Agents, Paywalls, and Browser Battles: The Race to Control AI's Digital Highway
Hi, everyone. Welcome to the monkey patching podcast where we go bananas about all things AI for good, enemies, and more. My name is Morello, joined as always by my great cohost, Bart.
Bart:Hi, Morello.
Murilo:How are you doing, Bart?
Bart:I'm doing good. I'm doing good.
Murilo:Crazy weekend. It's been a lot of news, I feel.
Bart:There's been a lot of AI and data news last week. Yeah. It's been hard to follow everything.
Murilo:Had to make some editorial issues. Browsers, regulations. Yeah. A lot
Bart:of different things.
Murilo:Pool rug pools. Yeah. A lot a lot of stuff. So let's get started. Let's get started.
Murilo:MCP proposes USB C for AI, letting agents call site functions instead of clumsy click automation. And I quote, MCP B gives AI direct access to your website's functionalities instead, promising a smoother bot web handshakes for everyone.
Bart:Yes. So so MCP stands for model context protocol for the browser, like you're showing on the screen here. We've we're hearing a lot of chatter on on AI enabled browsers these days. I think 2025 will be the the AI browser here. This takes a bit of a different approach.
Bart:So what they do is that they that they have developed a browser extension that allows you with a very minimal setup at the websites and to inject a little bit of JavaScript code and then through that, expose your API VM to model context protocol.
Murilo:And your API is the website content? Or
Bart:The website. It can be website content, can be back end servers. You decide a bit. Okay. And how how it works is that you need to to install this.
Bart:I think it's a Chrome extension. I think they only have support for Chrome at the moment. And then your MCP clients, like Clot Clot for your desktop or for your desktop, They or your or your coding, whatever. They can run you they can use this FCP server in your browser, which is basically this extension to leverage these functions from your website directly. So instead of having to what most AI able browsers, like, take is this approach where you have this more of an RPA setup, this robotic process automation where you basically have to simulate interactions with your browser.
Bart:You need to scroll down a page. You need to read the content. You need to click buttons, etcetera.
Murilo:It's like the the Puppeteer or the Playwrights. Exactly. Like, has its browsers. Right? Yes.
Bart:Indeed. Yeah. Indeed. And here, they basically, through this through this extension, you have basically this proxy directly to the functions. So your website becomes basically this model context MCP server that is then that you can then directly interact with, which arguably is a smoother integration, but you need to do extra work to enable it.
Murilo:Yeah. Indeed. So for example, Wikipedia. Wikipedia is also a static website. So now you need to have something that list listens to MCP calls.
Murilo:Maybe it's gonna be a more complicated setup. Maybe it's gonna be more expensive as well. But then you don't have this clunky automation, and everything is one place as well.
Bart:It's centralized how you how you expose the functions of these. Yeah. Yeah.
Murilo:Yeah. Because otherwise, if you didn't wanna do this, you would have to have your website and create an MCP server, like an external MCP server, and I try to publish it somewhere so people can find it and try to integrate it. Right?
Bart:You need to. Difficult thing of this, of of course, is that it's still you need to have this this extension installed. While with let's say, if you would use a sort of a headless browser approach to basically go through your web page, simulate things like you don't even have to have a browser visible.
Murilo:Yeah. Indeed. And I'm also wondering, there are some peep I mean, there are some people that don't want to expose things, like even APIs. Right? Doesn't need to be MCPs.
Murilo:So I think browser is, like, the only safe way, quote unquote, that you always will get the content. Right? Because some people are not gonna play this because they don't want to. Some people are not because they're behind, quote unquote. So yeah.
Murilo:But I think it's it's it's nice. Nice nice idea as well.
Bart:It's a nice idea. I think the the the difficult thing, and I think this is also like the more that's why they pitch this as a USB c for AI, like, with a bit of this this this universal protocol to do these things, these interaction with the browser, but at the same time to actually start using this like everybody that needs to start using this. So it's very big gap to fill to to make it actual actually be usable.
Murilo:Right? Indeed. Yeah. If you have a standard and no one adopts, then Yeah. What's the what's the point?
Murilo:Right?
Bart:And the standard today is it's like manual interaction with websites, whether it's a human or a bot. Right? But it's a it's a it's a manual interaction.
Murilo:That's true. That's true. Yeah. And I think it makes sense as well. Again, I think I think it's trying to solve that pain.
Murilo:Like you mentioned, people are using the headless browser to say click here, do this, do that. But, yeah, it's nice. Even though I coded a bit with, Playwright because I wanted to automate some clicks and stuff. And it feels a bit magical, you know, like, you put the MCP server and then you can kinda watch. Right?
Murilo:You can show me the screen, then you can use it, like, typing itself as well. It's like a little like a little dude right there, you know, doing stuff for it. So yeah. Cool. Have you have you tried this or have you do you know any website that actually implement this?
Bart:No. I haven't tried it yet.
Murilo:It's probably too soon as well. Right?
Bart:It's early. Cool. Related to this is OpenAI because AI is reportedly cooking up a Chromium based browser with a built in operator agent that can, for example, book tables or fill forms for you. And, quote, OpenAI is planning to launch an AI web browser in the coming weeks, teeing up a fresh duel with Chrome and comments. So this is really directly in browser wars news.
Bart:This is Yeah. This is not an MCP extension. There are rumors that have not been formally confirmed that in the coming weeks, OpenAI will come with a browser, the wrong version, which I guess makes sense. What do you think, Manuela?
Murilo:I mean, I think it makes I don't know. I feel like OpenAI is trying to to branch off a bit from the AI provider only, you know, we didn't like I think we I think we did talk about last week about the consulting business, you know, browsers. They're also we're in conversations to acquire IDEs. I think if you think like that, then I think it makes sense. Browser for AI, if it's not MCPB or something like that, I'm I'm I'm not sure.
Murilo:Well, maybe it's probably like automations. You can talk to the website. Right? Like, you can talk to open a site pane and say, what is here? Or book something for me or do something like that.
Murilo:I can see things like that.
Bart:Yeah. So the the chatter that resonates is that they will have their their operator agent, which is basically a smart agent. They can do stuff for you. For example, booking tables, very tightly integrated into this. So it will be able to very robustly simulate interaction with websites.
Bart:What I also understand is, like, it will probably have next to your website. You'll have this chat interface. And why I think it makes sense is if I just look at my own ChatGPT usage, like, often data actually comes from the worldwide web that I use. And behind the screens, they're being crawled, whatever the research is being done. But probably when you have this side by side, you can say, ah, stop, but also follow that link.
Bart:I have these these three tabs open. Make sure to keep them keep that in mind when you when you research this topic.
Murilo:I
Bart:think it's a very like, the the Worldwide Web is a very logical source of data for probably, like, at least 50% of what you do typically during a more extensive session. So I think there it makes sense to to have this integration be smoother, but also have the user closer to that.
Murilo:Yeah. That's true. I do think you solve a big chunk of problems that people probably have by being on the browser. Yeah. Like you said, if you have different tabs, you can manage the tabs, you can manage the context.
Murilo:So I think it could be very interesting. I wonder because there are there are other projects for this as well. There's one I heard Navado or something. There is a Google actually announced this a while ago, like, that it was had a project project Mariner, I wanna say. I mean, you also talked about AI AI browser, but you also said
Bart:you tried Comments from perplexity.
Murilo:Comment from perplexity. Indeed. So but actually, the thing as I'm saying these things as well, I've I have the impression that none of them really stuck. So which makes me think that it's actually a harder problem that I'm making it right now. So I'm curious to see what's gonna come up with with OpenAI.
Murilo:Right? I think so yeah.
Bart:I think the advantage that they have is that they they have the model. A lot of people already like, if you have a JetGPT license, you will probably be able to use all the power in the browser. And, like, they will also be able to optimize the integration with regards to their own model. So I think, yeah, like, you will have good performance out of the box, and you already have a willingness to pay because you're already paying for JetGPT. Like, you have a lot of these browsers where you need to bring your own key.
Bart:It's much more complex. Like, it's it's Yeah. Yeah. Yeah. You think about the extra cost.
Murilo:You need to be also a bit more developer y. Right? You know, API keys, what is this, how do I get it? I just wanna use it. But do you think they would just add it as a you think you need to pay for it additionally, or do you think they would just
Bart:I doubt it. I don't know. We'll see. Right? Or Not even conformally confirmed that it will be a browser, but I doubt it.
Bart:I think it will be part of the TCPT pro license. I think that makes sense adoption wise, especially because, like, the the data that people generate just by interacting with this will also be valuable for future model training. Right? It's not just an extra feature. It's also a a new source of data.
Murilo:Yeah. That's true. But
Bart:then I think it's also I think what you will what you will see a bit is that OpenAI is not only pushing for improving my models, also, like, pushing the whole agentic ecosystem. And this is like getting a step closer to, like, OS level automation. This is not operating system level, but this is for very close to, like, having very good automation of the browser from the moment that you have actually have a browser.
Murilo:Yeah. Yeah. That's true. And if you think like that, like, you're closer to the clicks. Right?
Murilo:You see more the Yeah. Environment. And I think when you think like that, it makes a lot of sense that they were interested in the ID. Right? Because it's a bit the same.
Murilo:Like, you're closer to the to the action, let's say.
Bart:Yeah. True. True. Yeah. Another another discussion, but it's maybe not one to have here is is privacy, of course.
Murilo:That's what was thinking as well. Because they said they're collecting this data to train. Imagine the simple example for agents and even the agent to agent protocol was, like, booking stuff. Right? Like, you wanna plan a trip.
Murilo:So one agent plans the trip, one agent buys the tickets, one agent books the hotel. And I can see if you have something like this maybe simplifies a bit because you can just have multiple tabs and then the agent has context with all the tabs and you can kind of, you know, bounce between one and the other. But then, yeah, if you have, like, personal information and financial information, if they're gonna be collect like, it's a bit it's way more sensitive. Right?
Bart:Yeah. I I fully agree. Although what we've seen in the in the last years is that data privacy is not very much a high priority, and I and I doubt it's a high priority. Like, it's the 90% level of the users of JGPT.
Murilo:Yeah. Which I fear I
Bart:fear that that true data privacy is only a concern of the people that really understand the technology and understand potential implications, which is which is not a lot of people. Like, if you look at the the general populations who who uses Chechibati or or browse the Internet. Right?
Murilo:Yeah. I think I think you will be like that until something happens to enough people that it catches enough attention and it becomes more more mainstream. Right? But, yeah, indeed. I I don't see a lot of people really concerned about even, like, having a a second thought, even people that are more knowledgeable, quote unquote.
Murilo:You know? So I think it's not even just, like, about the know how, but I also think people need to be be may maybe train the muscle of having a reflex of but this is private data, so you shouldn't use this. Maybe I should use that or maybe I should you know? So yeah. Let's see.
Murilo:Let's see. Hope I'm hope I'm wrong. Hope I hope nothing nothing bad happens, but I think the the easiest example I can think of is, like, your own, I don't know, and you ask, oh, give me the credit card information from someone or and then it's, yeah, there you go. You know? Or
Bart:Yeah. Yeah. Exactly.
Murilo:Like, you're typing code, and then the autocomplete just fills in, like, keys and stuff and credit cards. You know, that's the that's the fear. That's the fear. And So what else do we have? What else?
Murilo:Swiss reach researchers will release an entirely open supercomputer trained LLM fluent in more than 1,000 languages later this summer. Project led project lead, Emmanuel Schlag says, and I quote, fully open models enable high trust applications and are necessary for advancing research about risks and opportunities of AI.
Bart:Yeah. So this is this is a new LLM that will be released somewhere later this year. Late summer is stated to release. It's a collaboration between ETH Zurich, University of Zurich, and EPFL, which is the University of Lausanne, both Swiss universities. They are using a Swiss supercomputer called Alps to train a new LLM on, basically.
Bart:It will be open source. It will be Apache two licensed. It will have fully open weights. The data will be documented, if I understand correctly. I don't think it will be available.
Bart:It will be trained on more or less 10,000 GPUs. Oh, wow. If I understand correctly, they are using the Hopper h hundreds from NVIDIA. It will be fluent in 1,000 languages.
Murilo:You said it will be. It will be, but, like, I'm assuming that it's already at least partially trained
Bart:or fully trained. Assume it's in the process of training. Yeah. Because there's also quite some information available. There there will be two versions of the model.
Bart:The ones will be an 8,000,000,000 parameter. The other one, the large one will be a 70,000,000,000 parameter. So it's a very, like, a white type of user needs slash base that it could potentially serve. The it's also the idea of this is also to have a more sovereign model. Right?
Bart:Someone like a model that, at least from a Swiss point of view, is fully independent. And the the researchers, they also argue that this is key to have this kind of model, this kind of sovereign model for, like, high trust applications and also to be be able to be compliant with the European AI act. So it's it's will be interesting interesting to see the performance of this model.
Murilo:And maybe now if we had to put this next to Llama, so Llama is also maybe maybe one difference is that it's less open maybe, because I know Llama is you can use it, but I think the license is a bit different and the data is not well documented. Like, you don't have ex it's it's like you can use it and it's the closest thing to open source, but I guess this takes another step further.
Bart:This takes another steps further. This is this is this license is actually Apache two. It's it's an actual open source compatible license, and a much more open Asia because the data will be like, the whole if I understand correctly, like, they want process they they want to document the whole process of training the model so that you could potentially replicate it.
Murilo:Yeah. Yeah. Yeah. And I guess maybe it's like you mentioned the languages, so I guess that's also focus. People will be more proficient in more languages.
Murilo:Anything else as well? Any other differences that we can highlight between what we already have and know, which is drama in this?
Bart:Not really. I think we need to see wait and see a little bit. I think the the interesting thing thing is that I think it's the at this point, one of the only serious attempts at this. Like, there have been a few documented attempts also from, let's say, local governments to build smaller LLMs. Like, this is something that that should compete with the larger ones out there.
Bart:Right? I think it will be difficult to do, to be honest, and it will be interesting to see where it lands on the benchmarks. But if we have something that is at least a GPT four or maybe in four o
Murilo:Yeah. Yeah. Yeah.
Bart:Like, at that level of performance, I think that would be very strong. I think what we see here is that is that the way to build these models, train these models, like, it's still very much an an expert task, but it's something that is reproducible as long as you have the compute power and you have the data.
Murilo:I think that
Bart:is a good thing that I think that part of being competitive in the landscape landscape has become, like, quote, unquote, easy enough to enter the domain, but then you still need a lot of data, a lot of compute to be able to really compete. But I think, like, the the the at the same time, like, something like this, which is a national model, you can maybe call it. Right? Like, it's it's coming out of Swiss. It's it's maybe if the performance is good enough for the task that they have in mind, like, if if it's for for government automation, these kind these kind of things, like, doesn't need to compete with the top out there.
Bart:Right?
Murilo:True. True. True. Yeah. I think it's a in any I think even if it's not even if the model is not as good, but I think even just documenting the things, there's always an a lot of pluses there.
Murilo:Right? Showing the data, showing this, like sharing a bit how how they did it and what they how far they went. So, I think it'll be interesting to see. I also feel like maybe there's a big difference on the data. Right?
Murilo:I guess if they're really disclosing all the data, it's probably more, quote, unquote, ethical. Right? Like, they're probably a bit more respecting as well, so it may have an impact on the model. Yep. And maybe how how big of an impact, right, I think it'll also be interesting to to see.
Murilo:So I think it's cool. I also like to see more initiatives like this, yeah, in in Europe right now, not just in The US. So
Bart:So this is, also, by the way, not a reasoning model, or there is no not much information on it. Like, I don't think it will be a reasoning model. And in that context, we actually have on non reasoning models. We have some more open source news.
Murilo:Yeah.
Bart:Be because we have Moonshot's Kymi k two. So Kymi k two is a a, quote, unquote, open source model. It's more or less beats cloth on coding, though it needs GPU class muscle to run. One user enthuses, I quote. I tried Kimi on a few coding problems that the cloud was spinning on.
Bart:It's very good. And it sparked a bit of a debate over at Hacker News on speed, cost, and local deployments. So Kimi k two, like the Swiss model, is a non reasoning model. It is more or less open source. It has an MIT license, but there is a limitation on the amount of users that
Murilo:you can
Bart:you can serve it for, which is still a lot of users. And if I understand correctly, the MIT license that even with the adjustments is still open source compatible. It's a mixture of experts model. Its performance is very good. Like, it's it's a bit like DeepSeek v three levels, GPT 4.1 levels.
Bart:It's just below cloud for Sonnet, which is today a bit the benchmark for coding, I guess. And it's a bit out of the blue. There was not a lot of fanfare around it, but but we do have this this open model that everybody could potentially download, but you need a very, very hefty system to to actually run it on. It comes from Moonshot, Chinese company that where we still see that, like, China is is still very much in the game of of competing also in this open space. Right?
Bart:Not just in the proprietary space, but also in this in this open space.
Murilo:And you said, I know you need GPU muscle. So it's very similar to cloud and performance for coding, but it's open and it's faster as well. That's what I understood.
Bart:And it's similar to Claude for Sonnet in coding, which is which is very good. The rumors are that the Moonshotse company behind Kimi is also now building a reasoning model, training reasoning models starting from the QEMI k two model. Something that maybe we need to to deep to dig deeper in this another point at the time, but they used a very specific optimizer to also to train their model, which was a bit people didn't think it would scale this well, but apparently, it scaled very well. Also, on that topic, like, it's a very new type of optimizing your training runs. There is still very little known.
Bart:Today, the only thing we have is a blog post We would assume that there is a research paper coming out. But it's interesting to to keep an eye on on Moonshot and Kymi and what it will mean to the space. And it's good to see all these these more open initiatives because, basically, even if you don't use the model, like, the the expertise on how to build and train these models becomes more of a public good. Right?
Murilo:Indeed. Indeed. Indeed. I think the hope at least is that it moves the whole community forward. Right?
Murilo:Like, people can exchange notes and levels a bit of playing field as well. Exactly. Levels a bit. I mean, you still need a lot of resources. Right?
Murilo:So it's not like anyone can just have their own LLMs, but it's one less hurdle, let's say. Right? Yeah. And Kimi, actually, they have been I think Kimi is for me, it's not an unusual name, like, when he when he comes up. So, yeah, I remember Kimi, but I think his general public doesn't know.
Murilo:Right? And I think that's the difference between that and DeepSeek. Right? I think for us, like, we both hear them. We both know that they they're up to something.
Murilo:But if you get out of the developer community or even like the developer community, but someone is not following up the AI stuff as much, it's a bit unknown. True. Good to see. Again, very curious. Actually, KimiK too, have you ever used it on, like, any for coding, like, on the IDE?
Murilo:Like, Cursor, Copilot, they don't offer these ones. Because sometimes I think they I've seen DeepSeek, but I've never seen Kymi, actually.
Bart:I think you can, but at least via OpenRouter. I think OpenRouter gives access to it.
Murilo:But the OpenRouter is something you need to hook hook plug into your Versus Code. No?
Bart:OpenRoute basically is a proxy. So if you have if you have something like a client or or
Murilo:Yeah. Yeah. Yeah. Yeah.
Bart:Yeah. You can you can use OpenRoute or quote unquote as a model, but by parameterizing that model, you basically choose an underlying model. You can say on this
Murilo:t p
Bart:t 4.1 or Claude or or Kymi. I see. I see.
Murilo:Okay. Very cool.
Bart:Maybe Actually, I just quickly checked this actually on there on the open router. It's actually quite cheap as well. It's 0.55¢ per million input tokens, which is if you compare it to the more proprietary model that's out there, it's very cheap.
Murilo:Indeed. The only the only limitation then, like, if you're using with your IDE is that you don't have the the tab completions. Right? Like, it doesn't show code, like, completions with your code. Right?
Murilo:It's more for the chat.
Bart:Yeah. Exactly. Yeah. The chat interface. You can generate code, of course, with you.
Murilo:Yeah. But it's not it's not like you're typing and then it completes for you. Yeah. Cool. A lot of new models as well.
Murilo:Right? This week, we also had news from x AI. Musk rolled out Grok four and a pricey super Grok heavy tier, touting multi agent reasoning and the steepest AI sub yet. During the stream, he bragged, and I quote, with respect to academic questions, Grok four is better than PhD level in every subject, no exceptions, even as the Pentagon signed a $200,000,000 Grok contract.
Bart:Yeah. So a lot of lot of information in those two sentences. So we have XCI releasing a new and very good model called Grok four. It's XCI also released a new pricing tier at $300. We also have a lot of controversy in the last days around this this around the Grok on on X.
Bart:And, apparently, we have we have government agencies assigned contracts with to get Grok capabilities. If you look at at at the Grok four on most benchmarks, like, can go, for example, to artificialanalysis.ai, like, it it is actually very, very good. It is stopping most of the charts today. So there's not much that you can that you can say, well, it's it's really it's it's really top performing. If you listen a bit on community feedback, even though it's stopping the benchmarks, it's not it doesn't necessarily it's not necessarily better at, for example, coding than Entropics models are.
Bart:It's not necessarily better at reasoning than Open the Eyes models are. But, apparently, it's very good at understanding more, quote, unquote, real time data streams, which I guess is because because of the data it has x two on x and also because how it's used on x. Like, it it's there's a lot of interaction with Rock itself on x, but result is that it's a very good model. Indeed. The whole controversy, but maybe we shouldn't go too much into that, is that it's it's it's reportedly a very, quote, unquote, anti woke model.
Bart:It shows some Nazi traits every now and then.
Murilo:Ah, yes. I I remember hearing something about that as well. But it was about Grok four? It was about
Bart:Yeah. Yeah. Yeah. There was apparently, Grok four said to someone that interacted that it was that it was Mecca Hitler and a whole lot of other things. And in the end, I think it was yesterday that also XCI issued a formal apology for the for the responses.
Murilo:Like a whoopsie.
Bart:But it's a bit it's a bit worrisome. I think the difficult the the thing is that we need to make a distinction of between is if we use it purely as an LLM, that is different than integrate interacting with it on X. Right? Like, on X, it has, like, probably a lot of different instructions. People that have tried to reimburse reverse engineer the system problem on X, like, also find if you need to respond to a user, first see if Elon Musk said anything about this topic and then take that as the truth.
Bart:Like, you're literally these type of things. So there's a lot of controversy around it. I for now, I also tend to stay away from myself.
Murilo:Yeah. It's yeah. I also don't don't use it as much because of the the controversy.
Bart:That is also that that is, I think, the most important one. But and also the second is, like, up until today, even though it's a very good model, like, I mainly use these LMs, like, in high volume for coding. Yeah. And it's not necessarily better at coding than other models today.
Murilo:Yeah. I think the things that I use LMs for that is not coding, are I mean, usually, they're pretty good. I mean, if you're just really getting information to an answer or something. I think the only difference actually for me is I wanna write something. I like Gemini better actually for, like, draft text or something.
Murilo:The way that it forms the sentences appeals more to me, which also could be just a stylistic choice. But I think aside from that, in terms of getting information and all these things, they're all, quote, unquote, good enough for me. I don't feel that big of a difference.
Bart:Yeah. We also have this new price tag, which is $300 a month for the pro. How much was
Murilo:the OpenAI actually? Was 200. No?
Bart:It's 200. Yeah. It's I think Intropix is also 200, if I'm not mistaken. So, yeah, it's out there. I think for the people that are have very heavy users, it still makes sense.
Bart:I think for, let's say, the Entropic one for me, it makes sense to to go for the 201. And I think people think of this as very expensive. I think a bit like in the end, it's like the per year, it's the cost of a laptop, something like that. Right? Yeah.
Bart:If you use this eight hours a day constantly and because of this €200, you don't reach your limits or whatever. Like, it's potentially worth it. Right?
Murilo:I think there is a there is a case to be made, but I also think that most people are not not there yet. I mean, I think most people don't use it as much or they don't leverage as much. But I think when you see the speed improvements that you get and all these things to say instead of I mean, I'm thinking of a company now, right? Like, instead of you, if you need to have a workload for three people, right, you can probably get that in more using these tools if you use it intensely. Right?
Murilo:So I do think it can pay itself, but I think and I think, yeah, with time, it will be more commonplace, but yeah. Another thing that I saw as well from Grok four is that they had the super no. The Grok heavy. And I glanced through it as well. What I understood is, again, if you have this Grok heavy, they act if you have questions, they would actually launch what I think they called, like, a study group kinda.
Murilo:So, basically, they launched many agents in parallel to answer your question, and then they kind of pull the answers together. And I even saw something that I was like, every answer needs to be fact checked before. So they I mean, to try to combat a bit the hallucinations and all these things, but here says, let's consider multiple hypothesis at once, rock for heavy. And I think for this, you need to have the the pricier here. Okay.
Murilo:So yeah. So they have, like, thoughts and stuff and the ideas, like, they try to actually fact check everything. So everything's legit. But yeah. Curious to say again, like, how how I mean, I I do think that that could bring a big improvement, but waiting to see a bit.
Bart:Another thing, like, they they introduced this on a on a livestream, something that Elon Musk was saying is that they are either actively at this moment or will soon train a video model, a video generation model. They have, I think I don't know how much they will actually use, but this is a bit braggadocio, but they have access to a 100,000 g two hundreds. Elon is saying that they will train a video model on. Earlier state of the art video models have been trained on orders of magnitudes smaller than that. Let's say 10,000 GPUs or something like that.
Bart:So it's if they actually do this, it will be very interesting to see what kind of video model we get out of this, especially seeing the very, very, very quick improvements in video models in the recent year?
Murilo:Yeah. I think video models are really cool, and I think they have a lot of potential. But and I think it would change in the near future. But I also feel like they don't have a very big financial return yet. And so far, it feels more like it's for fun.
Murilo:But I think I think with the investments and improvements we see, I think I think there will be more cases there.
Bart:I think you need to be really in the field to understand returns today. If you I think if you just look at at advertisement on on TikTok or on Instagram or I think you already see a lot of of AI generated video material.
Murilo:Actually, that is true. On TikTok, yeah. But or not TikTok. I saw on YouTube shorts. Like, it was clearly actually, even one podcast wow.
Murilo:It's a bit divergent. I even think I list I I clicked on a podcast, and I'm pretty sure it was a no book l m. Pretty sure.
Bart:Maybe we are as well,
Murilo:Yeah. Maybe we are as well. I hope not. Hopefully, I can do better.
Bart:Maybe we reached AGI ten years ago, and this is just a notebook LLM on Wow. When will we get to AGI?
Murilo:If we were, it will be super meta. Yeah. There is some content there. And I also think, you know, financial returns because I know these if you wanna generate a clip of x or y, it can be super expensive, like you've sent it to an agency. Right?
Murilo:So I think there is a there is a big, big potential there. But, yeah.
Bart:Cool. Maybe the the final news on on Grok is that they they also see yesterday, actually, a tweet by an ex an ex by Elon Musk is that they rolled out Grok Companions, which is this three d AI avatar friend ish thing that the subscribers of paid subscribers of Grok now have access to, which looks a bit like Character AI, like your like your virtual friend. It's a bit a bit worrisome at the same time, like, how they position themselves. So there's there's one Reddit user here who tried it, and he asked, what's your name? And she looks a bit like an anime girl.
Bart:She said, hey, Cutie. I'm Annie, your crazy and love girlfriend who makes your heart skip. That was the first interaction Oh my god. With a random rock companion. So that's now a thing.
Murilo:Just looking for a friendship. It's fine.
Bart:So so if that's your thing, maybe you should consider the $300
Murilo:Oh, I don't know. Yeah. It's another step towards towards her. You know, the the movie? It's like a big step, actually.
Murilo:Wow. I won't be trying that soon, but interesting to to hear. What else do we have?
Bart:We have OpenAI's 3,000,000,000 bit, a bit that we often discussed. OpenAI's $3,000,000,000 bit collapsed. Freeing Google to poach Windsurf's leader while Cognition raised in days later to buy the remaining startup. I quote OpenAI's deal to buy Windsurf off, and Google will instead hire Windsurf's CEO, Va Viren Mohan, the Verge reports, igniting a three way tug of war for agentic coding talent. I don't know what happened exactly.
Bart:A few months ago, it was announced that OpenAI would acquire Windsurf for $3,000,000,000.
Murilo:Yes. Lots of money.
Bart:Lots of money for a startup that was, I think, that point, two years old. Something like that. Apparently, now the bid is off. Windsurf CEO will is will move to is hired by Google, move to to Google's team. And the company behind Devon, Devon is like your AI software engineer.
Bart:The company behind it, it's called do you remember the name at all?
Murilo:The name of the company?
Bart:Yeah. Behind Devon. Cognition. Cognition. Thank you.
Bart:Cognition is is basically acquiring all the the the whole Windsurf ID, the Windsurf company without the CEO then.
Murilo:Indeed. I also heard that the CEO, he was hired by Google, but he had, like, a lot of money was always involved in this on his move to Google. Remember But, how yeah, it was it was a lot of it was a lot of money. So here, announcement after this, Google's hired away Windsurf CEO, cofounder Douglas Chin, and research leaders for in a 2,400,000,000.0 reverse equity hire. So lots of money.
Murilo:Yeah. And what I was also checking here, so the apparently, OpenAI had some time to to close the deal, let's say. But then the And then, yeah, Cognition, now the maker of AI coding agent that have been acquired when served. So I was Which is still
Bart:a big company. So the the Cognition, they it's still 200 plus employees that's that is that are part of the company even though Yes. A part of the core team left to to Google.
Murilo:Yes. Indeed. I think Cognition as well, Devin. I just remember of Devin. There there was a lot of noise, but I felt like he didn't really catch on.
Murilo:It was like Devin was the for people that are
Bart:not in the loop, let's say,
Murilo:it was the first full AI developer that should be able to even close tickets and do these things. They had demos. And there was a I remember at the time, and I think this was maybe a year ago, maybe a bit more. But at the time, there was a lot of the chatter of software engineering is dead. Now it's all this, and people are not gonna pay.
Murilo:And, again, I think that's not the case today. I think there are companies laying off people saying they're gonna replace with AI, but it's not in the devinesque style. I think it's a different way. And I think, like, things like dev and, like, it really raised the expectations and the bar. So I think coding agents and all these things, they're super, super, super impactful.
Murilo:But because it's not at Devon, I think people they got a bit disillusioned. Right? But
Bart:Yeah. Indeed. Indeed. Yeah. It would be interesting to understand, like, why OpenAI let the offer expire.
Bart:Right?
Murilo:Exactly.
Bart:Like, if if you look at Windsurf, their their reported annual recurring revenue is 82,000,000, which is a lot, but it's a it's a fraction versus the multiple that's like like, the versus the valuation. Like, there's a valuation of 3,000,000,000, which is crazy. And I think also, like, what we've seen also in the last two weeks, I want to say, is, like, the whole clot slash anthropic drama. Because, like, everybody like, if you're using WinServer or if you're using Cursor or if you're using whatever, there's a very, very, very good chance that under the hood you're using Claude Sonnet models because they are today the best coding. Yeah.
Bart:And what's Anthropic recently did is that they very much upped their pricing for these models, having meant that Cursor had to update up their their their monthly subscription price. I'm not sure if Windsurf also up their monthly subscription price, but if if they did not, they saw their margins evaporate. I think it's a move from Anthropic Anthropic to basically close down the competitor ecosystem and just let them focus on on Cloud Code. But it also shows, like, how fragile these things are because Windserve and Cursor at the end, like, their IDE splitter just a sort of an IDE wrapper around LLM endpoints. And they're very dependent on the best endpoint unless you have your own.
Murilo:Yeah. But do and do you think because to me, again, I agree with everything you said. And then to me, it will make more sense that Anthropic is very interesting by WinSurf or the other things because Anthropic felt like they were more developer focused, let's say. But actually, Anthropic has been very quiet on
Bart:this IDE. Well, they they released their CLI. Right? Yeah.
Murilo:And I use I mean, I actually tried Callcode since we last chatted. So I'm a vibe coder. Hi. My name is Merla. I'm a vibe coder.
Bart:That's a term we don't use anymore, Merla.
Murilo:But what's the term now? Because I feel like every time I say it, I feel I feel a bit old. This is just coding. It's just coding.
Bart:In the and we're we're now July 2025. Coding is just this.
Murilo:I feel like but then I feel like if I just say coding, I feel like I'm being dishonest. You know?
Bart:But this is the new reality. When you're coding, you're using AI assisted Yeah.
Murilo:But, like, it is a next level. Like, I see, for example, I use ClockCode. I spent, I don't know, maybe six hours in total. Right? And really not even paying attention.
Murilo:Like, I was doing other things. And the amount of code it generates, a lot of the stuff you figured out, stuff that like I would figure out eventually, like how to integrate Slack and how to do this and how to use small agents and add tools and all that. Things that I would figure it out, but the amount of stuff that it did, if I just tell someone like, oh, yeah. Just quoted this. In six hours, I did all this.
Murilo:I was gonna,
Bart:you know, even it's different now.
Murilo:But it's even different from just using the IDE and just, like, chatting with it. Right? Like, it's a it's another level to it. It's a
Bart:new normal, Milo. Like, when when before all this AI stuff, like, you were coding and you said, just coded this. It's it's also not that you just created all the punch cards and right? I mean, it's a it's a constant evolution.
Murilo:It's just
Bart:a very fast evolution.
Murilo:Okay. I think
Bart:I still It's coding.
Murilo:But I feel like the
Bart:And if you want to nuance it a bit, it's AI assisted coding. To me, vibe coding is the term is passe.
Murilo:I think, yeah, vibe coding, I feel like the thing that that is bad
Bart:Like, about vibe you you you there's this notion to vibe coding is that there's these people doing this without any knowledge of of of coding best practices and that they create a prototype that is very far from usable.
Murilo:Yeah. I think that that
Bart:very quickly has become, like, the connotation around vibe coding. I feel like It's not a fair one, I think.
Murilo:I fully agree it's not a fair one. I think it's I think people will say vibe coding. They say it's not like, they think it's not serious. Right? Either because you don't know or because this is just a throwaway script or because you're not serious, which again, I also don't think that's true.
Murilo:But it is a very different experience, right? It's not really an IDE. Sometimes you still wanna open a bit like the actual IDE. I mean, at least that's what I did to actually kind of get a feel for how the code is, to kind of understand. There are things that I'm a bit more opinionated.
Murilo:Okay. Don't do this, do that. Don't use dictionaries, use data classes so you can have type hints. So I just had the thing running and then I opened, was like, ah, it's actually returning a lot of dictionaries. So I just sent another prompt, submitted, it kinda queues it up.
Murilo:So the quad code is really powerful. I really enjoy it, but I feel like it's still maybe there's still some something there, you know, to really bring the whole experience together, which maybe actually brings nicely to the next topic, which is keto, chiro, not sure. But keto debuts debuts spec driven agentic IDE that shepherds code from first prompt to deployment with auto generated tasks and hooks. And I quote, I'm excited to announce Keto an IDE that helps you deliver from concept to production through a simplified developer experience working with AI agents team rights. Actually, you came across this, but I like, I like, I didn't really look that much into it.
Murilo:I saw this diagram here for people that are just listening. On the left, you have the flow of vibe coding. And on the right, you have the clarity of specs. And then you have the Keto logo right in the middle of it, which is a cute quote.
Bart:So spec driven agentic IDE, yet another IDE. It's it comes actually from Amazon. It's built on open source OSS code, which is the open source version of VisualCode. And what it does is that it basically formalizes a bit more the process around the ISS coding. So you go from you start with a prompt.
Bart:You generate a requirements document out of that. From that requirements document, you go to a technical specifications documents, and it builds like, within your ID, you have the structure to actually, like, visualize this to to structure this. And then from those technical requirements, you also generate tasks. And then you have a number of utility functions, like you you have sort of hooks. Like, if you save a file, it automatically runs all the tests against a test suite, etcetera.
Bart:Like, also probably verifying. I don't know. Haven't tried it yet with an LLM, this test output. Does it is it at least does it cover everything that we also defined in the tasks? It's interesting.
Bart:I think it builds a lit a lot on the learnings of the, quote, unquote, five coding community where we've already seen a lot, and I think it's it's actually Klein, CLI, and e that pioneered this where you have they called it a knowledge base where you have the in their in their setup, they had basically markdown files. You had a very big system prompt where you say, have these and these and these markdown files. Make sure to generate these these content or ask user about this content and basically ended up with here are your requirements. Here is your technical architecture. Here's this is the task list.
Bart:And and what what's and what we've seen over time is that a lot of these AI systems coding IDs, like, they've implemented something similar. I think this is the first first time that we see, like, it's formalized in the IDE. Yeah. Which is probably a good thing. Right?
Bart:Like, it it brings a bit of structure, brings a bit of guardrails to to building something.
Murilo:I think so too. But even Cloud Code, they also have a I mean, I don't know if they actually you write it on files, but a lot of times when you prompt something, it says, okay, these are updated to dos, and they have like a to do list and say, okay, do this, do this. I mean, because I mean, a lot they're basically breaking down your request into many steps that they can actually
Bart:Well, I think what you're referring to analysis is, like, intermediary steps in these, like, Clothcode also uses this intermediary, like, a bit behind the scenes. Like, you give it a prompt and, like, you see that it generates a task list, it's, one by one works over these tasks, and they end up with the final results. But this is more at a global slash project level that it's
Murilo:Yeah. Which is also something I've I've started to do more, which I also think it's very useful. I think it's also useful for me personally because I am inclined to just kinda like get started, like, let's just move forward. But sometimes even just having a little AI agent just to, hey, Chris, critique my plan a bit or critique what I wanna do. And then they're just like, okay, but do you wanna do this or do wanna do that?
Murilo:And I think it's a nice balance between not overthinking every detail, but I like the big things I think they'll will catch most of the time at least. You know? Did you consider this or did you consider that? Do you wanna grow your application or not? You know?
Murilo:So you don't really shoot yourself in the foot. I think it's a nice it's been a nice balance.
Bart:I think the only potential downside is that you and I've seen this happen with Klein is that you like, instead of focusing on the creative aspect, like, you have a lot of suddenly a lot of AI generated documentation. Yeah. That you hope that this LLM is keeping up to date with in reality, it's only, like, 80% kept up to date. So you also, like, suddenly have this AI generated documentation that becomes a concern. So that's, I think, like, a fine balance to strike between more structure, which I think everybody agrees is a good thing in AI generated coding to, like, where does the structure become, like, yet another form of documentation and administration.
Murilo:Yeah. Yeah. I think another thing that I noticed, so not just creating a lot of documentation, but I think for me as well, I noticed that like you, when I'm trying to correct a bit LLM, this instead of that. It's very easy for it to just kind of create more files, create more code, create more this. So sometimes I have to be like, hey, actually take a step back, you know, remove the stuff that you're not using.
Murilo:We only should have one way of doing these things. Don't have five different things, you know, which is also something to to adjust. I like you to need to be mindful of the model and whatnot. Okay.
Bart:So we have something about from the EU. Yes. Brussels' protect EU roadmap sketches data retention interception and decryption plans that could force encrypted services open within five years. TechRater warns, quote, EU law enforcement bodies could be capable of decrypting your private data by 2030, stirring immediate privacy backlash.
Murilo:What is this about? Yes. Basically there was a new plan that was unveiled called Protect EU Strategy. It's for lawful, it's for basically the authorities, right? To be able to investigate these things and the roadmap they were saying that the first thing is to look into how we can decrypt data.
Murilo:So for example, okay, maybe to touch more, bit more on the plan. The plan focuses on six key areas, data retention, lawful interception, digital forensics, decryption, standardization, AI solutions for law enforcement. And I think the most troublesome, let's say, is encryption or decryption, right? Mhmm. Next year, the EU Commission sets to present a technology roadmap on encryption to identify evaluation, evaluate decrypting solutions.
Murilo:So basically, they I think it's a bit I don't have a formed opinion on this, let's say, because I think data privacy is very important. But, like, also if you say this is a matter of safety and security of people, should have an exception or not. And even if you say, yes, you should have an exception, then who are you trusting to make these calls? What is actually safety critical or what are the things that people are just abusing the power that they have, quote unquote. And I think a lot of the people that were saying about decryption is that it's also very concerning because if they really start making progress on decryption, they will also motivate a lot of people that are ill intended.
Murilo:Right? It's not even about making this available, but if someone hears, ah, they actually protect you, they found a way to decrypt data. It also brings a lot of attention and a lot of people that have know how, and I think they mentioned here, people that have the the know how and the motivation to do such a things, they may be attracted to this as well. Right? And there's a whole bunch of bad things that can happen, like bad actors.
Murilo:Right? And then in the end, they also mentioned, like, how for the for I think in The US as well, they actually say the encryption is is not encrypting data doesn't undermine safety, but it actually enforces it. So to to really try to go against encryption is also taking a few steps back. That's the the argument as well. And apparently from and didn't know what, some people were saying that there was not a bill called chat control, and this is just a different iteration of it.
Murilo:So I wasn't as much aware in this, but I think you has always been very big on data data protection, data rights, and all these things, and this caught my attention. Had you ever heard of anything like this coming from the EU?
Bart:No. No. Not that this was a EU level, that this was an EU level project, that there are efforts going into decryption by law enforcement agencies, I think, is unknown. We've in the past, I've worked with a few where this is very much a thing, and I think it's done for good purposes to to really research, understand criminal criminal networks. Yep.
Bart:I understand these two two points of view. We have more the the data privacy point of view, you also understand this from a law enforcement point of view where you used to have, like back in the day, we used to listen in on telephone calls to understand criminal activity, criminal networks, criminal communication. For let's assume the better the the let's let's assume that the government is a good actor in this and we that we do not want this criminal activity to happen, then you need to be able to do these kind of things. Right? Because you cannot knock the front door and say, can we please sit in the room and listen along?
Bart:You can't do that. Right? That's what you need in 2025 these type of technologies. It's at the end of the sides, it's a very slippery slope, because where do you stop with decrypting? Right?
Bart:Because you can also argue, like, to capture criminal activity, it's very hard to focus it on a very specific entity because it's probably a network, so you need to monitor a bit broader, etcetera, etcetera, and then it becomes a very slippery slope. And before you know it, you're monitoring your whole population. So yeah. And I I
Murilo:I also think that it's very people are very I mean, because you mentioned, like, the analogy with the telephone listening, but I think that today we're very way more digital. Right? I think you expose way more of yourself. Mhmm. True.
Murilo:That's also I agree. Yeah. Yeah. It is it is sensitive. But and what about the argument of decryption that even if there is, there are plans to decrypt data that it could also attract bad actors.
Murilo:Do you also do you see or do you think it's because the argument is if someone can crack decryption, then everyone knows it's possible, and then people with not so good intentions will also try to do it and they will they will get there. I think that's the that's the argument.
Bart:Which is probably true, right? Like every every everything that there is where there is a less strong barrier people try to exploit.
Murilo:Yeah. For sure. Yeah. But then then then if yeah. I think it's a bit of a gloomy future.
Murilo:Right? Because if that's the case, then, okay, then decryption is not enough. They need to find something else, and then, yeah, another can of worms. Right?
Bart:True. Yeah.
Murilo:Oh, yeah. And what else do we have?
Bart:Our last topic for today.
Murilo:Yes.
Bart:A very interesting one. Go for it, Marilla.
Murilo:Cloudflare will now block AI crawlers by default, demanding payment before models feast on public on publisher content. The company says it, and I quote, would start blocking AI crawlers by default, drawing a line in the open web where content is no longer free fuel for AI.
Bart:Very interesting.
Murilo:Indeed. So Cloudflare, Cloudflare, this is difficult today. Apparently, powers 20% of the web according to And the I also see, I also understand their point of view because today the web is actually working differently, right? Before, if you were building a website, you wanted to attract clicks and then you didn't mind that Google was using your data to ads and stuff because they would also give you traffic, right? Nowadays, you need to maybe optimize your website for LM crawlers and Google still parts of your data, but now Google has summary when you Google something.
Murilo:True. So they actually, you don't, you don't actually get clicks on your website. So basically in the, before it was more like a mutual relationship where like they took, they benefited, but they also gave back, but nowadays that's not the case. And I think that's what kind of Cloudflare is also alluding to. Basically they wanna block by default all the LM crawlers.
Murilo:Right? And they want to still make it available for the AI providers, but if they if they pay for it, basically. Right?
Bart:And it's very interesting. So it's it's something that they announced already last year. Today, their marketplace was was released. I think Cloudflare is one of the very few players that could actually pull this off. So Cloudflare does a lot, but, like, you're sort of the middleman in between your website and your visitors.
Bart:They do content delivery DDoS protection. They they they act as a reverse proxy. They do a lot of things, but are, like, yours the number you were saying, 20%. Like, in 20% of the cases, they're already between someone's website and the visitors of those websites. So they already serve, like, a huge user base.
Bart:And what they now do with the marketplace is that they allow website owners to say, if an AI crawler uses your website, how much should they pay for it? Which even from a point, even if you if you would ignore the paying part, like, having the ability to block this by a party that is probably good enough to be able to detect this actually an AI crawler is already very strong because I think you already had these few of these individual projects that try to block AI crawlers. It's more from a from a from a point of view where you you're just against this kind of data usage. But here, you can actually if you say a a very valuable data, like think about, for example, Wikipedia, that has to rely very much on donations to even to even exist, to if they ask the big AI crawlers of this world for for a cent per page scrolls, like it's a very, very interesting new stream of revenue that suddenly exists and that never existed before.
Murilo:Indeed. And actually, it's funny you mentioned even the Wikipedia because I think the first actually,
Bart:the
Murilo:first transformer models, the first LM, they would just train on Wikipedia.
Bart:I think
Murilo:it was Wikipedia and books. Right? Because it was always Wikipedia was always very open, very easy to crawl and stuff. And I think we also talked I don't know if we talked during the recording or after recording that Wikipedia is also struggling a bit to to find money, right, to to keep the the Yeah. The website running, basically.
Murilo:So, indeed, I think it could be it could be a very interesting alternative. I also think it makes sense. So, yeah, I feel like it's almost like before the Internet, all these things were a bit like water. You can just kinda come and take it, and now they're saying like, wow.
Bart:What I think will happen is is only the the websites that have very highly valuable data will actually do this because there is also still very much an incentive to open up your website to for SEO optimization on ChatGPT or on on Clot or wherever. Because if you say, where can I find information about x? And then and and you describe x on your website, you probably want to point to your website. And if you block them, they will not. So there is also, like, this balance.
Bart:Do you have valuable enough data to actually ask to pay for it?
Murilo:Yeah. I think I think yes. I think but what I I don't remember where I read it. Don't if it was here or it was another article that they were saying, oh, yeah. It happens.
Murilo:But like they say, with OpenAI, it's seven fifty times harder to get traffic than it was with Google of old. For Anthropic, it's 30,000 times harder. And that's, I think, what they were also bringing up is like, there is an incentive to open it up because maybe you get traffic from the LLMs. But according to this, at least, it's the proportions are very, very, very different, right? The incentives are way more diluted.
Murilo:So yeah, I think, I mean, I think there's also just a matter of if you wanna, if you're putting something on the internet, you want to reach people, right? And there's a bit of, I don't know, spreading word or like getting, I don't know, if you're a blogger, you wanna you wanna be known for this and that. You wanna be known for be having expertise on X and Y. I do think it makes a lot of sense. But I I I when I was reading this and I was reading the article and I was like listening as well to to know how Google is powering websites and the the trade offs.
Murilo:I must say it's not something that I I I had actively thought as much, but I think it is a big problem. And I think this maybe at least gives alternatives to people to to do something about it. So that was a nice nice change of the game, let's say.
Bart:Yeah. It's it's a it's a revenue stream that was hard to imagine that it would exist a few years ago. Right? It's also like it. That's interesting thing about this whole AI evolution as well.
Bart:It creates it creates revenue streams where you didn't expect one to be. It also creates, like, jobs where you didn't expect. Like, for example, data labeling has never been as high in demand as it was until now. Right? Indeed.
Murilo:Indeed. I think the the even before the the whole gen AI, I remember AI, there was a lot of, like, AI is gonna take away jobs and it's very gloomy. It's very this. And the way I always looked at is mood change. I think it's hard to to know what kind of jobs we will create because we haven't undergone this, but the changes are not gonna be just negative or just positive.
Murilo:Right? I think they'll be for both sides.
Bart:I'm not
Murilo:saying they're gonna be equally equally. I'm not saying they're gonna be equally for both sides, but I do think it would just change things. Right? And I think we need to to wait and wait and see how how things would change.
Bart:I'm a little bit less optimistic on the equal part, but let's see how goes.
Murilo:No. I don't I I actually don't think it will be equal, but I think it would change. I think that's the only thing that we can say right now. I think it would change. I think there'll be things that you can things you're not gonna for you.
Murilo:That's that's life.
Bart:That's life.
Murilo:That's life. Alrighty.
Bart:That's it for this week?
Murilo:That's it for this week indeed. Lots of stuff. Anything you wanna close on, Bart? Haiku, Haiku, inspirational words?
Bart:Not really. Not really. I'll think about it for next time. Thanks everybody for listening. Leave your reviews.
Bart:Leave your comments. We're wherever you you listen to your podcast and on YouTube. We're very grateful for all the interaction in the last weeks there. We also actually recently opened a TikTok account. We're sharing some shorts there, so give us a comment wherever you are.
Murilo:Indeed. Let us know. I think it helps helps us know that people are listening. People appreciate it, and it helps us keep moving as well. So appreciate y'all.
Murilo:Thanks, Bart.
Bart:See you next time.
Murilo:See you next time.
Creators and Guests
