161k GitHub stars in 4 months with 65 lines of Markdown

0:00Mehdi: Hello everybody, welcome to another episode of the Explain Analyze podcast. I'm Midi Devrell at Mother Duck and join again with another teammate of mine, Dumkey, which is developer experience at Mother Duck. Hey Dumkey how are you doing?

0:18Dumky: Hey Medi, yeah, super excited to be here.

0:20Mehdi: So as usual, we have a couple of links around 10th that we cover around data and AI news. And yeah, we'll bring the topics and share our experience. We are gonna start with a fun stuff. I know this guy he has one point three million subscribers on YouTube. they are ruined. Why did you bring that up, Dumkey?

0:43Dumky: Yeah, so this is very interesting. I actually I thought of this again because I saw this article that Spotify is now allowing basically AI slop music on their on their platform. Some artists were actually kind of fed up. And I I thought of it because what's interesting here is so this guy, if you haven't if you haven't seen him, like you really need to check it out because he mixes different

0:57Mehdi: Mm.

1:12Dumky: styles and different artists together. And at first you're like, this is just this is just AI, right? I tell the AI to take the Reddo chili peppers and let them sing other lyrics. And then he goes into his process of actually making the music. And you see that he is a full-fledged producer. He plays a lot of this stuff himself. He creates his own models for every artist to mimic their voice and to Do the lyrics and the intonation themselves. And so for me, yeah, it was just a reminder of how something can sound like, this is AI, AI can do this. And actually when you go and and peel back the curtain, right? It's there's so much more work to it. And it was a reminder of how a lot of these models these days work as well. Yeah, let's go.

1:59Mehdi: Okay, let's listen just a few clips. Yeah, it it's like it's pretty accurate. I've seen a lot of mix which are a bit more fun and I think so I have a lot to t to to do and I think it's a really good reminder that it's actually hard to draw the line on what is AI slop like and what's the effort of the human or the musician. So you mentioned that he's a he's a full artist and I think it did back then it was it

2:04Dumky: Mm-hmm.

2:29Mehdi: He did the golden music from the Korean you know movie show. and he did it if Queen ward Gordon. And he got a lot of backlash to say, hey, that's like AI screening, you know, basically getting backlash from Queens fan, I guess. But then he did another video, I'm not sure I see it here. maybe he took it out or it was not in a short where he show actually

2:31Dumky: Mm, yeah, yeah, yeah.

2:57Mehdi: the process that he did. So you sh we see that he's singing actually the stuff, the the lyrics and there is the model that, you know, match the the singer's voice. And then there is other stuff that he do around it. so it is interesting because well it it it is fun and it's entertaining first, but I think you know, from a music industry perspective

3:02Dumky: Yep. Yep.

3:22Mehdi: And what Spotify is gonna do is I think it's gonna be hard really to draw the line. And I expect that over the years there's gonna be less and less, you know, human work. Like you could say, okay, I just typed the text and then now I have, you know, and I have a song reference. I don't need to song sing it anymore. And then you're gonna have to question copyright and what's the you know, the work from humans involved.

3:47Dumky: Yeah, yeah, no, exactly. And and I actually think that a lot of this is still evolves around creativity, right? So it's using these AI tools or enhancing your skills or your possibilities with AI tools is just another tool in your toolbox, right? You still need to be creative and come up with like funny weird stuff to make this make this interesting. The problem, however, that I am worried about a lot and you see this in in

4:04Mehdi: Mm-hmm.

4:16Dumky: programming or or on GitHub as well is that the AI slop is just overwhelming the rest of it. So you need some form of curation or you need some filter mechanism to be able to filter out what's actually good and interesting for you to look at, or you'll be just overwhelmed with the the slop and you basically the the good stuff will drown in in everything that's out there.

4:43Mehdi: Yeah. Another example is Pass Windu. I don't know if you saw that one. so this is another channel. is he's smaller, it's I'm a Jedi. and as you see he has a guitar, he has the the helmets of the what's it called again? Jesus, I have I'm not a that much a Star Wars fans, but I I do have watched them.

4:50Dumky: No, I d I don't think I Mm. Yeah. Yeah.

5:11Mehdi: And anyway he has like a full fledged video clip where film music from if one of the Jedi was you know a bass player, but it's a full music that he produced and he does like guitar solo in it, and you can see him sometimes, like quite often. You see he's of he's here. So everything else is fake. And yeah. And it and he's over there. So and it and

5:21Dumky: Yeah. nice. Like he is in there as well, yeah. Ha ha ha.

5:37Mehdi: And so the the s like this created like a lot of creativity and it's still like as you said, like a musician. So yeah, it is interesting. I think it's also I think from an industry perspective, it is gonna be challenging for a musician and authorities to, you know, make draw that line on how much rights you should add and where is your creativity, right? But I think it's

6:00Dumky: Yeah, yeah, but but we've had that challenge before, right? Like yeah.

6:03Mehdi: Yeah, but it it's it's much more challenging today, but it's reassuring that no matter what's being you know automated and created, people find ways to be, you know, even more creative. So let's see how crazy the next part is gonna be. Okay, next link is from mine. so this is fun. So it's cloud it it's Car Party Skills, which is basically just a cloud MD.

6:28Dumky: Mm-hmm.

6:30Mehdi: it's sixty-five line of codes and as you can see there is basically you know not that much, just think before coding, simplicity first, and he said that he had you know dramastic better improvement with that with goal driven execution. I know some of the like clothes code included the you know the the goal slash now to have your your specific task specific goal to to test. But what's crazy here, I don't want to talk actually about the skill, is that it has 161,000 stars.

7:10Dumky: Yeah, I mean I mean he's a he's a rock star, right? He's a rock star of of AI, so

7:13Mehdi: Alright, alright, alright. Here, here is the trick. This is not the repository of Carpathi.

7:20Dumky: interesting. now I see. I I didn't check the URL. Wow. This is this is good. Yep. Yeah. Clever.

7:22Mehdi: Yeah. So this is just one guy, you know that he do has quite some following on Twitter. But the point is it took this from inspiration from a tweet. So he he stating that on the design. So this is the tweet of Karpati.

7:45Dumky: Yep. Yep.

7:48Mehdi: So that that's that's where

7:49Dumky: I'm I'm seeing the playbook here, right? So there was if if you were like doing anything web related in in let's say around the the twenty tens, right? I think every time a new channel w went viral, it was a new like there was a new search bar, there was a new list basically where you could get on top of things and you could game or or play with the results and how you manipulate those results and this happened on twitter this happened on on spotify even like there's there's a lot of people trying to get into like the the top lists and and artificially inflate their their play numbers, those kind of things. if you've ever seen like a phone, one of these mobile farms basically where you know they have like hundreds and hundreds of phones to manipulate all this stuff. It reminds me a lot of of that, right? I wonder how many of those stars are from from bots here.

8:51Mehdi: Yeah, I remember. Yeah. But so it feels like GitHub is turning into more a social platform. Like there is virility in terms of repo. And it there was always somehow, you know, GitHub has like GitHub trends and so on, but the numbers were not like this. Like if you look at let's look at the Linux kernel on GitHub, which is basically powering, you know

9:01Dumky: Yeah.

9:21Mehdi: all the infrastructure in the world is two hundred and thirty five thousand and this has years of experience y you know of developer and computer and the other skills is four months old.

9:21Dumky: Yeah, yeah, yeah. The world. Yeah. Yeah. Yeah, I wonder actually, do you have any numbers on I I know so I made this blog post on the number of repositories that's new repositories on on GitHub, but I I couldn't really find anything on the number of users on GitHub. So I wonder if there's like I would imagine there's a big influx of new users that just flock to this type of like here's a skill. It's

9:59Mehdi: Yeah. Yeah, but I I that does yeah, I think like you could have farming stuff where you have and we had that there is like site that buy stars, right? But here I'm not sure about it like that mostly so your hypothesis that there might be a lot of newsers that are just bot and open closed stuff and that star this. But I think like why would you start like usually you start something because you believe in the project, like kind of a bookmark?

10:09Dumky: Mm-hmm.

10:27Mehdi: But does OpenClaw kind of like need to start this? Like you can remember a neural, it's not as stupid than a human. I mean, hopefully. But yeah, it's a good question. I wonder. But I think the conclusion here is that we always said that GitUp stars are, you know, vanity magics and they should not be taken too seriously. And I think with AI, it's not just the GitHub star, but it's a lot of stuff and things get inflated.

10:32Dumky: No, that's fair, yeah.

10:55Mehdi: And GitHub is kind of like turning into a social platform. So you could t exactly as you said, you could take advantage of that somehow. but yeah, that for me that's pretty crazy that a skill file not created by someone which is not the author at the end, get like, you know, two two thirds of the number of stars than the Linux kernel. yeah.

10:58Dumky: Yeah. Yeah. Yeah, yeah. That's that's super interesting. I do wonder what the like w what does he get out of this other than maybe some, you know, exposure or something like that? It's

11:31Mehdi: I think he's it you know it's like under like a startup or something. if I click so he has a project, you see multicast. So I guess that's like that's huge marketing. Like if you have hundreds sixty one thousand people.

11:39Dumky: Yeah. Yeah, yeah, it's just marketing. No, I I always go into like full paranoia mode and so you know you can you can see the the people who star this and and so I'm thinking like is he scraping their their details or anything? Like like you have this this these sort of honey pots on LinkedIn as well, right? Like like you have a beautiful woman who wants to connect with you on on LinkedIn and then actually what they want is your email address and and or your details. Yeah.

11:56Mehdi: yeah yeah. Yeah. Go to market engineering. Yeah, yeah. Yeah. I'm always disappointed. All right. next link. up and down the ladder of abstraction. This is no stuff.

12:20Dumky: Yeah, so this is actually this is really old. Yeah, this is 2011 and it's becoming more relevant again. So just for context, right? So this is an article called Up and Down the Ladder of Abs Abstraction. it shows it's actually very simple. Like it shows you how you can think of an algorithm and and go into like this idea of a car that makes a turn and how you could visualize that and how you could visualize like multiple cars making multiple turns and how would all of the cars you so there's there's l levels of abstraction that you can go through and ways to visualize them and kind of convey the ideas like how do you convey a complex message to a user that's reading your your web page. And the reason I brought this up again is that this was very interesting to me when it when it first came out. So Brad Victor, the the author of this, is a spent a long time at Apple as a UI designer. I think he then started his own firm. he has some other very interesting talks and and articles as well. one of them being blanking on a name, it's something don't don't draw that fish, I think it's called. basically saying like, hey, what we're doing on the internet these days is we we just tell people stuff and we we're not actually showing them anything. They don't feel what's going on. They don't feel what they can do. It's not interactive and they're not learning anything because you're just putting walls of text in front of them.

13:43Mehdi: Mm-hmm.

14:04Dumky: So why is this interesting right now? Well, if you think about how hard this was in 2011 to make an interactive visualization where you can use sliders and make simulations and do this kind of stuff, I realized that these days, you know, I s I in 2011 I I spent a lot of time learning D3.js just to make this kind of stuff because I thought it was so cool. And then

14:14Mehdi: Yeah, yeah, that is true. Yeah. Yeah.

14:30Dumky: All of these things will take me like half a year to to build. And these days I just point clot at it and I say, like I wanna have this kind of visualization or or a kind of interactive thing and yeah, yeah.

14:43Mehdi: Yeah. I'm just I'm just looking at the at the tooling fr and actually I never I never heard of like that Moo Tools.

14:55Dumky: Moodles. I don't remember what this was. yeah.

14:58Mehdi: release date wow yeah two thousand and six modes and it's stopped already ten years ago. So so the latest release was ten years ago. So yeah that's that is super interesting and to see s

15:08Dumky: Yeah, yeah, I know it's super early stuff. Yeah, and and so my conclusion is like the the principles behind this are still very relevant. Like if you look at some of his other stuff, I put a link to the ten bright irteas. Just look at the first one here, right? So you what's interesting is that you can you can look at that twenty percent thing that he has there.

15:27Mehdi: Yeah. Brighter idea. Yeah.

15:38Dumky: And actually that 20% is a slider. And so you can see the context change based on your assumptions. You can change your assumptions and your whole story kind of wraps around that. And so that was what I thought was really interesting at the time, but still very complex. But I feel that today it's easier for us to create

15:42Mehdi: yeah.

16:01Dumky: compelling stories that appeal to us as as humans, right? Like we're very visually oriented. And if we can see the change of our assumptions as our assumptions change, then that's super interesting and super, super great tool for learning as well. And so I feel like whether it's it's with with our dives or examples or any kind of interactive visualization that you do, you can learn from these principles and the people that have done this for years.

16:04Mehdi: Yes. Visually. Yeah, we are in track to it. Mm-hmm.

16:30Dumky: and now you can apply this yourself.

16:32Mehdi: Yeah, that's great. I I really like the idea. Like I was a bit skeptic to see something from October. I think there is always foundation that stays, right? But I was wondering how's that you know, relevant. But I think I wa I I got tricked about the content itself rather than how it's done, which is the topic of today. And yes, like so it's super interesting to see that the framework is already ten years old.

16:39Dumky: Yeah, yeah. Yeah. Yeah.

16:56Mehdi: And there is a lot of interesting thing. I'm curious how many traffic is still getting there.

17:02Dumky: I'd be interest yeah, that that'd be interesting to see. I I wonder if it's I've seen it pop up. I was listening to Lenny's podcast, actually. I think you had it on last time as well. And someone mentioned Brett Victor. I was like, Hey, there there's other people that still care about this guy.

17:04Mehdi: T Yeah. Yeah. Alright, so check it out. we put that's a good moment to remind that all the links that we discussed today are available at mother dog.com slash explain analyze and you can get an email on each new episode and gets all the notes including the links. what do we have next? Christoph Pleferi. if you know him He's been writing for quite some time. The data news stopped, then he did other content. He's still doing now a podcast actually on his YouTube channel and he found that now which is ID for data people. I don't know exactly what's the The exact tagline, so I don't want to you know mesh over there, but it's called the Analytics Agent for Context Engineering. so I invite you to watch that, but but but but he writes an interesting article about how he got attacked via GitHub PRs to GitHub action. And the long story short is that someone had multiple proposition in the PR to change

18:27Dumky: Mm.

18:37Mehdi: some comment in the GitHub action to run those things. this is actually redirecting to a website where you can see all the requests happening and then printing the environment secrets and yeah other things and the thing that it's scary is that as soon as you have basically a GitHub open source project and you do a build on pull request target event, then basically anyone can do that. So there is also another attack that's been similar a couple of weeks ago. and and we had to to patch I mean some things on our sides. so it is the the it is interesting how I feel like the GitHub action is kind of like the new email scam.

19:05Dumky: Yep. Yep. Yeah, yeah, I've seen this as well. Yeah.

19:29Mehdi: And that we need to be more r you know, mindful that multiple agents can find easily vulnerability into those GitHub action and GitHub repo settings. Because here basically yeah.

19:42Dumky: Yeah, exactly. And it's it's easy to spin up a a like a hobby project, right? any and so even for these types of projects where like people who are experienced with with building building software, you still need to know so much about all these details that you rather only spend time on once, because it's not not core to your product.

19:47Mehdi: Yes.

20:09Dumky: the the C I and C D flow. and there's yeah, there's a lot of people trying to get your tokens basically.

20:17Mehdi: Yeah. So so yeah, be safe out there with Kitab Action. That's the TLDR. We need to be more safer about what we allow in term the preview build.

20:23Dumky: Yeah, yeah. And and you should probably you should probably tell that to GitHub as well, right? Because they leaked their their internal repositories. I I don't know if you saw this, but they had a similar issue where all of their internal repositories were leaked by I think if I remember correctly, it was a little bit of social engineering going on there as well. Someone I I know I remember this was this was another MPN.

20:48Mehdi: Okay.

20:51Dumky: Thing because Visual Studio Code, if you're familiar with that, I I know you're not a VS Code user, but you was for years. Okay, okay. Okay, so you can install all these nice extensions, and one of these extensions, they're basically just npm packages. And if you hear the word npm these days, you like there's a constant flow of attacks going on. And so one of these

20:56Mehdi: Yeah. I was, I was for years. I was for yeah. Yeah, for years I was, yeah. Yeah. Yeah, yeah. Yeah, yeah, yeah.

21:19Dumky: relatively big packages was compromised for a little bit. The extension manager in VS Code does auto updating and so it updates to that and then it leaked all the repositories of an internal GitHub user basically. And so now these attackers have much more they're not in the dark anymore, right? Like they can see where the vulnerabilities are. They can scan the actual repositories for vulnerabilities, which is even harder.

21:27Mehdi: Mm. Yeah. Yeah. And I think that's a that's a good point to mention. I've seen that also someone doing a blog, I don't remember or video, that auto update is therefore even more dangerous because things get spread even more quickly, right? In terms of like vulnerability than it was before. And so if you don't pin some stuff, you get, you know, exposed like pretty quickly because even the

22:03Dumky: Yep. Yeah.

22:11Mehdi: The trusted source, like, you know, V S code here won't be able to prevent those thing to to spread. So yeah, pinning and stop type thing.

22:21Dumky: Yeah. And and in general, I don't know if you feel this way, Matty, but but I was I was going through this old stuff on from the internet, right? Eve I was going back even further, actually. I didn't put all of that in, but there's a yeah, I won't go into details, but it reminded me of the time where the internet was just a nice place for people to have conversations, and there was a lot of fun, interesting stuff, and you didn't really have to worry too much about what you put out there. And then obviously that changed a long time ago. But I feel that even these days it's becoming more and more hostile. Like it's people are out there automatically scanning your websites, your repositories, your open ports on your computer. they're social engineering you.

23:10Mehdi: Your agent list yeah. Yeah, yeah. Yeah, no, a a lot of agents are are scanning things and yeah, there is question on like, you know, what what data do you want to be private? Like you just need to be aware that whatever you put out there can be, you know, fully exposed. A bit like when you do a social post and you're, you know, a a nobody person, but something can get viral and that's not always a good thing. I've been victim on that. So like I write some stuff really quickly. Sometimes like w how when it's happened to be viral and then you know you wake up and you have hundreds of comments like, Why did I write this in five minutes? so yeah, just be mindful that you know it can be in whatever data you put out there and especially for your GitHub Action Preview but approved contributors to restrict the the preview builds.

23:50Dumky: Yeah.

24:05Mehdi: Next, do you want to go next on this one? What is this? It's it's all vintage okay, this one is more recent. But but it's it just look a bit vintage.

24:09Dumky: Yeah, yeah. So this is no no, this is this is more recent. So this is it is kind of related to to the Brett Victor stuff I talked about earlier, right? So Luke W, Luke Roblevsky is another designer, UX designer at Google. I d I don't think he still works for Google, but he used to work for Google a long time. He wrote the book on the canonical book on web forms like back in the days, like web forms were a big thing when you know you're adding data. You're basically you're communicating through a website with their back end system, right? So every time you want to order a new t shirt or you wanna do something on the internet, you put in your details or they need some data from you. It has to go through a form. And so He's always been looking at this from a UX perspective, like what can we do to make the the friction between humans and systems less less of a a thing. How can we make it easier for people to add in their their details? And so recently over the last year or so, he started looking at this with his UX view on AI chat as well. And so

25:29Mehdi: Mm-hmm.

25:30Dumky: The the pattern that he started observing is that we used to one, two years ago, we had like a big chat box, right? And that that was it. Like you put in a question, you get an answer, and then maybe like you go back and forth a little bit, or you copy and paste your sequel that you got from the from the AI chat basically. And now you're starting to see that this pattern is shifting. So

25:50Mehdi: Yeah yeah.

25:56Dumky: the the chat is becoming less prominent. Instead, you're starting to see these like artifact windows on the right side or the left side, or even in your browser, like you'll have a a small sidebar next to your browser so that you can interact with the actual web page and do stuff with that. And I think we're only at the beginning of this new way of thinking about interaction between us with our agent and the the thing that we are looking at, right? So we need to figure out a way to we need to figure out all these new patterns that we're used to. Like back in in the 2000 era, we weren't used to web forms at all. And then web forms started appearing and we realized that we need input validation before we actually sent the form and We started developing these patterns that we as humans, as a society, started learning and started anticipating as well. And now there's this whole new thing that can actually act as a web form, right? So you could say, like, why do I still need to put in the details of let's say a book that I've read or a movie that I like or something like that, when I can just talk about it with my agent and It can basically ask me all the questions that I need. I could do that in voice. I could do that in on my keyboard. But we need to start figuring out these new UX interaction patterns basically for for chat.

27:23Mehdi: Yeah, user experience. Yeah, yeah. Hey you were mentioning like VS code and like yeah, you're not using and I think I just had a reflection this morning that my IDE developers tool journey you know takes years and so I was on JetBrains for a while and most of my time years on VS Code. Even before JetBrains a bit of Atom and also just terminals. so before going to full fledged IDE. And then this past like seven months, I changed like three times. I I was on Cursor for like months or two, love it, and then switched to terminal back for cloud code and now just you know pure terminal with the setup and T Max.

28:02Dumky: Yeah, yeah, exactly.

28:14Mehdi: but there is other tools like, you know, conductor cursor have changed their UI significantly now, kind of to adopt multi agent. So it's funny to see that all the tools providing like even from expert user experience, they're still they're still testing. They have no idea. It's not that they have no idea, it's that things are changing rapidly and the needs in time of user experience change. And yeah.

28:20Dumky: Yep. Yep. Yeah, exactly. And and you see this if you think about what your agent is doing now that they weren't before, right? So there's this whole thinking process and tool calling process, which, you know, sometimes you want to go into. Yeah, but sometimes you don't. Like you don't want to have all of that in once and then you have different sessions, and maybe those sessions are actually on the same repository or they're on the same repository but in different work trees, or they're on different repositories, but they're

28:48Mehdi: Yeah. And review. Review. Yeah.

29:03Dumky: still related so they need to connect in some way and there's this whole new pattern that's appearing that we we didn't account for before and our tools it takes some time to catch up basically for the for the tools and new tools start to appearing. So I'm very curious to see like where will we end up? Will the input still be a a a checkbox? Will it be more voice stuff? Will it be like some kind of hierarchy or tree form?

29:30Mehdi: But that that depends on what you do, right? Like I specifically know like for video editing, you need a specific in motion design. For example, there is new tool where you prompt and you have a video generated, but then you need to annotate specifically. You need to say, I want to move that frame over there. it's far more easier to show the frame where it needs to be moved, right? Than trying to write the text around it.

29:40Dumky: Mm-hmm.

29:57Mehdi: So and we discussed that for slide generation too, right? and so I think there you're gonna need specific dedicated UI, to review and give feedback. And those are those are being built right now. But yeah, I think we're yeah.

30:00Dumky: Yep, yeah, exactly. Yeah, and that's interesting because it is actually becoming easier to build your own UI, right? Maybe you can start from something and adjust it to the way that you need. Maybe that becomes the pattern where, you know, someone provides the building blocks and it's up to you or your agent to like make that work.

30:17Mehdi: Yeah. I think we're gonna have fle more flexibility for user to tune things rather than just the layout. Like let's say you open a Adobe Premiere for video editing and you're gonna change the layout. Maybe you're gonna be able to change more in the futures. But I do think the UX is really a significant work and you need domain expert from video editing to say, yeah, what's the best way to being productive and review for, you know, a video edit? Like I need to add FX, I need to be able to color grade quickly. and so so I think that needs to be worked from expert rather than a random person doing, you know, some edits. and and once we get there, then yes, user is gonna be able to tune more things. But yeah, it's being built. we don't have anything for the moment still the the Wild West. All right. I have a new another one. Which is from Simon S Simon Spatty. if you listen to it to this, hello. He's also writing for us at Mododac, but it's on personal blog. Yeah, he's he's a big, big writer, he writes a lot. So he's talking about snow technology and better productivity to slowing down because I've seen a trends on a couple of device. I don't know which one he's mentioning. bhum.

31:23Dumky: Okay, tell me. Tell me. Mm-hmm. Yeah, yeah. He has some very interesting blog posts, definitely.

31:50Mehdi: For example, I mean Kindle for example has been there for a while, right? It's a reader tablet. it's not like a tablet iPad. Like people read more on like E ink tablet because it's more minimalistic. It's it's it's like starring.

31:54Dumky: Mm-hmm. Yep. It's funny to that you mentioned the Kindle actually and 'cause we're talking about like twenty twenty ten, twenty eleven stuff. I got my Kindle in in twenty eleven and last month I got a message that finally they're like sunsetting it, there's no more access for new books. 'Cause I was still using it. 'Cause it it just works like it's it's kinda like a book and and you don't need something super advanced.

32:22Mehdi: Yeah. Yeah. So he's talking about the typewriter also like this. He used I think he has one of one of those here. was a Kickstarter where basically you can just write and you have a screen. now there is another product where you just have your phone as a screen and you can use any keyboard.

32:39Dumky: Interesting. Okay, that's that's pretty

32:51Mehdi: multiple options today for writing. one that I've seen also is the Feedbit Air. Did you you see that? It's a new product from Google. Yeah, from Google. Yeah, yeah. But so it's just there is no screen. so let me actually

32:59Dumky: The Fitbit error. No, I I do actually have a Fitbit watch, but what's the what's the Fitbit error? So is it is it like that, what's it called? The whoop, which is just just a bracelet basically?

33:16Mehdi: Yes, it is. So yeah, basically you have you have just no screen, it's just a sensor, and so Google launched that. So again, we remove some screens and some distraction and I just got the terminal, so it's a small

33:18Dumky: Interesting. Yep. Ooh, I d I I wanna I wanna go back on the on the Fitbit era a a little bit. Like I wonder how you think of this. So I actually got a Fitbit watch so I wouldn't look at my phone anymore for the time. and so I think it's interesting that people then also like go back to these types of like bracelets. and I think that just points to like how everyone has their own individual like workflow or or flow for these kind of things, right? But I do like the idea, and I guess that's that's what you're pointing at, right? To have a single task device, like something that solves one or two problems instead of like a big phone that tries to solve all of your problems.

34:18Mehdi: I think it's really about distraction here. So that's the the world trends that Simon is mentioning with slow tech. So it's really slow tech. It's just we have so much burden. We are being solicited by AI, chat, and so on. And I'm sorry, but like even if AI know my agenda and I can ask a question, I would rather look at this. Look at this.

34:42Dumky: Yeah, it's it's beautiful. No, I've I've been trying to get this terminal thing for I've been thinking about getting it for a while, but I wanna have a a purpose for it.

34:45Mehdi: So that's that's my agenda for people listening, I'm sure yeah. Yeah, so this is called terminal. I guess it's pronounced like this. It's just a E Ink tablet, think it of Kindle, and it yeah.

35:02Dumky: Yeah, and it's it's beautiful, right? So it it it's very simplistic, but it has some colors as well, if you want the colour version, I think.

35:08Mehdi: Yeah. Yes. And but the good thing is that you can you have a lot of plugging. They are really into open source. They have ev even a kit where you can do it yourself with your own component. Here I just buy their their hardware. It was really easy to just connect my calendar and then I have my calendar there. and the use for me is that I can show that to my wife and my kids specifically, which doesn't have a phone and I don't want to give him a phone, it's just seven.

35:34Dumky: Mm-hmm. Of course, yeah.

35:35Mehdi: So that it knows that if I'm in a meeting or not right now, I can put it there. And it's not like it's a e ink. And the beauty of this, I I just like really a big fan of how they design this, is that it lasts for weeks on Wi Fi and how it does is that it's you schedule the time that you fetch new information so that you preserve the battery. So you can say refresh every hour because for the agenda it's not moving, you know.

35:57Dumky: interesting, yeah. Yeah.

36:03Mehdi: that much. It might be, but once an hour is it's okay. So again, less waves also in your, you know, room because there is just when in time needed information instead of having a phone constantly pinging you and having so extending your attention. So yeah, big fan of those slow tech and I haven't got the feedback here, but yeah.

36:18Dumky: Yep. Yeah. And so just just to clarify, is it so this doesn't have a touch screen or anything either, right? So it's just it shows you one thing or

36:31Mehdi: So yeah, that was my that's my reflection. but I think their new bottle is I'm not sure the Terminal X. maybe I'm I'm saying something wrong there. but but yeah, so don't know the detail about specifically there. They d you can also jibreak actually your your Kindle to transform it into a terminal and being able to add their plugin. But the topic is that we are getting more distract

36:50Dumky: Yeah.

37:00Mehdi: you know, free tools and data are single purpose and I think that's nice. it's it's help us a bit focusing. All right. you want to talk about Vives to evals from more R

37:12Dumky: Right, yes. Vibes to Evals. So our our colleague Jacob has spent a lot of time on on evals. So what's what's an eval? it's just short for evaluation, which could mean anything obviously, but I think that the term these days is a little bit more tied to LLM evaluations, right? So how do you determine if a model is good? at a certain task. And so there's all these different types of benchmarks basically around programming questions, around let's say recognition, like image recognition, these kind of things. But the problem that we have to deal with most of the time is is text to sequel for one. but not just text to sequel, but actually I would say like text to insight. we wanna get A we want to write a question and then get an insight from our data warehouse that basically says, Hey, if this is your your question about your business, then I think this is this is the answer. And so we've been trying that quite a bit. Like we have our our MotherDoc MCP. obviously all the other big data warehouses have their own MCPs as well. and what you start to see is that there is Something very difficult about the specifics of your business. So the large language models are large and and generalized, right? And so there was this original trend. If you if you think of it from the perspective of what are the the different stages of going through that loop of getting writing text and and getting interesting data. back, getting insights back, right? At first it was just you were copy pasting SQL from an LLM into your your data warehouse or wherever you use it, and then you use that to get the data back. So you're basically the the loop. Now we're trying to do that with agents and and with Cloud or Codecs and MCP servers and tool calling. But you want to automate that as well. And you want to have the context of your your Data in your data warehouse as well. And so I think what what Jacob points out here, which is really interesting, is that as you start to try and improve this process, you need something to validate that you're going in the right direction. Right. And so he points out that actually what helps you as a data team, as an organization, is to just have a set of questions that help you evaluate if your models, if your agents, if your workflow is actually improving or not. And improving means that it could be either improving in terms of accuracy or in speed or in cost or in all three of these. so that's something that you have to decide for yourself. Like you could go throw a larger model at it and you'll be spending more money and maybe it improves, but not in a way that you think is is efficient for your organization. Right. So His recommendation basically a read the full thing, it's on our it's on our website. But his recommendation is like start with twenty to thirty questions that are specific to your organization and your data, and then just go from there and see if you can make improvements to column names, to comments, to adding in your your DBT manifest, all kinds of things that you can you can use to try and improve your your model basically.

40:52Mehdi: I think you talk also about how people build their model like as for SAS or whatsoever, but it's also just as a user individual, how do I evaluate if I need, you know, all the latest, greatest Opus four dot eight, we just would just get released, or another one cheaper. I heard for example that from people using Avely. Those kind of model that some of them the weight is changing, like it's not always constant, like in the back end from on tropic. and other smaller model are much more consistent. So depending on what you do, which kind of task you do, you might get much more accuracy and also just you know, much cheaper re results for for your task. So it's It's not just it feels like okay, evals is only reserved for people, you know, doing clump com building complex SaaS tools around the eye, but I think it's also just as an individual, how do I, you know, balance my task so that my manager is not knocking at my door because I spent three thousand you know, dollar just for a docs website. So yeah.

42:05Dumky: That's that's super interesting, actually. I hadn't thought of it that that way, but that makes that makes a lot of sense. So what I realize is that I've I've been adding in all these like skills and plugins and MCP servers, and then sometimes you're like, this eats up a lot of context, so I've got to remove some of them, and then a new model comes out, and then actually all of the time you're not sure like something is changing, and you can feel that the the quality of the responses you get is not as good, but there's no way to put a finger on it. And actually this is a really good I'm gonna try this out. I'm gonna gonna make my own questions basically for like different repositories that I work in or different data that I work with and see how adding in a skill or a different model like actually does the does the trick. I think that's a great idea.

42:32Mehdi: Yes. I think we've been blessed because we don't have that much push for the moment on our AI budget at Model Duck, but I think a lot of teams a lot of teams actually are in the same situation that that us for now, because they just want people to experiment and you know move forward. there is much more risk to not use AI you know than using AI and burn budget. but yeah. okay, so

43:01Dumky: Yeah, yeah, yeah, yeah. Ha ha. No, but but like I've I've seen other people who are in organizations where they're they're forced to use like an open source model from two years ago, and that's all that they can that they're allowed to use, right? Yeah.

43:31Mehdi: I feel bad for them. It's like using a caveman, you know. like you use a monkey and the others like use it actually at least. next is what is next? All right, let's talk about this one. So Lenny's podcast, AI paradox, more automation, more humans, more works. So then

43:39Dumky: Exactly. Yeah. I think I think in general, if like if people don't know about Lenny's podcast, this has been my discovery of like I think a year ago or something. This is such an incredible podcast around like deep conversations on on product and interacting with products, but also the whole AI cycle. Like he has had a lot of people from Anthropic on his podcast. and all kinds of other interesting people from from Notion, from all kinds of different places. And I learned so much from this just listening just listening to that.

44:30Mehdi: Yeah. Not sponsored, by the way, but it's a good podcast. So the the point here is that there is a couple of highlights that I like. SAS is not dead, is that's that's something they said that like

44:32Dumky: No, no, no, not at all. No, no. It's just good, but

44:55Mehdi: Agents are going to be you know consuming much more SaaS resource than a human did. and he also said that SaaS is gonna make will shift where every SaaS has been bringing AI features. They had been adding a chat bot on every SAS. But by the way, we didn't, we resisted so far to put them. And I think there is a big challenge with that when you had a chat bot in your AI software that now you're responsible for the tokens and for whatever user do with their prompts, like if they're not writing a good prompt or whatsoever. and so that would include basically SaaS margin if you say bring your own AI tool and you give more freedom to the user. You can use Codex, you can use and we have interface now for that, right? For NCP and so on. But I feel it's it's the change into product architecture. where you rely by default of you know people bringing their own AI versus having having a chat. And so I do I I wonder about the first point. first like what do you think about super agent inside the Slack instead of having everybody their own kind of like workflow. do you think it's gonna be a trend?

46:10Dumky: Yeah, it's it's a really good question. So I the way I think of this is that and you mentioned this before, right? If you're working with with some kind of video editing software, then the flow that you need for that video editing and the the type of AI that you need for that specific tool is gonna be different from just a generic model, right? So it makes sense to have something that's attuned to that. Now the The way that we interact with that can be very, very different. And I don't think it's like i my bring your own model versus using the the other person's or the company's model. There's a possibility where you you have an MCP or an API or a CLI that exposes an interface to your model, and that works in in certain use cases. And then there's other workflows where you want to be inside the tool and maybe you know have a sign into or or through open router or something use the model of your choice that you prefer but I think this is all part of discovering like what are our workflows and actually I think this is a very nice way to bring all of this together, right? So we're still discovering all these all these workflows.

47:28Mehdi: Yes, that's true.

47:29Dumky: And I think to your point about slow technology and and the terminal thing, it helps sometimes to have a dedicated window or app or workflow for one thing. So I find myself switching from cloud code to cloud desktop because mentally it's easier for me to do things in cloud desktop that are more like business related or related to our analytics and data warehouse than it is to do them in cloud code. And

47:40Mehdi: Mm-hmm.

47:57Dumky: That helps me visually be in a different mindset. And that's the the point that I think we're still very much on the lookout for. It's like there's no one big model that's gonna solve everything. At least that's that's my my feeling.

48:11Mehdi: Yeah. And what do you think about one super agent per company versus your own stuff? So we have the super deck, right? At MotherDAC. preset, you know, the dashboard company by Maxim Beauchavant. They have also an AI tool internally data engineered. They even made him write a blog. and so I do we have

48:18Dumky: Yeah, yeah, of course, of course.

48:36Mehdi: a lot of like specific individual flow and we don't share enough. And I think having one that we collaborate, we can see what others are prompting, you know, and and also just iterate together a bit like we do with scenario. We have scenario is a tool basically for dedicated for a design image generation together. So we have we generate our ducks over there and we can see you know which prompt

48:47Dumky: Mm-hmm. Yep. Yep.

49:04Mehdi: people have been using what we can fine-tune to not have the model drift and have weird ducks being generated. but it is a like kind of like a one single thing, right? It's not everybody has his own is on agent. So why do you think how this gonna go as a trend?

49:23Dumky: I I think of this as the way that you would think of of a person who works for a company but also has their own personal life that also has a professional life that goes beyond just the company. There's let let's say you draw the analogy of that super agent in your company is like a librarian, right? And you go to the desk of that librarian, you ask your question and they come back with with an answer. Now that's perfect for for your company, but what if I wanted to also remember all the the things that I did and the articles that I read and the conversations that I had with it before. And then some of those conversations might be more like around in the abstract what like the environment that we're operating in or you know it's not just I'm not gonna share all of that with the company agent, right?

50:16Mehdi: Yeah, but that that feels more like agent orchestration. So you could have, you know, like you get access to the GitHub org when you join a company, then you get access to their agent, right? Yes, exactly. Yeah, yeah, exactly. So that's why I'm saying it's it's different to say that today within companies, there is not that much companies that have specific dedicated agent that is being shared. I've seen like for coming back to data.

50:25Dumky: Yeah, yeah. Yeah, but I also still have my own GitHub account, right? Yeah. Yeah.

50:44Mehdi: a lot we had the customer sharing that they've been using heavily Slack as analytics request. So they ask this. Because they can the team can review together, say, no, no, that's that's not correct, right? versus today, for example at Motoduck, most people use our MCP to question data. And so if there is something wrong or miscalculated, even if we have great data a great data team with, you know,

50:53Dumky: Mm-hmm.

51:12Mehdi: semantic layer and all that, you're the only reviewer, right? Versus a common thing that, you know, has been has been shared. So I think that for me is basic today a bit a bit more like a human collaboration towards specific domains. But yeah, I agree. We will have always like you know personal personal agent for other things. But I do think, you know, company as a whole need to kind of like gather

51:16Dumky: Yep, yep.

51:41Mehdi: specific action that is being redundant. Yeah. I think we have gone to everything for today. I opened that. So yeah, what's great Dumkey to to have you again if you want all

51:42Dumky: Yeah, yeah, or per team or yeah, yeah. fully agree. Yep. Yeah. Yeah, super exciting. It's always nice to to go through all of this together and and get each other's perspectives.

52:02Mehdi: Yeah. If you want all the notes and the links that we've been discussing today, you go to mother dot com slash explain analyze. the body's available on Spotify, Apple Podcast. And if you like it, just you know, be a gentleman and thumbs up, a star, a review even more. That would be super appreciated. And I wish you a great weekend ahead. see you all.

52:27Dumky: All right. See you around.