes package it, cut clothes with the goal thing. I think there is interesting point where, you know, it's say you can have, for example, a stop criteria when you're launching a loop based on specific criteria that's met, or based on

46:57Jacob: Hm.

46:57Mehdi: the time. Like I haven't thought about this. Like look during five minutes, check my PR.

47:04Mehdi: to to see, you know, what's coming in and so and review it. So I think like there is different kind of way to loop things that I haven't think about it, which I found found interesting. Have you started creating loops for your agent workflow?

47:24Jacob: I've done a little bit. not a ton. I've just found that having you know, sitting at my computer and doing the work has

47:36Jacob: is helpful for me developing some sort of intuition for how my thing works, whatever I'm building. I do

47:40Mehdi: Yeah. Yeah.

47:42Jacob: find I I I did do some experimentation on my vacation where I had Chat GPT running on my my personal laptop and I was using the remote from my phone to

47:53Mehdi: Mm-hmm.

47:53Jacob: have it build stuff. And while that worked really good, and I got some good results. I also like I'm so far away from the process at that point that I don't have a good feel for like what trade-offs were made.

48:05Jacob: Or anything like that. So it makes me a

48:07Mehdi: Yeah.

48:07Jacob: little nervous. it also reminds me, it reminds me of like stored procedures in SQL, which are full of loops. you know, while

48:15Mehdi: Yeah.

48:17Jacob: this is true, do this until it's no longer true, right? so it's very it's very funny to see kind of that emerging. But you know, I've definitely built a few of them.

48:29Jacob: You know, are they critical? It's unclear to me. Obviously, like if you're building like fully agentic stuff where there's no human in the loop, I get why they would be. My my kind of perspective at the moment is that like I would like to have some sort of intuition about what the thing is that I'm building and know how it works. And I don't know how you do that with a fully autonomous agentic system.

48:48Mehdi: Yeah, yeah, yeah.

48:49Mehdi: I think it's depending on the task you're doing, but one I can give you recently I've been doing it is video editing. I started to do

48:52Jacob: Mm-hmm. Mm-hmm. Yeah.

48:55Mehdi: more editing with AI, so removing all the blanks and so on. So I build skills for that, do the

49:00Jacob: Yeah. Yeah.

49:02Mehdi: transcribe and do the editing. but then I realized sometimes it still missed something.

49:08Mehdi: So I ask it to do a loop to say, okay, do a version and then re-loop on that version, redo a transcribe and re-identify

49:17Jacob: Okay.

49:18Mehdi: the gaps. And I feel that that kind of things take much more time. and I see like there is like so it's it's it's more like creating loops for him to iterate over it, right? Rather,

49:30Jacob: Yeah, yeah.

49:31Mehdi: but it's the same goal, right? It's just redoing the task or slightly different that work well.

49:38Mehdi: and so so that that kind of things because there I you know I'm I'm not I don't need to to be babysitting I I I I know what I want in terms of results. so so yeah I do I do I do think this is kind of like the future on how we should work as we get better on what we want to do.

50:03Mehdi: so yeah, just interesting thoughts for food for thoughts. Even if you're not using cloud code, I wouldn't recommend a read. All right, that's it for today. Jacob, thank you very much. It was really fun.

50:17Jacob: Yes, of course.

50:18Mehdi: And all the show notes again are in mother doc.com slash podcast, and we'll see you in the next one.

50:27Jacob: All right. Thanks, buddy. Ciao later.