← Back to search

If AI is so good, what are we humans for?

Es la Hora de Aprender · 2026-05-01 · 63 min
relevance 51 12460 words Episode page ↗ Audio ↗
Show full episode description
Si la IA escribe mejor, programa más rápido y resuelve problemas en segundos… ¿qué nos queda a nosotros?En el episodio 10 (¡llegamos a las dos cifras!) nos hacemos la pregunta incómoda que muchos tienen pero pocos se atreven a verbalizar: ¿para qué seguimos siendo relevantes los humanos? Y la respuesta no es la que estás pensando.Hablamos de la aristocracia digital (un concepto que nos dejó Diego pensando), del request for startups que acaba de publicar Y Combinator y por qué literalmente cero SaaS clásicos entraron en la lista, del estudio METR que muestra que la productividad sube x2 pero el gasto sube x10, y de por qué la universidad sigue formando gente para un mundo que ya no existe.Además, un benchmark express de los modelos del momento: Opus 4.7, GPT-5.5, Minimax 2.7, Mimo de Xiaomi y el nuevo Grock 4.3. Spoiler: hay sorpresas.━━━━━━━━━━━━━━━━━━━━━━━━━━━🎙️ 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/rojoCursos de Rodrigo Rojo→ https://tienda.rojo.me━━━━━━━━━━━━━━━━━━━━━━━━━━━🔗 RECURSOS MENCIONADOS━━━━━━━━━━━━━━━━━━━━━━━━━━━→ Benchmark de Modelos por Cristian Tala:https://benchmarks.cristiantala.com/→ Y Combinator — Request for Startups Summer 2026https://www.ycombinator.com/rfs→ METR — Model Evaluation and Threat Research (estudio sobre productividad e IA)https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/→ Minimax 2.7 — Modelo alternativo con buena personalidad para codinghttps://www.minimax.io/→ Grok 4.3 (xAI) — Última versión recién lanzadahttps://x.ai/→ OpenClaw — Plataforma de agenteshttps://openclaw.ai/→ Hermes Agent — Nuevo competidor en el mundo de agenteshttps://hermes-agent.nousresearch.com/→ Blog de Cristian Tala (análisis del RFS de Y Combinator)[CONFIRMAR URL del blog]→ empresasaumentadas.com — Blog de Diego sobre IA en empresas━━━━━━━━━━━━━━━━━━━━━━━━━━━📌 TEMAS QUE TOCAMOS━━━━━━━━━━━━━━━━━━━━━━━━━━━00:00 - Llegamos al episodio 10 (¡y todavía no nos cancelan!)02:59 - "Si la IA es tan buena, ¿para qué servimos los humanos?"05:50 - Creatividad, liderazgo y la habilidad clave del futuro09:11 - Por qué la IA es estática y nosotros aprendemos todos los días11:58 - "Nos estamos volviendo más tontos por confiar en la IA"14:53 - Aristocracia digital: la brecha que la IA está agrandando18:06 - El futuro de la educación: universidades formando para un mundo que ya no existe22:08 - El 50% de las carreras en Chile no son rentables (y nadie hace nada)25:56 - Y Combinator Summer 2026: CERO SaaS en la lista28:41 - Adiós al SaaS, hola a los servicios apalancados con IA30:01 - El estudio METR: productividad x2, gasto x1032:33 - "No es necesario tener el Ferrari, importa que estés manejando algo"34:54 - Benchmark express: Opus 4.7, GPT-5.5, Minimax 2.741:53 - La personalidad de Opus (y por qué la gente la extraña)43:03 - Mimo de Xiaomi: el caballo oscuro a US$14/mes45:23 - Nano Banana vs GPT Image 2: cuál ocupar y cuándo49:41 - Por qué la gente no usa los agentes que tú construiste53:20 - Cómo habilitar IA en una empresa de verdad (no top-down)56:33 - Recomendaciones finales: prueba, rompe, controladamente
✨ Episode Outline — click any point to jump to it in the episode
Problem solved
If AI and agents are so capable, what value do humans still provide?
Benefits
  • Humans supply criteria, context and business understanding
  • Creativity and leadership of AI tool-teams remain human
  • Domain knowledge lets you spot AI hallucinations
  • Human-to-human trust matters for high-value deals
  • Continuous live learning distinguishes humans from static models
Use cases
  • Chile's birth rate is at 1.6, below two per family, illustrating demographic decline
  • A SaaS at $20/month can have AI fully automate sales, no human seller needed
  • A $10,000/month high-ticket sale still needs a human face, not a robot
  • Use AI on tasks you already know to judge response quality and detect hallucinations
KPIs / results
  • Chile birth rate ~1.6 per family
  • $20/month SaaS vs $10,000/month high-ticket threshold
Tools / build
0:00 / 0:00
🌐 This transcript was automatically translated to English from the original.
Very good morning, afternoon or whatever to those who are listening to us right now, and very welcome to our episode number 10 of It's Time to Learn. Everyone thought we weren't going to get to the third, I thought we weren't going to get to the first, but we got to episode 10. In any case, the turning point is 20, more than 90% of podcasts don't get past 20, so we still have a good way to go. Wish us luck. Wish us luck, click the bell, on all questions, continue, give us good vibes and everyone happy. However, we don't come to talk about that, we come to talk about other things. Very welcome guys, here I am with Rodrigo Rojo and Diego Arias, my name is Cristian Tala, and today I want to start with a question that has been on my mind for the last couple of weeks by people who have listened to the podcast and a couple of talks I have given lately. If AI and agents, LLMs in general, are so good, what use are we humans? And it's true, because they asked me. And in fact I think I did a great job with my answer, because in the context I was able to respond. But I would like to hear your opinion first. Guys, if all this is so wonderful, what the hell are we humans for? Perfect, I want to leave with a little bit of a light of hope. In this world where AI has more computing speed than us, it maintains information much longer than us, right? I'm sure you don't even remember that you had lunch on Monday. So, there are many things that we, because of our reptilian brain that seeks to survive, store only what is relevant, right? In fact, I'm sure that some even forget things that they once mastered perfectly in their previous jobs and that now they no longer even remember how they did it. So, sure enough, we have bad memory, AI writes faster than us, solves problems faster than us. In other words, in that sense, AI is superior. Isn't it to give a ray of light? But, but, here comes the ray of light. The issue is that AI in the current state requires someone to give it an instruction. It requires, if we go to AI, the classic chat-type assistant, GPT, Diclode, etc., it requires the human to go and give it an instruction. If it is an agent, it also requires that the human give it the instruction, only the AI ​​will have more autonomy to solve. You are going to make decisions on how to resolve a plebe, you are going to resolve it, right? And if I do an automation, right? I make an automated process that has it in between, I also defined what had to be automated and in what order. So, humans contribute two things. One, in the criteria of how to solve a problem, right? In having, in identifying that there is something and wanting to solve it. And two, in the creativity of how to deliver the key information to solve it. Hey, I know this, I think we talked about it like in the first or second chapter. You, cachai? That, I insist, I think I know 90% of the people I know, I am going to say 90% because for people to feel that they are in the 10% of those who are listening, that they are less creative than the AI, they are less autonomous than the AI, and I also have to tell the people I am thinking about how to do the tasks and what tasks to do. Clear. So, where I go, and that's a challenge that we have. But today, right? AI is going to replace the human who has creativity, and the human who understands the business, and the human who understands the context and has to hand it over to AI to solve. Good. And that there is a disclaimer. Yes, you. So, you have to be the person who understands the business and you are creative. Exact. And for me the challenge we have as a society today is that we are talking in the laws, in how we introduce AI, in the speeches, in the talks, etc., about things that are not relevant. What is relevant today is how we ensure that the people who are in the U, that the people who are in schools, develop those skills that are like power skills, critical thinking, creativity, systemic thinking, etc., to be able to live in a world where those skills are going to be key. And today, the university and everything is designed for another world that is going away, that is falling further and further down the wave. So, there is a giant challenge there. But, indeed, for me the human will continue to be relevant. What happens is that it changes the configuration of how they do things. You will no longer be the one who executes, but you will have agents and technology tools that will execute for you. But you have to give them the instruction. Then, also, and this is like the key skill that I think trumps all the previous ones, leadership. Because you have to know how to lead these tool teams. And there are also going to be jobs where, effectively, the human is still going to continue to be relevant. There are things where, indeed, person-to-person treatment, for example, is key, right? Or where, because of the nature of that topic, right? Having a human involved is going to be essential. So, there are going to be jobs. But there are others, especially in what we always talk about, which is the office, company work, right? Where, indeed, there are a lot of tasks that AI is going to start solving and it is going to solve them faster and it is going to solve them better. So the challenge is to learn those other skills so that you don't fall under the wave. Later I'm going to tell you a little about why we came to this topic, but first I want to hear Diego's opinion. I want to comment that I cheated and Cris had... Sometimes I commented to him as the answer to what the AI ​​told me. I told him, let's see, I'm going to ask you. And the first thing he told me is to survive and reproduce. That is what distinguishes us as humans. The basic foundation itself. And we are doing very hard to reproduce, because, for example, in Chile the birth rate is already below two people per family. Globally too. But in Chile in particular we are like at 1.6 and basically that means that if there are two of us, we are a couple, right? First of all we are having less than two. So in the end that is causing us to go to extinction. Well, Diego already made his shout of sand. I'm going there. I already have to catch up. Oh, how is it. Well, in addition to that, well, I agree with what Rodrigo says and that... Well, and also part of what you say Cris that, of course, it is key to know what is there, to lead critical thinking and, of course, that in the end it is something that is extremely relevant, but you also have to be aware that the majority of people do not have it as a direct and critical center. So it is something essential that has to be worked on and that, in general, the world in which we are training children, the schools, the universities are already there. It's not for that. We are still in the industrial age of memorizing things. Which doesn't make sense. And another point I think is very important is that what we have and I don't know if at some point you will even have the idea of ​​how... That's what we feel. Then it gets cold, it makes us cold and also understanding that also makes us have greater empathy and greater... a better solution to create a certain project thinking, I don't know, thinking this about people who are cold. When you are cold certain things happen. Then I will be able to better understand how to provide a solution to that. Be more empathetic with developing a solution but a given problem. Because I thought that this is how we are cold, we are weak and the idea is going to kill us. No, I was referring more to the context of understanding and being able to move forward to generating greater solutions and also being more... generating better relationships. I believe that this is the fundamental thing that is going to happen is that we have also seen the idea quite a bit and in particular with Open Club, which is a liar many times and if you didn't know it would be true, you are telling me. No, but it's not that he's a liar, he wants to please you. Yes, but that happens not only with Open Club, with anyone. But in the end that's why it's very important and it's advice that I read to people. When you start using the idea, ask about things that you know how to do to do them better, to do them in less time, to do more of that because you will have the ability to distinguish if the idea is hallucinating or not, you will have the ability to distinguish if the quality of the response is good or not. And then, in parallel, you can use the idea to learn other things that you don't know. But part in that part of the domain, why? Because you will begin to discern and understand the capabilities and limits of the tool. I'm still going to be devil's advocate on this. How many of the humans you have worked with before have lied to you? Everyone, po. I just say it to have it in consensus. Oh, I haven't lied to you. Chanta, now. I'll let you talk about that. In chapel. Hey, I have a comment regarding what Diego said because I remembered another topic where today the human is still relevant. And AI has two components by nature in how AI is created. The laboratories are starting to put together these models, right? To optimize them for certain tasks. For example, Claude at first was very optimized towards programming and they discovered that this helped a lot in the use of tools, in writing more creatively, right? Different things of that nature. But there is a concept called the Jagged Frontier. Because? Because they optimize for certain things but there are others where they are really rubbish. So, for example, if you ask the AI ​​to make you a joke, it will make you the worst joke there is. TRUE? Because it is not optimized for that. Is it optimized for what? To help you write, to help create programs, to help strategies, right? There are different skills that they focus on. So, obviously, there are a lot of things that are more niche perhaps in our daily lives where the level of quality that AI really has is not that much and where my ability to give it the information, the context, guide it to solve is going to be super relevant. So, for all of you who are listening to us right now and love to tell jokes, don't worry, you'll be able to keep doing it. Exact. And the second thing is that, given the same thing, why does that happen? Because these models are trained for months and when they are born the model is static. That is, the model learned what it learned with the data it had in training and left. And if we want that model to later improve its ability to tell jokes, I have to do training on a new model, a new version of the model or a fine tuning of the model every day. TRUE? Fine-tune it so that it can now tell jokes well. That means the model doesn't learn live. So, humans too, we learn every day. We were born and knew how to do anything, we didn't even know how to speak, right? And we start learning skills every day and every day we learn something new, every day we grow, hopefully for the better. Every day we adopt things. Not AI, AI is static. What they have done is that the AI ​​tools, for example, the Cloud application, now have the skills, now have memory that works super well and can connect with other things, but the brain behind it doesn't learn more, right? Until Opus 5 comes out, same with GPT, same with Gemini. So, in the meantime, because this will surely change in the next 5 or 10 years, in the meantime, another thing that distinguishes us is the ability to learn and find new ways to face a problem or seek to solve something, while AI is static until the next version comes out. Now, they are all working and have said it a thousand times about how we can make a model that learns live and when that comes out we will obviously put together another chapter to talk about the subject. Well, in theory the Minimax models, the 2.7 was created by the 2.5 so there is something there that can start to be done. But remember, we are going to continue with another one. But what I'm going with to make the distinction is that it's not that 2.5 improves itself but that 2.5 takes it over to create its heir. The same thing happened with GPT and with Opus in fact. And with Opus too. So, of course, that accelerates. I think what it does is accelerate the fact that we have faster and more successive versions, but it is still not like the version that was born now is Opus 4.7 and tomorrow Opus 4.7 is better and knows more things. So far there is still a replacement as of the next version. Diego, anything else you want to say about what humans are good for? I would only highlight that in this matter of the capacity to learn, there is also a capacity to learn. There is also a lot of questioning that many people are getting sillier about the issue of using the idea that in the end as the idea is important to see that it helps us and not that we trust everything and that is that the idea told me as it was not in Google and in the end we stop thinking because we rest on all the artificial intelligence and that is serious. Well, perfect, I agree with both of them, in fact, part of the answer I gave during a talk last week was that for the creative part, for the part that they can't make a machine, the example was that the person who asked me, according to him marking according to him, was a sales expert, perfect, I wouldn't hire a sales expert to sell a SaaS for 20 dollars a month, I would have the AI automate the sales process for a product for 20 dollars a month, but if the ticket is already 10,000 dollars a month It is unlikely that someone who is going to hire a service for 10,000 dollars a month is going to do it against a robot, an automated system or without seeing the face of the person who is buying it, so for those who are listening who think they are experts in sales, possibly the human factor in high ticket is still very relevant. The reason why I came to this particular topic and why I want to start with it. so absorbed in the world of AI technology we are testing a lot of things we do benchmarking we do tests etc. that we do not realize that the other 99.9% of the world is not doing it I am going to make a disclaimer there because I am I do training for companies every day and for people every day and I have to face people who are getting into this topic I am very clear about the distance there is between someone who is involved in the world of AI playing and testing tools every day and thus creating things and someone who is a secretary in a company who is in training and tells me where the copilot button is so I have to face every day how to see that distance and there is an issue that is gigantic and that for me is even more relevant which is that there is still a lot of lack in the use of technology in general so the distance between those who are occupying AI today and those who are still climbing in Excel the same this plus this plus thinking from the schools from the State unless we are going to talk about the State so I think we have to make efforts jointly among all those who are available to spread this and make people adopt it quickly because if not they are going to fall super low now but you realize the irony of this, right? AI was born to democratize the power to have access to information, now the agents have access to work that can help you get ahead and instead of lowering the barrier we are making it bigger. I just talked about that. I also think it is interesting and I assume that we at Sofielo are doing various things in training like in school and people. of the week and I talked about it and I felt uncomfortable like what word to use that people would understand because first I started talking about certain things like I saw that everyone was looking at me with a strange face and I said I'm going to talk simpler like simpler simpler and it was strange but well in that there is a topic that is very relevant that I sort of know it I was calling it like Digital Aristocracy that although there are many tools that are free also using tools that are paid makes the work very different although now several have come out that are like 10 times cheaper, for example, than Opus and they work well, you still have to pay and people do not like to pay, the company does not want to pay for several of these tools either and that is going to generate a very relevant gap and in particular, speaking at the level of schools, if the schools do not decide the decision, it is worth the abundance of getting involved with everything in this topic such as AI and seeing that or also in relation to what we were talking about that it is relevant to learn these new contexts, a very important gap is going to be generated instead of shortening it. to enlarge because if I don't know, because of certain schools, AI can be enhanced and the other skills will probably generate millionaires when I finish school versus the others who are going to be poor and not going to be able to find employment, as being very excess, one thing happens to me, sorry, there is one thing that happens to me with AI, which is like AI has a funny thing, it's like it's kind of dichotomous, on the one hand, we all have access if you know how to speak and have internet, and if you have internet, you can access AI from a free account, which is like step one or In that sense, it is democratic and she is going to do something that is going to be competent and that is going to be better than what she does, but of course, later on, for example, I dedicated a large part of my life to marketing. If I ask the marketing AI for the prompt, the information and the type of questions that I am going to ask the AI and the requests that I am going to make to the AI, they are going to have a much higher level of depth in terms of the knowledge behind the model and I am going to be able to exploit it even more, right? to make a better marketing campaign and get more out of it, then at the same time that it is democratic in access, it makes a huge difference between the one who knows how to use it and the one who makes their knowledge available to exploit that computing capacity and knowledge that exists within the one who does not then also generates this gap that Diego said where the one who knows how to use it is the one who understands what they are doing, right? It achieves a giant acceleration and that temporal distance between the one who is only asking as if it were Google to the one who is creating completely new things with AI in the short and medium term a giant competitive advantage, right? which means that a company, for example, today is producing much faster, releasing many more features, doing a lot of things, and that generates this gap, this aristocracy that Diego said, right? If things are not done to try to ensure that there is growth from here, perhaps not at the same speed but enough to stay competitive, what you will begin to see is a brutal change in the composition of the market in the distances between people, right? people who are like living a totally different life from the other. I want to get a little bit of that because this week I had a meeting with students from a university in Holland who are talking to me about finding startups where they can do professional internships. I have business and computer science students and I was talking with them about what things they are doing in the world of AI or if the fields have changed a little and the reality is that they are not doing internships in the same way as they did 10 years ago knowing that we in a world where 10 more years we don't know if universities or colleges are going to exist, I don't know, I just don't know, we shouldn't really be closing that gap today, not only for the people who can really pay or not, but also for the courses that really are long courses, that is, studying 4 or 5 years or 6 years for a course that doesn't serve you any purpose when you leave, it doesn't make any sense. How the hell, today we are not realizing that and we are making changes. I know that there are a lot of universities that are trying to see it, but those who are entering this year or entered last year they still don't have any changes there is a political problem because in the end, for example here in Chile, a study was done by the National Economic Prosecutor's Office that I don't know, 50% of the courses were not profitable, the people who were studying were going to finish something that wasn't going to earn money, they weren't going to earn more, yes, but that was the reality without AI, imagine of course, but and given that, it's like it was already identified that what should have been done is hey, it's over, they can't do more of those courses, but nothing was done, it's like done. go out to study bye if you want to get into studying that is like studying at least it gives you the option that you have the knowledge to be able to decide yes but when people go out to study, how many people read the study and see the detail? Like they keep the percentage of races because also and before that, how many people know that this study exists? Hey, I thought I would say how many people read no no no dramatic in Chile, luckily in Chile, luckily we have that basic part covered completely or 99.99%, but what I'm saying is, how much? Think about it, when you left school you made the decision to study super conscious and knowing all the information and handling all the data nothing or it was more by tincada because I think I like this not me not that chana well tincada I took the test I had no idea what to study I took high school because I had no idea what to study and apart from there I went studying things I was honest I said I have to go to university yes I don't want to take the test again I don't want to take the test one year like the year they took doing I try and in the end anyway after the test then I'm going to get into the water of the shot and there I was discovering my path little by little but of course how many people make the decision uninformed or out of ignorance and then because they are not prepared or they are not taught to be prepared at that moment so for me it is a super systemic problem and it worries me because of course we have an advantage Chile is a small country in general the countries of Latin America if someone from another country we are also small countries in general compared to other places then we have the deep down, if we make a change to act as if there is a faster transversal impact, the knowledge that existed or the way of communicating 20 years ago when we made the decisions to enter the university may be a little different than the access to the knowledge that exists today, that is, okay, I'll buy it from you, we had no idea, but there are 20 years of difference, but well, one of the things that the study also mentioned that in the end is also very linked to the part such as systems that Rodrigo says and the dissemination that institutions do. who have cared about what the study says, I'm not going to say like institutions, nothing is more than saying hey, study with us to have this employability, study with us because we added these branches and how it looks and to give it a nice name, study engineering with AI and with I don't know and in the end of course in Chile what has happened is everything now there is engineering engineering in marketing engineering in economics blah blah and now they are also putting everything in her bio but what is not being worked on is what the market really needs and that enhances employability and that is a problem that it creates because of course selling to potential students clients, hey, our most beautiful career is different from our most effective career, I must admit that at the time I thought I was going to study a career in an institute that I am not going to say the name because it had a super catchy song and I had been listening to the damn song all my childhood and I still know it by heart that you can easily tell which ones but ok let's spend a little time to continue talking about different topics because there is something that is the basis of Diego, you said that it is very important to have it. in consideration in the same way that it may draw our attention that a university student or someone from school is not relating to AI as they will be in a future job, the same thing is happening with companies and ventures and startups. Y Combinator, which is one of the largest accelerators in the world, published I don't know if this week or the previous week a request for startups final of last week final of last week and the list comes out of I'm looking for these startups to give them money to accelerate them and grow them and you know how many SaaS companies or something related to SaaS they appear on that zero zero list, why? that everything is AI everything is AI or first everything is AI which makes a traditional SaaS as we are all used to and when AI came out I said I'm going to make a SaaS if you make me a millionaire the barriers to entry are zero it is assumed since every company every startup every venture its technological base is not Python it is AI so everything you build to solve a problem has to be leveraged on AI and not necessarily be a SaaS remember that we have talked before that we are going to change the world where instead of selling jobs working in a SaaS we are selling results that really work and yet I keep hearing people who are going to build a SaaS I don't see a problem with putting together a SaaS but if you are going to see the Y Combinator list well in fact we could share it on the screen so that the people who are watching YouTube can see it I published the result on my blog but I don't know if I have the list at hand I have the list here just remember that not everyone sees us but those who are watching on YouTube share they are going to see it here but we are going to read it for those who are in Spotify, if you look at the call now, for example, it's the one I don't even see, so zoom in. It's summer 2026, right, this is the initial call, there's for low pesticides, pesticides in agriculture, right, with a new technology, right, boosting native AI search or discovery engines, native AI services company, that is, the service, what's happening is that they're at a stage that's intermediate, it's not like for the end customer, that's why I said SAS, remember that we've always said in this podcast in particular that the companies of the future are service companies. leveraged with AI but not only a SAS is not a platform it is a clear result so that the people who are listening understand what we mean by a SAS the SAS is the software as a service it is a software with a service that is normally either bought by people or bought by companies and pay to be able to access that service here what they are doing is like they are intermediate layers with AI, right? personalized medicine with AI brain of the company that occupies AI, right? Behind this is what we call the second brain, right? The second brain but with all the data, the knowledge base of the company, defense against, right? drones dynamic software interfaces that this I think is going to be super useful in the world of agents entering and that the software changes because in the end the service is going to be less static electronics in space hardware supply chain issues industry skills in space difference chip for agent workflow SAS Challenger competitors for SAS I think you already understood part of what we are trying to demonstrate in this podcast is also that if there is a before and after I think that yesterday I was the one who said it before and after Open Club why I would have to pay for a SAS that gives me a job to do something if someone else can do it for me and that someone else is the AI or an agent then the paradigm changes I insist it is for an agent that gives me the result instead of me having access to a job to do it myself but there is a point that is very important that we have also talked about it before but I emphasize it shit times 10 with AI or someone who uses more something does it is going to be a much worse result so garbage in garbage out if did you see? did you see? did you see? did you see shit? the same thing, this version is simpler than saying like when the user than the user's problem but well what I'm saying is that I recently saw a study I'm going to look for the shot so as not to say that just like the data that production is being increased I don't know by 2 but the expense was by 10 so it's convenient to do it with AI if you're spending 10 times more and that actually also happens because people don't make it clear I'm going to look for the study is that the incentives have to be well aligned or the incentive of a startup can be lower costs and increase productivity of companies that we have talked about before, such as Meta Apple and others, whether we spend tokens depends on whether we are productive or not, let's spend tokens, so I suppose that also dirtyes the results a little, yes, but that is still very interesting what you are saying because that has also made everyone win, not everyone, but in general, Google Antropic earns money and Meta does not because in the end because of this investment by Google in Antropic and Amazon, which of course also have a virtuous circle as a business, some of which, of course, more tokens are consumed than they can. They themselves are consuming more but that also implies that they buy more from the cloud and there is Meta who is losing and it is funny now Meta is losing Meta plays another game Meta their income has always come from another side of their Instagram networks what happens is that Meta has also tried to get into this competition right and it is a bit what happens to Apple that everyone questions it why you didn't get into the world of AI and it is going to occupy the Gemini models or build on that in other words they are companies if one sees it as that the big Big Techs are all trying to enter the world of AI but their score and backbone are different, it is true that from the Big Tech that entered the AI, none of them have been profitable, of course, I would give Google an asterisk because Google also has the TPUs that they make themselves, so it has some deficiencies that the others do not, but for example, Microsoft is in a salad of things to the level that they updated their agreement with OpenAI a little while ago and now OpenAI can occupy more things but Microsoft also, for example, this week I had to do a lot of Copilot classes and Now in Copilot I can choose if I want Opus to be the model behind the brain behind Copilot or if I want it to be GPT55, so that also gives Microsoft the solution. In fact, they have now launched agent modes in Word, PowerPoint, and Excel as the default mode because obviously they are more complex tasks, so in fact, Big Tech has not managed to solve how to get in because they have to maneuver, they have a lot of balls they are juggling, right? but in fact in the end what one sees is that there are certain laboratories that are dedicated exclusively to creating and those who are actually creating the top of the line models Diego something to add if not I found the study called METR Model Evaluation and Treat Research well, remember to keep giving us little stars and everything because Rodrigo is going to put in the description of this episode the link to the study that Diego is saying here below and it is 10 years not 10 chapters so that we have much more time to talk and spend it well until later the robots replace us and we continue talking here while the robots do the work for us guys I have a final comment regarding what we were talking about which for me has to do with the fact that when we are involved in this AI and the impact that it begins to have in terms of the developments of everything one thing that I think is key as advice for everyone is not to get so overwhelmed that is to say things are happening every day new news comes out every day new models come out new everyone comes out all that matters is not that you are Occupying the Ferrari, what matters is that you are driving a car, right? or a bicycle then as if you are occupying you start little by little like don't make an effort I have to be occupying the last thing modifying the processes automating everything but it goes little by little but the important thing is that you don't stop right? because in the end after that you are going to gain speed you are going to find something with which you are going to understand something new you are going to start incorporating it and the same with the processes in the company and with all those things for me the important thing is that if a new thing comes out it is not necessary to throw yourself headlong into learning it but it is important to know what is happening and what the trends are because if you don't catch a little bit of where this is going obviously you can be shooting for one side that in the end is not going to work I'm going to hang on to that because we also talk in another episode about how much You should be spending monthly on AI but the reality is that the most expensive AIs should not necessarily be the best for all types of cases. I recently published how to eliminate the $200 from the Antropic subscription to use CloudCode with other models and I said to myself, wow, we are still talking about $200, we are in Latin America, that is more than 10% of the salary for many people, including the person who told me and it is true, that is why one has to understand which models are very inefficient because I have been I have been struggling a lot with this and I have had both my teams, servers and APIs working 24 to 7 to do an AI benchmark and I am super happy with the result. With this I reached the stack that I am occupying, for example in my OpenClaw agent, but even so, a couple of things catch my attention: first, if you can have an agent up and running even from your home for 5 to 10 dollars a month, which is good enough for many of the things that you have to do, but on the other hand, what has What stands out is that models like Opus or 4.7, which should be the best in everything, have done quite badly in all my benchmarks until I think that my benchmarks are bad. I am not comparing the same thing that the rest are comparing. I am comparing for certain businesses operations that are for super important entrepreneurs so much so that it made me think ok maybe I should see another type of test to see that Opus really works well in these cases where deep reasoning is good and yesterday I started to test all the models of my benchmark that I hope I will also put here Below so that everyone can see which model theirs is. In fact, I have tested 87 different models in more than 60 tests of each one, so you can understand how long I have been doing this. I think that at this point Chris does more benchmarks than actually working. I can only say that the last three weeks I have been very inefficient because I really wanted this to be working well because it helps me make decisions about which model to use for each thing. However, yesterday, but it is horrible, I mean, every time I think the benchmark is over I have to add something clearer ok that was the point there is a test called hey hey hey which is like a needle in the haystack that I generate a sufficiently large context I put a text at the beginning in the middle and at the end I pass it to the different LLM and I ask them to return the information that I want that I put in between to explain it with Christian Spanish it is like if you had to take the Frankenstein novel in fact the one that Marto used at the beginning in these tests and they put the word pizza at the beginning they made another one in that in the middle they put pizza and in the last one they put pizza at the end so it had three versions and you asked the guide to try to find the word pizza which in the novel Frankenstein did not exist well and I started to run the test and strangely Opus 4.7 came out among the worst and I don't understand why and knowing that this test is standardized the only thing I did was change it to Spanish in fact all the tests that I am doing for content generation are being validated by a jury in Spanish because in English the model changes a lot. better it turned out bad however yesterday I was doing tests with Cloud Code and Open Code for certain things and I came to a problem that I could not solve in any way with any of the high reasoning models even with Nemo Tron 3 Super that I have running on my Spark and Opus it took very little time to find out what the problem was not to solve it but to find it there I have a comment but regarding Opus 4.7 that came out two weeks ago Opus 4.7 the verdict of the community is that it is worse than Opus 4.6 In fact, in my tests 4.6 is better, why? because Opus 4.7 has this adaptive thinking, right? So they kind of start to doubt a little and what people have said is that, for example, the code is very good, like in the part of finding what things are there, but then people pass it on to Codex to solve it and Codex is very bad, which is GPT, right? very bad to find but it is good to solve so in the end like with Opus 4.7 they managed to solve some things but in others like a regression and in general the despite if one goes to see those needle in the haystack tests what one finds is that even though they are models that have a larger context window that does not mean that they have the precision to find a particular data inside and if one is going to see the curves in general they are good at finding if the topics are at the beginning or at the end but if you put the topic in any Part of half of them find it difficult to understand the general idea of what is happening in that context but not in the particular details, so that is why other techniques are used to try to control that because sometimes you do need it to have a lot of context, so there, for example, what happens and where they work better, for example, agents today versus a chatbot with a broader context is that people start to have a work document next to them, so they look for something, find a small piece, and keep it next to them, so they always have it as a reference and that makes, for example, Cloud Cowork work better. with long contexts than the Cloud chat, then there you also have to change your strategy when dealing with these things so that we can actually get the most out of it, but in general today I would say that there are not so many models that handle as well as the context in an ultra precise way at least not available as completely good there we have tell us Cris what we are seeing in the end was to return to the issue of costs if you want a good model that for example for reasoning that does not have hallucinations in appointments and in long contexts and you have a maximum budget the month of for example 500 dollars you call 2000 times a month and you want a minimum quality you are going to see that Opus does not come out first Lama 4 Scout comes out not even the one with high billions the one with 17 billions for reasoning for appointments contexts where there is no hallucination this costs you one dollar per month versus Opus which is in 20th place and you get 234 is that Opus is expensive but you want the model that works better or do you want to hire the most expensive model depends on the task we have to go towards a world of orchestration where I am going to use different models for different things and for example that is what I am doing I am for example with I think I said it last month at this moment I am using the Minimax subscription Minimax is not in the top 10 on my list but as an orchestrator to converse on a daily basis and then load the corresponding model it is very good so instead of spending a lot of money to have a model that orchestrates I am with the subscription that is more expensive because it is faster than 40 dollars per month and within everything Minimax 2.7 Minimax 2.7 I like it also it works well the 2.5 is perfect I am happy with the 2.7 the 2.7 works very very very well one thing happens to me that of course when we who were occupying Open Clore remember that there is the Open Clore chapter on the channel the first one we launched we said Opus was wonderful because an attribute that has nothing to do with the resolution of problems that Opus has is the personality that is very like what you feel who understands the problem that this ability to occupy tools has so it was very useful and in fact many people on Twitter say today that since Antropic removed the Opus direct access the Open Clore subscription died in the sense that it no longer works in the same way and we have to fight more to achieve things and everything was not that Opus was better it is that Opus had a personality I understood better the problem it was like it had a different resolution and of all the models that I have tried to try to give a spirit to Open Clore at least it is in a percentage close to what Opus Minimax 2.7 was, it has worked spectacularly for me at the level of personality, perhaps at the level of use of tools, as it is a little shorter but as in that initial orchestration layer it works well and apart from that it has that entertaining personality despite the fact that some Chinese characters suddenly appear between it and GPT 5.5 with high thinking and with the latest Open Clore updates it is finally also working better but it took GPT to get to that point so they had to release a new update and optimize it a little more and do it with the harness the codex harness and not the GPT one then but Minimax was going to get to blackmail him and he already had something that was 80% of Opus or a little more even working relatively so I was very surprised by Minimax eye and we have models like the ones from Xiaomi that are in the top 10 or the top 5 for many things the problem is that there are no servers in the United States not the one from Xiaomi it happened like that suddenly it appeared no one talked much and suddenly one goes to see the rankings there is a classic ranking that shows that for example yesterday it came out 4.3 4.3 so they show typical is the graph like the performance and it shows how much it grew from one version to the previous one and you see like the Google models the GPT Cloud models etc and in between in the top 10 is Mimo Mimo was called, right? Yes Mimo from Xiaomi and it kind of slipped through and is in the top 10 as well as competing with the graphics and the subscription is 14 dollars a month and is almost unlimited just like Mimo or the same as Minimax's the problem is that the servers do not have any servers for the subscription in the United States and the closest is Europe and even so I have latency having my BPs in the United States so I assume that at some point they will have the servers in the United States or something a little closer so that we can use it in a better way but there are also problems with other things like Binance's API, I think it's very crazy for like at a company level like getting burned like making the decision to work with a single model thinking that I don't know we've talked about it several times in particular you said it with Claude that it was already very good that you got married but then a change happens that gets worse and either or a better one comes out but I, for example, in the image part I've thought about how I don't know I worked with Amante Day and Deogram then if not and then bye nothing works nothing but now I can't stop using Image 2 from GPT it's like very very very there it happens to me that the generation of images in general I think that in the middle of last year it was resolved with the release of Nano Banana and before that Image 1 from GPT when we all got the Studio Ghibli version it happens to me today and I have done a lot of tests on this and I have also seen some that Nano Banana from Google and GPT Image 2 which is the last one that came out last week are both at a super good level super Well, where there are advantages for one and the other, for example, Nano Banana still generates images faster than GPT and is very good with character consistency, is very good with understanding the prompt, etc., but GPT is for more complex images, right? As there is an exercise that is really entertaining, which is to give him your resume and tell him to assemble it with a character sheet from an RPG, if I didn't like mine, of course, the type of infographic there works perfectly as a type of infographic, the type that has to handle a lot of text and things there. GPT in that, which is complex, is also very good and there is a trick that I want to give to people who are at home because these two image models, unlike the previous ones, have different modes, that is, if you go to Gemini and ask it to create an image It is different if you have selected the Pro model than if you have selected the quick model and the same thing happens in GPT if it is GPT you put create image and at the top you have set GPT Instant it will work well but if you put GPT Thinking it will work better so if you want to create more complex images you are making a presentation or anything occupies the thinking models select the thinking model and put create image so that it occupies the model the flavor of creating image that will dedicate more resources to it and that will be able to work with something more complex then it also starts It's like you have to know where the button is that you have to press to achieve the result, which is something that really happens to me a lot in training, people say no, it's not that I prefer GPT chat, it kind of got stupid and what happened that they had it set to Instant and no one explains it to them, it's not natural, it's not obvious that you have to go up and select the other model and the same thing happens with Gemini and the same thing happens with Opus, that is, with Cloud, that is, there is a very big change in the quality of the response and in the reduction of hallucination and everything just by changing the model so on the subject of the images it also applies of course but important also to understand the contexts as if they always have the thinking model on for example of GPT it depends because I don't know the root it's boring I already want to take a better photo but I don't want to take so long if it's not that good it still works for me then it depends on the objective and there is a topic that I mentioned before about the tokens I remembered because on Monday I started my new version of the Cloud course they are already asking me to open another version because the cubes very fast it sucks like crazy Opus 4.7 and the issue is that it spends a lot of tokens but there were people who said but why am I going to ask if it is because it is going to consume all the tokens and I said go spend the tokens on something else like people especially the tokens like the subscription the limits of your subscription right you pay 20,100,200 dollars for a subscription and people say no what's happening is that I wasn't sucking my tokens ok but go ask for something else this is the task you want to do if it is like this spend the tokens that is the limits are reset every so many hours it has weekly limits it has limits on different things it uses the tokens and it is a resource that you have available I remembered people who when they play video games who start to earn like the in-game currency and they don't want to spend it and they start accumulating it and they have those rare items and they don't spend them and they don't use them and in the end it makes their game more complex just because deep down they are afraid of losing that resource cachai? So it's cool here too if you want to generate an image and it's good enough it occupies the model it doesn't matter but in this case the majority doesn't but the majority were afraid to spend them because they always got to the rolling window they ran out of tokens and they had to wait a week to continue working no the majority no at least the ones I catch who are in IT groups the good ones spent everything who are the people? Speaking yourself, you said the topic, the geeks, who are the ones involved in that computer group in your community that are consuming all the tokens like concho? people who are using AI to the fullest, right? people who know what they are doing and are creating code that giving you code is something that consumes a lot of tokens because they are creating applications and things but think about the office worker whose job is to make a powerpoint analyze an ex or it has to be something else my past in fact I am seeing if I get some people to use agents because they are not using it and what they say is my agent you yourself said that if they didn't use it you were going to remove them you said it in another episode of course but the thing is that they are using but they are using my agent their own agents are not using it so and that also happens because of course people in general I think they care about doing their job well not trying so much to reinvent what new things they can do of course I think it is but that is another problem what happens is that there I think there is a problem in the approach of the subject why? because of course we who are involved in this because we like it and we vibrate with the issue and we are reading all day and testing and benchmarking like crazy instead of working on the issue that we have, of course we are involved in the issue but the person on your team has to work and comply with you because I am not going to speak for Cristian who confessed that the last two weeks I said unproductive no that I was working I was not working no that I was working I'm not going to fall for that no but what I'm going for is that that person has to do his thing if that person feels that does well, you have no incentive to use a tool to do it any other way, why? Because learning to do it with the new tool with the agent who installs it with an AI, etc., means a learning cost that usually at the beginning has a greater investment in time, so the adoption strategy is not trivial, it is not just like here is the tool and use it, you have to give the enablement times and many times it happens that the strategy is defined top down that is, from the management of the leadership teams and says here is the tool but what we see of the work of the people who are there on a daily basis is the output of Their work is the result, it is the presentation, it is the commercial proposal, it is the analysis of the data, but we do not see how they do that, so many times we are optimizing for the result, but not for how they do the task, so here always when we enable and make adoption strategies in the company, we also have to think about how we enable from the bottom up because I have spoken to many companies, I cannot reveal names, but they know who they are who put together the megaprojects, they made a personalized assistant with all the knowledge of the company, and no one used it. Why? because it did not solve a real pain that the person had in their day to day life so we have to go see and enable and generate those spaces so that those who are working and doing the problem can recognize what part of their task of how they do things is lazy, it takes a lot of time for them to do and those are the things where we could look for projects where these people could actually take part in it, solve it faster and for them to realize that and have to learn that so it is not simple and on the other hand another thing that happens is that they have to take the time To do so, they also have to have the incentives, that is, if their boss tells them, and I have had a boss, what someone tells me when they go through a course is that my boss says that I would be a lazy person. What incentive do they have to really use it at work? If the boss doesn't make the valley, then it's not just how to define, there is a strategy from the top down, but I have to generate the spaces from the bottom up so that the projects that have an impact on how we do things emerge, but there also has to be a cultural change of acceptance that this is the new reality and We have to take it there then, as I didn't put agents on everything, yes it is a good idea as long as people really solve a real pain that the person has, then we have to do a survey and it has to be something that is not just you pushing but there also has to be other members of the group who are also pushing. He shares things because it also happens that even though I give them AI, they continue to work like heaven, so maybe the sales person figured out how to make a commercial proposal with AI and maybe what you already did changes two things and it helps the marketing person to put together the freebies for the agency, right? So if there is not the space to build and collaborate, what is starting to happen is that you are also spending more resources on solving problems that are very similar. So if you don't see it as a systemic problem, a problem that we have to, from above, take charge of understanding the problem well. First, we are going to spend a lot of resources on building things that in the end are not used, underused, or in truth they don't even really solve the problem. no, charges blacksmith knife stick that but well well a solution that we have found from scratch that I have to start applying for internal is that of course this is what Rodrigo says as from the bottom up to identify well what things are key to see work and in this to see how more than leaving it to them they are the ones to be built, you see how this construction is accelerated and they are taught well how to use it in their day to day and of course how it shows productivity and it is obvious that they are going to use it because it is something like they needed and asked for but Of course internally I build things but an issue that also happens in remote work to close that remote work has not happened this is the issue of silos that is more difficult to see how to work and I do not want to work in person I am not working at home either I see that I have had a field of half a hectare already but it's okay guys we are going to close I am going to close with a couple of my recommendations and I would love for you to give yours too for all those who are listening to us who have not yet jumped out of the pool to learn everything This may not be so expensive, you don't have to go with the most expensive models, we are going to leave you the link to the calculator here below the benchmarks. On the other hand, don't always depend on a technology. Remember that we are in fact, we are at the beginning. Many of the things we are dealing with are going to break. I said at the beginning but today is a holiday here in Chile, I think that in many parts of the world it is also the day of the worker where you have to work to put it forward, fight for your rights, yes, those types of things, but it is not an excuse that it is expensive to be able to leave. We are spending a lot of money among ourselves to learn, to get ahead, to do experiments. speed that these things are happening, don't assume that because you tried to do it once before and it didn't work, you're going to try again every two or three weeks, you're going to try the task again with your tool. I think that's vital because what could be done three weeks ago actually changed. things for example in the world of Gemini today you can finally create Gemini to give you presentations and things and give you the file to download which was something that in GPT and Cloud was already solved a while ago now Gemini also has it but new things are coming out so try it or if it occurs to you hey I could do this ask the guide try if it didn't work wait and try again in the next version right? And the other thing that I think is like a general recommendation is that for example with OpenCloud and with all these things they are still like solidifying things by adjusting things we have OpenCloud we have Hermes Agent that came out is in Cloud Cloud that now has Dispatch and is trying to simulate the same Codex that now can also do everything, almost everything they published so there is the option of testing at the border but assume that what happened to Cristian is going to happen to you that I updated and it broke, right? For example, I said to my OpenCloud, hey, I want to update and it says no, I can't update and I eat, but yes, I have always been able to tell you, update. It's not that they changed something now and I like, hey, but what's happening? And what happens? Of course, we read the release notes of a new edition of OpenCloud and review the documentation to understand what changed to adjust how it was doing before. No, we don't, right? We update and tell the idea to check that it broke and fix it, right? So obviously there is also our approach that is more breaking, I would say Tarzanesque, right? to jump in and see what happens along the way but because we are also doing it, it is a nice way of saying brute or idiot, yes, Tarzanesque, right? Well, I like it, we go ahead and try and break up, but because the risk for us is also lower, then if I break up my agent, what happens? I spend a couple of hours that I can spend, right? or more fix it or more clearly we can do that, right? Because of course we, who are in an independent crisis, have a different rhythm. I have to do the courses and the talks, but between classes I have space to fix things, right? Diego, on the other hand, is in the office and has to raise the issue and has to come here and has a team moving forward and the probability of breaking something that is working is higher, so that is why it is super important with this, I mean that if they are in a formal company, right? With processes and with things, yes or yes, they have to generate an experimentation space, a sandbox environment, have a team that what they are doing is like in a controlled environment, breaking things, right? Because in that case you can't always break it because you have to keep working, so we have to create those spaces to break because since this tool comes without an instruction manual, updates come out every day, there has to be someone in charge of investigating that and then finding out how, right? We do it, we promote it towards the rest of the team, so that's my invitation, try it, break it up under control and now thank you very much Yogi now Diego, if you remember Diego, last week he was going to get married and he got married, so congratulations Diego, thank you very much now, your recommendations, please, very good recommendation, don't get married, he says, can you imagine, don't podcast, a holiday that they are going to challenge you if you are married, I think the words add to me and I believe and I also highlight something that has been said that although many people, including myself, feel overwhelmed with all the new things and of course suddenly they try a tool and say wow it's great and I'm going to start doing more things and I see an audience and I want to do it. On the other hand there are also people who haven't even gotten involved in basic things about the models and it seems like you're speaking Chinese when you start talking about the topic of artificial intelligence, like they believe that artificial intelligence is something abstract that they don't know how it works but that it's important that they help their life and that they have to use it but they don't spend it comfortably or at the most busy is all this TPD and Yemeni doing basic consultations so what I'm going for is that the opportunity is the window is now and we don't really know how everything will evolve if later someone is going to have a job or not what are we going to do later so it's key we're going to do I have two I've thought about it a lot and I think we have to start thinking about having a self-sustainable plot with self-sustainable solar panels with like in the background like thinking that there's going to be a time when we go to leisure is going to increase, right? And we don't know that he's even talking a lot about universal income or things like that. I don't know what the remaining policy will be, but you have to find a way to get by on this alone, right? How do I generate that space where in the end it will still be safe, right? so we have to look for plots nearby that that that very well Diego you finished hey thank you very much everyone for listening to us very good long weekend for those who are going to listen to us today remember that the links are going to be in the description and anything that you think we should talk about in addition to talking stupid things during these episodes please ask us and we will always try to improve that they are super good and remember that there is still the idea of what to do in the face-to-face group so you have your ideas in the comments and subscribe right? subscribe now open close to let recording subscribe subscribe subscribe subscribe subscribe and subscribe