#52 Can AI-Generated Voices Fool Us? Insights from 'De Mol' TV Show

Speaker 1: 0:03

you have taste in a way that's meaningful to software people hello, I'm bill gates I would.

Speaker 3: 0:12

I would recommend uh typescript. Yeah, it writes a lot of code for me and usually it's slightly wrong. I'm reminded, incidentally, of rust. This almost makes me happy that I didn't become a supermodel. Cooper and Netties. Well, I'm sorry guys, I don't know what's going on.

Speaker 1: 0:34

Thank you for the opportunity to speak to you today about large neural networks. It's really an honor to be here. Rust Data Topics.

Speaker 4: 0:42

Welcome to the Data Topics Podcast.

Speaker 6: 1:06

Welcome to the Data Topics Podcast. To restart, that's fine, that's fine, sorry, we have exactly.

Speaker 5: 1:09

We have exactly zero, zero viewers. It's for the people that re-watch. Now people can join, tell your friends yeah ready now.

Speaker 6: 1:12

We started muted that's because, uh uh, we lack alex today yeah, maybe, uh, yeah you're gonna notice, you're gonna notice today where it's.

Speaker 5: 1:23

The quality has decreased a bit, so we apologize in advance.

Speaker 6: 1:25

If something goes wrong, it's because Alex, it's Alex, let's retry.

Speaker 5: 1:29

All right, let's do it. You have taste. In a way that's meaningful to software people. Now they hear it again. Hello, I'm Bill Gates.

Speaker 3: 1:43

I would recommend TypeScript. Yeah, it writes a lot of code for me and usually it's slightly wrong. I'm reminded, incidentally, of Rust here Rust, this almost makes me happy that I didn't become a supermodel. Cooper and Netties. Well, I'm sorry guys, I don't know what's going on.

Speaker 1: 2:05

Thank you for the opportunity to speak to you today about large neural networks. It's really an honor to be here. Rust Data topics.

Speaker 4: 2:12

Welcome to the data topics podcast.

Speaker 5: 2:17

Hello and welcome to Data Topics Unplugged, your casual corner of the web where we discuss what's new in data every week, from moles to human pins, anything goes. We are live on YouTube, linkedin, x, twitch, you name it. Check us out there. Feel free to leave your comment or question or, if you can hear us, feel free to also let us know. Shout out to Maria Constantinescu. Whoever that is, also let us know. So, thanks, shout out to maria constantinesco, whoever that is. Uh for the for letting us know. Um, my name is marillo. I'll be hosting you today, joined by the one and only bart hello, and we have a very small before the special guest we are, I think, if you I don't know how much you heard of that, but Alex is not with us today. So if you see Bart stretching to reach the soundboard or if you notice any other technical issues, you know why it's gonna go super smooth well, let's see from now on

Speaker 5: 3:18

if this wasn't live streamed, maybe I would say that, and then we could just edit everything. But now, let's see, the stakes are higher now and we're joined. We have a very maybe I would say that, and then we can just edit everything. But no, let's see. Let's see, the stakes are higher now and we're joined. We have a very special guest. We have Calisthenics superhuman human flag on his spare time and when he's not doing that, he teaches machines. You know we have Sene, so can I get a round of applause for Sene? I'm going to ask Bart to do a lot of these. All right, sene.

Speaker 4: 3:53

How are you?

Speaker 5: 3:54

Good Thanks for having me. Thanks for coming. Would you want to share?

Speaker 4: 3:58

a few words about yourself. Well, yeah, I'm Sene. I've been at Data Roots for two and a half years, I think, now working as a machine learning engineer here, and then, yeah, some calisthenics in my free time. You're right about that.

Speaker 6: 4:12

What is calisthenics? People don't know.

Speaker 4: 4:14

Calisthenics is a bit like fitness, but with your own body weight, so it's like pull-ups, push-ups, all that kind of stuff. For me, it's a fun way of doing fitness.

Speaker 6: 4:27

I think it looks very cool. Yeah, it looks very impressive.

Speaker 4: 4:31

Yeah, I can't do that. What is doing right there?

Speaker 3: 4:34

Not yet. That's like a planche.

Speaker 4: 4:37

It's like one of the hardest things to do. I was practicing on it yesterday actually.

Speaker 5: 4:41

Oh really.

Speaker 4: 4:41

I'm not nearly close.

Speaker 5: 4:43

I can do the plank on the right, sometimes on a good day that's like one percent of the yeah it looks very impressive, like calisthenics yeah, there's a lot of like stuff and everyone that you see is like ripped.

Speaker 6: 4:56

Yeah, yeah, I also have the feeling like when I think about calisthenics.

Speaker 5: 4:59

Like you hang with some friends in the park, yeah, right all ripped and like, yeah, just hanging from just off from equipment, you're just chilling you know, but there is a park here in Leuven, like at Sportoval.

Speaker 4: 5:09

It's like five minutes from the office.

Speaker 5: 5:15

And there you literally hang from stuff.

Speaker 5: 5:17

Sure, yeah, actually I asked Bart if we can install a poll here for a Senate to do a human flag, but Bart said no, I do a human flag, but bart said no, so I did not say no. So that's uh. If anyone is like, ah, we, we want the human flag, we want, you have to send it. Send an email to bart. Yeah, at dataio, um, to get the poll install hidden at our studio I feel like studio is a bit of a strong word, but our podcast studio our table with a few mics around yeah, our, our corner, our corner.

Speaker 5: 5:43

Yeah, did you know? Actually, bart, I think Sander was the first person I met from Data Roots. Yeah, that's true. Oh, wow. Actually there are a few asterisks here.

Speaker 6: 5:56

First person you met from Data Roots. Yeah, he wasn't from Data Roots yet.

Speaker 5: 5:59

Yeah, yeah, he wasn he wasn't from data roots. But when I was, when I was a student, I used to work out in the gym in the K-11 used to some time ago, and Sander was there and I was like man, this guy huge, like all the weights down, you know like like emptying the machine basically wow, and I was like not true, but yeah, man it's nice to yeah, man, so you have history yeah, we do, yeah, I

Speaker 5: 6:26

even, I think maybe even when they was hey can, can, I can use it. How many, how many more sets you have?

Speaker 4: 6:30

so we bonded you know, you're uh gym bros exactly you flex your packs at each other exactly yeah, and I think I actually know your loving buddy when you joined as a student here, like thomas ah, okay, you knew him.

Speaker 5: 6:43

Yeah, do you?

Speaker 6: 6:45

have uh interesting gossip are we no?

Speaker 5: 6:49

no, it's not maybe we can talk after um, but yeah, yeah, very true, very true, and uh, more than that, cine um you're quite famous. No, starting yeah what is I mean? Did you know this part?

Speaker 6: 7:15

I heard a few things over the last weeks what did you hear? I heard there is this famous television show series the Mole, the Mall in English. Let's see if I can put it up here as well. Maybe I'm not the best place to give a description of the Mall.

Speaker 5: 7:34

But yeah, tell us, Sander, what is the Mall?

Speaker 4: 7:38

Well, I can explain.

Speaker 4: 7:39

So, it's a popular TV show in Flanders and there are always, each year there are 10 participants unknown people, so not like famous Belgians and one of them is the mole, and over the course of a couple of weeks, they do all kinds of challenges. They try to raise money by completing those challenges, but the mole, the one person among them he tries to sabotage everything without getting discovered, raise money by completing those challenges. But the mole, the one person among them, um, is um, yeah, he tries to sabotage everything without getting discovered. Um, and in this year's uh, yeah, this year they wanted to do something with ai, so they reached out to us, um, and their idea was to, um, make a sample of all, uh, fake voices of the participants, fake AI voices, and blend that together with the actual voice of the mole and then, yeah, show that or make one of the participants listen to that sample that whole mix. And the idea was, of course, to make the mole indistinguishable from the rest, but that's, yeah, a tough thing to do, of course it's kind of like that.

Speaker 5: 8:49

The game that was very popular in the pandemic, the, is it imposter? No, the one that's like the sus, there's always one that's like infiltrated and it's like it's like on the your smartphone.

Speaker 4: 9:00

I think so yeah, I never played it, but I know well, to be honest, I the malls, I never really yeah, but it's popular in flanders, I think it's every episode has like 1.5 million viewers or something. So for belgium that's pretty. Yeah, it's quite a lot, it's like 90 of belgium, but this is.

Speaker 5: 9:20

this is really cool and you talked about so. You mentioned AI-generated voice, yeah, and last year at the RootsConf you also presented that.

Speaker 4: 9:30

Yeah, true, by that time we actually already. So I did that together with my colleague, sophie, who also had a big part in the mold. So shout out to Sophie.

Speaker 3: 9:38

Yeah.

Speaker 5: 9:39

Couldn't be easier.

Speaker 4: 9:40

Yeah, and yeah, we did a presentation at the RootsConf Afterwards. We also joined on this podcast for like 5 minutes, I think so he's a podcast veteran friend of the show friend of the show. Yeah, thanks, yeah, so we we have some experience with that now very cool, very cool by some people Vitaly and Issam.

Speaker 5: 10:03

Vitaly is Parachello and Issam Sadik. They uh, the name of the game is Among Us, so thanks for the support. Um, pretty cool. And you also were you on TV as well, or, yeah, jonas was on TV, but I felt like what's you know? He's just trying to steal the spotlight.

Speaker 4: 10:22

You know what I'm saying reframe from my commenting I don't confirm nor deny um, yeah, it was in a cafe, the mall, so it's like um, a show behind the actual show where they yeah, some famous people in belgium discuss what happened in that episode and also the one, the one candidate that was eliminated from that episode uh, he or she joins at the table and just talks a bit about the mall, and there we got like two minutes of uh time to talk about what we did so you generated deep fakes, deep fake voices from all the contestants, basically from all of them.

Speaker 4: 11:00

Yeah, including for the mall, because also we weren't allowed to know who the mole was interesting and top secret.

Speaker 6: 11:08

How uh easy is that these days?

Speaker 4: 11:12

well, um, like the voice characteristics, or like the sound of the way your voice sounds like the mannerism yeah, let's say like the I don't know the exact term, but um, the color, the voice color or tone of voice, that's um quite easy to reproduce these days. But like for, in this particular case, for the mole, uh, we had a lot of different accents from Flanders, so you have a lot of different accents, like West Flemish, limburgs, antwerps, and it's really difficult to go from text to speech in a specific dialect. So the way we approached this was by letting other people say the sentences and then transform their voice into the voice of the candidate we wanted to recreate. Because which?

Speaker 4: 12:11

is voice transfer is that right now voice transfer actually. Yeah, so that way you can, um, it's easier to get the dialect right, to get the intonation right, to get the way of speaking actually like, how would this person sound if you would have the local dialect? Yeah, okay so for each, for each candidate uh. Someone else spoke in the specific fragments, so one person did it in west flemish, another one did it in, in gans um, and then each of those was transformed to the uh, the target candidates for that dialect.

Speaker 5: 12:44

It's pretty cool, maybe. Uh, vitale sparacello again. Uh, does this mean that you knew in advance who the mole was?

Speaker 4: 12:52

no, we did not know, because we um created fake voices for all of the candidates, so including for the mole. Then we delivered all those samples to the production team of the mall and they, um, they made the mix and they sent it back to us. So if we really wanted to know who it was, we could compare the the samples we gave them with the ones we received back and then we could have, in theory, uh, deducted who the mall was, because in the ones we, yeah, we got back, the actual voice of the mall was used to they differed a bit from what we gave them. Um, but we didn't do that and we, we did, uh, we tried out some tests, uh, we listened to it, we uh, yeah, and also some technical tests but the result was always a little bit different, depending on what kind of test with it, so it's not like we're 100 sure of who the mole was.

Speaker 4: 13:56

And also you.

Speaker 5: 13:57

There's a blog uh article that you've uh wrote up, says Sophie, but uh mainly work of Sophie the blog post?

Speaker 6: 14:00

no, no, she's not here. She's not here, it's fine um, and it really was a team effort from uh usph, yeah, yeah indeed, yeah, I think so for people that are interested.

Speaker 5: 14:08

I think you talk a bit more about the technical stuff as well. Tortoise tts is like a framework, and give you more in the technical degree right yeah, so here we explain.

Speaker 4: 14:17

At first we tried the text-to-speech approach, where we start from like just written text and try to transform it into the target speech, but, as I said for it, you had to reach different dialects. With actually less than one hour of data per person was just not possible.

Speaker 4: 14:33

Um, actually we got really good results for the host of the show, for Jill the Costa um because for him he speaks like very general Dutch, uh, or like Flemish, uh, we had a lot of data for him, we had also good transcriptions for him, so for him that approach did work, but not for the candidates, because, yeah, they don't speak like nice flemish maybe a question as well.

Speaker 5: 15:00

You're talking about flemish. Is there a big difference from, maybe, research or from your experience between English and Flemish for these things, or the language does it impact a lot of quality?

Speaker 4: 15:11

Yeah, there certainly are some models that are purely trained on English. Some of them are multilingual, but then often when it's multilingual for Dutch, you have, like the Netherlands kind of Dutch, not a Flemish Dutch. So in terms of quality it does have an impact okay. I think I actually saw for a GPT-4 row right now. There was also I saw a clip of someone asking it to speak Flemish, and it was just Dutch.

Speaker 5: 15:39

Oh really so maybe should we play something very quickly here, just to kind of see the how it looks like a little sneak peek actually I'm not sure if we're gonna be.

Speaker 6: 15:50

You need to unmute yourself here, right?

Speaker 5: 15:52

let's see. Let's see, this is. This is gills.

Speaker 4: 15:56

That's his name jules jill, yeah, so that one was pretty good, I think let's see and it is the finale scene from end tests.

Speaker 5: 16:03

Sounds very like radio voice, you know.

Speaker 4: 16:06

You know, yeah, he's a presenter, so all the training data we had was also like in a presenting style. So that's.

Speaker 6: 16:12

Yeah yeah.

Speaker 5: 16:14

And this is I don't know. If you want to jump in.

Speaker 6: 16:17

No, I just was wondering like when did you actually build this, because it's been a long course.

Speaker 4: 16:21

I was I think it was we started in in september and then we finished in october, while they were actually already on the on their trip so it was really last minute. Um, there was a possibility that wasn't going to be shown in the show because until like one week before the uh, the actual when they were going to use it uh, we didn't have anything that was working.

Speaker 6: 16:43

So it was kind of last minute and you. So you built this, let's say, roughly a half year ago. Like, yeah, how is it a fast evolving field? Like, would you tackle it the same way? Would you do differently today?

Speaker 4: 16:56

I think, like the voice conversion, that approach I would definitely take again, because I don't think the the dialect problem is solved for for text-to-speech right now. And then yeah, one, yeah, I don't know exactly the state of the art right now in detecting fake voices, what models are available for that right now. But yeah, one precaution was taken in the in this in the mall as well was to add a kind of a filter on top of the total mix to make it a bit more, to make it harder to, uh, distinguish the actual voice for people analyzing at home.

Speaker 5: 17:36

Yes, indeed, yeah or uh because even, even though I'm sure data scientist yeah but can you like, when you say, add a filter on top of the voice, yeah, I mean like yeah, so actually they played it through, uh, through a car, so they uh, so it was like yeah, uh, how do you say like a cell phone filter?

Speaker 4: 17:55

like when someone is speaking to you through a phone it's a bit distorted yeah and that was the same with the samples we uh I see.

Speaker 6: 18:01

So when they say filter, you mean more like, uh like they use an external speaker and record that yeah, yeah, no, yeah, just because so that's kind of what I thought pictures that get re-shared and get more pixely every time yeah, yeah, yeah we take a screenshot of a screenshot and then you know, or when you take a picture of the picture and then it's like, yeah, no, but yeah, just to be clear as well for the people that maybe think it's a bit yeah, yeah, add a filter is a bit, yeah, cool, but like, I think also there is a bit of um.

Speaker 5: 18:32

I think, yeah, I'm segwaying a bit to deep fakes in general, and now we're talking about uh voice right but there is also a bit of a let let's say opportunity for bad agents. Let's say yeah, absolutely.

Speaker 6: 18:49

Right Phishing will never be the same again.

Speaker 3: 18:52

Exactly.

Speaker 5: 18:53

So what are some concerns there? Or are there any measures? Are there any things to safeguard these things?

Speaker 4: 19:03

Well, yeah, I think that's a big question, um, because, yeah, for people with a lot of audio data available online, basically anyone can recreate their voice almost perfectly right now. So it uh, and it becomes very convincible, especially in english and for people that speak in a in a normal dialect, you can almost perfectly recreate their voice and, in theory, you should always ask consent, which is what we did as well, just to be clear. But, yeah, it's a challenge going forward.

Speaker 5: 19:41

Yeah, and I also think so. This is not not I haven't read this thoroughly, but apparently scarlett johnson also. There's a.

Speaker 4: 19:49

There's a bit of friction with open ai yeah, so the in the, the gpd 4.0 model, um, there are like five new voices and um, yeah, scarlett johnson um has the opinion that one of those is freakishly close to her own voice, and it's a bit special because OpenAI asked her to do to record for one of those voices.

Speaker 6: 20:16

She's the actress in Her, which is a movie about an.

Speaker 4: 20:22

AI, which is played by her.

Speaker 6: 20:23

Yeah, indeed, and there is one voice that is pretty close to hers, a movie about an AI, which is played by her.

Speaker 4: 20:25

Yeah, indeed, and there is one voice that is pretty close to hers. Openai claims that it was.

Speaker 6: 20:30

It is a voice of a voice actress, but yeah, they want to keep guarantee her anonymity, so it's yeah but leading up to the release they contacted Scarlett Janssen to ask a couple of days which she may be okay with it and I think, but even before I think, then a couple days before again, I think it needs to still said no weeks or months before then, a couple of days before, and then they released gp24.

Speaker 4: 20:54

But I I saw today actually a um, an article of the washington post and they say they saw some evidence that indeed they contacted someone else, that there is another voice actress who yeah, they're saying now that there that some records show that there is actually a voice actress that was hired to do this, but it's all very vacant, like it's also it was released. Some altman uh tweeted her like just the word, her like it's very uh yeah, it's a bit shady yeah, I was trying to open up here.

Speaker 5: 21:25

That's why I was struggling a bit. Uh, because there's a paywall, so, uh, but that's the. I will link the articles there.

Speaker 6: 21:31

So for anyone that is curious, but uh so it sounds without any proof that there was actually an actual or some voice actress. Like it sounds very shady, yeah yeah, I think it is.

Speaker 5: 21:43

Yeah, it is. Uh, it is very shady. The likelihood is that there is someone probably out there that speaks very similar to me, right, yeah? If they really look for it especially if someone like scholar johansson is like okay, you have a like a voice double or something you know like uh yeah, but even there it's like if let us assume that there is a voice actress that they explicitly mimicked Scarlett Johansson. Exactly.

Speaker 6: 22:07

That's also phishing. Well, it's phishing. I think there's also a legal precedent that that is not allowed there was this case a long time ago, motor company, I want to say Toyota, I'm not sure they did a song advertisement, a song where the artist that they had pictured didn't want to do it and they had someone mimic her no, really and it got the artist got to the cuts of the the.

Speaker 5: 22:33

It was ruled in her favor which I think I mean do you really make sense, right?

Speaker 6: 22:40

it intuitively makes sense if it's explicitly to mimic this individual person, scarlett johnson, here like and everything like very much seems like some moment tweeting out her when it gets released, like it's a bit uh, it's hard to say that you're not trying to mimic it, right, and it's like even, yeah, like we reached out to her like a few times and like we really think about like her, the, the movie and her voice. Yeah, it's like really like when it ended up, like it's all, like like yeah, sorry, I guess there's some similarity there.

Speaker 6: 23:10

Yeah, it's crazy.

Speaker 5: 23:11

Everything can happen and that's the thing is like uh, people that just listen to the voice probably didn't even like. They immediately thought of her like, oh, scarlett, you're handsome right, so it's like.

Speaker 4: 23:23

But I've also heard people say that they don't think it sounds. Has your hints, right? So it's like. But I've also heard people say that they don't think it sounds like her at all, so it's yeah I can't judge at all. I have no clue I just heard two different fragments, but I think what the discussion on this is.

Speaker 6: 23:41

Also a bit like the broader ethical debate. Like we where OpenAI started from, we're the good player, we're an open research foundation and we want to do this for the greater good. And if you even don't have correct actions on something like this, you can very much question the whole broader ethical framework around it. Very true.

Speaker 5: 24:03

Very true. I think there was actually one criticism. I heard from Elon Musk that he I don't know if it was one of the founders, or he invested in OpenAI or something. Is this?

Speaker 6: 24:14

He invested. He was very early as an investor, Very early right.

Speaker 5: 24:16

And I remember I think I saw a tweet from him that he said like yeah, I invested because I really believed in open standards and all these things, and now the company is against everything that we started.

Speaker 6: 24:32

Not sure if that is the best source of information these days.

Speaker 5: 24:37

But why I'm bringing it up? Because there's actually XAI right, and I'm not sure how timely this is I think it's relatively timely yeah, 26 of may.

Speaker 6: 24:50

Indeed very timely that xai raises six billion from valor a16z and sequoia, which are big funds right a16z they also have what it's, what the name stands for and recent horowitz is like cuban netty's, like the 16 is 16 characters it's like something from his cuber and netty's but they also do.

Speaker 5: 25:18

They have a like a podcast or something as well they have those as well they have a podcast 16 news.

Speaker 6: 25:25

Yeah, okay, what is's a huge fund, huh.

Speaker 3: 25:28

It's bigger than the podcast, but what is this so?

Speaker 5: 25:34

XAI always exists, maybe. What is this about, bart? Just any more context you want to add?

Speaker 6: 25:40

It's an AI startup from Elon Musk. It raised 6 billion in the funding round, which is big, and I'm not sure at what valuation actually. Maybe we can see that as well.

Speaker 4: 25:53

I think 18 billion, according.

Speaker 6: 25:55

To.

Speaker 4: 25:55

Musk himself, I think.

Speaker 6: 25:56

That's crazy and well, I think it's Musk's attempt to have a competitor to the likes of OpenAI and Anthropic and stuff like that. He marked it very much as this is going to be the answer to get the truth on the table. This is going to avoid all the fake news and hallucination and this will get the truth to the people. I think that is Musk's rhetoric on this, but it's very much in Musk style that it's brought and everybody knows where we're at with X now. Right, yeah, but this is Xai.

Speaker 5: 26:41

Ah, it's Xai.

Speaker 6: 26:42

Yeah, this is Xai.

Speaker 5: 26:43

Man, I'm going to buy all the X domains, just just just have it, you know just like you can make some money off of it just saying and I think like what it looks.

Speaker 6: 26:57

So x, the twitter shitter, has grok, which is their llm model, and I think this is also a bit of an attempt to bring that ability to a broader audience than not X user, like that.

Speaker 5: 27:14

And it is tied to X like the Twitter, right Like the company or.

Speaker 6: 27:18

I don't know if it's actually the same model, but I would assume that they're sharing some resources, yeah.

Speaker 5: 27:25

Do you think, do you have any high hopes, that this actually will be different from the big players? Do you think it's going to add something new to the table, to this space of models and all these things?

Speaker 6: 27:37

There's going to be even more news on Musk. I think that is where we can assume from this I don't know, man, I'm tired of all the Musk news. Well, let's see, I don't think that, man.

Speaker 5: 27:49

I'm tired of all the Musk news. Well, let's see. I feel like let's see.

Speaker 6: 27:52

I don't think that this will make the change. I think so. But I'm surprised that there are so huge funds that clearly do believe in this, and I'm putting 6 billion on the table too.

Speaker 4: 28:04

Yeah, and also never he's able to create big companies.

Speaker 5: 28:07

Maybe they're like this one will take up as well, it's ai, yeah, yeah but I don't know I just don't know his fascination with the letter x. Like space x.

Speaker 6: 28:22

Change twitter to x, now xai like jesus, maybe you can buy x or space x, that is this dot space.

Speaker 5: 28:26

Is that a thing?

Speaker 6: 28:27

yeah, it's a thing. Yeah, um, there is a question that we have uh from the audience, uh one of the friends of the show who is this guy?

Speaker 5: 28:40

is he vital?

Speaker 6: 28:43

mr sparatello, is there a real uh open ai competitor? The models are always better and cheaper. I think it's a very fair question.

Speaker 5: 28:55

Sennabal. What's your opinion? That's a very vulnerable question. That's a tough one. What do you think?

Speaker 4: 29:05

I think, generally speaking, I think Vital is right. Openai and their models are really good, um very easy to use, and so far I I don't think so, but it's good maybe, to have some more competition and to yeah, I fully agree.

Speaker 6: 29:21

I think the the first time I had a feeling that something is on par with uh chpd for turbo it was was with claw 3 from entropic. So the first time I had to feel like this is something that is more or less comparable, but that is an also like from the moment that there is something that is more or less comparable, like they immediately released, in this case gpt 4.0. Yeah, I like they were again miles out of the competition.

Speaker 5: 29:43

That was impressive, which is uh which is crazy, entropic, the, their models are also closed and everything.

Speaker 6: 29:49

I think so yeah, but what they don't? Well, maybe I'm not actually 100% sure. I wanted to say that they don't have any vision models, but I'm not 100% sure what I'm saying.

Speaker 5: 29:58

I never tried it For me as a user. Let's say that Anthropic is the same as GPT. Why would I choose one over the other?

Speaker 6: 30:09

Choice, maybe also the platform that you're on. If you're on AWS, anthropic makes much more sense because of their partnership. If you're on Azure, you're probably going to go for OpenAI.

Speaker 5: 30:19

So it's more like the context around it, but it's like okay. No that's fine.

Speaker 4: 30:24

Is the pricing the same? I don't know by heart.

Speaker 6: 30:28

Let's say, at a reasonable scale. Scale it's more or less same, I think, from the moment you truly scale up. I don't know. So I think what about? Uh, I like on open ai pricing, when it comes to talking via the two models, via the api, is that they always, when the their newest model comes out, it always becomes the cheapest, which is very smart because it pushes people to use it to give feedback to.

Speaker 6: 30:54

The new models, the new models. Yeah, it's not cheaper than their let's say, than 3.5, because these are very niche, small models. But GPT-4-0 today is cheaper than GPT-4-TOPO, for example.

Speaker 5: 31:16

Which is interesting, but it is like gpt 4.0 today is cheaper than gpt 4 topo, for example, which is an interesting. I think it pushes people to try the next new thing. Yeah, that's true, and I think also it's like, if you think of, pricing is a two-way door, they can always adjust it. It's like, okay, get the feedback and now. It's okay now. And also it's good to get the word out, right, I think they, they want people to always be thinking of opening eye for stuff. Yeah, it's efficient. Uh, maybe more on open ai. Then, since we're, I feel like we try to not talk about them, but just it's hard yeah, it's harder, like these days it's very difficult.

Speaker 5: 31:42

Um open ai news corp deal accelerates paid data race. This is is from the 28th, which is I didn't even say the day today, did I? It's from today? Yeah, it's from today. Wow, okay, bart, I see you.

Speaker 6: 32:01

Yeah, this is yet another paid data deal that OpenAI strikes We've seen, this time with News Corp. News Corp is the owner of titles like the Sunday Times, the Sun, the Times, the Wall Street Journal, New York Post.

Speaker 3: 32:25

The.

Speaker 6: 32:25

Australian the Daily Telegraph, a lot of stuff.

Speaker 5: 32:27

News Corp. They own all this.

Speaker 6: 32:30

Yeah, oh, wow, exactly, and it basically again gives OpenAI access to their data. Probably I haven't looked into the details, but probably it's for training purposes, to give them training data but also to do more accurate real-time queries. So if you want to get the latest news on something like whatever topic, that it can also crawl their recent indexes.

Speaker 5: 32:56

New York Times had sued OpenAI, like a while ago. I think we talked about something like this. No, did we have a resolution on that, or no?

Speaker 6: 33:05

I don't think so, but this is not New York Times here, this is New York Post. Is it New York Post? I think they are.

Speaker 5: 33:10

It's a New York Post, I thought it was a New York.

Speaker 6: 33:13

Post.

Speaker 5: 33:14

Do you think? These are ripples of that, in a way.

Speaker 6: 33:18

Because now it's like Well, I think there's definitely a reaction to that.

Speaker 5: 33:23

Because I remember we talked about back then. We were saying like, well, it's a bit, people, people do they want to bite the bullet or not?

Speaker 6: 33:30

because now there are these deals happening but if the ruling favors open, ai right, it's a bit, uh, it's a bit like a moment of tension, you know we're a bit in suspension, like if, if there is ever a ruling like that using this data for free is considered fair use, then you're not going to get any money for it. Exactly like. But this is like where there are a lot of lawsuits happening and like opening. I probably also doesn't want to spend millions in uh in legal fees yeah, and do you think this is much?

Speaker 5: 34:05

because today data, like I mean, I think it still catches my attention as well, but data is already a commercial product, right, like Google and all these things. They were always selling data. Do you see in that lens, do you see this as a very different thing? The difference? Now is that we know where the data is used right.

Speaker 6: 34:28

Yeah, I think it's a different source of revenue for these type of companies like it's uh, um, and it's a new channel where news can surface, like you can maybe search for recent news on chat, gpt, where you typically always have the notion okay, this is trained on data and it's probably going to be outdated when I search it and like this. This allows you to to get that access and hopefully also get some transparency, like where does the data come from? Quotations, these type of things is it? Well, it's new. I think it's in that context it's new. It's it's new that it's become such a big new channel. Yeah, yeah, yeah, yeah, I feel like it's new. I think in that context it's new. It's new that it's become such a big new channel.

Speaker 5: 35:11

Yeah, yeah, that's true. Yeah, yeah, I feel like it's. You know, for me it's kind of like people have been using this for that. It's become like a pattern, let's say, and now they're investing on it, so it's more robust, yeah, but I think even like they also need to do this.

Speaker 6: 35:27

Like like this is like they give their the source of their data to something like open ai. In this case, I can very much imagine a future where you go from okay, I'm gonna open the, the wall street journal, I'm gonna read the article and instead of instead of doing that tomorrow, I'm gonna go to chat gpt and gonna ask, like, what were the five major headlines of yesterday? And it's just gonna generate a text for me, hopefully based on correct underlying data, but like it's a completely different use of of of the of the channels that something like like these new papers have available?

Speaker 5: 35:59

yeah, I think also you mentioned like using chad gpt to get a overview of the news but I think also people can like okay, give me a summary, or maybe I don't want to know everything in the article, I just want to know what happened to this person or that person, right, so you can also kind of um which gives you the.

Speaker 6: 36:15

But let's see, I think you have the same thing with, uh, with social media. You have the a bit of a danger that you create this bubble around yourself. Yeah, if you start querying the news very much based on what you want to know, not not necessarily looking at the front page.

Speaker 5: 36:29

Yeah, that's sure it's like it's a bit the I mean, I think social media kind of does it a bit right, like it gets a bit magnified in your own echo chamber. Yeah, you know it's like, even if you have, even I don't know, let's say I have a very peculiar with me.

Speaker 5: 36:51

the internet provides a platform in which I can surround myself virtually with all these people and I think that everyone is like this yeah, right. So yeah, I feel like the same way that social media and like the internet and like the fake news. I think people need to be a bit re-educated. I think there's also a bit of this there, and google also is not staying behind right on these things we mentioned. They're also using AI for their search, right. I think if you Google some stuff, sometimes they give like an AI overview.

Speaker 6: 37:20

I haven't seen it yet. I don't think it's available here. I haven't seen it yet in my Google. It's probably that they released it in the US first. Maybe, but what was the…?

Speaker 5: 37:29

Why am I bringing this up? Yes, sander, why am I bringing?

Speaker 4: 37:32

this up. Yeah Well, there are some pretty funny answers that the Google search provides. It's clearly indicated that it's an experimental feature, so people should be warned, I guess. But if you Google what you should do if your cheese is not staying on your pizza and how you can make it more sticky, then there is one answer that you should mix about one-eighth of a cup of glue in with the sauce and then the non-toxic glue will work. And apparently it's based on a Reddit thread from more than 10 years ago where one specific user advises exactly that.

Speaker 6: 38:15

User called Fucksmith. Part just likes to mention that Nice detail, yeah, but I think this person is very happy now. This is like this he planted the seed 10 years ago. It had to get to google signing a deal with reddit to crawl their data, to train their model on yeah and now he's the authority on. How do you get cheese to stick to your pizza?

Speaker 5: 38:42

yeah and I think also like, so this I pull up this is something to aspire to but it's just gonna go on Reddit crazy, just like one day how do you know that this is not me? Sure Bart's like no, but the guy should be proud, really no really he should be. He did a good job, he did a. Actually he's even deleted.

Speaker 6: 39:02

No, the subreddit is deleted, I guess really right it's just deleted the user, the user that posted the right.

Speaker 4: 39:09

This is the user, the user that posted the. I see it there.

Speaker 5: 39:11

No, no the folks meet this year.

Speaker 6: 39:13

Right, but like I thought, our pizza was a you know the deleted, there is just just the user that posted the original question.

Speaker 5: 39:21

Thanks, thanks, part, just um. Yeah, so they put like. First of all, I think that the question itself is super weird. Right, my cheese lies off pizza too easily. It's like okay, uh, but also like the. This is the most upvoted uh answer, as well, yeah, probably that's.

Speaker 4: 39:41

They made a selection of like the, the best answers and only use those yeah but, also the funny ones from afar, that's that sounds smart. Huh like I have a question already, like the most upvoted answers probably want to be okay, but that's when you're going to the sarcastic.

Speaker 5: 39:55

Yeah, but that's what the internet is great, right, like there's always like sometimes, like people, they just kind of find each other. They just come up with this like outrageous thing.

Speaker 4: 40:05

So most, most comments I see that are uploaded are just really funny ones actually.

Speaker 5: 40:09

So yeah, no, I yeah. I think red is also a bit of a space for that, and there were other examples as well.

Speaker 4: 40:15

So, right, yeah, yeah, some other, but a bit less harmless, I think. Um, I saw wait, uh, someone who graduated 21 times from a university instead of once. Uh, a python is apparently a mammal now. Um, so, yeah, there was a question what mammal has the most bones? And it was like, oh, it's, uh, it's a snake, um, which is, but it's harmless, but and, yeah, it's experimental, but shows that it's not perfect yet, although, uh, what's harmless.

Speaker 4: 40:43

You can debate it yeah, yeah, I mean the python one, the the glue one. I mean I assume 99 of people knows it's. They shouldn't put glue in their pizza, but just for the one percent.

Speaker 5: 40:53

But you know, I like, I mean I will be, I'll be cautious, because I even heard like stories that people were following the gps and they drove into the ocean because the gps was saying something like that, like, even things like that is like don't cross people, yeah that's also what I was thinking with, like google.

Speaker 4: 41:11

Normally you think like okay, google, uh, if I do my search I will get some decent sources and there, if you see a non-toxic glue, yeah, well, maybe that works. I don't know.

Speaker 5: 41:20

Yeah, it's like, oh yeah like I mean, yeah, glue, of course, if you had a bit of uh details to it, I think you can make it very uh believable and uh, yeah, I see here also that a dog has played in the nba.

Speaker 6: 41:30

That's it yeah, so they just switch from there. What's the like? If I search on something, what is the highest value based on page ranks, versus, and they just switched into what is the highest upvoted on reddit.

Speaker 5: 41:42

That's what it seems like, right you know where their attorney that has come from. They're like we're not paying. New York Times we just do that, you'll be fine.

Speaker 6: 41:53

Batman is a cop. Are you both Reddit users?

Speaker 4: 41:58

not frequent, but just occasionally I scroll through.

Speaker 6: 42:00

You're pretty much do you have like? Do you post stuff on Reddit?

Speaker 5: 42:04

no, I feel like I'm too insecure.

Speaker 6: 42:07

I sometimes post stuff, but now I'm wondering like I'm posting stuff and like reddit is getting money from this from, for this from google, like it feels a bit weird are they getting money?

Speaker 4: 42:20

yeah, yeah, they close the deal, I think uh but I thought reddit data was uh free.

Speaker 5: 42:24

No, because the api there was a change, but I think I've accessed a ready data a 60 million dollar deal per year that's some money for my data for you and user fixments data yeah, but I think that yeah

Speaker 4: 42:44

yeah I was looking for one other funny example. I finally found it was was like a reporter Googled whether you could use gasoline to cook spaghetti faster and they replied no, but you can use gasoline to make a spicy spaghetti dish, and that was the recipe.

Speaker 6: 43:01

That's risky right.

Speaker 5: 43:01

Yeah, that is risky, that's risky, that is risky. So yeah, there's also here.

Speaker 6: 43:07

So before sharing the merch, but you would expect a bit more robustness testing before something like this goes live yeah, they claim it's like really rare, it almost never happens like this kind of answers yeah I know yeah, almost never, but I think that's also what is almost never for you, right? Yeah, yeah yeah, but I think uh what if your cheese slides off your pizza right like you end up there?

Speaker 5: 43:33

or like, if you're like, a young child at home warming up your pizza, and then that's true, or like you're, you're like you're a kid home alone first time. You know how I want to make spaghetti pizza.

Speaker 6: 43:44

Uh, spicy ones, yeah oh yeah, my dad has some gasoline in the garage I like it spicy.

Speaker 5: 43:50

I mean. But I mean yeah, like I think. Yeah, I mean I don't know, I think it's scary this is a bit scary. It's scary yeah it's scary, uh, but I do feel like I have this.

Speaker 6: 44:01

I share this sentiment and I think for these things the data deals that are happening with, uh, respected news outlets, like where there's a bit more fact checking, where you expect there is some actual journalism with fact checking- at least because it's also debatable these days. I think for those type of things like that is very valuable to move this forward when it comes to having trustworthy data coming out of it but I think like so.

Speaker 5: 44:26

Google has a deal with reddit open. Ai has a deal with the News Corp.

Speaker 5: 44:31

You see the difference, yeah yeah, yeah, what I'm going to say is, like you think it's because Google they're not investing as heavily as OpenAI, or do you think they have? Like, why is this? Because I mean, as like, you're in charge of Google AI search, right? Like would you do a deal with Reddit? You know what I'm saying? It's like maybe the person is not a Reddit user, right? So it's just like oh yeah, Reddit, there's a lot of people there.

Speaker 6: 44:56

Yeah, user base, like it's like Stack Overflow Q&A style.

Speaker 5: 45:00

Yeah, upvotes, exactly, it sounds like it fits up but for everything like wow, this is great, you know, it's like and I think stack overflow would be more trustworthy than reddit.

Speaker 6: 45:10

Yeah, but like but if you don't know reddit, it can look like that stack overflow for everything you know, it's like, if you think of the system, people ask questions, people reply, there are upvotes.

Speaker 5: 45:18

It's like, yeah, there are moderators as well. Yeah, this is the same thing, but the one is not know.

Speaker 6: 45:26

Yeah, I think it's also it's gonna be like if you, if you don't have a way to actually query like up-to-date sources, like knowledge that is very uh uh, like um, sensitive to evolutions, is always going to be very hard to capture in an lm, like, for example, um, if you have to depend on training data, like stable knowledge, like what, like pythagoras's theorem, like it's been, like it's been very stable for a very long time, it's probably like if you answer today what it is like, it's going to be the same answer in 30 years from now, I hope. But like who is the president of the united states? That's a very like every four years if you're, if your training data is outdated, you can't answer it correctly and then you need to have this way to for recent stuff, to to exit out and to query actual news sources to have a good answer yeah, but then I think, well, I don't know if I'm bridging a bit much, but uh, then there's also the strategies.

Speaker 5: 46:25

Right now, you rely on the training data to give you these answers or like are you gonna do like a rag kind of?

Speaker 5: 46:29

thing right, yeah, indeed, but uh, I fully agree. I fully agree because, arguably, even if you do buy a news corp data but you still have like a rag system kind of thing and you don't vet your sources, you may still end up with what's true has right. So, yeah, but I think, yeah, I think the thing was also makes a bit uh, surprising is that first is google, which I guess they used to be the standard in ai, right, they used to be like, oh, google, deep mind and all the things right. Today, I don't know if that's the case, and it feels very half-baked like it feels like they're, and the reason why I'm bringing half bakedness um, we talked a while ago about the human pin.

Speaker 5: 47:16

We talked a while ago about rabbit r1, um, and now, like, I think, for the human pin, we had already said that like humane humane, yeah, humane. And eye pin, humane AI pin yeah, humane, an AI pin that we said like I think that it was a hot take. I think that this AI gadgets excuse me, this AI gadgets are like the future of AI gadgets, are basically smartphones, right. And then, um, the joke was that you have this, uh, humane ai pin and I'm gonna put here on the screen for the people following us in the live stream. So basically, the idea with this device is that you clip, so I I'm not super into it, right, but like you clip on your shirt or whatever he sees everything, you see, if he, if he needs to give you information, like I think he projects stuff on your hand or something, yeah right, I think he has a microphone. And then they were even joking with those foldable phones that you can get that right like and like sees everything.

Speaker 4: 48:14

I think you have to touch it to make like a picture or a video. So I think it's not.

Speaker 5: 48:20

It could be yeah so, but uh, yeah, that was this whole, this whole hot take like that. These are just like, okay, this is cool, but really the future is just gonna be phones, right. Um and lately. So this was on may 22nd, so a bit less than a week ago. Um, humane is looking for a buyer after ai pins underwhelming debut. What is this about Sennett?

Speaker 4: 48:49

Yeah, so apparently the direction time is quite slow. There are some hardware issues and they're now seeking to sell their startup for about $1 billion, so it's already quite expensive.

Speaker 4: 49:03

Yeah, it's like that and, yeah, I looked a bit at it. I think the, the main penis sells for six or seven hundred dollars right now, and one, one of these things, yeah, for one of those, and, as you said, I think it's a cool thing, but I'm not gonna drop my smartphone. So if I'm having my smartphone, um, I think all of the the big models are coming with ai functions as well now, so I don't see why I would buy that AI pin.

Speaker 6: 49:35

It's interesting that they're now for sale. I mean, it's clearly showing that they're not doing well, yeah, but they're still a billion, right?

Speaker 5: 49:41

Well, that's what they say. They want a billion.

Speaker 6: 49:43

but I think when you're not doing well and you have to sell.

Speaker 5: 49:47

it's a very hard yeah there even if you sell it and it's like, if they get a billion, like it's not a failure, right, like you have a startup or something for a billion, that's pretty good I very much wonder whether or this will, I will go I don't think it would be a billion as well, I think maybe also the but I think this comes also comes from after, um, for like halfway april that got uh trashed by like a very, very famous uh youtube reviewer, marcus brownlee I want to say his name.

Speaker 6: 50:18

Uh, no, but I really really trash him. Like he has a huge, huge, huge following. Um, and I think a lot of the reasons why it got trashed are also fair um, maybe a bit harsh, but apparently like there was a huge drop in demand, which was already not impressive. Uh, after the uh, it had a huge impact on the community. Um, so this is probably also. Is it this guy? No, this guy no partially influenced the one on the on the right?

Speaker 5: 50:47

yeah, it's the same, yeah, yeah, yeah, yeah, it's uh. So, yeah, this guy, yeah, the worst product I've ever dot, dot, dot, let's see what. This is the worst product I've ever reviewed for now. Yeah, it's a bit so strong language.

Speaker 6: 51:00

25 minutes of uh, yeah, harsh language but let's say that a year from now this still exists, or that their production gets sold for parts which is what I would assume actually maybe uh, we don't have a phone.

Speaker 5: 51:19

I was. I wanted to follow this up as well, but I haven't had time. I think humane ai pin also gets bundled up with the rabbit r1 like. Actually, the previous video for people on the live stream was actually about RabbitR1. This guy also reviewed RabbitR1. And the title is like barely reviewable as well.

Speaker 5: 51:39

This video I did watch. I did watch some things. There's more drama behind that RabbitR1 stuff as well, but apparently it was very underwhelming as well. I think this guy actually mentions that it feels like these AI companies. Drama behind that rabbit r1 stuff as well. Uh, but apparently it was very underwhelming as well. I think this guy actually mentions that it feels like these ai companies. They have a different strategy. Usually people would build something, they would test it and then go to market, and now it's like they build something, they go to market and they try to make it work after you know. So it's like the rabbit r1. Apparently. They talked about the. So just a reminder for people, it was like the little orange device that you have a school or you have a camera.

Speaker 5: 52:11

Very nice design, nice design from Teenage Engineering. There was this big like Steve Jobs-like. How do you call it? Like a launch or a keynote?

Speaker 5: 52:20

I think they called it, and they had the large action models which there's a bit of drama there, but I don't know enough to talk about it large action models, which there's a bit of drama there, but I don't know a lot enough to talk about it and, um, basically they said like, okay, you buy this, but there are like three apps that you can use. So it was like you buy but like there's Uber or like DoorDash, which is like delivery stuff. Uh, they said that a lot of times. You got it wrong. Uh, yeah, uh, yeah, and there's not enough apps, I think spotify, uber and doordash. So it's like, yeah, you release this product, but you cannot do anything with it. And then they were saying like, okay, yeah, we have this and and, uh, we're working on it, but it feels a bit backwards. Yeah, you know, you release something that doesn't really cater to the, to your audience, and now you're working on it, right, so it's a bit weird. Um, also, maybe vit Vitale has another question.

Speaker 5: 53:09

Let's say, did you see the CoffeeZilla video about Revit R1? He claims it's actually a scam. Yeah, that's the I think. I haven't watched it. It's more recent though. Yeah, so that's the one that I still wanted to take a look. I think they mentioned a scam because. So someone told me that the scam is because the large action model was not actually a large action model, it was just a regular foundation model. They're just using it. But I need to look into that and maybe I'll bring it up next week. So hold me to that, Vitaly Sparacello, and maybe we can discuss more about it. But yeah, so I'm not sure if it's a trend now, looking also at the Google stuff, but it feels like these big tech players they're a bit releasing stuff prematurely. Maybe I'm not sure, but it's a bit. It's a weird time, it's a very quickly evolving domain.

Speaker 6: 54:07

You can also imagine time is of the.

Speaker 4: 54:10

Yeah you want to be the first, and maybe quality is not as important, like, but that's. That's what I think is weird.

Speaker 5: 54:17

I don't know but that's what I think is weird. Like quality is not that important.

Speaker 6: 54:19

I think it's one of the first times that I actually yeah, I don't discuss this, you know no no, I know, but it's like I mean, I think we're discussing this now because the first time we discussed this was on hardware. We've seen this software already happening a long time ago with games, for example, where you get these very early versions released and then it gets improved over time. You just now see it with hardware.

Speaker 5: 54:40

Yeah, I'm not sure. For example, google with this AI thing. It's a bit weird.

Speaker 6: 54:44

No, yeah, but that's maybe just the Google thing. Yeah, maybe, but I mean even that like we discussed a bit the Google thing and, like we had the GPT-4.0 presentation. It was a very cohesive presentation the day before Google IO, very cohesive presentation for OpenAI. Like this is a product, this is what you can do with it. And then the day after, like you have these hundreds of news items from uh presented at Google IO. It's like like all the engineers sat together for a week and had a, had a, had a hackathon.

Speaker 6: 55:17

I'm like these are the results but like what is actually going to be a, what is going to be a product, what is not and what will end up on google graveyard after a few months and like yeah, but even the gemini stuff.

Speaker 5: 55:28

You know, when they did the announcement, they just like oh yeah, this is amazing and I think what google?

Speaker 6: 55:32

has always been super good at is page rank, yeah, which which they're now trying to replace, and ads, of course, yeah yeah, but yeah, it's a bit weird.

Speaker 5: 55:41

Huh, it's a weird uh when you look at that. It's a bit weird time, but yeah, yeah yeah.

Speaker 4: 55:48

So yeah, I just wanted that, like with the pins. I would. I would imagine, if you um want to quickly improve and get user feedback, uh, you could sell it at a low price and maybe, like that, it's good for a user, like okay if it upgrades, you got it for a low value, for low price, you give some feedback, but it's like six, seven hundred dollars and a monthly subscription of, I think, 25 a month on top of that. So it feels a bit much for uh, not getting a lot in return.

Speaker 4: 56:16

It's good for people that want to try it out for that yeah, well, if it would be cheaper, it would be nice, like like a deal, okay, you get to try it out and then after a while get, gets upgraded and get better. But yeah, I'm, I'm still much more a believer in.

Speaker 6: 56:30

I have my iPhone, I have my earbuds and we need Siri 2.0. Yeah, yeah yeah, and that will cover a lot of these things, that's the R1 and the Humane.

Speaker 5: 56:40

Yeah, you know, it's a funny thing.

Speaker 6: 56:41

All of these, like AirPods and iPhone, are very much. It's normal for everybody.

Speaker 5: 56:46

Yeah, the R1, there was one part on the video that he's talking about. So basically, like there's a scrollable thing, and he said that the scrollable thing is very slow, which makes the experience a bit, you know not great yeah, and I think the way you go back, the navigation, so he goes a bit about that or typing or something.

Speaker 5: 57:04

And then he was like, yeah, but uh, what if you could use the, the screen as a touch screen? And it's like, well, actually you can, but they only allow at a certain moment, right. And it's like, oh, why would you only allow certain moments? And it's like, well, maybe because if you did that it would look too much like an iphone. And I think in the end it's like I think the iphone or whatever smartphone, it is the most flexible design, right. And I think to try to go away from that, because now you want to do like a ai specific hardware, it's a bit like it's just trying, it's just marketing, right like, if you have this, I think there is something to say for that are for people that that need a more distraction environment.

Speaker 6: 57:46

There is something to say for that. The modern laptop and phone are not that. But I think there may be something like that potentially can play a purpose.

Speaker 5: 57:58

But then it's not like an AI-focused device, it's more of a focus-focused device. You know what I'm saying.

Speaker 6: 58:04

But maybe AI can very much empower that.

Speaker 5: 58:06

No, it could. But then I think it's like the rabbit r1 is like it was a uh, like a phone, like a redesigned phone, thinking of these large action models, right, like that was the the big claim, and I think for me it's like yeah, okay, that was just, that was purely just marketing, yeah claim we just buy it for the design.

Speaker 5: 58:27

Yeah, and for the cool new gadget, how do you say Like hype? No, not hype, but you know, like the shiny new thing effect. Yeah, the RBR1 is actually not that expensive. I was tempted to buy it just to give it a try, but I was like no, and I'm actually glad that I didn't do it, but in any case it was actually interesting.

Speaker 6: 58:50

It's also relatively recent when it comes to distraction-free devices. There is the Daylight Computer I think you can go to daylightcomputercom and they are trying to build sort of an iPad-ish device with like e-paper that you typically have like on your kindle or on a kobo e-reader and like they're trying to marry these two type of use case where you really have this real-time, fast, refreshing e-paper on an ipad as a device this is the put him on the screen yeah, it's actually the the headline is the computer de-invented, designed for deep focus and well-being, but like with the same set of more or less functionality.

Speaker 5: 59:33

Yeah, so I think your description is good for people following. You can put the link on the show notes, but people following just on the audio. This looks like a iPad versus a Kindle.

Speaker 6: 59:44

Like something in between an iPad and a Kindle and a remarkable tablet like yeah, like they like marriage, these things.

Speaker 5: 59:50

Yeah, it looks like a ipad, but with the screen of a kindle yeah, but more, uh, kindle is like it's the.

Speaker 6: 59:59

It takes a long time to go for. Yeah, a new screen, right, like, yeah, more and more, but the screen isn't like the look, like it seems like it's black and white.

Speaker 5: 1:00:05

You see all these things very interesting. I like that they say de-invented. It's good that you read it out loud, because I just glanced I was like reinvented, uh, yeah what else? Do we have? Oh, we have quite some stuff. Do we need to find the time? But do we have, uh, something from the tech corner? You know, a wise man once told me that a library a week keeps the mind at peak yes, uh, and I added a library.

Speaker 6: 1:00:34

Um, didn't have a lot of inspiration, to be honest, but I read this. Uh, I think it showed up somewhere on my, I want to say, twitter. Maybe massad, I'm not sure, but it's a PostGrest, postgrest, postgrest, instead of PostRest, postgrest, starting to become difficult, postgrest, the database that everybody knows, one of the best databases out there, not opinionated in this at all.

Speaker 3: 1:01:03

Let's not go there.

Speaker 6: 1:01:15

But this is PostGrest, which ends the suffix is REST, like a REST API, rest interface and how? I didn't try it, I tried it out myself, but I read a little bit of the description. It's like a kind of a web-ish server that is in front of your Postgres database. That allows you to talk to your database like if you would just do a REST call. Like if you do a GET, you're going to fetch stuff on the database. If you do a POST request, you're going to create stuff on the database. If you do, I would assume I'm not sure actually on the terminology to use if you do a PUT request, then they're going to update and you can also do authentication like you would do for an API, like if we have bearing token or these type of things. So it feels very natural it's cool, Maybe for the people.

Speaker 5: 1:02:06

just a bit more concrete, I think.

Speaker 6: 1:02:09

Yeah, so maybe also like what you typically, when you talk to talk to a rest interface, you're showing some stuff on the screen like you use uh, typically use json structures. Yeah, so when you create something in, when you do a post request, you send a json structure there and the json structure ends up in some way in the database. If you do a get, um, it's going to give you back a JSON structure which is typically very easily mappable to data classes in whatever language, like dictionary in Python or object in JavaScript and these type of things. I'm not sure if I I'm not 100% sure on like what is the main use case of this.

Speaker 5: 1:02:48

I actually have a very nice use case for this.

Speaker 6: 1:02:50

Okay, interesting.

Speaker 5: 1:02:54

Well, maybe just very quickly for the people. Basically, the rows that will return from the query is like a list or an array of JSON structures with the key is the column name and the value of that column right. So also, this is not. I didn't know, this was not possible at all with Postgres, if you don't use this.

Speaker 6: 1:03:19

You can just talk in SQL to Postgres.

Speaker 5: 1:03:21

That's what you typically do In Snowflake. So here you have something called external functions, which basically exposes stuff as a recipe. Actually it's a bit different, right, because, um, um, actually it is different, so I'm not sure if I have a good use case, but bear with me, bear with me, um. So in snowflake just finishing this part here the idea is that you can have like a udf kind of something that have like a UDF kind of something that looks like a function in SQL. User-defined function, user-defined function. Yeah, but that actually what the function does underneath. They just make an API call. Okay, right.

Speaker 6: 1:03:59

So from Snowflake SQL you do an API call to somewhere and you integrate it Exactly right, but it's like it looks like a function.

Speaker 5: 1:04:07

I'm going to explain what the use case I thought for this? Yeah, cause I think that is actually slightly different. The use case for this is for machine learning deployment, so there's also another. I think actually you posted this minds DB that they they actually have a. That's the first time I had seen it, this approach. Should I share this tape? Instead?

Speaker 5: 1:04:31

The idea is that a lot of the times, models are deployed as a REST API, because then you scales and all these things, right, all the goods, and then the database actually makes a REST API call for that machine learning model. Then it's hosted, it can scale up, even if you have like a batch thing, right, and you can have specialized infrastructure. So if you need a GPU for that, whatever you know, or if you have even LLMs or something, it's all there and then, like, everything is just SQL for you. Like, even if you really want and that's something that's an idea for a little poc and maybe do a presentation about is like doing a full end-to-end pipeline just with dbt and these external functions, you can have like a dbt project that one step it makes these external function calls and then in the end you have your. You have your predictions there.

Speaker 6: 1:05:24

It could be a little bit right, and I think this is something that, like minus db, also helps with right, but in the end it's just like a rest api it's a rest api in front of your database and like I'm thinking like typically not always, of course but like if your application lodging and you have a database behind that, it's like your typical use case yeah and like this to me then just seems like another extraction there in between.

Speaker 6: 1:05:49

That's what it feels like you can directly dock in SQL in Postgres native protocol and you don't need this at all. But you can maybe say for that specific use case you maybe say for typical developers they will be accustomed to.

Speaker 6: 1:06:08

HTTP and maybe not. Maybe not to postgres, but it's like between. It feels a little bit like a reach. If you're going to develop this product, maybe you can say the use case like I have this application like tens of different databases that I need to talk to and like then then hdp becomes a very nice like like rest becomes a very nice abstraction layer.

Speaker 5: 1:06:22

Yeah, I'm also thinking like if you have, like you again think your machine learning world right, like you have I'm not sure how fast this is this would be, but if you have like a real-time application. So, for example, recommend your system by real-time Whenever you make a request. I probably want the history of your purchases in the past, and then maybe I can make an API call for the database instead of, like actually having the integration, all these things. Maybe that's fast, maybe that's easier to set it up.

Speaker 6: 1:06:55

It's a big abstraction, it's just to do that. Why don't you just do it in SQL, like I mean, that's what I would ask you.

Speaker 5: 1:07:02

But then, like you would like, I mean you would you, because I still have the like. Okay, so my, I have an endpoint.

Speaker 6: 1:07:10

It's python, because I'm building my model in python and this is okay, so like your, your model is served like it's a python-based model and it's behind an endpoint. Yeah, okay, it's a bad endpoint.

Speaker 5: 1:07:19

You click on the, you log in to what to the platform. I need to give you recommendations, so the moment you log in, I need to get you I'm a user on your platform exactly. Okay, so I need to get your purchase history. Okay, in my python endpoint you would use like sql, alchemy or something to just fetch the stuff from the user, for example. Yeah, I guess you could.

Speaker 5: 1:07:37

Just you could, yeah you can yeah, I'm just wondering if there is a well again. I guess it's like a sql alchemy versus a rest thing. I'm not sure if there's a but it well it's.

Speaker 6: 1:07:50

It's postgres native, because sequel alchemy translates postgres native protocol. Yeah, which is always there. It's always going to be there, right? Yeah and the the. The question is like are you going to put something in front of that, just to not speak? Postgres native.

Speaker 6: 1:08:07

That's the question, right yeah, yeah, that's true yeah even though it may be feel from a developer experience. If you're talking, that could be an argument like rest feels more. It's easy, understandable more maybe more transparent, like it's easy to understand what is happening yeah and I guess, but I think, yeah, I think.

Speaker 5: 1:08:26

The other thing I'm thinking is like, if you have the prediction you want to share with consumers, I think, if you say rest, most people would expect them to know.

Speaker 5: 1:08:35

Um, I'm yeah, I'm thinking a bit out loud here, but I'm thinking like, okay, so I have predictions there and I want to serve it to people, and maybe your team is only using rest for everything, or maybe your team doesn't use python, or maybe your team doesn't use this or that. I guess maybe exposing a rest, but then it's just an abstraction layer, right, and most people, I guess, would be familiar with sql well, that's the question.

Speaker 6: 1:09:00

I think it's cool. It's cool library yeah, it's cool.

Speaker 5: 1:09:08

Yes, yeah, cool, indeed. Maybe if someone has an idea for this, for how to apply this. This is curious. But actually, just to be clear as well, this, the UDF thing, that is actually a REST API call. This is not covered in this library, right, bart? Like what I showed from Snowflake, this is not covered at all. Like the other way, from SQL to a REST API call.

Speaker 5: 1:09:36

Also, I know this is the last thing I'll say about this, bart, I promise I know you're eager to move on. This also, like this reminds me a bit or made me think of, like SQL model from Sebastian, and he also has the FastAPI, and I remember, like the benefit of using SQL model is that you can. Well, maybe I'll put it on the screen as well, very quickly. The SQL model is a Pydentic model, which is what FastAPI is based on, and he even says that the benefit of doing that is because then you can specify and you can get back the object, which is very similar to what you're saying but like then, it's like you're building this interaction layer, right, all right, all right, all right.

Speaker 5: 1:10:18

I hear you, bart, I hear you.

Speaker 6: 1:10:21

Okay, I have a hard stop in a few minutes.

Speaker 5: 1:10:23

Yes, so very quickly. Yeah, we mentioned domain names. I mean, I joked about buying xai, xcom, xeverything, um, and it pays off, apparently. It pays off, apparently, indeed, and I think the reason why the for people that are not super familiar with all these things is that, uh, once you buy a domain, it's kind of yours and as long as you keep paying, it's yours, and if someone wants to use it, they need to buy it off of you, right? But what happens if there is a government that bought a domain and apparently there was, I think, an ethical hacker that he was.

Speaker 4: 1:10:59

Yeah, so domain typically expires after one, two or three years, and then if you don't renew your purchase, your purchase then, uh, it just becomes freely available. And there was indeed an ethical hacker from belgium who um bought, I think, like expired domain names from police regions, from like hospitals, uh, government agencies I don't know exactly what I think he bought yeah, a lot, yeah, a couple of hundred or a couple of hundred domain names, I think.

Speaker 4: 1:11:28

And then I just um, he watched actually for males coming into that domain, um, because yeah, people were still using those email addresses and he got a lot of sensitive information. So just by buying those domains, painful, yeah, and actually I uh I've worked. My thesis was, uh, when I did my master's was on, um, the detection of fake web shops in the dot b zone, and that's actually like a common practice to just buy all domains. Uh, I remember there was one one site on, I think it was called wandeldrumbe at the time, so it's like a walking dreambe and it was like just a fake web shop. They buy random domains, um, but yeah, here, especially if it's like trusted or seemingly trusted domains, that's really uh, wow that's um dangerous dangerous, yeah, yeah indeed yeah

Speaker 4: 1:12:21

yeah, super easy. It's also just like a couple of 10 euros, I don't know. Uh, yeah, 10, 50 euros for like three years of a domain name.

Speaker 5: 1:12:29

So yeah, yeah, domains are very, very cheap.

Speaker 4: 1:12:31

Yes, you know you can just buy a hundred of them.

Speaker 6: 1:12:35

So I'll tell you now what is available quickly doing like a domain name check, like override dot gantt is available that means government again. There's a big city here is available.

Speaker 4: 1:12:49

That means government. Again, there's a big city here it's available for uh database?

Speaker 6: 1:12:54

yeah, yeah, yeah, right, like put your credit card here and wow, that's, uh, that's yeah.

Speaker 5: 1:13:03

She never thought about this, this security risk. Yeah, well, if anyone from ghent is listening, let's make a petition. I'm going to check if tomorrow's still available.

Speaker 6: 1:13:05

Anyone from Ghent is listening. Let's make a petition. I'm going to check if tomorrow is still available.

Speaker 5: 1:13:10

Okay, let's do that, and then we can follow up.

Speaker 6: 1:13:11

I see the glint in your eyes, wow.

Speaker 5: 1:13:14

you looked at me like I'm sure he's one of those. All right, on the interest of time, are we okay to move to the hot take?

Speaker 6: 1:13:23

Yes. Oh, hot, hot, hot hot.

Speaker 5: 1:13:34

Hot, hot, hot, sizzling. Um. What do we got, bart? We have here that humans now share with the web, equally with bots. Report warns amid fears of dead internet, and I think that internet is the what catches our attention here well yeah hot take is the internet is soon to be that.

Speaker 6: 1:13:45

That is the hot take. You're making it a bit harder, huh. You're just like adding some. This is an article by the independent UK news site where it states that humans share the web equally with bots and that there are fears of a dead internet, an internet ruled by bots, bots chattering to each other.

Speaker 5: 1:14:09

But, like when they said, bots is just like crawling the web. Is it generating content? Generating content, yeah.

Speaker 6: 1:14:16

And especially now with uh, yeah, yeah, yeah, it becomes very easy to generate authentic looking content. Now.

Speaker 5: 1:14:23

But if there's a person behind that is like prompting or doing these things, it's still considered a bot. It's still considered a bot. It's still considered a bot. Okay, yeah, but then okay, then I agree that like well, I'm not sure if I agree that share equally the web with bots, but I do agree that there is a bigger amount of bots, but to me I wouldn't define it as bot, though because then I think the question is like do what is it?

Speaker 6: 1:14:45

is there a danger of the internet being dead? Is it?

Speaker 4: 1:14:48

truly a problem that there are a lot of bots like, if you're aware of that well, but are you? Are you though?

Speaker 3: 1:14:55

I think the issue.

Speaker 6: 1:14:55

The issue is this like if we talk about red, like if we look at the top page, how many of those are actually either posted by humans or upvoted by humans? I think that is a bit the question, like how many is still authentic these days? Yeah?

Speaker 5: 1:15:09

and I also think it's like open ai. There's a deal with news corp. News corp authors use chat gpt to their content and it's just like, you know, a loop of ai feeding on ai and then we're all gonna kind of like converge to noise eventually, you know. So, um, yeah, I mean to be very honest, like ai feeds on the internet, if ai is ruling the internet and feeding on it, something's not gonna go. Well, right, sounds a bit like a race to the bottom.

Speaker 6: 1:15:42

Yeah, right, but I feel like at one point people are gonna be like well, this is, this is not getting better and what is also in the article is that it's it's not necessarily only content, but also and I'm not sure where they get their their numbers from is that also the the proportion of internet traffic will soon be bypassed by bots versus the internet traffic by humans yeah I mean also ask like, like, what does it mean to your infrastructure?

Speaker 6: 1:16:10

because the scaling of the amount of humans, yeah, and maybe if we ignore the technological progress, I mean it's not exponential.

Speaker 5: 1:16:21

Yeah, but very much have the potential for exponential growth yeah, I do think that at some point it's gonna happen. To be honest, I feel like, yeah, these things get more and more accessible. Yeah, I'm not sure. It's a bit unsettling to think about this, like if you have bots and you have more and more and more and it's gonna be crowding the internet in a way yeah, it's a, it's a like to me.

Speaker 6: 1:16:45

Like the example is a relevant one. Like you don't know, like what was created by a human or a bot, and like or what was it? And if it's a bot, like what were the intentions? You also don't know this by human, but like you can do this very, at a very large scale. Yeah, like I can today. I know you're a big fan of flip-flops. You told me yesterday about yonas. Like I can intimidate you and create an automated bot on reddit that really goes wild on all flip flop related posts, or maybe non-flip-flop related posts, like nike shoe related post, and like on all nikes.

Speaker 6: 1:17:19

Really bash nike and really try to push what's the name of the brand that just havayana havayana and like if I as a person do that, like it's super hard for me to get any noticeable traction, but I can tomorrow launch thousands of bots that do that on reddit yeah, it's true, if I can get around the the security checks and don't get blocked.

Speaker 5: 1:17:41

But I think to me like it's just emphasizing more something that was already a danger, or like fake news and stuff you know like.

Speaker 5: 1:17:49

But yeah, it's been on people's radar for a long time like the russian trolls in the us election and stuff but what we're seeing here is, if that is, 50 of internet traffic yeah, that means that we never were able to stop it like it's on our radar, but it's only growing yeah, I think like it just kind of puts this on the top of the priority, yeah, but I think the fear is not like the issue is not.

Speaker 6: 1:18:07

No one really knows what to do with it.

Speaker 4: 1:18:09

Yeah, that is true. Can you stop it?

Speaker 5: 1:18:14

I'll leave that for the listeners If anyone has any. If you agree disagree, send Bart an email, bartadatorsio, with the poll request. Thanks everyone for listening. Thanks, Sene.

Speaker 6: 1:18:29

Thanks Sene, thanks Sene thanks everyone.

Speaker 5: 1:18:30

Friend of the pod and you wanna hit the outro music, I'm gonna hit it.

Speaker 1: 1:18:37

You have taste in a way that's meaningful to self-love people hello, I'm Bill Gates.

Speaker 3: 1:18:47

I would. I would recommend TypeScript. Yeah, it writes a lot of code for me and usually it's slightly wrong. I'm reminded it's a rust here Rust, rust. This almost makes me happy that I didn't become a supermodel. Huber and Ness. Well, I'm sorry guys, I don't know what's going on.

Speaker 1: 1:19:09

Thank you for the opportunity to speak to you today about large neural networks. It's really an honor to be here.

Speaker 3: 1:19:15

Rust, rust, rust.

Speaker 4: 1:19:15

Data Topics. Welcome to the Data Topics. Welcome to the Data Topics Podcast.