Guess who’s back? Back again. Valentin’s back tell a friend. He was on Joinit Online before as a guest, and now he’s back for this episode on the mic again.

Zobraziť prepis
kubista
JoinIT.online. Takže čauky mňauky poslucháči, toto je JoinIT a máme tu dneska, dneska špeciálneho hosťa a budeme to hovoriť po anglicky, budeme riadni džentlmeni. A, a týmto vítam Valentin. Hello Valentin. Say hi to our listeners and to my right side, to my right side I have the OG join3r. Join3r, say hello to our listeners.
join3r
Yes, uh, hello, hello.
kubista
And again in the background-
join3r
It, it was a very nice intro.
kubista
Yes.
join3r
You should probably also, uh, uh, tell, tell us about yourself, what, what your nick is.
kubista
Uh, but, uh, we, we, we, we forgot about drunkez. drunkez, say hello.
join3r
Oh, of course, of course, drunkez, yeah. As always, recording, right? Always.
kubista
The man behind the scenes.
join3r
Yeah. Oh, yeah.
kubista
Pulling strings.
join3r
Uh, again, if someone is, uh, m- cannot hear, uh, drunkez, um, it's probably a problem with your premium subscription, subscription, uh, and you need to renew it, uh, so you get his, his, uh, audio back essentially. Hello, Valentin, how are you? Long time no see you. No, no hear.
Valentin
Fine, thank you. Uh, I'm good. Uh, yeah, we haven't talked for a long time, I guess, uh, one year more even. I'm not sure exactly, but yeah. But it's probably because, uh, yeah, um, it's either that we are too, too, let's say, uh, obsessed with the PCs and we don't have friends. Or-
join3r
Yes, yes
Valentin
... or, or, uh, we have too many kids. Or-
join3r
No, no
Valentin
... it could be also the case that we, uh, we don't have time, like enough time.
join3r
No, no, it's, it's, it's, it's the first one. It's the first one-
Valentin
The first one
join3r
... that we have no friends and we, we just, we just talk to the computers.
Valentin
Yes, we just stay all day inside and that's it. Yeah.
join3r
Yeah.
Valentin
Yeah. Um, but yeah, um, last time when we had this kind of, uh, setup, yes, we also had drunkez, but he just hates us and he doesn't want to join and talk with us. Yeah.
join3r
Uh, yeah. Un- un- unfortunately, he had some, uh, some family emergency, so our hearts go to him and, yeah.
Valentin
Okay, okay.
join3r
Otherwise, he would be here, so that's... That is the, that's the o- only pardon I can give him, right? If something, something bad happens. Otherwise, kid has birthday, I don't care. I don't care.
Valentin
You will-
join3r
You have to be here.
Valentin
Yes.
join3r
You are sick? No, no, no. You will be here.
Valentin
Yes, just call him from the toilet and it's okay.
join3r
Yes, yes, yes. Uh, no, the-- He always records from the toilet because that's where the-
Valentin
I see
join3r
... where the best audio is essentially.
Valentin
It's just the default. Yes, yes, yes. Yeah, this is exactly why I'm now located into the, um, how is it called? Like a closet room. Like a dressing room. So there's no window, nothing. But it's-
join3r
Mm-hmm
Valentin
... uh, but it's quiet. Yes, yes.
join3r
Yeah. And, and, and it's really the best, uh, place for an IT guy. No windows and-
Valentin
Yes, yes
join3r
... just one door
Valentin
No sun touches the skin. Yes.
Yes. Yes. Uh-
kubista
You get your vitamin D from the screen.
Valentin
Yes, yes, yes.
join3r
Yes. Yeah, yeah. And, uh, best perhaps would be if, uh, in the same room you would have also a toilet. So
Valentin
Maybe next year we can upgrade. Yes. Yes, yes.
join3r
Okay. So, uh, we have, we have a lot of news to go through. Uh, apparently you no longer are a proud house owner, but you moved out, right?
Valentin
Yeah. No, no, no. So I sold the, the one which we had, uh, which was a vacation home. And yeah, um, we still have the flat, but yeah, uh, we sold the vacation home, yes. Yes, it was a bit of, uh, let's say headache, because the state which it was in, it was a bit too, too crappy.
join3r
Mm-hmm.
Valentin
And yeah, we didn't have the time to travel all day long, uh-
join3r
Oh, okay. I see
Valentin
... between the, the, the homes, let's say. And yeah, it wasn't making sense from a financial point of view.
join3r
And how quick was the internet? Was this a-- that a deciding factor? For example-
kubista
Starlink
join3r
... internet was too slow at the house?
Valentin
Not at all.
join3r
No?
Valentin
Not, not at all, no. Because, because of course I can't live without internet, and of course the family is also too, too obsessed, uh, getting access to Netflix, YouTube, and all of this. Um-
join3r
Mm-hmm
Valentin
... uh, I did a setup, um, using an LTE antenna. So there was a dish which, um, was used for, uh, connecting a router to the, to the, to the LTE. And I was using a SIM card, and, uh, I was getting the maximum, uh, which was on the plan, like 100 down, 50 up. So the latency was of course not, let's say, optimal, but it was much more than what we needed, so it was totally fine.
join3r
Mm-hmm.
Valentin
I was able to work from there. We were able to use all of us, um, the internet without an issue.
join3r
Mm-hmm.
Valentin
It was working.
join3r
Okay.
Valentin
Even IPTV was w- was working. And the wireless itself, I, I had a, like a kit from, uh, Google, this Nest Wifi, and I think I had like-
join3r
Mm-hmm. Okay
Valentin
... five, five routers spread, uh, throughout the house.
join3r
Mm-hmm.
Valentin
And it was working pretty much perfectly.
join3r
So it was like a mesh setup?
Valentin
I- it was a mesh setup. It was the one from Google. Works okay.
join3r
Mm-hmm.
Valentin
Um, yeah, we were using it really, really nice. So we even had, uh, IP cameras and stuff like this. There was no kind of issue.
join3r
Mm-hmm. Okay.
Valentin
Yeah, so i- i-- so internet was, was really not, not a problem. Yeah.
join3r
Mm-hmm.
Valentin
The, let's say, main issues were that you have, of course, maintenance, which you have to do on the house and all of this kind of, uh, let's say, actions. And we didn't have the time, because traveling from Vienna to there, it was always taking, like one hour. And, um-
join3r
No road. Yeah
Valentin
... yeah, so back and forth, uh, it's, it's, yeah, it's too much, uh, too much effort or too much time spent on this and I was just grilling, I don't know, let's say once per month if this was the case and it wasn't making, let's say, sense. Yeah. Um, so the, the kids of course still liked it, yeah.
join3r
Mm-hmm.
Valentin
But, um, yeah, we were thinking that maybe we get something else in the future which is either a bit closer or a bit maybe in a better shape. Yeah.
join3r
Mm-hmm. And, and didn't you think about, uh, moving there?
Valentin
No, no, no. No, the, the plan was never to move there. It was just, um, to have it as a vacation home because y- once you work in Vienna, it's better to live in Vienna because, um, the traveling will kill you. All the traffic jams, the, the time wasted on the, on the car, um, it's not making sense for me. But yeah, there are people, like a lot of people ... I, I have just seen, like, a statistic yesterday, one million people are traveling, uh, to and from Vienna. So V- Vienna itself has two million people living inside, but one million is traveling all, all the time between the, the, their home and the city, yeah, so which I find crazy, yeah. Like half of the population-
join3r
I was one of them.
Valentin
Yes. Yes, sure. But, but you were of course, like in a different case because you travel also outside of the country, yeah. So it's a bit different. But you ... But one, like this one million, I believe they were, like, living in Austria but traveling to Vienna.
join3r
I see. Okay.
Valentin
So yeah, so, so it's a bit strange, but it's making sense, I guess, from, let's say, a capital point of view because I guess all of these, uh, big cities are having this kind of, um, this kind of use case, yeah.
join3r
Yeah.
kubista
I was just in Vienna this weekend, and I have to complain- ... that the metro- The metro, the metro in Vienna, it's like the, there is no automatic doors, which I don't understand because you have to manually the, pull the, the lever, which I mean for Vienna-
Valentin
I-
kubista
What the hell?
Valentin
No, you, you ... So you have the old ones and you have the new ones. The old ones where, where you have to pull the, the, the lever, uh, that's the, that's the really old ones, yes, which you have to open them like that. The other ones you usually just have to press, uh, like to press the button. They're ... Yeah. The red ones are the new ones and the, uh, gray ones are the old ones.
kubista
Oh, okay. But still-
Valentin
Yeah
kubista
... I was, I was, I was very disappointed.
Valentin
Yes, yes, yes. No, so I also ... There are, let's say, more, uh, minus points for the old ones because i- in the summer you don't have cooling , in the winter it's too much heat. So you will, uh, you will hate them. But I'm not sure why they don't replace them because, yeah, I, I, I'm sure it's expensive to buy new, new ones, but, um, there are ... It's a lot of people which, uh, which are using them, and they're always full and stuff like this, so they should be able to change them, yeah.
join3r
So one question. Which is better, um, according to you two guys, right? Uh, is it metro in Bratislava or metro in Vienna?
Valentin
I don't know.
kubista
Okay.
Valentin
I ... So I'm not sure if you have one in, in Bratislava.
join3r
No.
Valentin
Okay, no. Okay. Okay. But if you would have one, it would be probably newer, but for sure not better-
kubista
No
Valentin
... because-
kubista
It's-
Valentin
Yeah. Because the, the, the, the way you can move in Vienna with the subway and all of this, uh, connections, it's really good.
join3r
Yeah.
Valentin
And it's cheap. A- a- and it's quite cheap. It's like ... Well, now it's more expensive. They increased the price with something like 100 euros per year, but yeah, it's still cheap. Yeah.
kubista
If we would, if we were to build a metro, we would just build, uh, from Bratislava to Vienna.
Valentin
Probably, yes.
join3r
Well, the metro in Rožňava is quite good actually. Um, uh,
so-
Valentin
From the name I'm sure it doesn't even exist on the map, so.
join3r
No, no, it's, uh ... You don't know Rožňava. It's, like, it's like Rome essentially-
Valentin
Mm
join3r
... but, uh, better. All the roads lead to Rožňava. I'm sure you, you've heard that saying. Yeah.
Valentin
And-
kubista
Sure thing.
Valentin
Yes. Yeah. Like all the Slovaks are living in Kitze. Like, it's the same, uh, kind of case, right?
join3r
Um, I, I don't know about that. Probably, probably drunkez could, could talk about that.
Valentin
Yeah. Yeah.
join3r
Um, anyway, let's talk about computers. So, so you, you had a plan to, to buy a new MacBook, right? And so-
Valentin
I was looking into it, yeah. Yeah
join3r
... and we started talking about why you need so much essentially RAM, and, uh, if AI is any use case there. And instead of writing you the short answer, "Don't do it," I, I decided let's, let's do a one-hour podcast about it.
Valentin
Mm-hmm.
join3r
So basically you wanted to ... You were looking at M2 Max 46 gigabyte or was it 128 gigabytes of RAM? I'm not too sure.
Valentin
I was looking at M2 Max, I believe it was, yes, with 96.
join3r
96. Okay.
Valentin
Yes.
join3r
So, so I, I've fallen into the same trap. I have, uh, M4 Max with, uh, 64 gigabytes of RAM, and I was, I was really thinking about going 128 gigabytes of RAM, but I didn't go there because, uh, that would, that would mean that I can run 70, uh, billion large models, 70B, and that would be nice, but the, the performance would be under 10 tokens per second, and that's really slow, right? It's, uh, it's barely good for chatting essentially. And right now since, since, um, you have agentic AIs, right, and, uh, the, the way you use those models is quite different. You don't chat anymore. Or, or okay, y- you still chat. You, you, you can obviously chat with it, but a lot of the things go in the background, like reading files, right, and, uh, checking previous work, planning and stuff like that. And, um- Essentially, if you have anything below forty or thirty tokens per second, you are basically, yeah, you won't use it. Opus right now on Open Router is forty, uh, a-about forty tokens per second, I think. I was looking that up recently. And that's not a really quick model, right? But it's a, it's a, it's a model that, that you can reliably use. And the problem with self-hosted models is also twofold. So first one, you need performance for it, right? Which is something you not-- won't really get on a MacBook because it's, um, you, you... The, the inference is, is quite good, I would say if, especially if you use, use MLX and, uh, and stuff like that, the, the, the Apple, uh, stuff around it. But, um, processing of prompts, for example, you have input. You have an input file input, right? He has to ingest that before that, be-before he can enter, and that's like painfully slow. And the context sizes are also growing as more as, as we, as the change of AI, uh, as the usage of AI changes, we, we, we push more context into it, and that makes it... For example, on my M4 Max, essentially only some mix of expert, uh, local models are useful. GPT-o6-20B. Right now it's, uh, and, uh, right now it's, um, Gwen three point five and twenty-seven B. And even though I have more RAM, I could essentially run, um, larger model. I don't do it because of the speed. And, um, yeah. Then you have those, those really large models like Kimi, uh, or GLM, um, the, the, the really big ones, and yeah, you, you don't have enough RAM for that. You c- and you cannot get en-enough RAM for that in a, in a MacBook, right? So.
Valentin
I have sixteen gigabytes of RAM, so I don't have this problem. So, uh, I would like-
join3r
Yeah
Valentin
... to be in your place.
join3r
No, it's a, it's, it's, it's not really that much different place from someone who would have, uh, thirty-two gigabytes, for example, or forty-eight, or forty-eight.
Valentin
Mm-hmm.
join3r
I, I can, I can, I can put more context, right, uh, into the model, but I'm more limited by the, by the speed itself. And if you are looking about, yeah, I would like to play with it, right? I would like to train. I would like to do, um, anything, anything that's not, uh, not the, the base LLM, right? Uh, the... For example, if, if we are, if we are talking image generation, if we are talking local video generation, if we are talking o-out audio generation, right? And I'm talking music, hey, um, it's you, you essentially have a lot of stuff that doesn't work at all because it-
Valentin
Mm
join3r
... it requires Nvidia or is painfully slow.
Valentin
Okay. Yeah. So first of all, like the disclaimer is that I'm, I'm a noob, and I'm not, let's say, versed enough and experienced enough to really give out good, good opinions of or feedback. But I'm interested because of, uh, multiple reasons into this, uh, let's say discussion, uh, regarding the AI models and so on. So I already have an M2 Max, right? It's with sixteen four gigs. And yes, uh, it's, um, it's, yeah, uh, it's like pretty much like you said, right? You can load these models into it. The problem is one, the loading. Uh, when you-- let's use some LM Studio, like the classic way you put some model, you select the MLX, uh, variant, and yes, it will take some time, it will load, and it will stay in there. But yes, I also feel that you, you pretty much hit the nail on the head. Um, one, it's slow. Then second opt- the second thing is that nobody really uses it now in this way, like chatting with this, with the AI's and say, "Yeah, do this, do that," and so on. You-- From what I have seen, you give it pretty much this spec file, or you give it an empty file and say, "I want all of this stuff. Do it like this, do it like that," and so on, and communicate then with, with some MCP servers and-
join3r
Yeah
Valentin
... do what I need. And afterwards you will check it, you will, let's say, adapt it or go the next steps, and then you will end up with some result. But yes, when you need probably like a couple of agents or three or four agents working together, uh, yeah, local, I guess would be a problem for multiple points of view or multiple reasons. And this is why I actually have, now I have, so I activated the subscriptions for a couple of them so I can test them out to see how they would work. So I have Claude code, I have a ChatGPT, and I also have the Google one.
join3r
Mm-hmm.
Valentin
And I'm trying to get an idea which one would work, let's say, for my use case, and then would there be any benefit going one or another? Like, because for now, for example, the, the ChatGPT one which I have, it's the, I don't know, twenty euros per month one, the pro version I guess, or something. Then the Claude one is also the pro one. Al-
join3r
Mm-hmm
Valentin
... also twenty something per month.
join3r
Yes.
Valentin
And the Google one is also s- I know, I think it's a bit cheaper. It's a bit cheaper, like fifteen-
join3r
Yeah
Valentin
... or something like this, but it's pretty much the same, the same plan as they all are on, like on the same level. What I have discussed with some people from work, from other, uh, parts, they all like pretty much Claude. They like pretty much all of them Claude code. They say that that's the best, that's the most exact, the most, the faster one when it comes to getting to the target which they put to these agents because they say, "Okay, with ChatGPT, I have to go over the same stuff multiple times in order to get to the same result." And yeah, Google one seems like it's good for image or video generation, but not so good on the code side. But it's also like a, like a constant race, right? Like they always like someone, uh- Like we'll release like a new model, then the other ones are also doing like, uh, like, uh, this follow-up. Uh, so I guess now we're in the stage where it's like, uh, a race pretty much. They will all of them catch up at some point and unless someone really comes out with some crazy performance or crazy feature, you can't really say that, okay, that's the best or that's better than the other one. I guess it's just a time, time thing. Yeah, please.
kubista
Uh, so I have like two questions, but let's start with the hardware. But you told me you have M-M2 chip.
Valentin
M2 Max, yes, yes.
kubista
And you are, and you're looking forward to buy another M2 Mac?
Valentin
Uh, well, n-no. I, so I'm-- I have this M2 Max 16. So it's, it, it's a MacBook, uh, uh, Pro, um, uh, 16. It's with M2 Max 64 gigs. And I was looking into the newer stuff, uh, because-- And now we even have M5, right? So M5, I'm not sure-
kubista
That's a new, that's a new-
Valentin
Yeah. I'm-
kubista
The newest, yeah
Valentin
... yeah. So I don't need, let's say, a new one. I don't need more memory. I guess it's, it's just something messed up with my brain, and this is why I always have this need of looking to some newer laptop or whatever. I'm not sure why. I guess I'm just messed up. But, um, but to answer the question, uh, from joiner, uh, related to the amount of RAM, yeah, I also don't believe you need more than 16, to be honest. Because even if you say, "I want to run containers, I want to run what..." It's not a good idea from the beginning. So just get a tiny PC, whatever, put a Kubernetes cluster, um, on it, uh, put a Proxmox on it, uh, deploy, uh, uh, K3s, and you're much more safer and better than doing it locally. Of course, locally will also work because you have it there. It, it's okay, but yeah. But you-
kubista
But then because-
Valentin
Yeah, yeah, go
kubista
... yeah. Because I was, I was looking at the, the MacBook because if you go on the Apple website, you can compare the models.
Valentin
Yes.
kubista
And honestly, I have M1, and I'm looking at the M5.
Valentin
Yes.
kubista
And honestly, there is like the, the performance or the spec difference is not that big. I think it's like, uh, just 20 or 30% better and so I think if you have three models, as you said, you have GPT, Claude, and Gemini, Gemini.
Valentin
Yeah.
kubista
Uh, you can just buy MacBook-- Uh, you can buy the, what is the, the box? Mac-
Valentin
The Mac Mini
kubista
... Mini. Mac Mini-
Valentin
Yeah
kubista
... and then you can, you can maybe cancel two subscription and it will-- subsc-subscriptions, and it, it will, it will pay for itself if you want to play like with the agentic coding or...
Valentin
I'm not sure. I'm not sure. That-- So you, you-- Yeah, go.
join3r
Uh, you-- So the, you can, you can basically run those subscription on, on mobile, right?
Valentin
Mm-hmm.
join3r
Uh, so-
kubista
Probably, yeah
join3r
... that, that's not where the limitation is. But for example, if you wanna, wanna play with it, do some reinforcement learning or training yourself or creating LoRAs or whatever, right? Then you essentially need something, something stronger. And that's-
Valentin
Yeah
join3r
... where, where, where good, good hardware could come in, right? But that, to get good hardware is probably still an NVIDIA card and, uh, and not a, not a MacBook. You can do something, but I'm, I'm, I'm really feeling quite, um, limited. But by-
Valentin
Yeah
join3r
... what can I do and how stable it is at, at the end, right? Uh, I can, I-- For example, I found some workflow on how to, um, do some, uh... I mean, reinforcement learning and stuff like that and, and learning your own, um, LLMs, that's quite and good. But, uh, essentially, if you go outside of it and you tried some, uh, some image, uh, generation or video generation or audio generation, that's mostly where you f- where you fail. Um, and you also fail on some LLMs if, if, if you wanna be really specific about the way you train, you are also limited there, but, but not in a way that you cannot learn, right? You can still learn stuff around, uh, transformers, essentially.
Valentin
But this is a good point, what you mentioned now. So for example, I have a use case, uh, where I would like to train a model, yeah. I-- It's for, let's say, um, so I can go in some details that, uh, so to give, to make the long story short is that, for example, my parents are suffering from memory loss, yeah. So they are, uh... My, my father had a stroke, and my mother, uh, has some issues because she's, she has some, uh, degradation of her brain, and sh- they are suffering from memory loss, yeah. And, uh, I already, I'm thinking of starting a project which would, would essentially help people with memory issues-
join3r
Uh, okay
Valentin
... uh, because of medical, uh, reasons or actually like, um, genetic also, uh, reasons or they would like pretty much to improve, let's say, their memory capacity from a point of view of, uh, retention and building, uh, let's say a more, um, flexible memory, right?
join3r
Okay.
Valentin
And I would like to be able to train, let's say, a model which is pretty much just focused on this, yeah. So this is one of my goals. And I was looking, okay, what would this require, right? And this is why this NVIDIA route is probably the one which you need, right? Uh, in order to go and say, okay, we start training this model, and we come up with, with the stable one, and you put it in an app, and then you go probably the mobile route because if you go the PC route, yeah, not-- So people usually after a specific age are not so keen on using a PC. This is just my opinion. But the phone is usually more, more, more available or more spread Than a, than a PC, yeah. So yeah. So in exact this case, I'm, I'm interested to see what would this actually entail to, like what do- how much money do I actually need, let's say, to invest? Do I need to buy, uh, to build a PC or a server or whatever with, I don't know, five cards, four cards, one card? I'm not sure what this would actually be, let's say, needing.
join3r
Yeah. Depends on what you want, right? If y- if you wanna train model from the start, you are li- limited to, let's say, 250 M. That's a million, not billion, right? If you, if you wanna do really something. And at that scale, that's basically not good even for summarization. Uh, there, there are, there are better free models on Open Router, uh, that are essentially always free. And you could route through all them, and you would get a, get a something much better than, than you could essentially train from the start. And if you are talking about some, uh, some fine-tuning or, or reinforce- reinforcement learning, that's something where you can do something, right, with the, with the models that you have avail- available. But depends on what exact use case you are. Because I, I, I strongly suspect that anything that you could, you could do with fine-tuning or reinforcement learning, you would essentially get a better result if you would just use some generic model that's ChatGPT, that's, that's Claude, Kimmy, something like that, right?
Valentin
Yeah, but, but there's also the licensing stuff, right? Because, uh, so can you, let's say, use an open source model and then go next level and improve it, let's say, with specific, uh, information or specific habits or specific steps in order to help these people and be, let's say, just focus on the memory issues? If yes, can you use it, uh, in the target which we have? Because, because if, let's say, this app, you p- you build this, you build an app, uh, you make it free. I don't know, you make it free until some level, because after that one it would probably cost you too much money to run it. So you make, I don't know, some sort of plan. You say, "Okay, it's five euros per month or per year or whatever," and you get access to, I don't know, unlimited time usage or whatever, right? Because I believe that from the beginning, if you start and say, "Okay, I want to make it free. It's for people. I want to help them," and so on, but you will never be able to make it free like Google free, like, like Microsoft free, because they, they, they can afford to lose money. They don't care, because they win money from the other angles, right? But you as a, uh, as a one-man show or a company, you will never be able to keep it free because the economics are not going to work out. So-
join3r
Yes
Valentin
... can you use one of these models in this kind of way?
join3r
Depends on, on what ex- because you are s- you are saying something to help with memory.
Valentin
Yes.
join3r
What does it mean? Do you wanna, do you wanna ... So if we are talking about elderly people with some cognitive disabilities, right-
Valentin
Yeah
join3r
... or memory issues-
Valentin
Yeah
join3r
... you probably want to have, um, voice-to-text, right?
Valentin
Voice-to-text and also, um, also, uh, let's say input from the people. Like, not only talking and say, "Okay, uh, build me some schedule f- uh, today to train my memory on, uh, two words or three words," right? I'm talking about also input. So because the, the, the, the age range could be quite big, right? You, you could have people which are affected, uh, because of an accident. Let's say you had some sort of crash or you had some sort of fall and you, yeah, and, and you have some damage, uh, in your brain which actually impedes your memory capabilities. Or you could have people suffering from COVID-19 with this long COVID, right? This long COVID stuff, which is also affecting, uh, your capacity to memorize stuff, yeah? So it de- it depends. You have the old people, like really old, which of course are c- are not so keen or, or able to push buttons on the screen. And you have other ones which are, let's say, able and willing because they always use their phone. But yeah, um, you want to cover all of them pretty much, yeah.
join3r
Yeah. V- voice input is a solved, solved thing unless you are, you are talking about Slovak or, or some, some language that, that's much smaller, right? Whisper and open source Whisper is, uh, is, is really something that can give you in, in English it's really great. You can do it o- on, on a mobile phone itself, so you don't incu- incur any costs, right? Because it's not running on your, on your servers. For German, I'm not sure I would, I would, uh, I would assume it's really good also. Uh, it wouldn't be too far behind be- behind English. And, and what would you imagine, for example, could I say to the model, "Hey, uh, remember, uh, uh, if, if I ask next time where I put the milk, tell me it's, uh, it's in the second fridge," or, um-
Valentin
No. No, no, no. Uh, it's not like this. You would have to imagine yourself being, let's say, someone which has, uh, which struggles to remember stuff, right? Like, uh, remembrance.
join3r
Okay.
Valentin
Well, this can be on different scales, of course. Like, uh, from forgetting where your keys are, like you said, right?
join3r
I-
Valentin
To, uh... Yes
join3r
I have the door already. I have the door already, so, uh, yeah.
Valentin
Yes. I, yes, I also suffer from memory issues. It, and it's mainly from the COVID stuff. After COVID, yeah, because I had COVID like three times or four times, and yes, afterwards the, my memory went really down. Yes. Yes.
join3r
Did you, did you go to Wuhan or where did you get the COVID four times?
Valentin
I'm not sure which, which, which variant it I had, but it doesn't matter actually.
join3r
No, no, it, it, we were just going around Vienna and doing the stuff we normally do.
Valentin
Yes.
join3r
Licks everything, uh, in the metro.
Valentin
Yes. Yes.
join3r
In, in, in the trains and, uh-
Valentin
Yes
join3r
... yeah. Um-
Valentin
Yes
join3r
... if, if, if he w- if he wants to ask someone something like, "What's the time right now?" Or anything like that, he doesn't do it from a meter away. No, no. He goes in. He goes in on you, right?
Valentin
Yes. Yes.
join3r
Like five centimeters.
Valentin
Like very close.
join3r
"Can you tell me please what's the time?" Um-
Valentin
Yes, but this can-
join3r
I, I also he-
Valentin
Yes, go. No
kubista
I also have the memory issues, but they started before COVID.
Valentin
Mm. Okay. But yeah, so, uh, j- it's like joking aside, this is something which is, is of course affecting, let's say a large, a large, uh, percentage of people. Some because of, let's say, distractions, uh, some because of medical reasons, depends, right? But on this scale, like I said, something like I always forget when I, like, where I leave the keys to the scale of like I forget the name of my kids. Uh, like-
kubista
Right
Valentin
... uh, it can be, it can be quite, uh, quite, uh, quite, uh, drastic, yeah. And this would be, let's say if you go into the app, it pretty much will build a profile based on the initial input which you will give it, and based on the interaction which you will have with the app. But the target, the main target of the app is to build you like some sort of schedules daily or depends how much you want to use it and so on. And from there, building you some sort of habits which will also help you to, uh, strength, so to have a stronger memory, to have a more flexible one, to be able to recover from something which you believe you forgot and so on. Because from the research which I did until now on this topic, it's not that you actually, let's say, forget where you left something or where you put something or so. It's actually buried under, let's say, some other levels which you are of course always, uh, replacing because your brain tends to, tends to, uh, replace actions and items which you stored in order to, of course, uh, keep something fresh which you are doing right now, yeah. And yeah, so these habits and these schedules which you are, uh, working with this, uh, app daily, the target would be to, to make you more, uh, resistant, let's say, uh, to, um, forgetting in quotes, uh, these things and also have a faster response on getting the, the, the memories from deep down the brain, yeah. So this would be the, the, the, the targets, yeah.
kubista
Okay. Uh, and, uh, just by the way, I think we lost join3r, so I don't know.
Valentin
Not sure. Oh, no.
kubista
Oh, no? Okay.
Valentin
Uh, so-
kubista
But, uh, if I, if I, if I can go, uh, back a little bit, like still about the, the subscriptions because I just-
Valentin
Mm-hmm
kubista
... like my sub-subscription to Gemini just expired last Friday and I was... I, I also wrote, uh, down or I wrote to the Discord like which subscription should I take, and I was essentially thinking about Kimi, uh, Gemin- Gemini, Claude, and GPT. But I would like to exclude GPT or O- OpenAI because of the latest news that they, they are going to-
Valentin
Mm-hmm
kubista
... collaborate with the Pentagon. So for like for me that's a hard pass, so I would like to-
Valentin
Mm-hmm
kubista
... exclude GPT for any, or OpenAI in any-
Valentin
Mm
kubista
... further, uh, I don't know, discussions. So I am thinking-
Valentin
Mm
kubista
... about Gemini or Claude. And my problem was that I would like to work with the, like some agentic stuff.
Valentin
Mm.
kubista
And I, so I need some cheap, cheap, cheap token, cheap tokens, cheap tokens. So I am thinking about Claude or Gemini. So I don't know what do you think, like which should I... Since you have all, all three or maybe all, all two, G- Claude and Gemini, what do you think?
Valentin
I'm, one thing here which, uh, join3r already mentioned is this usage of, uh, Open Router, let's say.
kubista
Yeah.
Valentin
Um, because Claude... So most of the peoples, uh, which I discussed with they say, "Ah, okay, uh, don't use Open Code," let's say, right? They use a Claude Code directly because that doesn't go over the API whatever and the usage doesn't, uh, counts against the API and Open Code would actually call the APIs, right? Because they did some changes, if I understood correctly, in Claude in order to, let's say, limit the usage of third-party tools, um, instead of going the Claude Code route, yeah? And then a l- some o- some other ones are also saying that Claude Co- Code, uh, so yeah, Claude Code is much better than Open Code. Some are saying no, Open Code is quite cool. They like it. They like the interface, the shortcuts and so on. So yeah, I'm not sure exactly, but why, why would, would you actually use Open Router via Open Code I suppose, right? Instead of, uh, instead of just using Claude Code in this, in, in, in this example.
kubista
I don't know why always when I am podcasting someone is calling me. No one ever calls me, but just, uh- ... just when I need this-
Valentin
Did you hear the question?
kubista
... some idiot calls me.
Valentin
Did you heard what I said, uh, re- regarding this why would you use Open Code, let's say, uh, instead of using Claude Code?
kubista
Uh, yeah, no, I don't, no. There was the-
Valentin
Okay
kubista
... there was the right, right moment like-
Valentin
Do you go via the CLI? Do you, do you go to Gemini via this Antigravity stuff which they have? This, this, this-
kubista
No, no, no
Valentin
... I guess VS Code Fork which they have.
kubista
No.
Valentin
Uh, how do you use it?
kubista
Um, like through Gemini, uh-
Valentin
We lost you. You're muted.
join3r
We, we, we cannot hear you.
kubista
I just, I go, I go to gemini.com like through a browser, not, uh-
Valentin
Okay.
join3r
Yeah. That, that, that's the first thing, right, you need to stop. But, but the problem is that, um, the thing is, um, I, I, I for example I had the $200, uh, Claude Code for one month. Now I'm on the $200, um, um, ChatGPT, uh-
kubista
Whoa
join3r
... f- f- for, for this month, and next month I will probably do something, something else, right? So I, I just know the, the difference and I have no limits on What I can, can do essentially. And, uh, what you said about, about open code and, and Claude is, is, is very much true. If, if you go... The thing is, before I always thought that, uh, the, the AIs and the, and the LLMs will get l- it build, build become a commodity. And it won't matter what, what at the end you use because, uh, you can, you can change the API endpoint and, eh, f- now this, this week I will use Claude and this week I will use ChatGPT, and now I will use this Chinese model that's half the price and stuff like that. And it's, and it's turning out that, um, those companies are much, uh, much better at, at locking you down into their, um, environment than I initially thought would, would be possible. So even, for example, right now if you use Claude through API and you use OpenRouter and you use whatever else, you essentially... Or you are a company that's building on top of, uh, of Anthropic, uh, models. You are paying 30 times more for the API cost as if you would go to a route of the subscription. And that way you are essentially, yeah, you, you, you, you cannot deliver the value. You can still deliver the qual- quality if you have, for example, better tooling, if you have better, better things than, uh, than Anthropic does. The problem is that because of the price they have all those users. Uh, those users have more experiences. They create skills. Uh, Anthropic sees which skills are better, uh, which, uh, which they'll deliver better things. So they have all the, um, all the benefits, right, that, that you as a small guy or, or a smaller company, uh, don't have if you, if you implement on top of them. And the same is true for ChatGPT. It's not as, it's not as apparent. And for the other companies that are still acting like, "Yeah, the same price, API subscription, come to us," you know, uh, it's only because they, they want to gain some, some traction, some part of the market, and then they will do essentially the same. And right now I... For example, you, you think that it shouldn't matter, but a lot of people that use AI really like in a pover- pover user way are using Claude. And it makes a difference. Because, for example, I have, I have Codex right now, and I'm using GPT 5.3 Codex. And, uh, it's like, I would say the model itself, itself, it's Opus level. Not on everything, but on a lot of things I would say it's sometimes it's better than, than Claude Opus 6 point, uh, 4.6. Sometimes it's worse. But the things around it, for example, th- there is a really, a really famous, uh, thing that most of the Claude code users use, and that's called, uh, uh, superpowers. And you also have superpowers for Codex, and I use it in Codex. The model starts behaving in such a way that it delivers only, uh, scaffolding. I, I tell it, "Create me a webpage of this and this," right? So we can... We have a tab, and it- the tab can be, can be a terminal or a web view, right? Created this app. It created a webpage which was all blank white. There were just buttons. The, the, the, the HTML no, no, no changes there. It was j- just those buttons. And you can press them-
kubista
You are talking about GPT, right?
join3r
Yes.
kubista
Yeah. Okay, okay. So I, I had the, I had the same issue when I was doing the, the, the joinit.chat app.
join3r
Yeah. So, so this is what you get if you use some tools that are working perfectly with Claude code and you get much bigger value out of them. But in Codex, I don't know. People don't use it that way. They use planning and then don't use, uh, any skills on top of that. And bas- and, and you, you get... You, you are limited by that. And because people are not using it, it in such a power user way, and there are not as many users, um, nothing like, uh, superpowers comes out of it.
kubista
Uh, but I, I, I like small question for join3r. Since you have the most expensive, uh, GPT subscription, the one for 200, I don't know, euros or dollars.
join3r
Yeah.
kubista
Can you maybe share your API key with me or... I would like to-
join3r
There is no API key.
kubista
No?
join3r
The, um, I can, I can definitely share at the end-
kubista
No, no. It's, it's, yeah
join3r
... when I'm no longer using the ChatGPT, uh, web because I'm asking questions a- also through ChatGPT, their application. And that's where-
kubista
Okay
join3r
... you gain also the access to, to those questions and to the, to those memories, which are private. Um-
kubista
Yes
join3r
... at the end, uh, I will so, I will so, I will have only, I don't know, five days left. I will delete all that. Just as I was leaving Claude, I d- did the same. I will delete everything, and then you can, for example, go in and use Codex, uh, CLI and, and try it. Yeah. That, that we can do.
kubista
And so maybe if I also reflect on what you said earlier, so it's better the OpenRouter offers some subscriptions, subscription or-
join3r
No. You, you, you get i- on, on OpenRouter or on, on e- on Cursor, on everything, you get the added benefit is that you have multiple models.
kubista
Yeah.
join3r
It's not the added value that you, you, you get more tokens because you don't, right?
kubista
But I mean like-
join3r
Uh
kubista
... if I want to use like multiple models, so how, how does it manage the, the cost of API or something? I know the models are free, but a- all models are free on OpenRouter or no?
join3r
No, no, no. Not at all. You have, you have all those models. You have Claude 6 point f- 4.6, but you also have some free ones. Essentially the best you can do is, for example, if there are some new models that, that they want to promote, uh, you can use them, uh, them. Uh, there was, uh, Groq 4 Fast, for example, for a long time. Uh, there was also, uh, the code variant from that free on... I, I think it was not through OpenRouter but on, but on Kilocode, I think. So there, there are options, yeah. I will be back in five minutes. Sorry, guys.
kubista
Sure. The host of the podcast will be back in five minutes.
join3r
Yeah.
Valentin
But, but because you actually said this, um So in my case, right, in my case, it's making sense. So it's making most of the sense to pay this high, high amounts per month when you have the CLI, right? Because if you just go the chat route and ask stuff, it's not making any sense to pay so much. Um, but if you have some sort of, uh, CLI or tool which you integrate with, then of course you get the most benefit. Because I also have colleagues from, from work, for example, which are paying like this 200 per month, right? Which for me personally, I s- I find it as a large sum for, uh, let's say just for chat usage. Yeah. But if-
kubista
Okay. But, uh, for chat, chat usage, yes, but if you maybe like now work with, uh, agents-
Valentin
Yeah
kubista
... and you want to be like you are the power user then, and you are working so then maybe-
Valentin
Yeah
kubista
... uh, it's not, it's not, uh, like the worst use of, of the, of money.
Valentin
Yeah. Yeah. Yeah. I, I think it's good that you have, let's say all of these different types of access to it. Like you have the CLI, like you have the chat, like the web interface. Um, but you really need also like this good, good, uh, CLI in order to really get the most out of it. Like, like he was saying, okay, Codex is not so good because of this missing, uh, extra features like the superpowers, like, or the skills. I, I haven't used yet the skills, but I have seen that people are building skills, but I have also seen that, um, you have now, uh, also security related issues with the skills. You have to be careful what those skills are executing, what they are doing in order to understand if it could pose like a problem, right?
kubista
Yeah.
Valentin
And-
kubista
I mean, what-
Valentin
Yeah
kubista
... what I am, I, what I'm thinking about doing, I have like a old laptop, like eight, nine years old, and I want to completely like, uh, reset it and I would like to have some subscription with, uh, that gives me cheap tokens and like also with a good model, and I want to give it like full permissions for the laptop and I would like to try to build some like agent army or something. But for that I, I, I cannot do local model because the laptop is old. It would, wouldn't handle it. So that's why I was asking about the open router and about the subscriptions that which one would be best for this.
Valentin
Yeah. Um, so as I said, um, it depend, it depends what you want to do, right? But you, like in your case specifically, you said you're a developer, what do you develop in?
kubista
Um, uh, whatever. I, I, I ha- I have like a couple of ideas. I mean, at-
Valentin
Okay
kubista
... at the moment I would like to do the joinit chat app. This is, uh, like self, self-hosted chat app. And then what I'm thinking, I, I have, I have couple of ideas.
Valentin
So you're talking about TypeScript, Node.js, what, what kind of languages or frameworks?
kubista
Uh, what, what, what, what did we use? I would, I would have to check on the GitHub. I already forgot what I what I developed.
join3r
With vibe code you don't care, right?
kubista
Yeah.
join3r
It's something, right?
Valentin
Hey, if, if-
join3r
It's in Ada. It's in Ada
Valentin
... if, if Jensen said that you just have to vibe it to feel it in order to make it work, if the Nvidia CEO can say that, then it's all good. Um, but okay.
kubista
But I think, I think like the, the name vibe coding is not accurate, I think. I think because it, I think it loses all the seriousness about it. Like you just vibe coded you-
Valentin
Yeah
kubista
... so I think I-
Valentin
I guess this is one of the-
kubista
I am, I am... My agents codes, I don't code.
join3r
Mm-hmm. Okay. Um, I, I, I feel it should, it should be, it's the, the value of software that you can do through vibe code and then you do through vibe code is even though it might be more useful than, than real code, I think the value is lower because of the way you got to it. And in, if, if, if you got to it in some, some way through vibe coding, someone else can do it, and it's not, not that much skill needed for, to, to be created, right? And, and at that point, I'm, I strongly feel it's not worth as much because any- I'm not saying anyone, but a lot of people can do it, right?
kubista
A lot of people can do it better.
Valentin
But do you-
join3r
Sorry?
Valentin
But do you care about that? But do you care about that? Like-
join3r
I do
Valentin
... do you just want... Okay, then why? Because for me personally, what, what you said, it's making sense. Totally making sense. Like we appreciate or give values, a value to, to items or to actions based on the effort which is required in order to, right? Because if I need, let's say, a house, and if I can build it on my own in two days, I will not pay, I don't know, more than, I don't know, 1,000 euros for it, right? But if a guy comes or a team of people comes and need one year to build it, and I see or I can understand a bit the amount of effort, then I will say, "Okay, I will pay 10,000 for it," right? But at the same time-
join3r
Still cheap house. Okay
Valentin
... st- still cheap because I'm cheap.
kubista
Slovakia.
Valentin
But still, if you say if you want something, you need a CLI, right? You need like a CLI to, uh, give you the weather, uh, from Kitze every morning at nine o'clock, and you can vibe code it in, I don't know, one minute, and then you can pay me 10,000 euros to code it in one year. Would it make it for you? Like you know that I will invest, let's say, this effort for real in one year, and I will give you a fancy package and a nice app, one million lines of code, but they will give you the same output pretty much as the one which you vibe coded in one minute. For you, would, would it make a difference?
join3r
No, but that was not s- that was not my point. My point was that, for example, if you- Vibe coded that thing, right? It, it ... I'm, I'm saying it, it, if it could be vibe coded, vibe coded, for example, in one week, right? And you created an app, and you will come to me and tell me, "This is an awesome app. Look at it. Ju- just see the features. I can sell it to you for some amount of money." And I will say you, "Oh, this was vibe coded. Cool, I will make it myself." I won't buy it. Your code has no value. Your power has no ... B- before you, if you were a coder, I could say, "I could never have done it in a million years," right? You have those skills that you honed. You, you went to a school, which was probably a li- a huge mistake, right, kubista? And, uh, you, you studied programming and, and, and you do all the stuff, and you can program. Uh, that, that's something that, where I would need to get to your level, right, which would take years, and I would have to create that app, and it's a, it's a large investment of, of money and everything. But you show me your app, and I would, me being me, I would immediately think, "That's nice. I would really like this line to be a bit up." And I would go to you, "Can you make it up just for me?" No, I have, I have thousand of those, or I have hundreds of users, and someone wants the line down, and someone wants the li- line up. I like it more up. I think most of my user think that's best too, and I would say, "Okay, I don't care. I will go. I will vibe code it myself, and I will put the damn line just one line be-befo- under that, and I will, I will even, even take screenshots of your app, and during vibe coding I will say, 'Okay, make it look like that.'" And at that point, the, the value that you can, that, that your application has, has, has diminished greatly, right? And, and th- that, that, that was my point. If, if, for example, if, if you do something and I view it and I say, "This cannot be vi- vibe coded yet," right? Yet might be the, the important word there, but it cannot be vibe coded. Then it, un- un- ... then it has value because I cannot do it.
kubista
Okay.
join3r
I cannot do it better.
kubista
So I have to, I have two solutions for this problem. So either I don't tell you I have vibe coded it, or what, what if I develop application, I put, uh, like a ton of effort into develop it, into the development of the application-
join3r
Mm-hmm
kubista
... and then even after that, you say that I can vibe code it. Well, that's-
join3r
Yeah, that's, th- th- th- that's unfortunately you shouldn't be coding. I, I don't think ... There are a lot of things-
kubista
What?
join3r
No, I, I really think there are, there is a lot of things that you do, you should no longer code. I think there are a lot of things that you should just vibe code, and it doesn't matter.
kubista
But I mean, yes-
join3r
Those are some simple things that are not critical that, uh, yeah, if, if you, if ... As I said last time, if I went around and, uh, created a script that, that was cleaning up images in, in a r- in a registry, right, which were used in production, yeah, it would be really painful if something would be missing which is still needed in pr- in, in, in production. I would need to go and build it for two hours, and I would have two hours downtime. I will code it myself, right? That, that, that's something that, that's critical. I will probably still use AI to do it, but I will not vibe code it. But there are things, yeah, I will happily vibe code, and I think not doing it in the vibe code thing, uh, way is just, uh, is it's, it's ... First, it's waste of time, but it also might be a hobby. If you like doing it, do it, right? Code it. I, I like coding myself. I no longer do it because there is vibe coding, right? It's, it's, uh, it, it's, it's no longer fun for me. But as a hobby, yeah, go for it.
kubista
You see, uh, this is why I believe at the end of the day it's like, uh, it's always like a point of view thing or like a perspective, right? Like you said, I like coding. I will do it because I like it. This is a fun activity for me, and it's making sense. But at the same time, it kind of spoiled, let's say, the fun or took the fun out of it because you say, okay, now everyone can do it. I can do it like in one minute, uh, compared to 10 hours. It's not making sense. But for me personally, it depends what you want to achieve, what you, what the target or the, what the end goal is. Like, because, you know, we, we spent hours and hours on writing Ansible playbooks, for example, or scripts, when in reality you don't care. You just want to reach the target of deploying something on 20 servers, 1 million servers, whatever. And if you can trust it, let's say, that it will do the action which you need on all of those items, it's fine. You don't need to spend w- two days on testing and retesting and changing a small, uh, string in order to see if it will r- run now or not. And at the same time, it's also like a beginner and expert level thing. Like y- you say, "Okay, I can vibe code it like you. You vibe code it in one week, I will do it in one week or two weeks, and that's it." But in reality, I don't believe it's going to be like this because-
join3r
Why not?
kubista
... people underestimate also the, the effort which is necessary, the energy which they need to put into something. Because I also did like, oh, I, I have seen some of vi- some of the videos, and I will say, "Okay, well, I can vibe code it, right?" And then I started doing it, and I realized that, okay, wait, it's a bit more effort. Oh, okay, I, I have to be careful how I word this. I have to be careful how I write the, the descriptions. Oh, okay, I have a MD file now. Oh, wait, I also can connect some MCP servers. Oh, okay, now I can define skills. What's, what's a skill? There is a, there so, there are lo- so many levels and layers, and it will just get more complex with time. It's not like you can say, "Okay, I found, uh, Twitter. Now I will vibe code Twitter like in one, two weeks, and it will be great, and I can sell it," right?
join3r
No. There, there, there-
kubista
You, you, you cannot vibe code it, yeah
join3r
... there will be ... Yes, there will be cases where you can vibe code. It will be to some
Valentin
maturity level to some level, let's say to something which you have seen. But at the end of the time, it's also an energy input which you have to put in to make it work. Like, and you also need, need some sort of access to, to maybe to some hardware, maybe to some subscription. Maybe you can Vibe code, let's say something which you said, okay, you're, you're coming here to sell me this app, which I can Vibe code, right? But maybe you will need a 200 euros per month subscription, uh, to Claude or something for six months maybe. And you say, "Oh, I don't have this money. I don't want to invest it," whatever. So I don't believe people will be really so, so at the, like at the beginning, they're really like overestimating their capabilities and their capacity on doing something. But at the same time, I don't believe it will happen that much as we overestimate now, yeah. So- Because let's say y- one last point. If a guy... If, if you say, "I want to build here a wall," okay? And you say, "I can do it in one week, but I will, uh, call a guy," right? He comes and does it in three hours, right? And you say, "Oh, that was nothing," right? "I can do it then in two days." But when you start to actually do it, you realize that, oh, wait, it will actually take me one month because I underestimated the effort, the skill, the labor, the tools, the, uh, there's much more. But you don't see all of those levels or layers because you're not experienced enough, let's say, uh, in that, uh, thing, yeah.
join3r
So I consider myself being really expert Vibe, uh, coder-
Valentin
Coder
join3r
... Vibe dev- dev- developer, right?
Valentin
Yes. Okay.
join3r
I know most of the stuff. Um, I have-
Valentin
Mm-hmm
join3r
... much wider knowledge than, than I would say most of the people that I-
Valentin
Mm-hmm
join3r
... that I looked up on, on, on YouTube or at the-
Valentin
Mm-hmm
join3r
... where I was on training, right? So I, I would say there is some amount of expertise that you need. There is some amount of money subscription-wise that you need. But the subscription, right, I will have it because I need it for my work. I will need it because I have other small apps that I would like to use, that I would like to have. And adding your app at the end of it, right, uh, is just another project where I have to see, okay, oh, so I see, uh, here is a highlight on this window. I have seven windows open, seven projects, right? And I see, okay, oh, here it's asking something. I need to do some manual test that, um, uh, I will do or, or, or, or tell it to just, uh, run the Chrome MC- MCP and do it yourself, whatever, right? And the, the, the, the value of the software, right? If, if you, if for ex- if for example, right now, if you look at Reddit, and if you check the pages that are, uh, about self-hosting, uh, the, the self-host, uh, subreddit, for example. Or if you, for example, go on a some, some, some, some Reddit that's, uh, s- some subreddit that's, that specializes in, uh, in, in piracy of, of videos or in, in those radar, sonar things, right? You will get... You will see a huge flood of Vibe-coded stuff. And that Vibe-coded stuff has, has, uh, has v- very little value besides for, for that person who created it, it could be the, the best thing in the world, right? But for me, I see an unknown developer creating a project. I don't know how long he will do it, right? I don't know how, what everything is in the code, right? There are, there might be features that I don't want. There might be, um, security issues because he didn't run the second pass, uh, for, for security stuff, right? Uh, for example or, or, or Edisar in the, in the, in the plan, it, it told him, yeah, we will just do it, uh, this way. We will use no salts and yeah. Uh, uh, we will not use OT because I had problems with it and I didn't... I used something else, right? Um, uh, just, uh, Opus stuff. Um, yeah. And, and at that point I'm, I'm, I'm looking at it and the, the age of personal software has already started for, for, for people like me which, which pay for that subscription because they need it and they need some software because they are really opinion- opinionated, right? They, they know what they want. S- sometime, sometimes when I Vibe code something, I, for example, created a, a browser which had, um, which had folders, and there were also smart folders, and those smart folders would have like, uh, uh, something like SQL syntax. Uh, so it, it was really like emails, right? But, uh, instead of, uh, emails you had, you had, you had pages, right? And that was something that I really wanted. And once I made it, I, I, I found that I don't want it, right? It's, it's, it's, it's not good. But it's, um, I, I think I, I already forgot what, what I wanted to, to, to talk about, but Vibe coding is really fun. That's, uh, that's what you, you should take away from it. The, the, the major, um, barrier of entry is, uh, is right now price, right? Because if you have those 20 euros model, uh, 20 euros subscriptions, you can with, with Opus sometimes I've, I've heard stories that people couldn't go through the planning phase because during the planning phase their, the, the subscription ran out or ran out, right?
Valentin
Yeah.
join3r
Um, so yeah. It's, um, it's, it's like that.
Valentin
But, okay, so but if I understood you correctly, pretty much so this slop, right? Because this is one of these-
join3r
Yes
Valentin
... terms which is used. This is slop, and you generate pretty much garbage which is useless.
join3r
It, it doesn't have to be garbage.
Valentin
Yeah, yeah. But yeah, but it's like, well so it's not garbage like you said. It's valuable for the creator, for the persons, for the person which does it. But for 99, let's say, percent or 90% of the people, rest of the people, it's garbage. Or they would consider it, let's say- Useless to some extent, right?
join3r
No, it could be, it could be-
Valentin
Or not, not valuable
join3r
... no, no, it could be useful for me. It could be useful for me. But, and I- when I see it and I see, okay, it was created, uh, two months ago, two months ago, uh, the, the, the first, uh, commit was huge, right? I smell AI. I see the AI, uh, um, things in it. I immediately say, "No, it's a new application. You are a new developer. You have no history of supporting your project," right? If, uh, if something breaks, if, if you did something incorrectly, where is ... I, I don't trust you. It's, it's a, it's a, it's a thing of trust. And I trust myself. So if I see something that was Vibe Coded, I know I can Vibe Code it too-
Valentin
Ah
join3r
... because I'm, I'm using the best model, right?
Valentin
Yeah.
join3r
We have all the same harnesses. We have-
Valentin
Yeah
join3r
... essentially the same skills. Ev- every- everything, everyone has some custom shit, right? But it's, um, it's, um-
Valentin
Yeah, but I-
join3r
... it's not a big deal
Valentin
... I, I, I believe you're not in the, let's say, in the majority, uh, you're not in the majority category of, of, of people. You are probably like, like you said at the, at like this 1%, 5%, whatever, of the peoples which are really know what it in- what, what it means, what, what how they can do it, they will do it. And for you, it's just a choice, like from the point of, okay, do I have the fun, the time and the fun now, the energy just to Vibe Code it myself or is it cheap enough to buy it, and that's it, I don't care? Uh, for you, I guess it, it's, it's something like this. But if you have the time, if you have the money, it's not going to generate cost for you because you already pay for it. You don't ... It's nothing. It's just a copy-paste what you said pretty much, and some change and execute. Then yes, then of course you, you will probably never pay for it because you know you can do it in five minutes, 10 minutes, one hour, and it's okay. For, and for, and from what you said, I believe it's also like, um, cultural thing, some of this, like because we're used to, like we are, we are also old enough to know a bit what it, what it means, let's say, to create software. We did school, we went to computer science. We know, okay, wait, uh ... Well, not you, me. Uh, but-
join3r
Exactly
Valentin
... uh, but I, I went to the university. I did computer science. I did operating systems. I did programming languages. I did logic. I did, like you, you start and like your brain is saying, "Wait, wait, wait, don't ... This is a hard problem," right? You see a piece of software and say, "Oh, wait." You're thinking already, "Ah, okay, I need to build it, I need to compile it, I need to build classes, I need to build functions and so on." And which of course now you still have to do it, but you don't do it yourself. And this is why you say, "Okay, wait, this is not so much effort anymore." But when we started programming and we're think- oh, Pascal, FoxPro, C, C++, Java, whatever, and it was much more painful, let's say, to learn s- to learn a language and then to switch, uh, to look at architecture patterns and to look at, let's say, the good ways of programming and so on. Uh, so now it's different, but for most of the people which are not programmers themselves, and will never be, they don't care, in my opinion. They will never care how you wrote the function, who wrote it, is it coming from AI, is it coming from you. They don't care. This is just my opinion, but this is what I feel. Because, because as I said, if, if the entry now is really low and you can just ask for something and get it, maybe you don't care if it's written nicely, if it's secure, if it's, uh, really following the best practices or not. And you will get it anyway better than a guy in India, uh, which, uh, just started his profile on, on, on freelancer.com like last week, yeah? So maybe you, you, you don't care if it's, uh, AI generated or not. But yeah, now-
join3r
Um, I, I, I care-
Valentin
Yeah
join3r
... only from, from one standpoint, and that's if someone else Vibe Coded it, I can Vibe Code it too. And-
Valentin
Yeah
join3r
... it, it, it will, it, it will, it will, it won't take me five minutes, but it's a back- background task that you have to check once from one, one, right, right? You, you do the planning, which I find fun, right, where you plan how things should look. You do the, the version one. You read the, the first plan. I'm only reading the, the first plan.
Valentin
Okay.
join3r
And then you execute it, uh, you view it, you do few remarks here and there, change this, change that, and yeah, let's go to version 1.5. And at that point I, I'm no longer need reading the plans. I'm just Vibe Coding, really. And-
Valentin
Okay
join3r
... the, the, the, the thing is I don't mind if something is Vibe Coded. I just don't think you can ... You, you either have the trust of, of people like me, of, of IT professionals like, like you, right? Where you have a name and you can say, okay, uh, if for example you would Vibe Code something and you would, would, you would, uh, show me f- for free, right? We still, right now we are talking about free. You would te- tell me, "This is a good person. I Vi- Vibe Coded it," and I would say, "Okay, cool, thanks." And I would use it, right? I would have no, no quarrels about it. But if, if it's someone, uh, I don't know, I don't know if he's being, uh, if he supports it, uh, what, what choices did he make, right? I'd, I don't trust it basically on the, on the, on the, on the simple thing that it's Vibe Coded. And I will try and, and will trust myself better to do it. Um, if it's, if it's a real code and he, and that single developer was working on it for two years or three years, okay, you put the eff- effort in. You have my attention. Right now I'm believing you are, you are in that project to To stay essentially, and I can, I can trust that you will not drop it like a hot potato one week later. And that's something that we also see on, on Reddit, the self-hosted. There was a guy who came, he had nice Vibe coded thing. I don't know what it was, uh, of c- unfortunately anymore, but he... A lot of people jumped, uh, jumped on it. He created his own subreddit. He had a commu- community around him, and, uh, then some guy came in and checked, did the basic thing that I would do at, at, at first, right? He, he just do- uh, did a slash security review, uh, on, on, on some of the PRs, right, and all nasty stuff, uh, showed up and how, how you can get essentially API calls extraction without authorization and stuff like that. Really nasty stuff. And that guy disappeared.
Valentin
Yeah.
join3r
Right? And, and, eh, it's, I, I, I think it's... That, that, that's the problem right now with, with Vibe coded things.
Valentin
Mm.
join3r
And, and I, and I don't think it will get better. I think it will get worse because, uh, I, I think the barrier of entry, right now you still need skill, and I, and I think there is still difference between beginner Vibe coder and an expert Vibe coder, but I don't think it's a, it's a, it's such a large difference than it's be- between a non-coder, coder, or expert coder, right? Th- th- there are... Th- those are huge, huge steps that you have to make, and they are made much, much smaller thanks to, to AI. And right now we, we, we, we, we still talk like my mom won't be creating an app with AI, right? She, she, she won't be, she won't be doing that. But if you are talking 10 years in the future, what's an app, right?
Valentin
It could be. Yeah, yeah.
join3r
It could be like, "Hey, um, ChatGPT, um, please..." No, not ChatGPT. Um, "Hey, Mistral," because Mistral will win. I'm, I'm, I'm sure of it. A European co- uh, company. Uh, "Hey, Mistral, um, I wanna," uh, I don't know what. "I, I, I wanna e- every, every, every evening I wanna see the, the, the new series, and, uh, the series just doesn't come out anymore, so create me a new series based on that series, and I will watch it forever. And, uh, I also want an app where I feed my dog, and, uh, it, uh, it, it does those things," right? And she won't use the word app. There won't be-
Valentin
Yeah
join3r
... an app anymore, right.
Valentin
Some names. Yeah.
join3r
So that, that might be the future, but that's, uh, doom and gloom and, and also optimism and everything all together.
Valentin
Yeah. I also, I also, I also believe this app thing, yes, it will change, yeah, because it's something which we are currently using because of historical reasons, right? Like, like I said, we started into something which we inherited pretty much, right? Software is, for us, a l- legacy, let's say, term which we used and used for some time, and we're not the first ones which came up with these, uh, terms, right? With software, with application, with whatever. We... They came out of history, and we moved further with them and so on. But it will probably change because people are the ones which dictate, at the end of the day, what's relevant or what's used or what should be used. Because even if we, let's say, as software developers or software passionates or nerds or whatever, like using apps, like using some terms or strings for something, it doesn't matter. If we are, like, going to be in the 1% or 5% or 10%, nobody really will care pretty much because the mass or the majority is the one which is actually deciding. And this is why like, people are, are, are always look- uh, looking is this the year of the Linux desktop, right? And it, it was zero point something, and it was one, and it was two, now it's five point something, right? But it's still not the year of the Linux desktop. It will, it will be the year of the Linux desktop when it's going to be probably 60%, 70%, 80, 90%, right? So yeah, I... But to be honest, although I work in IT, I have, uh, let's say, uh, education which is based on the classical IT stack, and I'm... I really like IT, and I like my job and, like, at the same time, I can see the, the, the, the, the, the, let's say, the fun in doing some of these Vibe coding projects, right? Because you can say, "Okay, I just want to test like you did your browser," right? I just want to see if what would the browser look like and work like if I have this function into it, and you just Vib- Vibe coded it. It worked. You realized quite fast, okay, it's just not making sense. I wouldn't go further with it, and you dropped it. But what would this take from an energy and time point of view to implement on your own, write it by hand, uh, forking some project and then extending it, whatever, right? Much more effort. Much more effort.
join3r
Of course. No one is arguing against it, right?
Valentin
So, yeah, so-
join3r
Vibe coding is much easier. Uh, and, and I think because it's easier, it's, um, it's, it, it inherently has less value.
Valentin
Yeah. True. True.
join3r
Yeah.
Valentin
But, but, uh, but this means also a much cheaper, cheaper entry point of view because y- you pay for a, for the subscription, I don't know, two- 200, you will be able to do, I don't know, one project per month, let's say, full, f- full project, which is really cheap because if you would pay someone or do it on your own, the effort and the money required are much, much, uh, much different, yeah.
join3r
So for 200, uh, euros, I'm doing, like, a lot of projects. I don't have a count, but there is a lot of folders. And I'm not hitting anywhere near, near the limit.
Valentin
Wow.
join3r
And I'm using agents of teams that work at the same time, right? So, so they go to the... They, they, they burn much more tokens, right? Um-
Valentin
Wow. Okay
join3r
... it's, uh, it's, it's really... I wouldn't say it's, uh, unlimited Because once you start doing like for example, I want to do security audit of this, uh, network, uh, every day, right? And at that point, i-if you have multiple things like that, you start, uh, seeing the, the, the percentage rise. Okay.
kubista
Mm-hmm.
join3r
It's, it's higher. But as a single person, you basically for two hundred euros, you basically never hit the, the limit. And even with Claude one hundred euro, um, it's, it's still Claude Max one hundred euro plan. I'm... I still think you are-- you, you, you, you will hit the limit not that often.
kubista
So we moved from, uh, software as a service, like because two weeks ago it was the, the, the news that it's dead, and so now we've moved from software as a service to function as a service or model as a service. So all the subscription will move from, uh, I don't know what to, to the LLM providers.
join3r
Uh, it's, it's, um, if, if... So humans are terrible, uh, about predicting stuff, how it, how it ends up, right? And, and I'm not, not telling you I'm, I'm different. I made a lot of mistakes in the past, and they are on the internet thanks to the, to the podcast, right? Um, but still right now, if you are talking ten years, it's hard for me to image-imagine that, uh, that things are going to change. And, and I, and I think you, you would need to be blind or really close your eyes really, uh, really strongly to, to not see that happening, that things are changing. They are changing fast, and the how they can change is, uh, is also dramatic. The, the... We are talking web can essentially really change the, the internet itself because a lot of people and their own-- their only entry will be their ChatGPT subscription. Um, they, they won't go to Google. They won't search for pages. They won't search for cooking recipes. They won't search for news. They won't search for, um, uh, for stuff. There are still things that are not catched by LLMs right now. Some of them I'm working on, right? So for, for example, you, you can, you can have this one, um, I will now, now give you one, my one million dollar idea. Uh, you can, you can have this one, um, window into the internet through LLM, and you can have it to some, some extent right now if you don't go and watch, watch movies, series, YouTube, and stuff like that. Video is still something that's not being done that well, and audio I would, uh, I, I would say is, is better, but it's not that, uh, that present. You, you, you can do it with, with, with tools, right? And, and it's really easy, but it's not through that LLM. It's not integrated yet. But, um, for example, if I want to know what other people think, that's where I need to go to the internet. Or I will need to ask the LLM, "Okay, what's the, what's the idea about this on Reddit," right? And it can go there, but it probably will get limited sooner or later. Right now, for example, on YouTube, you will have a really difficult time to getting, uh, getting stuff off, o-off, off of YouTube because they are blocking LLMs right and left. I, I implemented something a week ago and, uh, for my Discord bot, and it's already gone because yeah, they, they blocked it. Uh, and there are services like that which are really hardcore, and they will go after you, and they will break your stuff. Um, but, um, yeah, I, I think that that change is coming and that software might... The, the, the software, the small software I think is already dead. It, it just that, that doesn't know about it, that there is, there is influx of new software which is so large on self-hosted Reddit that they said, "Okay, from now AI, uh, AI vibe coded stuff only on Friday," and they, they have to limit it. And, and, and if you see the, how the, how people react, it's not just this is AI, therefore it's, it's shit, right? No, this is just too much. There are too many things, and there will be too much software as, as it is now. Too-- It's too much music, for example, right? And, and still you, you find you have trouble finding something good, and it will be too much software, and you'll still have trouble finding something that you like. But at, at least someone like me can still create it and, and yeah, I, I think the really big change is coming, and I'm afraid for my job. Honestly, I'm, I'm really afraid, and I don't really have a plan on what to do. I should have become a, a dentist.
kubista
For, for sure.
join3r
Yeah.
kubista
So some, some let's say points here. One, maybe you can share with us the setup which you have and show us how to do this, like in a, like in an expert way like you are doing it. Because as I said, I have this project and I want to do it. It's not from a commercial point of view. It's for, so as I said, for my parents first and then of course other people which are affected from these memory issues. But I would like to, to do it, to learn how to do it in a, in an expert way because me as a one man show, I will never be able to do this to reach, let's say, the point for this app or model and when I say, "Okay, this is good enough. I will now put it in beta or something like this." No, it will-- It's too much effort. I will not be able to do it, yeah. And vibe coding, let's say, uh, would probably be able to bring it closer to, to that point. Of course, the quality, as you said, probably will not be the best, but I believe we, we have an advantage here because we have the classical knowledge, let's say, cl-classical view on this tech stack, on the architecture stack, and you can combine and of course probably the end result will be better than a simple, let's say, uh, guy vibe coding, uh, in one evening and that's it. Maybe. Maybe. Um, and then, yeah, I, I agree that the, that the industry is changing
Valentin
We have been working in IT for some time. I have-- I'm not sure how long you have been, but for me, it's twenty years, I suppose, more than twenty years. Um, but, um, so I'm not saying that, uh, jobs are not affected or my job is not in danger. It could be, it will be probably changing for sure at some point in the future, and it has changed actually. We actually started using free, for example, are using the Claude model via OpenAI services in Azure, and we integrate via the, so you, you can do it via the, the open code, or you can do it via even the code, I guess, via Claude code. And you can use the OpenAI APIs, uh, and that's it. So this way you can use it in an enterprise way by, uh, generating subscriptions. So you could have, you could create subscriptions, you could do, uh, APIs in each of the subscription, and you could link the subscriptions and APIs code per project. And this way you could see, okay, this project is, I don't know, uh, consuming per month twenty K, whatever. This money will be taken from the budget, uh, which is, uh, which is allocated to this project, and that's it. What we would need, we would need probably to have a context switcher like you have for Kubernetes in order to switch between the, between the projects, so that you will be able to switch between the API keys. So when you do diagrams, let's say Claude code is really good at generating, uh, diagrams, like architecture diagrams. So instead of you spending a lot of time creating these diagrams and linking between the components and the integrations which are ne-necessary for some systems, it could generate it quite fast, and it does a good job. So I can tell you this from firsthand. Uh-
kubista
Yeah
Valentin
... so this is something which would take my job, let's say, right? But your job is much more complex than just some diagrams. This is one example because you're going to say, "Ah, okay, but this is the main." No, this is not the main, uh, the main part. The main part is actually that you are an enabler for the projects which you are involved in. So if, for example, I go in some project and they say, "Okay, we want to do this, and what do we need?" Right. You will come up with all kind of, let's say, requirements which are going to be put, let's say, uh, black on white. You're going to come out with analysis, which is very important to understand what might be some sort of technologies. Yes, this could be taken by AI, let's say. But then you also have the connections which are necessary between the teams, the people, and so on in order to make sure that they are able to reach their, their release dates. You're going to-- So maybe you sound like some sort of manager at the same, uh, like at the, like, like in this project, but you're actually like an enabler, actually. You're going to act like a glue between all of these people and teams which are good at working small parts, but really bad when it comes to deliver projects. Yeah.
kubista
So, so you will become a Scrum Master for your-
Valentin
No
kubista
... agents.
Valentin
No, no, no. No. No, actually not be-be-because the, the Scrum Master will be still there. The Scrum Master is actually the one which is trying to make transparent the work which is done by these teams. You're actually going to know all of the people, all of the systems, all of the platforms, and then you're going to be able to, uh, let's say, enable the project or the other people which need some component from all of this in order to make sure that at the end of the deadline, you will be able to deliver something, yeah. And for now, at this point in time, a Claude or ChatGPT or something else is not able to do it, no. I'm not saying that it will not be able to do it, I don't know, in one month or one week, but at this moment in time, it's not. Yeah.
kubista
But there is a high chance that it will.
Valentin
It will, but even if it did, so let's say it will, and I actually hope that it will, because maybe then I will have more time to use my brain for something else more interesting. Now-
kubista
Be careful, be careful what you, what you wish for.
Valentin
Yeah, yeah. It still remains, of course, the question of how do you get paid or what do you get paid for. But this is also like a political question, right? And it, in my experience until now at least, there is no technical solution to a political problem. So if you say, "Okay, AI will take my job," it's okay. So I'm not cla- I'm not going to oppose it or whatever and say, "No, no, no, AI is shit," whatever. But it will be someone else problem, and mine, of course, to figure out how do I continue to be able to live, to have some sort of standard of living, to be able to feed myself, my kids, and so on, right? So if, let's say, AI would come and take all of the IT jobs in the future, you will have a mess of people which are unemployable, and then you will have to come up with options like, I don't know, do you re, uh, re, um, um, do you train them re-- So yeah, do you retrain them in order to work something else? Is that even doable? Or do you put them on some sort of minimum income, whatever, and you just give them free money because they need to live, you can't kill them, or whatever else-
kubista
Wow.
Valentin
... whatever kind of options do you have? And this is, of course, one of these political, uh, political questions. Austria is really famous, let's say, for having this social approach to, to, to people and s- and, let's say, society. So you have a lot of, like a lot of, uh, programs which are, which were designed, let's say, to have equality between the people and so on. I'm not saying I'm for or against it, but
join3r
It's of course not always working as it was, uh, designed. Yeah. Yeah
Valentin
Slovakia is known for not having this approach, so we have the
join3r
The, the, the depends, right? Depends on what you compare it to. If you compare it to, for example, I don't know, Nigeria, then we are, we are really, really good-
Valentin
Successful
join3r
... ab-about, a-about taking s- the taking care of, of the, of the lower social classes. But, but I think the... I don't want to get overly political here, but, um-
Valentin
Yeah
join3r
... the, we still live in a cap-capital world, world-
Valentin
Yeah
join3r
... and, uh, uh, capitalism forces, um, will, will, will, um, I'm not... If y- let's for example say 50 years in the future, right? And, and I, and I think, and my, my idea of 50 years in the future is much rosier than is some, some point between that where the AI hits, where the, the, the, the, the number of people that are affected is, is, is small enough that it doesn't need a global, um, solution-
Valentin
Yeah
join3r
... but it's still large enough that you cannot do it via the standard stuff that you do the, the n- normal with, with, with lower social classes, right? And, and at that point, I think that, that, that if, if, if, if it happens, right, I don't know. I'm, I'm speculating way more than I ever did. I don't know. But, but at that point it, it probably will be really bad, at at least from, from what I can imagine. Um, and from the, from the, from the technical point, right, um, there you, for example, said, uh, e-enablement. I, I, I think you would be really hard-pressed to tell me one thing that you can do and, uh, which couldn't be a skill for Claude or for ChatGPT. I think where, where it, where it, where it, where it starts getting into the trouble is where you cannot name something or it's a complex thing of multiple... B-because my thing is not just going in and writing Ansible scripts and wr- and, and running them.
Valentin
Of course, yeah.
join3r
My thing is not just going in Kubernetes and, and running that, right? My scope is much larger. I'm doing a lot of stuff that I couldn't name but for example-
Valentin
Yeah
join3r
... which, which, which the AI probably wouldn't do. And, and but the other point of that key, of that coin is that everyone feels he's like, not everyone, but, but a lot of people think that there is this delusion that they are way more important than they are for the business itself. I'm, I'm-
Valentin
Yeah, sure
join3r
... I'm not saying, uh, about, uh, how important you are to the team, to the, to the, to, to something, right? But for the business itself, once you start being in a not, not, not start- startup, not small c- company, but a middle-sized company, when you are there, you're ... You, you can think you are un- uh, irreplaceable, and many people thought that and were fired and the business went on. The, the, the part that we're ta-taking care of it would, it was much worse. It wa- People were complaining, whatever, right? Business went on and, uh, yeah, I think it's the same way with AI. But right now the, the major stumbling point of AI, for example, the adopt- the, the adoption of AI is, is difficult. It requires change of how we think, think about things. Um, first part, for example, when I was doing, um, training for, for people in my company regarding AI, the first part I was, I, I did was let's install its, uh, GSD, right? It's a, it's a workflow for, for, for Claude code and also for OpenCode. Let's install GSD. But the first part I did was open the, the GitHub page and you have three commands that you have to run to get it running on your computer, right? But that thinking is already something that, that you should, you should keep track of because, um, much more AI native thinking, I'm sorry that I'm using such, such, so, uh, such buzzwords, but the AI thinking here is, no, don't install that. Go to your Claude code or to your OpenCode or Codex or whatever you use, and right there install GSD and press Enter and let it take care of it, right? And nobody from, from our, I don't know, we don't think like that because it's not, we are used to 20 years of going in, reading at least the basic stuff on how to install it. Okay, it's, uh, just a Docker Compose, right? You just do NPM commands, right? I just edit the config here. But yeah, I think that's, uh-
Valentin
Yeah. Be-because we come with some legacy, right?
join3r
Yeah. Yeah.
Valentin
Like we come with some baggage and this is why old habits die hard, and this is why we also have this need of understanding what's happening behind the scene, and this is the main difference between us, let's say, and s- and like my mother, which will not care what's happening beside, behind the scene, what commands are executed, what Docker is actually running, which pod, how many, and so on and so on. She doesn't care the limits which are set on the pod. She doesn't care if she runs out of memory, whatever, right? You understand what happens or some part of it, right? You know, ah, okay, okay, there's a Docker Compose. I start the services. These are the exposed ports. These are the limits. Okay, these are the load balancer in front. Uh, this is an API. You know what happens, let's say, be-behind it. But the, the difference is do you care or you don't. If you don't care, then yes, you just go and say, "Claude, install this," and that's it. And that's the native way because it's actually making sense. If you really have this tool which is enabling you to do whatever, then why would I need to do some stuff manual, right? And this is why we have some package which we will probably never be able to, to shed off because it's just the way we were, uh, raised. We have, uh, got used to this and, uh, it's hard to change habits. It's hard. And so it's also hard to build habits, but it's also hard to change them, yeah. And yeah, uh, but from my job because you said, uh, okay, name something which you-
join3r
Mm-hmm
Valentin
... wouldn't be put into some skill. I'm sure it could be, let's say, so you could probably cover 99% of them But the main question and the most important one, I believe, is who is going to work with this AI? Because if it's people which are working with AI, then you will probably never be able to change me or not one hundred percent. But if you have AI working with AI or AI working with people, then of course it's a total different discussion. Because if you say, "Okay, I-" It's a big difference because if you say, "I have people," and you-- let's say you change the architect with some AI system, right?
kubista
Mm-hmm.
Valentin
People are always going to be able to work better with other people than with AI because you will always need some sort of input into that AI which is conforming to something. You have some rule set, you have some sort of, uh, uh, structure. You will have some sort of limits which people will probably be the ones which they reach and they say, "Okay, this is taking too much energy, too much time. It's too stupid. It's not understanding me. I'm not able to get what I need out of it," and so on. This will be the case, not because the AI is limited, probably because the people are limited. But when you say, "Okay, I have an AI and then I want it to interact with another AI which is doing the architect role," let's say, then it's okay. You-- If you are able to make them work as you need, then it's totally fine. You can exclude me or you or whoever. But if you're going to have on the other, uh, end of the stick people, then no. You're-- I'm always going to be more efficient if I want to than the AI system which you try to put in place of me. Because you can, you can go in multiple ways and flavors, uh, on your, uh, work style, on your speed, on your approach and so on. So this is what I believe that this is what's the-- going, going to be the deciding factor. Do you have people or do you have AIs or systems which are interacting with each other?
kubista
But, uh, sometimes I would even choose to interact i-I would prefer to interact with the AI than maybe some asshole colleague, you know? So...
Valentin
Yes, yes, I agree. But the-- but there are so many angles to this problem. Like you can say, "Okay, he's just an asshole and yeah, I don't like working with him." But you can also have the other coin, the other face of the coin, like you have a good, good, good, um, let's say colleague which you really like working with and you learn from him or she and you have a lot of benefits, let's say also on the social side because we are social, uh, social animals, let's say, and you need this, this interaction. Even if you are okay working on your own, even if you say, "Okay, don't need anyone," this will just go for some time, for one week, one month, one year, ten years, and then you're going to see that you've-- are facing issues because of this lack of interaction with people. Yeah.
kubista
Yeah, but then, uh, like maybe let's say for work you do most of it like with AI, but then if you want to socialize, it should be maybe something extra or let's say you are now so efficient that you, you don't need to work eight hours a day, but you can do all the work in seven hours and so let's dedicate that last hour in the company for some social event or something. So that's also-
Valentin
Yeah.
kubista
But I don't think it's going to happen.
Valentin
It's-
kubista
Because like different people have different, uh, social needs, I would say.
Valentin
Yeah. Yeah, yeah, yeah. So, so many characters and personalities and so on, it's different. Uh, people are very different. But, but as I said, at the end of the day, it comes who do you work with and then, uh, who do you give input to or output to and it's, uh, it's going to be, yeah, uh, depending on that because if it's, if you're working with AI all the time or most of the time, then yes, then you could probably also be replaced with AI at some point because it will be m-more efficient, yeah, and that's it. Um, but yeah, but it's for now at least it's still, uh, one of the problems which is a bit, uh, tricky to, to change or to, uh, solve. And yeah, companies have tried, uh, doing this kind of stuff. Like, like you have this Salesforce which, I don't know, fired most of their customer service, uh, reps and then six months later or three months later they realized, "Oh, it's not working that, that good and we have to hire them back." And, um, because, because when you call, when you call a number for, for example, a support line and you hear this, this like a robot or something like a, like an answering machine, you realize you, you, you, you know instantly, "Okay, wait, I'm not talking to a real person." Okay, then you press some buttons and so on and most of the times if you're not able to describe to someone the problem which you have and what you want to receive in order to be resolved, it's not going to be a good experience because it's just not. It's just not. Uh-
kubista
But let's say now you are working with your app and like do you feel the need that, uh, like l-let's say you, you are doing well, you get, you get some clients, you get some first users, you get some like revenue and so do you think you need or you have, you have some feel that you will need to hire some people that maybe that or you need, you feel that there is a work, there is so much work that you need to hire some people instead of like training some AI or trying some a-agent or something?
Valentin
Well, I guess the wet dream of all of this, uh, one-man show, uh, like entrepreneurs and developers is to be able to do it all o-on their own right and not, not hire anyone and have it as lean as possible and be able to do it with agents and so on. But it depends what you actually do, right? Be- because if you have, let's, let's say a product or software, whatever, which is not needing any kind of, uh, interaction with, uh, with, um, let's say with people from a point of view of calls or images or whatever, then you might be able to, to get away with only you or with a small number of people. But just taking, uh, just, just keep in mind the fact that when you interact with such of this, with software like this, like with Google or Facebook, if you have an issue, and if you ever tried contacting Google or Facebook, it's awful. You were never able to do it. You end up into this one thousand forms, one thousand, uh, contact forms, whatever, and you're never able to chat with someone or call somewhere because they make it impossible for you to reach them, and it never works out. It never works out.
kubista
I mean, yeah, also from my, like, experience, if the customer relation or customer support is automated, and as soon as I see the bot, I honestly, like, I've... I can see the, like, some need for this that, uh, maybe a lot of queries, a lot of questions get repeated. So maybe there is-
Valentin
Yes
kubista
... a need for the, the au- au- automation. But still, also when I call someone or some organization and I hear the bot, I get like I get, like, not disgusted, but I get like, "Why? Why?"
Valentin
Yes.
kubista
I would like to hear a person.
Valentin
Because you can foresee from the start that, oh, I'm going to have a hard time to get what I need, right? You, you know that basic- based on experience and the... But there was also this, this kind of, uh, mentality that, like at the beginning when this AI wave started, that customer service will be the first ones which they get-
kubista
Yeah. Yeah
Valentin
... kicked off and, and they will be replaced and so on, which was of course stupid, and it's not the way it actually works out. But you see, people were under the impression that because, okay, a customer service rep is somehow not so important or whatever they were thinking, then they were, "Okay, we can just put some AI in front and that's it, and we don't care, and customers will also not care," which is f- not, not true. This is totally the, the, like the, the opposite, yeah. So yeah, I, I still believe that if you need people, then you will need people. If you don't have to work with people, you can put some AI in, that's it.
kubista
But then it can be that it'll be like all the traffic or maybe all the payments or all the interaction will be done through agents. So maybe it's-
Valentin
Yeah
kubista
... it, it is the future.
Valentin
Yeah. It, it might be, and to be honest, I believe there will be also a lot of benefits from this. W- You don't need, let's say, someone to be touching all of these, uh, items or actions. But, uh, yeah, some of them are making sense, some of them not. And yeah, we shall see what the future brings. Yeah. I think most of the mistake here, mistakes here happen because people wanna go, you know, the, the bright-eyed manager that wants to make name in his company and he says, "Okay, we replace this whole team with AI."
kubista
Yeah.
Valentin
Right? And, and that's a drastic approach that obviously, um, it's too soon. It's not going to work.
kubista
Yeah.
Valentin
Right now, what's, what's reasonable from fr- at least what I think is reasonable, is pushing everyone to start working with AI.
kubista
Yeah.
Valentin
And doing it, like, step by step, right? Right now, AI can replace, uh, uh, me editing, uh, horrible, huge YAML Kubernetes files. I tell it what I need, and it does it, right? I just scroll through it and say, "Okay, looks good." Then the next step comes, right? And I think it's, it, it, it won't be like, uh, "Hey, you dear product, uh, owner, you are now gone," and here we will be, uh, th- there will be AI which we'll call Paul, and this Paul will do everything. He will, uh, he will ask the customer for, for demands. He knows the, the business, uh, because he's an AI. He of course he knows, and he will create Jira tickets for the developers, and, uh, developers will be n- next agents and tho- tho- those will be, they have different names. They will also be AIs. It won't be that. All right? But I think it's, uh, it's much more gradual, and it's much more, um, devious in, in that part that it's not-
kubista
Mm
Valentin
... you don't see it immediately, right? You see that, okay, this, what I did before by hand, is already being done by AI. And you think it's just a tool. It's just a tool like any other tools that I used before. I don't have to do calculation on hand. I have the calculator. I don't have to draw something on, on the paper. I have Excel, right? It's just a tool. But I think that there is a difference between those tools and using AI as a tool because I think the, the things that AI can do are, are much larger. It's, it, it won't stop at the thing that you are using it for. Uh, uh, I mean, it will stop if you want it, right? But, but the company that doesn't want it, the-
kubista
Yeah
Valentin
... the,
it's, it's much better than you because it can work twenty four/seven. It's always in a good mood, or it's always in the bad mood, whatever you want, right? And, um, yeah, it, it, it obviously has issues. It obviously has hallucination. It can do horrible stuff. It can remove a database just as a human can, right? And, um, if, if, for example, I like, really like one benchmark that I, that I like to follow, and that's how long can an AI do some human-related task or some human task before the error rate reaches 50%, right? So you are like, yeah, he will break it or not, 50/50, right? And it's, it's always, always growing that, that time, right? And it's, uh, b- before it was like, I don't know, half an hour, right? And it was immediately a bad, don't do it at all. But it's, uh, it's growing and it, and if it don't, it of course could be
join3r
We limit, we, we reach a point at which AI won't get better anymore. Let's say this is the point right now where we are now. This is the best we get, right? There will, there will be side grades, stuff like that, that there will be no more upgrades. You still are looking at the huge workforce displacement, at, at least in my opinion. And it's not because ri-right now AI is not used everywhere where it could be used because it's too dumb or it cannot do that. It's because processes are not there yet, our thinking is not there yet, business processes, right? The, the, all this stuff wa-was built for humans, not for, not for AIs. And you can see it from, from really small companies that are one person, one people jobs that are e-essentially doing agentic stuff, right? And the, and the employees, like they are-- It's, there is a software called, uh, Gastown, Gastown. And, and they can use stuff like this to essentially create skeleton companies with only them and some, some, some really optimistic people are not even the CEOs of their own companies, right? They are one point below that. AI is the CEO, right? And, yeah, I'm, I'm really, I'm really pessimistic about, about that stuff, and I'm really afraid about my job and what I will do in, in 10 years. I'm not afraid that I won't have a job. I think that the, the, the value of my job will diminish to such a point that my salary will be, will be much lower, right? If I will be employ-em-employable at all. And, and that's, that's something that I'm really afraid of. Maybe it's not 10 years, maybe it's 15 years, but I plan to live more than 15 years still, so, eh.
Valentin
Mm. But I, I, I guess it's also like, like this saying, like boiling the, the frog, right? Like you, like, yeah, like maybe we don't realize how bad it is going to get or how much better it's going to get. But at the same time, although it's going to be, let's say, affecting us, like replacing us, like at, at the same time, does it mean that it's going to be worse for the, for the topic of, let's say, IT or for the topic of software? Like, is this going to be better for the, uh, for the greater goal of having IT and having software, or is this going to be worse? Because that's what maybe matters, right? Because if it's going to be replacing all of us and it's going to be really good from, for the, for the, for the, yeah, for the whole IT field, then maybe it's okay. Why not? Maybe, maybe it's okay. Now, of course, as I said, I totally understand losing the job part because I would also be affected and so on. Um, but then it will, someone, I say lightly someone, will have to come up with some plan or some ideas what we could all do and what we could all work, yeah.
kubista
If I can like make a joke, like I see you guys are worrying that you are going to lose your job and I don't have any job, so, uh
Valentin
So you're happy actually. So you're actually happy.
kubista
Yeah.
join3r
You are ahead. You are already ahead.
kubista
Yeah. I am, I am behind, but I am actually ahead.
Valentin
Um, so you're the future .
join3r
These, these new guys-
kubista
I am a trendsetter
join3r
... Valentin, it's, uh, they are, they are stealing our thunder. They are too ahead.
Valentin
Yes.
kubista
No, but, uh, no, no, no, but, uh, if I, if, if I can also like add, I don't know how long this podcast is going to be. It's already two hours. I think it's the longest podcast in, uh-
Valentin
Yeah, probably
kubista
... joint history, and I have like something to do in 15 minutes. But if I can-
Valentin
Yeah
kubista
... uh, like say, say something, uh, that people say or people say that Warren Buffett is the Oracle of Omaha in investing. So I think Jensen Huang would be some oracle in AI, AI stuff because what he said in one interview that you should be asking as a developer two questions. That one is how can you make AI better at your job, and also how can you use AI to be better at your job. So there will be maybe like two, two ways to... How can you make AI to make you better at your job?
Valentin
Yeah.
kubista
So it would be maybe like two ways.
Valentin
Well, I guess this, this, this, these smart guys are always coming up with these cool phrases, which of course makes everyone else just hate them more. And they s- they sound smart for the shareholders because, yeah, they make them a lot of money anyway. They can say whatever shit they want. But for you, like exactly, you said you have no job. Did you ever had a job?
kubista
No, I Yeah, yeah, I, I-
Valentin
Okay. Okay
kubista
... this sounds like I am homeless or something . No, no, I, I just-
Valentin
No, no, no, no. No, no
kubista
... I just finished like a university and I, I had couple of interviews, and I have interview, one interview in 15 minutes, so that's why I-
Valentin
Okay.
kubista
I am-
Valentin
Then good luck for the interview.
kubista
Yeah, yeah.
Valentin
And second is don't worry about this AI for now. Worry about the interview and have the idea in mind that if they called you, they either just want to call you so they can make the, the headcount, uh, like, uh, to, to count the people or probably the other reason, they are probably interested into, in, in knowing you and probably hire you, yeah. So don't worry about us two old geezers because we are anyway too old, uh, let's say to, to follow from the professional point of view now because our past is, uh, not relevant in this, let's say, new times. But I wouldn't worry about losing or not having a job, um, for now, yeah. Because you-
kubista
But the-
Valentin
... es-especially you're going to probably work on something much more interesting than, than we did. Yeah.
kubista
Because like w- if, if I can like leak some information, like this is the job because I applied to multiple, I had a few interviews, and this opportunity is called something like forward deployed engineer. So what you are going to do is you are going to work on customer-facing solutions. And what it seems that you will be just talking with the customers, and you will see how can you deploy some agents or some, uh, I don't know, like some AI system to help the customer. So this is what I-
Valentin
Interesting
kubista
... like, this is what I think is the future, and this is what I would also like to do. So it will not be-
Valentin
Okay
kubista
... a lot of code. It will be maybe just doing some API, some integration, and some agent deployment.
Valentin
Okay. Okay. Yeah. So- So as I said, uh, worry about, let's say, what you will work on and less worry about, uh, let's say losing the job. At this, at, at this point in time, you don't care anyway about, let's say, losing the job, uh, because you will want actually to lose the job after some time or move on your own. That, that would be the best, and it's okay. It's totally fine. Yeah. So big boss, I guess you will take us out.
kubista
Yeah. Um-
join3r
I will take you behind the Bob Barn, bar and
Valentin
The, the pub-
join3r
Okay
Valentin
... and show me, show me, show us when you have time this setup because I, I need it actually, so I want to start using it. Yeah.
join3r
The, the trouble is with-- I already had something planned because I was, um, I was asked, uh, about that a lot, but I was planning to do, to do it in Slovak. You said you know Slovak, so is that okay?
Valentin
I will just ChatGPT it- ... and it's going to be okay.
join3r
That's, that, uh, that's, that-that's great. Okay.
Valentin
But-
join3r
Okay
Valentin
... but, uh, but yeah, but you don't need to do it as a, maybe as a video if it's not really needed. You can, uh, give me some Git repo, and I will take care of it.
join3r
Uh, we can, we can definitely have some work. Uh, how, how do we call it in the, in the Discord? What you do? Work... I, I, I don't know how we call it.
Valentin
Some work session.
join3r
But some work sessions-
kubista
Work flows?
join3r
... or something like that. I don't know. I, I don't know. Uh, I j- I just made it up. We never did it on Discord anyway.
Um, uh, so we can, we can do something like that, but, uh, there will definitely be something in the, in the podcast itself, like from zero to hero and yeah, the, the general overview and-
Valentin
Of course.
kubista
Zero to joiner. Kubista to joiner.
join3r
No, no, no. From, from zero to kubista because he will be the new AI-
Valentin
The future
join3r
... in the, the new company. In, in, in 15 minutes, he will be the head of AI.
Valentin
Ah.
join3r
Okay.
Valentin
One, one last advice. Ask for as much as possible money from the start because you will never get the chance to, uh, get more in the future. So yeah, so keep this in mind.
kubista
Okay.
join3r
With, with the, with the current trouble of, of, of getting money, just, just don't ask for money at all.
Just say, "Yeah, yeah, I just want my-"
kubista
Pay me bit in Bitcoin.
join3r
Yeah, yeah. Could I pay you a little bit? I could send you 100 euros. I don't have much more every month.
Valentin
Yeah. Yeah, true, true.
join3r
Okay. Okay, guys.
Valentin
See you then.
join3r
I really have to go. See ya. Bye.
kubista
Yeah, me too.
Valentin
See you.
kubista
Bye-bye.

