The Monkey Patching Podcast: Going Bananas on AI, Data, LLMs & Tech

Bart: 00:01

Hi,

Murilo: 00:08

everyone. Welcome to the monkey patching podcast where we go bananas about all things soya, shopping, and more. My name is Murillo, joined by my friend Bart. Hey, Bart. Hi, Murillo.

Murilo: 00:17

How are you? I'm doing very well. Very well. So you snatched the last days of the fleeting days of summer?

Bart: 00:25

Yep. Yeah. Exactly. Yeah. Caught a few days of sun sun in the Ahostla Valley in the North Of Italy.

Murilo: 00:33

Nice. Oh, wow. All living in Italy. Really, really cool. Hope it was good.

Murilo: 00:37

Hope you're It's a very good time. Refreshed.

Bart: 00:40

A bit of running in the mountains.

Murilo: 00:42

That's very nice. That's very nice. Cool. What do we have for today? Would you like to start or should I start?

Bart: 00:51

I'll kick it off. Fivetran is reportedly in advanced talks to acquire DVT Labs in a multibillion dollar deal bundling pipelines transformations. The pair already share 1,000 joint customers and integrations, Fivetran just bought Sensus and Tobacco to bolster ELT. If it lands, expect a one stop data stack gear for AI prep and fewer handoffs across the modern data platform.

Murilo: 01:20

Fivestrain is the extract load tool. Right? And now they're they're talking about buying TBT. Where should Yeah.

Bart: 01:29

So Fivetran is a big big cloud solution. Been been around for quite quite a while. I want to say at least ten years that does a lot on the data pipeline in general. I think they're where where their their original focus is when you look at ELT, extract load transform. I think their initial focus was very much on the EL part.

Bart: 01:55

And they bought Census recently. Census is a reverse ETL tool. Basically, to get data from your data lake or data warehouse back into your operational systems. Tobago is is a tool that focuses very much on the t part, transformation part with, for example, SQL mesh, which is actually a competitor to to dbt.

Murilo: 02:18

Yep. I was gonna

Bart: 02:19

say And now they're apparently in talks with dbt to acquire DBT and their solutions. So which would be a big consolidation in the space. Basically, the whole ELT space becomes $5.

Murilo: 02:35

Exactly. I was gonna say. And what do you think of that actually? How do you do you think that's do you think there are any positives that can come from this?

Bart: 02:47

That's a good question. I think competition is always a good thing. Right? I mean, this this removes a bit of competition. Of course, it's great space for new competitors to pop up.

Bart: 02:56

That's one thing. I think DBT, it's maybe a bit of a hot take, always had a bit of a trouble to find a good market fit with their DBT cloud offering, which is basically their commercial offering. Also, because their open source offering is very good. And, actually, like, what Fivetran and a number of players already did is that they had their own own dbt core executors, which you could already run. So you didn't really need to use dbt cloud even if when you when you were on on something like Fivetran.

Bart: 03:30

Even though Fivetran did support integration with DBT cloud or does support. So I think DBT cloud was under a lot of pressure to have a good offering, and I've been struggling a bit for for for a while to to provide that. But Fivetran clearly has, like, I think for when I think these are twenty twenty four numbers, if I'm not mistaken, where they reported that they had something like 300,000,000 in ARR. Let's say it's a it's a huge customer base for us. Yeah.

Murilo: 04:04

Yeah. Yeah. Indeed. Indeed. Yeah.

Murilo: 04:06

But I I again, I'm not as close to the DBT labs and all these things, but I also have the similar feeling as you that DBT is super popular, but DBT Labs struggle to get a piece of that, you know, the usage, the the profits, and all these things. So, yeah, maybe for DBT Labs, it's a it's a happy story maybe for them. Like

Bart: 04:31

Well, let's see. I I I don't think so this this is what I what I understand is these are rumors. This is not officially been acknowledged by either Fivetran or DBT. So let's see what it gets. Right?

Bart: 04:44

But, of course, it is creates a big roll up into in the space, like a big consolidation, like what what happens to DBT Core, for example.

Murilo: 04:53

For me, the bigger question was Sequo Mesh. It is already owned by Tobico. Right? And Oh,

Bart: 05:04

actually, it was a product from Tobico. Right?

Murilo: 05:06

Yeah. Indeed. So Sequo Mesh, I've always seen it, so I I haven't used it. But from what I've saw, it was like an alternative to DBT. Yeah.

Murilo: 05:16

Yeah. Yeah. Exactly. Yeah. So for me, that's also a bit okay.

Murilo: 05:20

Fivetran Fivetran owns both, then is it are they gonna focus more on one? Is one gonna be dropped? Do we still have the competition? To me, that that was a bit the biggest. That's a good question.

Bart: 05:32

And I think from a commercial standpoint, it's just like when you look at the ELT space, like, 2Bico and and DeepDs are both on the t, the transformation part. And, like, to me, a bit of a what they're doing is a horizontal integration and, like, you're you just buy up the competitors in this space, and then you become the only solution.

Murilo: 05:51

Yeah. Indeed.

Bart: 05:52

And what that means for the individual competitors, like like like and products they have, like like SQL mesh and dbt core is a bit remains a bit to be seen towards the future, but I think it's unlikely that we will see active as active development on the two as we've seen in last years. Yeah. If this comes to fruition, of course, it's still early days.

Murilo: 06:11

Yeah. Yeah. Yeah. Also heard from colleagues they were discussing this and they said they were surprised that five trend was the one that entered in talks with DBT labs that they thought that Snowflake was gonna be the one to sit down with DBT let's say I think also again not as in the loop with all the things that all the discussion that DBT Labs has but I know that DBT Labs had some like nice integrations with Snowflake for like the Python models and all these things right so I think they do collaborate but they were they mentioned they were a bit surprised by if Chen being the one to initiate the talks. Mhmm.

Murilo: 06:48

Mhmm. Were you also surprised by this, or you didn't really think about much, or you think it makes sense?

Bart: 06:55

Don't really have an opinion on it. I think it's a I think Snowflakes already has a very strong, strong name in the market

Murilo: 07:05

Yeah. True.

Bart: 07:05

As a one stop shop for everything around your data needs. It already has its name. And I think with especially in the at least in the European market, I can't really speak for The US market. Fivetran is more here here is this tooling that you can do to ease a bit, like the how you extract stuff, and now you upload it in a bit of a clicker drag and drop point of view. Like, it doesn't have this strong name in the market as this is your one stop shop.

Bart: 07:33

And I think it makes it from a strategic point of view, like, they acquire DBT, like, it's there suddenly like, it adds to their their recognizability as the solution for data.

Murilo: 07:45

I see. I see. I see. I see. No.

Murilo: 07:47

It's it's a yeah. True. True. I didn't know Fivetran was that big, actually, to be honest. Because I think it's like what you said.

Murilo: 07:55

Every time Fivetran is mentioned, it's not giving as much emphasis. Right? Like, I don't immediately think like, woah. This is a

Bart: 08:02

Yeah. I don't think we see it very much here domestically. That is true. But like like I said, like, they already exist for a very long time, and I think people that looked at it ten years ago and maybe didn't go for it also haven't really checked it out in the meantime. I think that is the the danger.

Bart: 08:19

And, also, like, from a more if you're an enterprise customer and you make this choice, you're gonna you're gonna make a choice for a number of years. Right?

Murilo: 08:25

Yeah. Yeah. True. True. Alright.

Murilo: 08:29

What else what else do we have? We have testing Veditas, an employee only chatbots to trial series AI upgrades with no public launch planned. Staff use it to chat through tasks like searching personal data or editing photos while Apple leans on Google's Gemini for AI search. If it stays internal, does Apple seed the consumer chatbot moment to rivals or double down on on device Siri later. Yeah.

Murilo: 09:03

I thought it was interesting. Like, Apple is still Apple has struggled with the whole AI boom. Right? They've always been yeah. It's a bit it's an interesting story.

Murilo: 09:17

Right? Because, like, Apple has always been, like, elegant and and, like, the the still on the forefront of technology, but Siri has never been a really product that stuck as an AI product. The AI boom came. They didn't really upgrade Siri. They never historically, they didn't really buy stuff from competitors much, and then they tried Apple Intelligence.

Murilo: 09:39

After a long time, there was a lot of expectations. Apple Intelligence fell flat a bit. Right? I think there's a lot of stories of, like, hallucinations and stuff. Even on the Apple Intelligence, they also mentioned partnering with Chi GPT, which was also it caught some attention.

Murilo: 09:58

Right? I don't know if it was the first time, but it's not something they do often to to delegate things. Mhmm. Then there's a whole discussion. Like, should they just buy stuff?

Murilo: 10:05

Like, right now, they rely on GoAI Gemini Google's Gemini. Should they build something themselves? And, apparently, they're still working on this. Right? For me, once you work on something for a long time, the expectations naturally grow.

Murilo: 10:25

Right? And I feel like they're just they're just digging a hole for themselves, it feels. Right? Like, if they they said that it's not planned to to announce to to consumers. I don't know.

Murilo: 10:39

Like, maybe we better to just say, okay. We partnered with Gemini and now Sirius, you know, powered by Gemini instead of, I don't know, trying to keep keep building your own thing and staying behind and, like, announcing it or putting expectations. And when it comes out, the expectations are so high that you're never gonna fulfill it. How do you feel about this?

Bart: 11:05

I so I'm still hopeful that we will have a good Siri within the year. So this but how I big again, a rumor more or less. Right? That we're seeing, like, the the rumors that they're that they have this internal chatbots that they're testing out to basically, like, interact with the Apple ecosystem on your phone, on your device. And indeed, like we've seen, things like this being released like Apple Intelligence, we've there is an existing Chateappity integration with Apple Intelligence, which I've tested out with, and it's really horrible.

Bart: 11:44

I understand, like, maybe they should just go for Google Gemini or whatever. But to me, Siri is something that should be so core to how you interact with your Apple ecosystem. Like it needs to be an Apple product to really have this deep integration to to do these things. And it is really, really a pity that it can't do it today. And I think also from a security point of view, it will always be very hard to have an external provider do this unless they say, please build our app for us.

Bart: 12:18

But I doubt that they will do that may maybe as an as an in the in like an intermediary solution, but going forward, this will be really key, I think. No. I think this is just like like serious shitty today. Yeah. Like but you should be able to just say to your phone, make this note and then also plan this in my agenda and send off this message.

Bart: 12:40

But you can't say that in one sentence. Right? Like, maybe you can do it, like, and then you need to do fifteen minutes, need to talk to your phone and say, no. No. Not that.

Bart: 12:48

No. And then please Yeah.

Murilo: 12:48

Yeah. Yeah. Yeah. But it's almost like you're describing clicks with speech. Right?

Murilo: 12:54

Like, do this. Now do this. Now do this. It's not very it's not very fluent. Right?

Bart: 13:02

And I think they will need to do something, and I think they're big enough. They're like, they have enough of the mobile market share to not be the far first mover on that, but they will have to come up with something that is very nicely integrated.

Murilo: 13:18

Yeah. I think so. I think it's just to me, it's a bit I think it's more like the expectations. Right? I feel like it's been long enough that people are gonna expect a lot, and I feel like I don't know if they were the impression I had.

Murilo: 13:32

So, again, I don't follow Apple announcements as closely, but the impression I had is that they were waiting for something quote, unquote perfect to really wow everyone. Yeah. But but what they also did in the and this is

Bart: 13:41

a few months ago, and I'm a bit scarce on the details by now. But they did a whole change in the leadership team Yeah. On the product development behind Siri and Apple Intelligence not

Murilo: 13:53

that long ago. And I hope

Bart: 13:55

that but but to get to Marcus for after a change of leadership team on a product, like, will take you probably six months. So I hope to see something in the coming in the com

Murilo: 14:04

Yeah. Because I actually I vaguely remember something like that as well, that Siri was a product that within Apple it was a different division or something so we felt it like it was a bit neglected in terms of how integrated it is with this and that and the development, a lot of details missing but I I do remember that the the Siri story at Apple was was was bigger than just Apple intelligence like for for years it was a bit it wasn't a new problem let's say it was something that kind of built up to this point right But, yeah, you're a big Siri user. No?

Bart: 14:37

I wish I was a big Siri user, but You're a big Yeah.

Murilo: 14:40

Speech to text or voice AI user.

Bart: 14:46

I do use, ChatGPT a lot in voice mode.

Murilo: 14:50

Yeah. Okay. And use it, like, when you're, I don't know, you're home doing something and you just talk

Bart: 14:55

to So for example, when we prepare for the podcast, like, have the list of our articles that we're gonna go over, and I use the voice mode and chest GPT to really interrogate a bit, like deep dive on some some stuff and to understand really, like, bit in a very easy way when I'm just hands hands free doing something else and to have a bit better understanding of the article, to, to do some deeper research on some some topics that are not described in the article. I see. I see. So, like, you're you're using you talk to to to prepare for this to get a better, like, say this and that. Cool.

Bart: 15:31

And

Murilo: 15:33

now you're not afraid of hallucinations anymore. Right? I feel like 2025, we're not really concerned about these things anymore. I feel like one year ago

Bart: 15:42

You're not?

Murilo: 15:43

I mean, I feel like like we mentioned, like, when you Google something, see the summary. I mean, most of the times, like, yeah, that's pretty good. Right? Like, you don't see the things like I don't know. Like, I feel like we are trusting more.

Bart: 16:01

But I think GPT five, maybe they actually noticed because I use audio mode quite a bit. So GPT five was really a step up in terms of, well, or a step down depending on your point of view, in terms of hallucinations, avoiding hallucinations.

Murilo: 16:17

Mhmm. I see.

Bart: 16:19

Voice mode still uses four o, and it's like, you really need to be aware that this can happen.

Murilo: 16:27

Oh, really? So, like, you you do talk to it, but you still verify everything that's saying, or you have a critical year to it, like, oh, that doesn't sound right. Exactly. Okay. Okay.

Murilo: 16:37

Interesting. Interesting. Very cool. So let's hope that Siri can be a contender at least then. I also heard that, like, this was a while ago as well, but I think Alexa, Amazon's Alexa, they were partnering with Claude as well for for for the voice stuff.

Murilo: 16:56

Yeah. Oh, yeah. Cool. And what else do we have?

Bart: 17:01

OpenAI launched instant checkout, letting USGS Jetsupty users buy from Etsy and soon Shopify without leaving the conversation. Payments run through Apple Pay, Google Pay, Stripe, or cards, and results are organic and unsponsored with a small merchant fee. Open sourcing the agent e commerce protocol could shift product discovery power from search engines and marketplaces to chat agents. What do you

Murilo: 17:25

think about this, Ruehlo, the ability to buy through your Chattypitty app? I don't know. I think it's, I think it's interest well, I know that maybe it's interesting for OpenAI, right, to tap into a new revenue. I mean, they're already doing ads. Right?

Murilo: 17:41

So maybe it feels like a logical next step. Are they doing ads? They were. Like like, if you like, ads in the sense, like, if you say, what's a good protein powder? And then they say, oh, these are some protein powders that I can you know?

Murilo: 17:56

Like, you can go to the website and have the pricing and this and this. Yeah. They had some stuff.

Bart: 17:59

But you're not paid apps. Right?

Murilo: 18:01

I don't know if they're paid. I don't think I they

Bart: 18:03

don't think so.

Murilo: 18:04

But I think like if you're starting to show products to people that they can go on the website and buy, I think it makes sense that now you're directly there right I don't know if I trust it enough that I would just kind of talk to JGPT find a product and just buy it I'll probably want to do my due diligence right on I mean most of the times as well like I already do my due diligence like looking at reviews is this is the actual the best price or something like that so I think it makes sense not as surprised maybe but one thing that I thought it was interesting from this announcement is that they started they are releasing another protocol the agentic commerce protocol another so right now like there is MCP that I know of there's MCP there's A2A so that's from Google right agent to agent There was an ACP which was agent context protocol I think that was from IBM and they merged to A2A. There is one ACP from Zed that is called the agent client protocol And now there's another ACP that is the agentic commerce protocol.

Murilo: 19:16

And apparently, built this in partnership with Stripe.

Bart: 19:20

Yeah. To me to me, like like, it's a bit of a maybe a bit of a like, the comparison doesn't really make sense. Like, these are these are protocols meant for different things. Like, in there Mainly in those, they're overlap.

Murilo: 19:30

But to me, it's just like, it's always the same name.

Bart: 19:34

What is interesting, like, the Agenda Commerce protocol, it actually does compete with Google's. So Google does

Murilo: 19:40

have its own and it it it let me look. Jits payment protocol.

Bart: 19:43

Yeah. Exactly. Yeah. Yeah. Yeah.

Bart: 19:46

So there we will probably see a battle on what becomes No.

Murilo: 19:49

For sure. For sure. I think it's what I think is in I mean, they are for different things.

Bart: 19:55

Yeah. And this is very much based on on transactions. Right? Which is something very specific again.

Murilo: 20:03

But I also feel that since they saw how big MCP got and all these things, now they're doing protocols for everything related to agents.

Bart: 20:13

I think the difficulty on this is that you need to because why so shopping on Jet GPT, what I already do is that, let's say, I need I'm not very brand loyal when it comes to running shoes, for example. Okay. But I do have something, some requirements in terms of specific features. And then I ask because typically I buy shoes, like, once a year. And then a year later, ask Chachipity, like, do some extensive research which brands have shoes with these features, which is a lot of work if you need to do it yourself.

Bart: 20:47

Right? And then I get to a top three and I do some reviews on that. Yeah. And I think it would be nice if we can then just click say, okay, please also send it to my house. But the problem with this is a bit like with this very early days.

Bart: 21:02

It's like, what I will be hesitant to do is like, because they're onboarding Etsy now than Shopify. Shopify, of course, is way, way, way bigger than just Etsy. But it's still just Etsy and Shopify. So will we get a bit of a bias on what will be provided to you as as buying options? Right?

Bart: 21:21

Just because these are integrated with ChatGPT. Yeah. For I would worry about a bit.

Murilo: 21:26

Well, but I think if they do this protocol, they're trying to also nudge.

Bart: 21:31

Well, exactly. And the the that's a bit the point I was I was trying to make. Need something like a very dominant, open, transparent protocol, which makes it very easy for everybody to just say, okay. I'm gonna opt in to also sell on test GPT. Because to me, like, this buying on something on test GPT stands and falls a bit on, like, how many how many companies offer this stuff on ChatGPT.

Murilo: 21:57

Yeah. Yeah. I can like any protocol. Right? Like, if people use it, then it's valuable.

Murilo: 22:02

If it's a protocol that no one uses, it's not. Because also, like, by having this protocol, Claude could also do the same thing technically. Right? True. Gemini as well.

Murilo: 22:10

Right? And I think it's then you're really nudging people to to adopt this. Right? Which is true. Just I mean, maybe I mean, again, even if there is more would nudge it more towards Stripe platforms.

Murilo: 22:28

Even if Amazon implements this. Right? And they say, this is available. But I don't know right it's a bit I don't know I don't know I don't know if I trust enough let's say I mean I think there's also a few things like if you don't care like if you just get a shoe that you know meet your requirements and you're happy with it like you know it's going be a good shoe, know you're not going to be disappointed maybe that's fine but maybe if you want the best shoe then maybe also this is not the best tool for it right but then at the same time you cannot bypass the actual work of vetting all the things and reviewing all the things right it's kind of like I read somewhere like if you do interviews what you want, want people that are good, you don't want to sell the best, what you want when you're interviewing is like you don't want to have someone that is not qualified to get in but someone that is qualifying to get out it's okay and I'm wondering if it's kind of the same thing here right like when you're shopping maybe for you right if there is a qualified shoe or even a better shoe that falls out of the scope of search maybe that's fine because the other ones are also acceptable and I think for and if that's the case maybe using AI for going shopping all these things it's it's fine.

Murilo: 23:43

Right? Like, maybe you won't get the best price, but it's the shoe you want. It's a decent price and it's

Bart: 23:49

Yeah. To to me, major thing is is more like what I don't want is that that I just get suggestions of the things that are integrated

Murilo: 23:57

to buy What if the things that are integrated, like, meet your needs, though?

Bart: 24:01

But that that's a that's a that's a very biased view on the world then because then suddenly, like, your options to buy is just what is integrated with OpenAI.

Murilo: 24:11

Sure. But I think it's

Bart: 24:12

very hesitant towards something like that.

Murilo: 24:14

But would you like yeah. No. I agree. I would too. But I'm I'm I'm being a bit dev advocate myself, but you would never really know.

Murilo: 24:23

Right? Because OpenAI is not first of all it's a it's a large language model and the other thing is it's not open right so if OpenAI says yeah we have from different vendors but they're still routing 90% to one vendor you won't know right Like, if they bias more towards Stripe than Amazon because they want to, because they have a partnership or whatever.

Bart: 24:45

Yeah. But that's something else. That that to me is is like but then you're then you're actually talking about, like and that is not part of the discussion today. Then you're saying, like, I'm gonna pay more merchant fees from my shop to OpenAI for you to promote me when someone asks. That's not part of the discussion today, but you you do have that, of course, on on something like Amazon.

Bart: 25:06

If I ask if I pay extra, I get on the number one of the listing when you show search for shoes. But that's like, I'm I'm a bit making abstraction of that of the the paid promotions because today, they they do not have that. And, like, we'll see how how it's how it looks at the future.

Murilo: 25:24

Maybe have you have you ever bought anything from MS from in the open no. Actually, you never bought anything from OpenAI because it's very new. Have you ever gotten recommendations from OpenAI that you're actually very disappointed with?

Bart: 25:35

No. It's not that I buy so much based on recommendations from from OpenAI, But I do use

Murilo: 25:42

it to export stuff. But you never been

Bart: 25:45

I have this product, what our competitors, what our competing products why do they compete? What are the features? So when this comes out, you'll be

Murilo: 25:53

inclined to, at least giving it a try.

Bart: 25:55

If it's also something that I would need, yeah, definitely. I think it's easy. It will probably take a while for it to to land here in the in the EU. Right? It's just US based.

Bart: 26:05

It's just Etsy rolling out to Shopify.

Murilo: 26:08

So Etsy.

Bart: 26:09

I actually do quite frequently buy stuff on Shopify shops, so should be relatively easy to try.

Murilo: 26:17

True. Yeah. There's a lot of like, Shopify is pretty it's pretty big, and there's a lot of I think even if you integrate with Shopify, there's gonna be quite a lot of stuff that

Bart: 26:25

won't be huge. Yeah. Shopify is huge. Online shops are so many online shops that run on Shopify.

Murilo: 26:32

You know something else that is not available in The EU or at least not in Belgium? Sora. So OpenAI unveiled Sora two and an invite only iOS app that generates videos with synchronized audio and even self insert, quote unquote, cameos. The model nails tricky physics, missed basketball shots, rebound, not teleport, and an API plus Android app are quote unquote coming soon. If the feed catches on, TikTok style creation shifts from filming your life to prompting it.

Murilo: 27:08

Have you seen the videos from so sort two is out? Right. They had a Video generation model? Video generation model from OpenAI. Oh, sort of one was the first one, now they released the second one.

Murilo: 27:20

I haven't tried it yet. I wanted to try it actually, but I couldn't because they're not available in in Belgium. It does feel like a step up in terms of the physics, in terms of, like, the different videos. I mean, we don't know how much of this is edited. Right?

Murilo: 27:39

They also have a lot of funny videos, like you insert yourself or you insert famous people inside the video, right, and with audio as well. So I think you need to upload and you need to I think I read somewhere that you need to prove that you are the the person. Right? So you're not creating deepfakes about someone else. And and then it creates, like, some some very realistic videos.

Murilo: 28:02

I think a lot of it is from the videos that I saw. I imagine a lot of it is comes from video game data as well because the way that, like, the camera was moving and when they were running. But it actually looks really good.

Bart: 28:13

Yeah. And it looks very good. The angle that they're taking a bit with this is that it starts to look a bit like they're making a social network app. There are also some rumors on this. Is that it's that it will be a social network app and that it's really gonna focus on short, let's say, few seconds AI generated content that you can prompt and then share with the world.

Bart: 28:43

And that indeed you can also you can insert a Cameo of yourself. You can share that with people, and you can also tag other people in your prompt. So then it will use their Cameo in your generated video. Apparently, they try they are using it internally. So it's creating a lot of engagement.

Bart: 29:06

I'm not sure how to feel about this, like a purely AI generated social network?

Murilo: 29:13

Yeah. Yeah. I generate yeah. I don't know how how to feel about it. Maybe for people following the video, there's a quick ten second video of this is one of OP AI researchers.

Murilo: 29:23

They had searching for Bigfoot. So you see, like some some funny stuff I think it's fun but indeed like if it's a social network for this it maybe feels more like of a virtual art gallery kind of thing like people just sharing what they created with this tool rather than, this is my life. Right?

Bart: 29:46

I don't know. Because we already see, like, on Instagram, on on YouTube Yeah. On TikTok, there's already a lot of AI generated content. Often not very personal. This makes it easy to also make it personal.

Bart: 30:01

I wonder if there is what what if there is momentum to make something like this big. I don't think I would be happy about it. We also have, like, there's a there's a big uncertainty on on how TikTok in The US would look like going further. Maybe that's also part of the momentum for them to to introduce something like this.

Murilo: 30:25

Yeah. So, yeah, I'm

Bart: 30:28

the videos look impressive, but I'm wondering how they're gonna what they're gonna do with it. And I also but this is just I haven't tried SORA two. The videos have looked very impressive for the last, I want to say, eighteen months of Sora, of their competitors like Vio and but when you actually, a few weeks ago, I tried out Vio three, the Google competitor. And it's actually when you actually try it out, it's a bit disappointing. Like, I think on these introduction announcements, like, they always pick out the best videos.

Murilo: 31:08

Yeah. Yeah. Yeah. So it's not like lying technically, but it's like you generate 10, you pick the best one, then you get something impressive, but like the other night I'm meh. Yeah.

Murilo: 31:18

The other thing I tried Vue actually the other day and it's only like ten seconds or less.

Bart: 31:25

Mhmm.

Murilo: 31:25

Yeah. It's short. Right? And I I was short. And then

Bart: 31:29

then they have so for example well, but then in the Google ecosystem, you have Google Flow, which is bit their video production workflow, and you can use the vo three there. But you can only generate, like you say, like nine seconds or something. And from the moment that you want to extend that, you can't use the vo three model anymore. Only vo two supports it. And then the quality becomes way worse.

Bart: 31:58

So, like, there's like, the actual, like, the the actual implementation, like, lags a bit the the impressiveness of the

Murilo: 32:06

I see. I see.

Bart: 32:07

I see. I see.

Murilo: 32:08

But then it's like if you just follow-up on that same video. So again, like, I I tried vo three as well. It like the first prompt. It can be vo three, but then if you could continue the conversation to say generate more shots or something, it's only vo two. Yeah.

Murilo: 32:19

Exactly. That's a bit that's a bit disappointing. One thing that they also mentioned, I don't remember where I saw it. Maybe it was on the announcement that they also have a bit more consistency on the on the different shots, so sometimes they have the same image showing from different angles. And now Sora two will be more consistent of what's the position of the person, what's the face of the person, kind of the same thing that we saw with the Genie three with the real world models.

Murilo: 32:50

Right? You can look away and look back, and it's the same. It's consistent. I think also they they had worked on this. So, yeah, I I think it would be very cool to to give it a try, but I think as you were discussing it, it is maybe bittersweet.

Murilo: 33:08

Right? Like, if you see what what's gonna happen with this. Right? Is everything gonna be just now AI generated? Right?

Murilo: 33:15

But I also saw that they're gonna release they also did put some watermark for the videos as well or somehow to try to distinguish between what's AI generated and what's not. So yeah. Yeah. Curious to see. Well, I wanted to play with it.

Murilo: 33:31

That's that's more realistic. Like, I wanted to play with it once, but wasn't available. Alright.

Bart: 33:38

We have Entropic released Cloth Solid 4.5, pitching it as a top coding model with major gains on reasoning computer use and math. It leads OS world benchmark at 61.4%, hits 77.2% on the SWE bench verified, keeps VOXX for thirty plus hours, and stays, $3, per input, $15 per million output tokens. There are some new goodies, plot code checkpoints, Versus code extension, and an agent SDK that aims to turn those scores into real developer speed ups. Yet a new model, I want to say, only like two or three months after they released, the 4.1, I want to say.

Murilo: 34:22

Yes. I think so.

Bart: 34:27

Again, improved benchmarks, and this is the today, the the top model for coding, which I think Anthropix models have been for quite some time now.

Murilo: 34:39

Even though I heard when when it was, like, the 4.1 and GPT five, I have heard from some people the GPT five, they had more success on GPT five for some things. But I think now that's a bit my expectation as well. The 4.5 is the

Bart: 34:57

Yeah. So maybe next in a new model. There's also some updates to Cloud Code. Cloud Code is the recording CLI. It now has integrated checkpoints.

Bart: 35:11

It's it also has an agent SDK, so you can basically like like, ClotCode is, you can also basically make make your own agents very easily. It has a native Versus Code extension, which works quite well.

Murilo: 35:23

You tried it?

Bart: 35:25

You tried it. Yeah. Solace 4.5 so it was on on the previous version before. I'm building an electron app. It's maybe a bit specific.

Bart: 35:37

I can't really say that I've noticed the difference, but I'm also, like, just working on this one specific use case. Okay.

Murilo: 35:45

And using Cloud Code for this? Or

Bart: 35:47

Cloud Code. Yeah. Yeah? So I haven't yet had the wow, this is so much better effect. But Yeah.

Bart: 35:57

But you use it. But I'm using it.

Murilo: 35:59

Yeah. We I mentioned earlier, right, like the agent client protocol. And I wonder if this is how they integrate it as well because agent client protocol something from Zed which is another IDE out of work from Versus Code written from scratch And they also introduced this agent client protocol, which from what I understood is like, maybe that's why I wanted to check with you. So it's like now with integration with Versus Code, you just have like a chat pane but then like when you when you talk to it the changes are you can easily see the changes on your terminal right so it's maybe more synced so before you have your terminal open on one window and you have the your IDE open another window and then as code is making changes on your code base you would just have to like go and click and see where it did the changes and review them and now it's it's easier I guess did I understand that well or what does it mean to integrate with Versus Code?

Bart: 36:59

I think before, it it just didn't have a native extension themselves.

Murilo: 37:04

But what does the native extension do? Like,

Bart: 37:07

why a chat pane.

Murilo: 37:08

It's just a chat pane.

Bart: 37:12

But I also heard it's like these these we we just said it like there. It how it keeps focus for thirty plus hours that it's very autonomous. It, can, make decisions on what what what's next. What should I focus on? Should I focus on test?

Bart: 37:29

Should I focus on integration test? Should I focus on setting up a database? Sure. Like, there are very impressive talks about this. But like I said, I haven't really seen this in practice.

Murilo: 37:42

But to keep focus is, like, focus on the CICD, and then they'll only touch CICD stuff. That's that's what I'd say.

Bart: 37:49

How do you exactly mean it in the in the the formal presentation? I don't know. But but but what mean is that they've heard that they've let it code autonomously on a certain code base for a very long time.

Murilo: 38:02

To me, it's also a bit weird that they use hours.

Bart: 38:05

Yeah. Yeah. It's a bit of a weird

Murilo: 38:08

It's like if you're just thinking for you're thinking for for twenty nine hours and you're actually changing for one hour, it's also a bit, you know, like. Yeah. I don't know. Maybe it's a bit. Maybe it's just me.

Murilo: 38:20

Maybe it's but cool. I'm actually very curious to to try this, especially with the Cloud Code. So did you did you notice any difference on Cloud Code, the the app itself, not the models? Or because now this is the checkpoint. I thought it was more like if you wanna go back to a state or something.

Murilo: 38:38

But

Bart: 38:38

Yeah. It's actually it's that. Yeah. I haven't used that feature yet, but it's in the app in the in the in the CLI and detail.

Murilo: 38:45

Cool. Another model's out there. And what else do we have?

Bart: 38:49

Somewhat related. Right?

Murilo: 38:51

Somewhat related. We have Google launched a public preview of a Chrome DevTools MCP server so AI coding agents can debug real web pages not guess. Agents can start performance traces, inspect the DOM and CSS, or check LCP. Try quote unquote please check the LCP of web. Dev.

Murilo: 39:16

If widely adopted, MCP could give every dev's AI assistant real browser eyes. So dev tools. You're using Electron app? What was the Electron app? What is it doing actually, Bart?

Bart: 39:30

It's a sort of a text editor. Text editor. Okay. But this this news, the the DevTools MCP, I'm I think this will mean big step up for AI assisted coding. I've tested it out yesterday.

Bart: 39:46

So, basically, what it is is is that it is a MCP server that you very easily integrate with whatever you use. Right? Like, if it's ClotDesk or if it's ClotCode or whether it's Gemini CLI or Codec CLI. And what gives access to tools like start up this browser window and allow me allows it to interact with it to really with basically not just the web page, but more the dev tools Pane so it can see what is in the in the console. It can actually execute stuff in the console.

Bart: 40:26

It sees how long it takes for it to render stuff like this. And it makes, which I think is extremely important for good performance of these tools. It makes that it gets a lot more feedback on change that is done on the code immediately by testing it out in the browser. And of course, that also immediately limited to stuff that happens in the browser. So this is very much something that is related to a bit to the web ish development scene.

Murilo: 40:55

When you say what do you mean by it it reduces what happens on the browser?

Bart: 41:00

No. It's linked to what happens in the browser. Like, it's limited to what happens in the browser in a sense, like, if you're if you're making, like, a native iOS application like it is.

Murilo: 41:10

Yeah. Yeah. Yeah. It

Bart: 41:13

doesn't help there. Like, it really helps if you're making a web app, basically.

Murilo: 41:16

Yeah. Yeah. Okay. No. No.

Murilo: 41:17

I get it. I get it. I get it. Yeah. No.

Murilo: 41:19

It's cool because before you had to use something like headless browsers, I guess. There are MCP tools for that, like Playwright and all that. Yeah. This goes a step further. Right?

Murilo: 41:30

Like, because you have the the dev tools there, which is something that the other MCPs don't don't have. Correct?

Bart: 41:35

Yeah. Exactly. Yeah. Exactly. And dev tools, like, it gets very immediate feedback.

Bart: 41:44

Well, what you saw in these, like, with other MCP tools like Playwright, like, it works by taking screenshots of the rendered page, and it's typically also very slow. So I tested quickly out yesterday, and it was I was very enthusiastic about it. Think You can

Murilo: 41:59

already use it. It's already out. Yeah.

Bart: 42:00

You can already use it. And it's just like, adding a few lines to your MCP server setup. It's made also from that point of view, like the developer experience is super simple. This is really cool. This is really cool.

Murilo: 42:13

I can see a lot of, it works really well. It would be very easy to say automate this by, like, create a script for me that will log me in with this and this and then do this and that every day. Right? So maybe there's not an API available and you want to, like, now

Bart: 42:30

To really automate stuff for you. Yeah. And to really thought about it like that. So so it was but I think I guess you could do that because you can interact with the console. What I what I typically use it for is is really, like, to just when you're working on a web app to really get very fast feedback on changes to the code you've done, what are the results in?

Murilo: 42:54

Mhmm.

Bart: 42:54

And it, like, it makes your context when coding through an AI tool much richer because it immediately, you know, it's like I I changed the line of code, but it's also like, it give the right result or not? Normally, need to wait for the user to say, oh, no. It doesn't look good. Or I see this error and you need to need to use it to have it pasted in. Yeah.

Bart: 43:19

So you get these much faster feedback loops. Thank you. But I think in theory, you could use it to to automate stuff because you do have access to the to the console. Right?

Murilo: 43:29

You have access to everything. Right? The console, DOM, everything. Cool. I'll definitely give this a try for sure.

Murilo: 43:36

What else? What's next part?

Bart: 43:40

Spotify is cracking down on AI slob and clones and pushing a DDEX standard to disclose when AI touched a track. It yanked 75,000,000 spam uploads in twelve months and will roll out new impersonation and spam filters within weeks. Clear labeling and enforcement could calm artists and shape how AI music gets paid and playlisted. It was about time. It was about time.

Bart: 44:04

Yeah. Yeah. Interesting. We we discussed, though, already a few times. Right?

Bart: 44:15

There is a lot of AI slob being generated in the music scene. If I remember correctly, we were saying that there were as there were estimations that, like, 18% of new 18, like, 18% of new uploads were AI generated music, which is a lot, which is, I guess, a significant problem, especially when it's about cloning existing artists. I would say then it's definitely a problem. Yeah. I think the ethics on the other one, like on, quote, unquote, original and music are maybe a bit different, but when it's about cloning, it's it's definitely a problem.

Murilo: 44:50

Yeah.

Bart: 44:52

And what they're now doing now is that that they are trying to build a standard. It's called DDEX, and I'm trying to find what it stands for.

Murilo: 45:07

DDDX. So it's here, like, the standard setting organization. Actually, I'm not sure what it means.

Bart: 45:14

And it's basically sort of a sort of a watermark that says AI music. And this music is AI generated.

Murilo: 45:21

Mhmm.

Bart: 45:22

Stands for the it's a digital data exchange standard. That's probably some metadata to to add in whether or not there's AI involved or not.

Murilo: 45:35

And then the people that upload it, they will need to add the metadata or they're working also with the music generators like Sunu and Wudu?

Bart: 45:45

Well, what I would think, but I'm not so sure what I would think is that if you upload music that you are responsible to mention whether or not it's AI generated, what you will probably see in the future is that tools will do that for you.

Murilo: 46:00

Yeah. Yeah.

Bart: 46:01

Alright. So recently actually actually, Sooner, so with Sooner, you can generate AI, tracks, but they actually released a Sooner Studio, I want to say, last week, which is really a digital audio workstation to really go to next step in in generating music based on Gen AI. But I would assume that tools like these will start doing this for you going forward. Indeed. And then also with things like if you abuse it and if you still upload stuff that is AI generated, it's not correctly attributed, then you probably get banned or something.

Bart: 46:40

Right?

Murilo: 46:40

Yeah. Yeah. Yeah.

Bart: 46:42

But I think the difficult thing is here is also like that it's difficult to to detect whether or not something is AI generated.

Murilo: 46:52

Yeah. I was thinking the same thing. That's why I was wondering if Sunu and Udi are also in on this to try to at least add the metadata to the stuff that they produce, like, kinda like a watermark as well on the on the music. Right? It would be interesting to see if this goes through to see actually how many are flagged as AI generated so yeah I will also discuss not that long ago how there are some actual AI generated artists and I generally music artists I guess So, yeah, I think the whole whole space is is interesting.

Bart: 47:28

But the impersonation is also very it's very hard to solve because it's very nuanced. Right? Like, if let's say you're an, quote, unquote, AI artist, you make everything with Suno Studio, it is original content, quote, unquote. Yeah. Yeah.

Bart: 47:46

But it is very close to, let's say, Moby's style of

Murilo: 47:50

music. Like,

Bart: 47:51

from when on does it become impersonation versus not versus original content. Right?

Murilo: 47:58

Yeah. Yeah. Like, if I go to Sunu and I say, write this song in the exact same style as Taylor Swift and do this and this and this, and it's like Yeah. How much of this is actually you and how much of this should be attributed to the artist? Right?

Bart: 48:10

Interesting that you immediately jumped to Taylor Swift.

Murilo: 48:12

I mean, I know. I just just what came to mind.

Bart: 48:22

But but even, like, if you don't have the AI tools, you could still do that. Right?

Murilo: 48:26

Yeah. True.

Bart: 48:26

And then you're just gonna say, I'm inspired by. That's that's probably okay. But so it's a very it's a very difficult and nuanced line there, I think. But, like, to be true impersonation when you say there's a cloned voice, I mean, that needs to be banned. Right?

Bart: 48:43

Yeah. Yeah. Yeah. For sure. Simple.

Murilo: 48:44

Yeah. I mean, it can yeah. It's a bit of a spectrum. Right? Like, on the impersonation end of the spectrum, okay.

Murilo: 48:50

It's clear. No. No. But, like, at some point, you need to say, okay. That this is okay from this point on, and where is that point, and what are the arguments for it?

Murilo: 48:59

What you said about now, like, about being inspired is the same it's the same kind of argument that I like, well, not the same argument, but what it made me think was the the Anthropic case. Right? How if you bought the books and you read it and then you spit out information based on what you read, that's fine. So if Anthropic had bought the books it would have been fine so the mediation is just downloading legally you know and it's like you know the thing that gives me a weird feeling is because we're also comparing AI to humans and I'm not sure if the comparison should be made like that you know maybe we're comparing apples to oranges but but we're okay with it kinda you know I don't know

Bart: 49:43

to me it's more like when we talk about this year the music generation is more about correct attribution, which is very hard to do in AI. Like, take take, for example, beat making in hip hop. Like, they use a lot of samples that are from existing

Murilo: 50:01

Mhmm.

Bart: 50:02

Musical other artists and reuse those samples. But from the moment that you want to monetize that, you need to also give the correct attribution through to the original songs that the samples came from. And, also, there there's a there's a flow of payments to the original source. And that today, we haven't solved for AI. Because I I think if we would have solved that, it would be much easier.

Bart: 50:30

Right? Like, make a song in the style of Moby. And and some You do have a and you do have this clear and transparent attribution and something and and you get you have a a top list listed song. Someone that should end up with Moby. But that's that's not what we have today, and that makes it very difficult, I think.

Murilo: 50:50

Yeah. But but I think, like you said, even that is a bit of a spectrum. Right? Because if I just say, describe the music of Moby, and then include

Bart: 50:57

me, like, the description does again, like, is it is it inspired by versus

Murilo: 51:00

Exactly. Exactly. Yeah. It's a Yeah. It is a difficult one, but at least we're talking about it.

Murilo: 51:08

At least there are steps towards this. I think like we said we can maybe all agree that there are cases that definitely should be flagged and stopped right or at least be transparent so this is a start. Alright and last but not least, definitely not least, we have Databricks struck a minimum 100,000,000 multi year deal to bring OpenAI models including GPT-five natively into its platform AgentBricks. Models will be callable via SQL or API, and Databricks pays the 100,000,000 even if usage misses the revenue target. It's a bet that putting frontier models where enterprise data lives will unlock real adoption and justify the tap.

Murilo: 51:50

OpenAI with Databricks? Is it the first well, I know Databricks runs on other cloud providers, but is it the first time that data that OpenAI is moving away from only Microsoft?

Bart: 52:07

What sense Microsoft? Like, Copilot?

Murilo: 52:09

They had, like, a partnership in the big mean, part Microsoft also owns part of OpenAI, or am I mistaken?

Bart: 52:14

They've I don't think it's all direct ownership, but they

Murilo: 52:17

have a big stake in those. Yeah. They're staking it. Because for a while, and I think now maybe now that I'm thinking about it, don't think it's true anymore, but for a while we would see more OpenAI on Azure. Right?

Murilo: 52:30

I think now we're starting to see more OpenAI in, well, Databricks now. I would imagine it's on AWS as well. So for me, that was the first thing that I was thinking, like, okay. OpenAI is now just OpenAI. Right?

Murilo: 52:41

It's less less linked to Microsoft.

Bart: 52:44

Yeah. I think the trend we've seen in the in the last years.

Murilo: 52:48

Indeed. And, actually, have you ever used AgentBricks? I haven't used it yet.

Bart: 52:54

No. I haven't. I haven't. But I do think because what what they're trying to do here, and I think that is interesting, is that they really want to give users of the Databricks platform, like, very direct access to to OpenAI. And that means that, like, you have a SQL query that you've made where you could say something like, select, this text column and create another column with a summary of that directly in your SQL.

Bart: 53:24

Yeah. And that does look like oh, refrain from going into the robustness or that testability with that. But it looks like like a very useful feature.

Murilo: 53:36

Yeah. Yeah. For sure.

Bart: 53:37

Because normally it takes basically to do that, which a lot of people are doing these type of things. It takes a lot of extra steps, and they're taking these these steps away from the from the developer than to just go directly to the source.

Murilo: 53:50

Yeah. They're short putting

Bart: 53:51

the tools that you're already using. So that's looks like a from a developer point of view, looks like a useful feature.

Murilo: 53:59

Yeah. One thing you mentioned, like, using natural Android on SQL, like, think maybe already some years ago, I remember that Spark had the English SDK. Yeah. That was, like, either spark dot a I dot or parentheses or something, and then you just put something natural language, like, when you say, like, create another column, and then this when you run the Spark code, the transformation would also happen.

Bart: 54:22

Yeah. Yeah.

Murilo: 54:25

Different flavor of this, I guess. Right? And from what I understand as well, I'm also just looking here, like, agent Bricks is let me share again the screen. Production AI agents optimize on your data. So I think it's also maybe like a local thing to do data exploration and all that and, like, bringing these models closer, I think, would make it easier, like you said.

Murilo: 54:48

So I think it makes sense. A 100,000,000, to be honest, like, all these big numbers, it's hard for me to quantitate

Bart: 54:55

big numbers. A 100,000,000 doesn't sound like much anymore.

Murilo: 54:58

Yeah. It didn't, but, like,

Bart: 55:00

I feel like it's

Murilo: 55:02

I feel they don't sound like a lot because I feel like we're a bit numb to to these big numbers, but there's still quite a lot of money. Oh, yeah. So so, yeah, let's see. Let's see. And I remember when I shared this I don't know if I I think I shared this.

Murilo: 55:17

You were also saying you're bit fatigued of AI everywhere, AI embedded everywhere. So I think you have you had a little story about this as well.

Bart: 55:25

Well, no. Well, I when I when I sent that well, I just created a new account on Airtable, and you need to go through this whole AI assistant onboarding where I like, Airtable is a bit like people that don't know it, like, it's a bit like Excel on steroids and, like, your base. So it's basically the the your spreadsheet. And I'm just like, give me a new give me an MPP spreadsheet, but I'm, like, forced to go through this whole AI based conversation to get to something. And then they want to also not just give you the empty base.

Bart: 55:59

They want to set something up for you. So you need to give input on the rows and then, like and this whole conversation style, well, I just want to do three clicks because I knew what I wanted to do. Yeah. Yeah. And then it becomes a bit annoying.

Murilo: 56:11

Right? Yeah. You have to. You have to go through the to the to the AI bot. Right?

Murilo: 56:16

Yeah. I know what you're saying. I also had the

Bart: 56:19

Like, the optionality is a good thing. Right? Like, you have needed the option to do it, but you shouldn't be forced to do it.

Murilo: 56:24

I think yeah. What would I want? What was I trying to do? I think I was trying to change the email for the invoices for a tool, and I was like, just wanna do that. I, like, I went to my settings, and it's like, There's no email for invoices.

Murilo: 56:37

Yeah. Yeah. And then I was like, okay. Let me let me Google. Like, okay.

Murilo: 56:40

Frequently asked questions. Like, no. There's nothing. And then it's like, turns out I needed to talk to their AI chatbot. And then they have to say, oh, can I help you with something?

Murilo: 56:48

I was like, no. You're like, I wanna do this. Like this is not possible, I want to talk to a human, it's like, no, but you have to do this, like you have to talk to me first, it's like no, but want to do this and then like there's no contact email, there's nothing you can just reach out to them and then you have to go through the steps or like is this what you want, no it's not, is this what you want, no it's not, I need this, I okay then I'll escalate to a person and then that person will will look at the conversation and and help you with it and in the end it worked but like I still had to to do the little AI chatbot dance. Yeah. Okay.

Murilo: 57:16

Yeah. Yeah. So and I at one point, it frustrates me a bit. On the other, I also understand if you have like, you don't have a big support team. Right?

Murilo: 57:24

Maybe you wanna force all the users to go through the the chatbot, but to me, I was a bit like, I know exactly what I want, and I know I cannot do it. Just, give me a person or just send this to someone or just they can just send a message. You know?

Bart: 57:35

So Yeah. Very recognizable.

Murilo: 57:37

Yeah. Yeah. But I think, yeah, it's the world we live in now. This is the new normal, I feel.

Bart: 57:45

This is the new normal.

Murilo: 57:49

Alright. That was it for today. Anything else that you want to say before we No? Call Alright. Then thank you everyone for listening.

Murilo: 58:02

Feel free to leave a review. Reach out to us anywhere if you have any ideas, if you'd like to be here, if you have any feedback. Thanks everyone for listening. Thank you, Bart. Thank See you all next week.

Bart: 58:13

See you next week.

Murilo: 58:14

Bye. Bye.