← Back to search

Karpathy to Anthropic: the signing that breaks AI ✨

Es la Hora de Aprender · 2026-05-21 · 63 min
relevance 62 11946 words Episode page ↗ Audio ↗
Show full episode description
¿Qué pasa cuando una de las mentes más importantes de la IA elige Anthropic en lugar de OpenAI o Google?En este episodio analizamos el fichaje de Andrej Karpathy por Anthropic, una jugada que sacudió a Twitter y que dice mucho sobre dónde está el verdadero centro de gravedad de la IA hoy. Además, repasamos los más de 100 anuncios del Google IO 2026 — desde Gemini Spark (su respuesta a los agentes personalizados) hasta Gemini 3.5 Flash — y desmenuzamos por qué KPMG y PwC se pasaron de Anthropic a Gemini de un día para otro.También: Cristian lanzó una campaña de marketing 100% con agentes en Claude Code, Rodrigo explica el concepto de "esquilificar" todo lo que hagas con IA, y conversamos sobre la teoría de las restricciones como el verdadero cuello de botella para que las empresas saquen valor real de la IA.━━━━━━━━━━━━━━━━━━━━━━━━━━━🎙️ HOSTS━━━━━━━━━━━━━━━━━━━━━━━━━━━→ Rodrigo Rojo: https://www.linkedin.com/in/rodrigorojo/→ Cristian Tala: https://www.linkedin.com/in/ctala/→ Diego Arias: https://www.linkedin.com/in/diegoarias/━━━━━━━━━━━━━━━━━━━━━━━━━━━📚 NUESTRAS COMUNIDADES Y CURSOS━━━━━━━━━━━━━━━━━━━━━━━━━━━Comunidad Cágala, Aprende, Repite — Cristian Tala→ https://www.skool.com/cagala-aprende-repiteComunidad La Patrulla Roja — Rodrigo Rojo→ https://www.skool.com/rojo━━━━━━━━━━━━━━━━━━━━━━━━━━━🔗 RECURSOS MENCIONADOS━━━━━━━━━━━━━━━━━━━━━━━━━━━→ Crea tu VPS en Hostinger (link referido de Cristian)https://www.hostinger.com/es?REFERRALCODE=GBD6IAMGM6NA→ Crea tu VPS en Hostinger (link referido de Rodrigo)http://www.hostinger.com/rojo10→ Andrej Karpathy en Xhttps://x.com/karpathy→ Gemini Spark — Agente personalizado de Googlehttps://gemini.google.com/→ Google Antigravity — IDE/CLI agénticohttps://antigravity.google/→ Hermes Agent — Nous Researchhttps://nousresearch.com/→ Minimax 2.7 — Modelo alternativohttps://www.minimax.io/→ N8N — Automatización con agenteshttps://n8n.io/→ DGX Spark (NVIDIA) — Mini PC para agentes localeshttps://www.nvidia.com/en-us/products/workstations/dgx-spark/→ empresasaumentadas.com — Blog de Diego sobre IA empresarial━━━━━━━━━━━━━━━━━━━━━━━━━━━📌 TEMAS QUE TOCAMOS━━━━━━━━━━━━━━━━━━━━━━━━━━━00:00 - Intro y el grito de Cristian02:29 - "Esquilificar" todo: el consejo de Garry Tan06:16 - Diego bajó la cuenta de GenSpark Claw de USD 500 a 10010:44 - Adiós OpenClaw: migrando a Cloud Code, OpenCode y Hermes12:51 - Cristian lanzó una campaña con 10 agentes de marketing18:27 - "Si se puede, no me reten" — la moraleja18:48 - Google IO 2026: más de 100 anuncios en un solo día22:59 - Cuando Cowork dejó de funcionar con Meta26:43 - "Necesito producir, no solo aprender"30:32 - Cómo Google te va a meter Gemini en el bolsillo sin que te enteres34:12 - Gemini 3.5 Flash: ¿modelo nuevo o reentrenamiento disfrazado?38:01 - La guerra entre Anthropic, Google y OpenAI39:17 - Karpathy ficha por Anthropic: el bombazo del año41:58 - Por qué Anthropic juega en otra liga (Constitutional AI + foco)46:01 - PwC y KPMG abandonan Anthropic: 3.500M de usuarios Google51:09 - ¿Sobrevive OpenAI? El blitzscaling gringo explicado56:08 - Agentes + contexto + cuellos de botella: el futuro real
✨ Episode Outline — click any point to jump to it in the episode
Problem solved
Building personal, ultra-flexible AI agent setups by skillifying workflows and migrating off OpenClaw to Hermes and other models.
Benefits
  • Skillify any repeatable workflow into reusable AI skills
  • Combine per-tool skills into higher-level skills
  • Hermes excels at research and content creation
  • Mix models (GPT-5.5, Minimax, Xiaomi) as engine and backup
  • Self-hosting workflows on a cheap VPS
Use cases
  • Skillified Granola across five or six categories (meetings, classes, D&D, prospecting)
  • Lowered a GenSpark plan from about $500 to $100
  • Migrated both agents (local Spark and VPS) fully to Hermes
  • Rebuilt all processes from scratch with GPT-5.5 in a whole day
  • Spent 200 million tokens in one day talking to Hermes with Xiaomi's model
KPIs / results
  • GenSpark plan cut from ~$500 to $100
  • 200 million tokens spent in one day on Xiaomi's model
  • Minimax context only 200,000 tokens vs Xiaomi 2.5's 1 million
  • GPT subscription at $200, supposedly multilingual across 36 languages
Tools / build
  • Per-tool and combined skills (e.g. Granola skill)
  • Hermes Agent (Nous Research) on Spark and VPS
  • n8n workflows on a Hostinger VPS
  • GenSpark Club slides and workflow setup
  • Two-agent Telegram/Discord group chatting via ACP
0:00 / 0:00
🌐 This transcript was automatically translated to English from the original.
Very welcome to a new episode of It's Time to Learn Very welcome to Time to Learn, It's Time to Learn, It's Time to Learn Very welcome to Gary Tran from Y Combinator who said everything you do, if you solve a problem, skillify it. Turn it into a skill, an ability for your AI. You are doing a topic, researching a topic or processing a document or something and you manage to come up with the result. The question that comes next is how do we turn this into a repeatable, consistent skill? So I started doing skills for everything. I connected a platform to it, I gave it the documentation, I want you to learn to use this well, but with my rules. You say, I use this, for example the other day I connected Granola, which is the tool I use for meetings, so I started to tell him, look Granola I do this, I use it for meetings, I use it for classes, so that I can transcribe the entire class, I use it for my Dungeons & Dragons sessions, I use it for work meetings, prospecting meetings, like five or six categories. Then I started telling him what I wanted him to do with each of those things. So later when I tell him to review Granola something, he is not only reviewing, but he is also acting depending on the type of content that I asked him for. So I started doing that with all the systems I occupy. I said, I deal with all these things. And it started, ah look, we could do this here. And I started converting everything into skills per tool and then skills that combined those, like a skill that called the others to do things. So that it would remain as well as everything skillified. And that was my weekend. I think I've always done that. Yes, but Diego, you tell us first. What I have been doing is deducting expenses, because the last thing I told you is that we are using GenSpark within the company, GenSpark Club. And he's spending big on GenSpark Club. That's where most of the credits are going. The good thing about GenSpark is that when you pay, with the subscription plans you have, you have like two free agents who call you. One that is from the chat, that any model can be used, even if it is top of the line and without limit. At most he says that, I don't know, you talked too much, you have to let him rest. But in general it doesn't cause that problem. And also the creation of images. But the creation of slides, which is like one of the things we do most, is eating up our top. But especially, and what actually killed me was a big plan I had, I lowered it. I had it for about $500, now I have it for $100. It's the Sklo, like the GenSpark Sklo. And the workflow surprised me. Like in GenSpark you generate a prop and create a workflow. And the workflow is consuming us about the same amount as the GenSpark Sklo. And that's super strange, because it wouldn't have to. Because the workflow, in theory, should be deterministic. And you shouldn't be thinking so much, regardless of whether you have intermediate agents. Yes, very strange. So... I think that's called, as people say, it's called layer 8, Diego. Well, but those who have not heard the term layer 8. Layer 8 is the layer that is between the chair and the computer. Because the OSI model, of layers, is from the physical layer, the network layer, to the application layer. There are 7, therefore 8 is the human. And here Rodrigo is saying that it is Diego's fault. Basically, yes. But it's still strange, because in the end the model... Saying layer 8 was nerdy, but giving the explanation with the OSI model and everything, redoubles our conviction that in this podcast we are pure nerds, that we love talking about technology in the run-up or long weekend. It's true that it's a long week. I'm going to review it, but I don't think so, because in the end you see how the points go... Like the step by step of Workflow. There's nothing strange about it, right? But hey, I'll see if I don't have to change it. Diego, well, Workflow is TN8N and bye. That is, in the end you have any productive VPS. Most of the Hostinger ones can even help you. Come on, if we see you on TN. Remember that below are Cristian's and my Hostinger affiliate links. How to put both? We put both and each one chooses the one that he likes best of us. Yeah, but that's cheating. You are adorable. I'm shit. It has been good for me. Well, we put the tunnel at the top and mine at the bottom. And let them choose if they want to study more. Okay, good. Let's put this parenthesis. The Grinch of the podcast. I was fine. Yo no he hecho... Yo no sigo hater hoy día. I pretend not to be today. For now. For now. Yeah, bro. Diego, is there anything else you want to tell us about what you have done? I have doubts. That is, more than what I have done is see how I can make this a cheaper and simpler version of Open Club. Thinking that I have traumatized. I didn't have a good experience with Open Club, working on a local server. I saw it and also shared it with several people who spent more time fixing errors than taking advantage of it as they should. And I don't want to pass by when you're in full season. Do you want to hit first or should I hit? I go first. Because here the simple answer is you can't. Not because of cost, but because it will actually fail. I know that... Oh, yes I was a hater. And in Spark Club it did not fail. Yes, but in the end I know that there are many Open Club services. Open Club has also had many updates. I am one of the people who, when I first started working on it, the level of productivity was incredible. But it was never the same since the updates. And since I couldn't use the Antropic membership again as a brain. The reality is that this week, or the weekend I think, I stopped using Open Club. And the two agents I have, one local to my Spark is Hermes and the other one I have from the VPS is also Hermes. I migrated everything. Everything should be working. And I am now saying many things that I did with Open Club with Cloud Code or with Open Code. So, I'd love to tell you that there's an easy, inexpensive way to run Open Club productively so that it really works for you, but I don't think so. But Hermes, are you differentiating? Two things happened to me there. Uno es que Open Club, efectivamente cuando partió, como todavía nadie estaba preocupado con la seguridad, podía ir a hacer de todo y era maravilloso. And Opus, in particular, which was the model we were using at that time, is also very good in the use of tools, computer use, in terms of how it wrote, right? He had a lot of things. And once they take away the model, which is like the engine of that car and we have to put a GPT or a Minimax or another, it was difficult to find one that worked as well. And on the other hand, in their search to enhance security, they began to make many changes to the project, so they began to break things. So you had created a Chrome or a process with Opus, with Open Club, and GPT didn't use it so well, so you had to do it again. So those transitions were costly. Today what I did, for example, on the new mini PC that I have, I installed Open Club from scratch, from scratch, a new installation, after all these security changes and everything, and it worked wonderfully. Of course, without immigrating anything from the old. I didn't migrate anything from the old one, I started to clone the processes. I mean, like, hey, I was doing this on the other side, how could we do it? But without explaining, without passing on the files, how did we do it? Cachai? To create the processes again, a whole day. And one of the issues you are working on? With GPT-55. With GPT-55 and it is working super well. GPT-55 is a bit wordy, but... And why isn't the 2.7 a minimaxe? It does the processes, because I have the, I have the, how I do the classes, I have a GPT subscription that is 200 dollars, that I have to use something on a daily basis, and on a daily basis I use Clot as my main day, so like, now, let's use the resources that I have, but I also have minimax as a backup, which I am, I was trying the two things minimax, it seems that on a personality level I like it more than GPT, but when I ask you to design systems or processes, like it got tangled. GPT manages to do more, it's more, it's like, more Matthew in that sense. But of course, you kind of have to build the system that works for you. And what's the fun of packaged products like James Park Club, or things like that, that in the end you don't have to worry about the tools behind it, but it's their team that is connecting things behind. And it did happen to me that on the other computer I added Hermes, which is Hermes Agent from Nose Research, and it has very interesting things. There are some that work a little less, perhaps, than Open Club, like the other one already has many plugins that people have created, and it has a lot of things. But you can do everything that has to do with research, which I kind of have there, make these sources that I'm saving, classify them, get me information or patterns or things. It's working super well. And probably for everything that is like creating content, I use Hermes more. And everything that is more like Chief of Staff, the one that helps me review the notes from a meeting and convert it to something else, etc., I left in Open Club, which has more connectors. What am I? And what I want to do, Próxima Semana project, is leave the two of them chatting in the same group, and let them talk with each other and fix themselves, and fix my life. And that on what channel? On what platform? In a Telegram group, I think, or in Discord. Either one works well. And they both handle it well. Remember that they can still connect via ACP. Yes, but what more entertaining to have a group where we all live. Where you see what they are talking about, it seems perfect to me. Sure, sure, sure. But, but of course, the problem is that still these projects, that is, if what we were talking about, the last class, the previous chapter that the adoption of guidance is still low worldwide, imagine that this, which is the tip of the ultra-personalized, ultra-flexible agents, that connects and everything, is still much greener. So, things are clearly going to break, clearly things are going to have to be validated, that the device is on. After your list, I left Spark doing a test of local models versus GPT versus GROC, which has now launched the subscription, it can occupy the GROC subscription as an agent. And since I had Twitter Premium, I said, I'm going to use it. I left it doing the test and suddenly the machine crashed and turned off. And I didn't find out until in the afternoon I told him, hey, it's there, and he didn't answer me. And I went to see the computer and it was like at rest. So, of course, there are still things where you have no way of measuring and that requires building systems upon systems, alerts and things that one discovers along the way. So, on a productive level, I believe that it is not there yet and you will always find that something can break, that something can fail. But the cost of... is that it works and the added value it adds to you is much higher. Igual vale la pena. Hey, Diego, I'm the host this week. Please. Just in case. But don't worry, later we can make an instance where we can answer all your questions about how to do things. Already. But it's up to me, what I was doing last week, and I also did a lot of migrations. I stopped using OpenClaw and was also trying other models. I am a fan of Minimax, but there is a problem that I have not been able to solve. Sometimes you forget to use tools. It has no problem in Tool Use. In fact, when he does it, he does it extremely well. But I think, and I did experiments, but it turned out very well for me, that, as you know, most tools or skills have a description and if I wrote that mini description, it kind of loads it. Clear. I think it is added to me with Spanish. Could be. So, as I already said, I'm going to make the changes, so I changed the skills, I made them in English, but since I communicate in Spanish, I didn't mache them either. So, talk to him in English, po. There is the solution. Yes, yes, that would be the other option, but I haven't done it. But maybe also in English, why is the model Chinese? Yes, po. The thing is that the supposedly independent training is multilingual with 36 languages, both this one and Gwen or Xiaomi's, but I can't get that type of Tool Use to work well for me when I am in a conversation with the agent. In OpenCode or CloudCode with Minimax it works amazing. Really, very good. But I never managed either with Hermes or with OpenCode to really grasp all the times I had to use the tools. Then I started to occupy the Xiaomi one. And I bought the Xiaomi membership which was even cheaper than Minimax. However, I think I'm going to go back to Minimax because Minimax has something that none of the others have given me. I'm not running out of tokens. And regardless of the fact that I was using exactly the same amount of work that I used with Minimax with Xiaomi's, with Xiaomi's I spent 200 million tokens in one day. Talking to Hermes, importing skills. I don't know what he did thinking that generated so many tokens. And the other thing I have done this week... Remember that when you are passing on all that information and it is the same conversation and you do not change sessions, it takes all the additional process as well. No, that's very clear to me. What is good... Another thing I don't like about Minimax is that the context is only 200,000 tokens. The one from Xiaomi, the 2.5 is a million. But I know they are adding up. But you're not going to spend 200 million tokens. And on the other hand, they challenged me last week for our latest podcast. Because I came out saying that they should stop going around and using agents instead of looking for money to hire people. And they told me, but well, have you done marketing agents? No, but it's full. Yeah, bro. Have you done marketing agents? No, bro, well. It's a matter of lifting a stone and someone is selling you a set of marketing agents for $9 on Instagram or for you to create your own agency. But have you used it? No. And these two previous days and today I dedicated myself to generating a... I love that Christian's motivation to do things is that... It's a challenge. Well, challenge accepted. And I created 10 digital marketing specialists. From Copi, one who is viewing the CUA, UI, UX, the one who is viewing the images, etc. And today I launched the first campaign 100% without doing anything myself. But nothing. I mean, I supervised. Which is very different to do. I don't know how the campaign is going to do because it is very different to promote a product that I think would have been much easier for me. In fact, I thought about generating agents who developed a product so I could test the campaign. But I have the community, so I'm testing what I can do with the community with this marketing team I made to see how it goes. Well. I don't know if it will be profitable. Oh? The agents where did you do it? I made the agents in CloudCode to test. Because since there were multiple agents and they had to talk to each other and I wanted to be supervising, it was much easier for me to do it with CloudCode. The idea would then be to transfer what I learned to something like Crew or with LengChain so that it is only 100% autonomous. Because, for example, on Saturday I want to see how the campaign is going and I'm going to ask, hey, check. And that's fine, but I shouldn't say, hey, check. I shouldn't even worry about reviewing or optimizing. That is the world we have to reach. But I wanted to tell you guys, yes you can. So don't worry, really, there is knowledge, it can be done, you can get creative. and not hold me back. I would add one thing, that in addition, many of those agent templates are on the web. That is, you can go to repositories and you will find people who already, some people from that line that you can reach and import the instructions, import how it works. Cloud itself, in Cowork, has the plugins. The plugin is a bundle, a group that already comes with skills and applications for how they converse. And there's a marketing one inside. You connect it and now Cloud Cowork has all the marketing tools. You don't necessarily have to connect all the services and you already know how to operate very well, you understand the briefs and you can take action. The goal MCP came out. So, if you want to run meta campaigns, you can connect it there. So it's getting easier to even do it. What it will not replace for you is what you want to sell and what you want to tell and where the value proposition is. It can help you a lot, but in the end that is a decision that you have to make. Like the decision to use agents. Cachai? That they are available there. You have agents like Manu, Jens, Parque, others that we have talked about that you can also use for this type of things. Open Club Hermes you can also connect things about the services and make them do marketing work. I mean, in the end it's about sitting in front of the computer for a while and doing it. In the end that's it. I mean, if you don't have someone to tell you it can be done, maybe it wouldn't occur to you. So, thanks for challenging me. Yes, but on the other hand, a lot of entertaining things have happened this week and well, today is Wednesday that we are recording, it is most likely that we will release it tomorrow, Thursday, which is a holiday in Chile. Again we did it before a holiday. But he's going to do super well, he's going to do super well. Yeah, good. In fact, part of the learning they launched a podcast on a holiday, people on their days off don't listen to us. Look, what do they think? Yeah, but it doesn't matter. Yesterday Google came out... What's happening? Clear. What do they believe? How are they editing videos and doing things on a holiday? Yes. Yes, but the Google I.O came out. where they are announcing all the new Google products for... for this barracks, it's not barracks, it's for the semester. It's just not for the year, because they are going to release another one later. No, but the I.O. It's once a year. It's like... It's the milestone of... This is what we're doing during the day right now and what's coming up for the rest of the year. And they will continue releasing things in between and then there will be more ads and everything. Approve the big event. And they took out a lot of things. From their new model, which now I'm going to have to add to my Benchmark because new models came out, they even got agents. And I was reviewing the video they showed. Well, the talk, because in fact Rodrigo told me look at this and I started watching it and I was stuck. But I don't know if the fun is as much as an Open Club or a Hermes or as it is, here it is connected to the entire Google suite and I work 24x7 for you without having a computer. That product you are talking about is Gemini Spark. Yes, it's Gemini Spark. It is Google's answer to 24x7 agents connected to your personal context. Well, then, in fact, he made me a mini comparison because he wanted to be prepared for this, but the reality is that I can't prove it yet, in fact, I think it is only, according to Rodrigo, working for the ultra plans in the United States. It's a matter of a couple of weeks before the rest of us can see them. In theory, of course, it will not work locally and it will not be multi-agent, but it will have computer use. And it works, the fun thing is that it works on Google Cloud, that is, it is on their server, but they assign you a space for your agent. So, it's not like it's living with each other's agents and data is leaked, no, it's in your space with an amelized part and it connects to your Google services, it has connectors, which is something that Google didn't have even in Gemini yet, in fact, it still doesn't have it in Gemini, which is the ability to connect to third-party systems and do things there, which is where there is great value, as we have said, I connect it to my application, to Granola for meetings, to Hubs for the CRM and then I connect it to my website, it has to connect and that is very good and it has computer use because it has its virtual machine, so in the end with this tool they are trying to, the funny thing is, they are fighting like in a different lane, because you have Cloud that you have with Word Code, right? that comes in the application and is focused as in work and you have codex in the world of ChatGPT that is also like open, focused on code but you can also use it for other things and here it goes like instead of going towards what is like a more professional context Google says no, I'm going to be a personalized agent for the person with their context so it goes like also playing in a different field I don't know if I can answer or I don't know if Diego wants to say something about it I have a question because I wanted to try it and I couldn't if it can also be managed from some platform like WhatsApp Telegram Discord So you have Gemini on Android It is from the Gemini application it will appear in the bar on the left as a Spark button that has its URL or in the application and it works the demos they made look pretty good yes, if you live in Google context and you have everything in Google Yes, yes I think that is going to be the limitation just like Antropic Today we have talked a lot about not marrying a single provider and in fact yesterday when I was doing the tests to do the marketing campaigns a few weeks ago I had used Cowork so that Cowork could take the power of my meta business suite and my meta developers per website and it worked quite well for me but yesterday I realized that they blocked the meta developers so every action that developers do I have to accept it and each action is a click or a scroll so it is no longer useful for me to use Cowork to help me with the management of meta applications or even with meta campaigns well I understand that they have an app and they have a lot of things but I don't know if Antropic removed that limitation if meta asked for it if something It really blocked I don't know why yesterday I was saying that I'm saying wrong in fact I was about to see what other tool like Cowork I can use at this moment because I didn't want to do it through API I had to create my own API to go directly to the goal campaigns there were a lot of MCPs that I didn't see the goal audience so I'm going to have to review it but they charged me at least 50 dollars a month to use them and in the end I ended up not using any going straight but thanks to you to earn money that's like that for me That's what's causing me pain in quotes lately is that there are very rapid changes, right? In all the tools, that is, you open Cloud, for example, today the desktop application and every other day you restart to update the application because there are new changes and suddenly they are super small things and suddenly they are mega functions, right? So you have that but sometimes you do that and a connection breaks or something that was there stops working and the same thing with the agents, like deep down there is a cost for going fast, right? Because so many things come out that you actually have to be continually rethinking how you are working, how you connect things, what the processes are that you are designing with these tools, the other things that they released, Google already had Antigravity, which was their development ID, they also have Gemini, it's called, it's not Gemini CLI, what is it called? Google or Gemini had the Gemini CLI and now they turned it into Antigravity CLI, yes, now they grouped everything together in Antigravity and the funny thing is that the Antigravity application was an ID, right? which was a development environment with the project bars and the whole thing they turned it into a Codex type version which is a people's application where you are no longer seeing the files directly and if you want the ID that still exists you have to download another application instead of updating the ID so that it would continue to be ID they turned it into something different and now you have the Antigravity CLI the normal Antigravity the Antigravity ID and there were others there were four I don't know but in the end I don't know at what point we committed to having a podcast where We'll have to keep people updated because I don't know when I'm going to program the new Antigravity or the four Antigravities the SDK that was the fourth oh my I'm not going to avoid it but it's a very entertaining world we're living in that's the truth no but in fact I don't know if I told you the weekend or yesterday or the day before yesterday we have a group the three of us on WhatsApp that I don't need new things to keep coming out I need to dedicate myself to producing that is I really like technology I'm passionate about reviewing I'm passionate about trying new things but there is a moment in which one has to also create, not only study and learn, that is, it is called the time to learn because the idea is that at least in one hour a week you learn something but I think that of my week more than 60% is only learning and 40% is with family so at some point I work at some point you are working Cristian but I don't know how you see it because in the end we all have to do things and the amount of things that have to be learned from what we are experiencing is enormous and that's where I understand why the majority of people have not used AI or the majority of people are not using these tools because it is a lot of time that one needs to be able to learn everything that is happening. I always give the clear advice. We are always talking about the latest, the new, we are trying. Hey, look, this always happens to me. This is a super repeated story. Diego is a witness. I send them a link through the WhatsApp group we have months ago and about 3 weeks later, Cristian says, hey, look, this is what I sent you 3 weeks ago. That no longer happens and there are times when I send it to you knowing that you had sent it before I thought yes I know it is the tradition but clearly it is something that takes a lot of time and in the end it is not necessary to be testing the latest it is not necessary you have to test it that solves a pain that you have in your daily life you have a pain because there is a process that has a bottleneck do an automation learn about automations see how to solve it it is not that I want to connect this with that I do not know what hey maybe it is an open club maybe it is a Hermes maybe it is another SDK there are ways but like don't start as I'm going to learn by learning that There is still value in that, if you have time and can do it, do it, but usually identify a pain you have and see how to solve it and in that solving you learn, right? so that it makes sense and not necessarily the best solution, that is, now that Google has come out with its agent, it can be very good but you don't necessarily have to leave the agent you were using to change to that one because the one that works is already working so it is good to know what is happening in the trend but not to do this as a quick jump from one tool to the other and all because in the end it wears you out, learning does not consolidate it does not translate into results, cache? in actions then I think it is not necessary to be testing the latest but at least understand where everyone is pushing that now with Google ads that Google releases its agent for example personalized is a validation that the Open Club project is pointing on the right path and that all those who got on when Juanito our friend Jensen Huang gets on stage and says that Open Club is like the new paradigm shift, right? equaling it when DeepSeek came out and when GPT came out, what were the changes? Like I have a chatbot, I have a reasoning module and I have, right? Now this personalized agent, of course, the only thing he is doing is validating the trend. Google did not present anything yesterday, wow, it did not present, did you not find out in the lens? Yes, but who of those lenses does not have Meta with their lenses or with the other competitors? I loved the beautiful execution. I'm thinking that my Ray-Ban-Metas are going to live in the drawer later when I bought that, but do you have Ray-Ban-Metas? yes, but the sun ones, the sun ones, but the issue is that it didn't present anything that couldn't be done with other tools or that other competitors didn't have, no, well, if not, if you're not reinventing the wheel, what it has is a fucking ecosystem that's also profitable, and that does shit, everything else is because while we're talking about how many billions are you talking about, Antropic or OpenAI is worth Google worth trillions and Google is earning billions while the others are spending billions, yes, but the thing is, of course, but in the end the funny thing is that Google didn't present anything new. to the ecosystem like these developments that we made and no one has it but the great grace it has is that it validates that the trends where the others are is the way and they come with solutions that come, as Cristian said, super immersed in the Google ecosystem, very well packaged functional that for the common mortals who have their Gmail it is already enough when they tell me hey hell yes chat that's awesome Gemini I don't know what Gemini if you have Google and you have the photos and you have an Android cell phone, go with Gemini and don't get complicated and it will have a button What are you writing to him, hey Gemini, I don't know what and it's already on the other side and it's included and it's included so it's not so necessary to jump around and the last and all the nice thing is to see that Google is finally catching up in terms of features in various tools that Dede was, right? compared, for example, to Cloudy with OpenAI and which also packages certain new products, for example Gemini, which is this model that is new video, right? which is omnimodal video image everything so this is what I want from Google, a model that works wonderfully, right? Nano Banana is still the king of images the last year and a half of Petitio so ah? I think Petitio beat him with Image, which didn't beat him, he didn't pass it, but it works well and basically, of course, if it's there in GPT, why are you going to change to Nano Banana? If it is there in Nano Banana, why are you going to change to GPT? But in the end it consolidates Google's positioning. That's what was entertaining about yesterday, as if it validates that the AI ​​strategy is clear that it has the products for the developer, for the ordinary user, for the mother, for the photos with me, right? In other words, he has the entire pool complete and that for the mortal, the average man walking down the street who is not thinking is not listening to an AI podcast but what he wants to do is send a meme to his friend, right? because it has other problems on a daily basis it is wonderful because basically Google is going to put it there in your pocket and you didn't realize it so you are going to start using it it is going to help us with the adoption it is very good thinking of me for the most mundane things of the day to day life of the company not so mundane either that is I did not see everything they announced in fact if you go to the Google blog there are more than 100 ads 100 we talk about 2 or 3 but there are 100 that is from their IDE now that the best AI will be included directly within Google, which was already there, but now it will be more embedded, it will have Google Beam, which is the competition of integrated agents in group meetings, for example for Zoom and others, and obviously, in Google Workspace, each time it increases, it will be better and it will help you do things within Workspace, that is, I don't even have to use Gemini or Omni, I'm not directly on a Google chip and I'm going to tell it to generate content for me or summarize this and it has the same function as AI. It works wonderfully and in general I think the trick is to serve the dish massively and beautifully so that everyone can eat. Did you eat this? In fact, we also talked a couple of episodes ago about the One Man Army that the reality is that they are generating so much technology, so many agentic processes that we can have entire teams that it is most likely that even this change in Google will make the same companies realize that they are much more efficient, nor go out and buy their Open Claw or their Gen Spark Claw or generate an agent in a BP, no, that is, with the same Workspace subscription that they already have, they are going to be more efficient, they are just going to have to see how to implement it of course. Yes, and the only thing I think I didn't like so much about the announcement is that they released Gemini 3.5 Flash, which is the new model and it is not much better than what they had before. It has a couple of good things, mainly in the use of computer use tools. In other words, they trained it to improve the agentic characteristics that are requested, which is why the agentic AI event, but the suspicion is that this is just a retraining on the model they already had because the cut-off date for the data is still January. 2025, that is, Gemini 3.5 has a year and a half empty, which is very rare, normally the models have 6 to 8 months of empty time since they start training, so the suspicion is that this is not a new model but rather it is a retraining of the model that was there to improve the use of things that it was missing to be able to get into this agentic world. I want to see what happens when they launch 3.5 pro or 3.5 with thinking, to see if it has the same cut? to see if it has the same cut or really and made a new model the 3.5 thinking if the pro is there no no they did not release thinking what happens is that thinking is important no if it is 3.5 thinking what happens is that Gemini copied the playbook from Claude and what it does is that the thinking is on the flash model is flash with thinking ok it is cheating chat GPT had the fast model and model with thought Gemini what he did is that he has the 3.5 flash model 3.5 flash with thinking that he dedicate more thought token to answer so the quality of the response improves but it is still the same model and pro that it is a separate model that is more capable and heavier that they are yet to launch and that they are going to launch supposedly during this month and in addition they launched the 3.1 flash light which is a faster version that for me is like getting a haiku from Claude so it seems that at the model level that it lacks a little improvement where it has to improve for Spark to work so that it works more agentically for example in all applications agents that they showed for example in Google Search in the search engine that they showed in Antigravity and things like that are cool but at the level of capacity of the model it is not that it has really improved as in intelligence or in information or data it improved the two levers that had to improve us so that left me with pleasure to hell Google I want you to take out your model that breaks it and that you take out a model and that at that moment it is surpassing Opus that is surpassing GPT I am going to give it the benefit of the doubt because In fact, of the things I was seeing in the morning, many said that it was better than Opus in something or better than both in something else, so I'm going to wait to make my own Benchman to be able to give you that answer because I need more information to see if it is better or not, however, of course it is designed to be agentic of course hey Diego, you are quiet Benchman I was also reading how in the latest Google news they found an interesting story and about the models and I also had what things I had activated in my Gemini personal and company account, which I have both. And of course I also have a personal allowance but I don't know if it's the ultra or the ultra is expensive or in fact the most likely thing is that you're paying like 20 dollars a month but for the ultra I think it's 100 right? Yes, I don't think I'm paying 100, I don't have that. Well, taking advantage of the fact that you're talking, Diego, I have a question for you, tell me, what do you think of the war between Antropic Google and OpenAI? very interesting I think I like a comment that Rodrigo made in a post as he was talking about the news and I don't remember if it was that way but what stuck out to me is that Google was saying a lot of things and I'm not sorry I'm sure this wasn't the words but that was the point that I'm not going to be worth making all those announcements and like with a player with a tweet from a player that draws much more attention than because of him I don't remember the specific name against the song by Andrés Carpatti of course I think I don't know, it really made more noise thinking that the news is also kind of super nerdy but I think it's a super important niche, it's from the people who did the research and development at OpenAI at the time and now he left for Tropic and beware Andrés Carpatti is like the Messi of the guide, I understand it the other way around if you tell me Messi Carpatti of the AI but basically what happened was that Andrés Carpatti, who was from the founding team of OpenAI, who was the one who designed the autonomous driving system at Tesla, the original one He was doing an education company and yesterday he announced that he went to work at Antropic, right? Then it started to generate noise, but it was brutal because basically, first, one of the people who best explains AI worldwide has created super interesting open projects, right? like the auto research that works wonderfully that basically makes it seem like the AI begins to optimize based on a goal and begins to iterate during the night which is like the precursor of this Slash Goal command that came out now in the tools and has always been sharing a lot of things and is launched with Antropic yesterday in fact I am looking here to make the comparison in Google Trends with the Google AI to see if indeed who set the most tone yesterday the news of the Google AI generated more news in Trends than Carpathia yesterday but I am sure that on Twitter it was On the other hand, that's where the niche is and well, if you're wondering why Rodrigo knows so much about what's going on, it's because he's a tweeter. I'm a tweeter. I'm a tweeter. Yes, I'm always reading what's happening on Twitter, from the vinete field to the hiring of Andrés Carpathia, but good news. I think the interesting thing is that beyond him, the only thing is that several stars have left as well, and that's what I think. I think he was also fed up when this thing happened in the United States of saying no to the government for issues of use of its platform, such as making Make a Tropic and seeing himself as the good guy in the movie who carries it in the culture and who also wants the guide for the good of the people. I don't know if they agree, but I think that has also moved people. To no one, if I say no to a government, it is not because I am a good person, because something could be generating conflict in my company. If you think that Antropic is still growing, not at the expense of the tens of billions that OpenAI had, but it still has them, that is, it can generate certain levels of conflict in the same way that OpenAI. users but it is growing like foam, it is marking an opposite point because it is concentrated only on one task, one thing, cachai? No, I mean, if Diego didn't change the subject and we continued talking about the war between these companies, we could talk about where each of them specialized and within that, the same Opus that you said, hey OpenCloud stopped being the same when I no longer have Opus inside, it's for a reason because Antropic at the level of the models does do different things, right? They raised the Constitution in LA and they have a document that tells you that it is like a letter to Opus that in the training data they put that document so that you have, as you imagine, the spirit of Opus and Carpathi went to work as in the pre-training team to put together a pre-training team of the models, that is, what they are going to do there is not minor because in the end it also talks a little about what Antropic is looking for, which focuses on productivity that focuses right? in a model like in this thing that they call it the safe EI, right? Like security and the issue of telling the Pentagon no, I don't want to work with you, it is a declaration of principles of theirs that one can agree or not or suspect or suspect that there is marketing but it is related to what they are communicating or what they seek to generate, right? On the other hand, the others are much stronger, that is, a GPT chat and a Gemini are competing in other leagues and therefore they cannot innovate as quickly, that is, that is where we could start to talk, for example, in which leagues are they competing? I mean, after yesterday, yesterday Google took on KPMG and PricewaterhouseCoopers as clients, of course, but well, the others also had large consulting companies within it, that is, and they also occupy the cloud and have agreements with other things. It happens to me that of course Gemini, for example, plays the Google playbook. What is the Google playbook? This works wonderfully with my tools with my ecosystem and if you are fully involved you will have all the benefits, so Gemini on its playbook is the entire massive wonderful Google ecosystem, right? It is not focusing so much on the professional, it is focusing more on the individual, so the Gemini focus goes a little there, it has things for companies of course, but all the functions that were there presented products that one does not understand the nomenclatures in the end because you have Omni but you have Veo, you have Laira for music, apart from that, in theory, Laira and Veo enter Omni, they are supposed to enter there, but we have seen so many times that Google has had parallel projects that live together for a long time, so Google's problem is that it has an ecosystem with many pieces. I think they get paid a bonus for the guy who invents a new brand for Google, so of course Google plays, if you have all this, it will work and some of those tools we launch are for professionals, but it is practically focused on the GPT mass market of the 900 million users or more than 900 who declared last year, which is estimated to be around one billion, 90% are free users of free accounts, 150 million are supposed to pay something like that, so that means they can't innovate much either because If the cost of changing a tool is no longer so high, I create an account on the other side and continue working. In fact, that is one of the points until the day before yesterday. In fact, literally the day before yesterday, it seems that Coopers was an Antropic client? Yesterday Gemini changed, we are talking about not companies with two employees, we are talking about thousands, the cost of change is not that simple, that is, the cost is not that much. The technology continues to operate. I have to follow the tasks forward. If I change the projects from one side to the other and so the context, I am ready, so was it changed? from Antropic to Google but it's probably because it already had Gemini. I don't think the work environment I think it already had. I'll just say what they said. We can ask them why they did it. If it's important to double click sometimes, also why Gemini also has so many people using its model in this adoption race, it's because people already had it. So it's not clear to Google that the distribution is always easy. If you have YouTube users, you have Gmail users, you have Android users, so in the end, between Those three things, in fact, the number they presented yesterday was that they have about 3.5 billion, that is, 3,500 billion people who use Google services in the world, so for Google they make a decent product because Gemini has never been at the level of ChatGPT or Cloud at the level of the application tool. I think that only yesterday with the update of the application they became on par at the model level. It was competing well but at the application level it was lacking but it already had many users because the companies that work with Google are well linked. Workspace and they have the plan, not the cheapest but the next one, they already have Gemini inside, so Google's way of introducing the Playbook, hey, how did they make Chrome become the number one browser in the world? the same Playbook, right? So that's the great grace that Google has and it knows how to play that game very well. In fact, they will never be trying to be the first. They don't have to solve the pain well enough for it to be useful for the average user. I have had to live it. I have been in Cloud since April 2023. When they launched the application, I used to use it before by console, but when they launched the application from the first day that I am there and of course I have seen all the advances that have their focus, it has always been business, productivity, code, business, productivity, code. They do not have image models. They do not have video models. They do not have music models. They have been constantly innovating because of course they are the youngest and they have to stand out in something and they concentrated on that and in that they have done extremely well at a level that, for example, and it is worth more than OpenAI in the last university, yes, but to finish the idea, the issue is that of course I think that as information that I give to everyone who is listening, I am dedicated to doing AI training for companies and I have the data from my courses in Platzi that I have the Gemini and Cloud ChatGPT course, ah? Do you get Cloud's better? The Cloud one is doing better in Platzi and in companies today everyone asks me for Cloud courses, no one asks me for ChatGPT courses anymore and of course the companies that are in the Workspace world ask me for Gemini, cachai? And those who are at Microsoft asked me for Copilot and are asking if Cloud should be added, then the change in practice of what is being seen in companies and all this wave like this Cloud trend that there is lately is real and in truth we are seeing how it is entering massively and not for nothing the evaluation that Chris said that Cloud at the moment apparently with the hiring of Carpath with the agreement of Elon Musk's Colossus 1, right? of Space Space AI as they now call it Space not generating money may be betting more because it generates money in the end through the cloud because of course OpenAI and Antropic are making a lot of money from the cloud so that's also why they invest in these platforms, it's a virtuous circle that it's good to invest in you because in the end you're going to buy but not like that, it's not being profitable, OpenAI is not being profitable until when are they going to continue investing and taking the risk also that it can go to zero thinking what you also say Rodrigo that it's going to be a gamble risky, a lot of people will lose maybe investor confidence and it will go to the floor now now yes there give me a second there be careful the strategy of very large companies burning capital exists because one will work like this it is not that they are going to compete until they kill each other one is going to end up acquiring the others and become profitable and by acquisition they are going to make money that is to say this model of burning money like an idiot is the gringo blitzkilling model and it does work for them for other reasons in Latin America we can't, but because we don't have money to burn like idiots and the number of people is much lower than users, so the capacity for growth is much more limited, no, and the investors who are in Silicon Valley made their assets with Venture Capital since the 70s, that is, they are industries that began to generate assets by investing in other companies 60, 70 years ago, that is, they go further forward, they can do this but it is not that one of these companies is going to reach zero, but it is most likely that one will dominate the entire market of some future if another actor does not arrive and integrate it vertically, only Rodi comes and there is an issue that I think is super relevant because it is not that they do not have capital, they have investors and they announce rounds and they announce all that and what happens I think it has to do with the Antropic growth strategy in terms of profitability is better because it focused a lot on the company so there is the big ticket it does not concentrate so much on the sale of individual plans even though it is doing better and better the strategy was a company one and one saw the curves when they started with all this IPO that they want to go public both GPT and Antropic and it was seen that in that sense the scale of Antropic is much more efficient for the size that it has ChatGPT grew massively it was the one that inaugurated the category if we do not have to detract from it in that either and it has the support of a lot of companies it has commercial agreements with a lot of companies it has a gigantic massiveness and also one sees how the same accuses the blows that is for example Zora who was the video model a month and a half ago They retired, they had made a whole application, a whole issue because they focused there, but if you go to see, for example, the conversation today between Cloud Code and Codex, it's true that it's like where everyone was saying no, Cloud Code is the best, and now they're all changing to Codex with the latest versions of the application, and in fact the Slash Goal, which is as a mission, is based on Auto Research, which was launched first by GPT for the first time. They were like innovating in something, so they added Computer Use in the Codex application on the Mac, so it's like me I would consider it dead or as fast as Diego said, just as they are not going to and they fell down the entire House of Cards, I believe that it still has a lot of power left to compete but in fact the market or how this competition between the actors is taking place is obviously forcing OpenAI to try to direct the darts better and not lose because it began to cover a lot then the water began to seep in so now it has to dedicate itself to some lanes because it does not have Google's back and it does not have the aggressiveness of Antropic to go to the ni cho full as well as obsessively then you have to find what your path is and where you have to stop now the interesting thing I think is that more than what is happening now between these three large companies is that it is how it will look when each of them launches the next model because in that leap is where I think the difference is going to be made, that is, when the GPT 5.6 or the GPT 6 comes out the Gemini 3.5 Pro or the 4 Cloth reaches version 5, it seems like a leap. There it will be as clear as how much each one jumped versus where they were and for me that indicator is going to be super relevant in terms of how effectively they are really training the next great model because if we are used to the fact that one jumps the other jumps the other and they are generally similar and they pass by just a little bit but now it is calculated that the growth should not be so much more so the one who jumps the most I think they are going to cry more and they are going to start winning the competition more because they are going to have trained the model in a different way I think that The fact is going to be better now, guys, you're just going to give me the pass so we can finish today's episode. It's been super interesting. It's been super entertaining. But I would like Rodrigo to do it in part. We'd close by telling a little more about what you think is coming now and it's an abstract question. By the way, you can say what you really want. I think that one of the things that I don't know if it's abstract is what comes from the world of AI, which is what comes with us, what comes with the world, what you want to answer, what Your heart tells you what my heart says is that now the issue of distribution and in particular the issue of adoption within the company is going to become fundamental, of course, that is why sometimes they have launched as their own companies as well and have bought to increase their distribution and in that lies the great challenge also of what we were talking about, hey, a crisis as a desperate means, but why is adoption not advancing so quickly and the company is not automating in that, I think there will be a very big opportunity in how it accelerates and helps companies to accelerate their processes and also as our friend Juan did about how to have these AI strategies, your friend Juanita from sends these AI strategies and agents and in that we also have to see how companies and well and we also take advantage of that window of time thinking that all this is also evolving so quickly that one does not know how far we are going to go humanity versus the space that androids are going to occupy and the use of artificial intrigues and to sit and work while one sees everything that is happening with some goats and executes everything learning I know that we have focused a lot on AI but AI in real machines like androids and robots is aquatic I also hope that the next episode we can talk a little more about that because this week incredible things also came out Yogi the robotics thing is very very very good to me what I see that is coming a little clear that the war has already been declared that personalized people is true and we have talked about it our podcast started with the first chapter talking about it with Open Club so that is already there and what I think is coming now is the consolidation of that category in the sense that what Diego said doesn't happen, I don't have to go around patching it, it doesn't work for me anymore, as it doesn't work, that is, they are more consistent models that are tools with a better harness, it's true that it works better, that it's more like the connections are simpler, we've been seeing a lot now that this one is coming out as the slash goal so that I can work a little more autonomously, these are coming out like the Cloud plugins or now Hermes has released one that were the skill bundles that basically come out as several packaged skills. with certain functions, then it will be much easier for my agent, someone, a marketing agent, to install this marketing bundle, this meta bundle, and it will automatically work, so it will simplify how we connect and how we create the recipes for how the AI has to operate. I believe that on the one hand, that is going strong and that means that it will increase the number of collaborators and we will also have AI employees, certain agents working there collaborating in the company, which should unlock many labor bottlenecks then. For me, a big trend has to do with this category that is becoming more professional, putting on a collar that is true and going to work in the office, and the other one that I think is growing a lot, luckily, is that given the above and the topic of AI, all this is emerging from documentation or the second brain, which before was like a practice as well as the productivity of the nerds as well as the clubs, and now the question has become serious that we actually have to document, things have to be in writing, they have to be the information because that is going to mean that the context of the company or on a personal level can be exploited with AI and there is a lot of captured value that currently is 50 folders and you don't know where the file is etc. and that AI is going to be able to capture and exploit it much faster, right, and that is going to unblock many current bottlenecks and within that I believe that the latest trend that I have been hearing more and more about has to do with the bottleneck because until now of course one said we are going to adopt, to what? to do more we are going to add to it so that it does more good but today what we are seeing is what is called the theory of restrictions or the theory of constraints which basically is that more productivity is unlocked if you are able to find that bottleneck put AI to unlock it and normally that bottleneck is a person who is not paying attention to the email is a person who said I have to send something but he is going to send it between hours because he has meetings and that has already delayed the process elsewhere so there is a lot that has nothing to do with making the problem better but rather has to do with making it more efficient, right? Not because the person doesn't want to do it, no, no, it's because they don't want to do it and do it in a timely manner, not to do it three hours later. The key is that in the end one comes back to the topic of business processes, how to draw the process, identify where it is, why we are delaying a task and how through the agents, through the context of the company, which are the other two trends, it's like you connect two cables and it unlocks automatically. Cachai? and that for companies is going to have a much bigger impact than having an agent who does everything because only that is going to unlock the productivity of the entire team to grow so there is an issue and reviewing the company's processes is going to be vital in the short term because you don't have time to do it you have to do it now because later the most advanced technologies are going to come out and if you have not prepared well by identifying those bottlenecks that as soon as the technology arrives you can exploit the difference in terms of competitive advantage between companies is going to be very big so the company that does be prepared you are going to have like a is like a how you are going to shoot like a rocket and the other is going to go there on a tricycle so for me that is the combination agent context and identifying the processes the bottlenecks Rodrigo Diego thank you very much for everything many thanks to everyone who listened to us talk nonsense until the end of the episode I hope we see each other in the next chapter all I can do is say the bells again follow us where appropriate for those who are undertaking and want to undertake with technology follow me for those who are starting out, that is, for those who want to know about technology Rodrigo, for those who want a course, some training or something a little more corporate with Diego and I think that I am not missing anything in the inkwell, take care of yourselves, make sure you are super well, see you in the next episode, see you in person at some point, we have to see each other in person soon, yes, take care of yourselves, guys, be super well, it is always a pleasure to share with you the preview of a holiday.