Nate B Jones
Gemini 4 Argon es el número 1 en una clasificación. Esto explica por qué aún no puedes usarlo.
Pregúntale al tutor sobre este vídeo
Entra con un email verificado y pregúntale al tutor lo que quieras sobre este vídeo.
TL;DR
Resumen de investigación
▶ De los benchmarks a la obsesión por el cliente
El narrador afirma que los rankings y benchmarks no bastan para saber si un modelo es útil, especialmente si casi nadie puede usarlo; propone como medida el uso constante del producto y la corrección de sus fallos. Presenta a Anthropic como enfocada en empresas, a OpenAI con una actividad empresarial creciente y una gran apuesta de consumo, y a Meta como centrada en consumidores. Cuestiona cómo Argon, el nuevo modelo de Google que dice estar en lo alto de un benchmark, se conecta con los usos de Google, y pide ver con claridad su vínculo con los clientes. Expresa que quiere que Google tenga éxito y que haya más actores en IA, pues considera que más competencia puede hacer las cosas más seguras y mejores para los consumidores.
▶ Crear productos mientras la inteligencia evoluciona
El narrador plantea a quienes trabajan en empresas cómo ganar dinero con IA cuando los modelos subyacentes seguirán evolucionando. Contrasta esa situación con el software tradicional, que podía actualizarse cada mes o dos y seguir una hoja de ruta, y opina que las hojas de ruta quizá ya no funcionen. Describe el reto de crear productos que acompañen la evolución de la inteligencia y ayuden a los usuarios a entender sus efectos de manera positiva; dice que es un desafío difícil en el que se están invirtiendo literalmente un billón de dólares. Insiste en empezar por entender a los clientes y sus actividades. Considera que la noticia de los benchmarks no es lo central: los mejores productos surgen, en su opinión, de quienes los usan intensamente, detectan fallos y se ocupan de que esos fallos importen; lo presenta como una disciplina empresarial además de técnica.
▶ Uso de Dots y Muse para gestionar suscripciones
El narrador cuenta que, tras usar Muse para revisar sus suscripciones personales, quiso incluir también las empresariales, cuyos datos estaban repartidos entre cuentas y servicios. En un vuelo conectó ChatGPT Dots con fuentes como Mercury y otros servicios financieros, reunió información de suscripciones y correos electrónicos, y usó Dots para crear un modelo financiero de sus gastos personales y empresariales. Después pasó la hoja de cálculo a Muse para actuar sobre las suscripciones: dice que Muse era más rápido que Dots y podía navegar por la web para completar tareas como cancelaciones. Aunque alguien le sugirió hacer esos trámites manualmente, responde que le resulta útil delegarlos cuando tiene poco tiempo. Compara ambos productos desde su experiencia: Dots es más lento, destaca en conectores y en reunir mucha información de forma coherente; lo usa para trabajos complejos que pueden avanzar con el tiempo. Como ejemplo, Dots estaba procesando más de 20.000 elementos marcados con me gusta o guardados en su cuenta de X para reunir publicaciones sobre libros; el narrador dice que llevaba días trabajando y había procesado 741 libros. Concluye que probar y llevar la IA a tareas reales permite ver dónde funciona y dónde no.
▶ Productos de IA para necesidades concretas
El narrador recomienda que las empresas usen los productos que venden para conocer los casos concretos en los que funcionan o fallan. Opina que hay espacio para muchos ganadores y que, si un agente cumple una tarea, puede ser útil para algunas personas; menciona que hay 8.000 millones de personas y muchos usos posibles. Aboga por más actores, como Google y otros, que acerquen la IA a los clientes de forma comprensible, y anima a quienes conocen un sector a crear algo que ellos mismos quieran usar y que evolucione junto con la IA. Como ejemplo, describe un producto nuevo y simplificado que utiliza reconocimiento de imágenes para que los niños identifiquen árboles y plantas durante paseos por el bosque. No sabe si tendrá éxito, pero valora que esté pensado para padres y niños, mantenga a los niños conectados con el entorno y limite el tiempo de pantalla sin renunciar a los beneficios de la IA. Pide más productos con soluciones útiles para clientes específicos.
▶ Salir del diseño de chatbot y aclarar el papel de Argon
El narrador cree que hay más oportunidades de las que se perciben y que la industria está atrapada en el diseño de chatbot. Afirma que ChatGPT ha llegado a 1.200 millones de usuarios y que otros productos han tomado esa filosofía y añadido elementos como barras de herramientas o espacios para artefactos. Sostiene que apenas empieza la exploración de lo que puede significar la IA y que hace falta una visión de diseño más amplia y más comprensión de los clientes. Sobre Argon, pregunta en qué parte de los productos de Google aparecería, para qué clientes y si iría a Búsqueda o a Gemini; también pregunta por qué tendría sentido dentro de Gemini, dónde sería más útil que el modelo actual y qué capacidades lo distinguen. Usa el caso para pedir más atención y optimismo respecto a los clientes, y dice que el discurso pesimista puede reducir la esperanza sobre lo que la inteligencia podría hacer si estuviera ampliamente disponible.
▶ Una respuesta empática en un momento difícil
El narrador relata que una persona conocida usó Muse para pensar qué hacer tras un accidente en el que un juguete golpeó el ojo de su hijo. Aclara que el niño estaba bien y que el ojo también, y que la persona sabía que Muse no era médico, pero buscaba perspectiva en ese momento. El narrador dice que él también habla con IA cuando necesita ordenar sus pensamientos, no por formar un vínculo parasocial, sino porque le ayuda a pensar en voz alta y comprender mejor. Según la captura que recibió, Muse tuvo en cuenta cómo podía sentirse el niño y sugirió que probablemente no necesitaba una reprimenda sobre los juguetes en ese momento, sino un minuto para calmarse. El narrador describe esa respuesta como un momento de empatía y alivio emocional que evitó dar consejo médico, y lo presenta como ejemplo de atención al cliente.
▶ Los problemas de experiencia y los usos cotidianos
El narrador dice que las historias sobre experiencias de clientes le dan esperanza sobre el futuro de la IA y que prefiere hablar de usos diarios a centrarse en benchmarks, investigadores o avances matemáticos. Cuenta que compiló 26 usos distintos de agentes en los últimos 3 meses, relacionados con preparación de impuestos, planificación de viajes, horarios escolares de sus hijos e historial médico, entre otros, para descargar complejidad y centrarse en lo que importa. Reconoce que la IA aún no es perfecta, pero opina que los problemas de experiencia y los puntos de fricción ya no provienen principalmente de los modelos: cree que el cuello de botella está en el producto y en la atención al cliente. Invita a compartir ejemplos que conozca. Menciona que la startup Instinct lanzó un chat grupal con IA para ayudar a coordinar encuentros, películas y cenas con amigos; dice que habrá que ver si despega, pero lo considera una propuesta dirigida a un caso de uso particular.
▶ Fable 5.5 y el próximo año de producto
El narrador pide que la audiencia comparta ejemplos de productos de IA que estén funcionando para sus clientes. Luego menciona rumores de que Fable 5.5 está por salir y de que sería un modelo notable; advierte que las impresiones que circulan en Twitter no siempre son ciertas y que el modelo quizá lidere o no los rankings. Expresa confianza en que servirá a empresas porque, según él, Anthropic está enfocada en resolver sus problemas, y considera que Fable 5.5 será otra herramienta para aportarles valor. Dice que hará una reseña completa evaluando el valor para los clientes, no la posición en un leaderboard. Cierra afirmando que el próximo año de la IA será un año de producto y no de benchmarks.
Transcripción completa
[0:32] the different labs and the different models out there, we see a really clear map of lab to customer except for one instance. So, I'm going to go through. Anthropic, they focus on enterprise. They focus on how you can professionally use AI to succeed. Open AI, they have a growing enterprise arm, and they also have a very, very big consumer play. If you look over at Meta, Meta is deep on consumer, and Muses the latest AI release there, and their whole motion,
[1:02] their whole push from Zuck is all about AI for consumers. What about Google? Well, Google makes a ton of money from compute. Google obviously serves a lot of people with search, and Google has Gemini. What I don't understand is how Argon supports those use cases. Why do we need that kind of cutting-edge model? And that got me thinking about the importance of having tight customer-product relationships in the age of AI. And that customer obsession
[1:32] is something I would love to see more of from Google because I'm someone who thinks that the world is better off with more AI, not less. And especially in a world where we're talking about powerful AI, you want that power in the hands of more people, not less. So, if you're one of those people who's worried about AI, and you're like, "I don't know that I trust Anthropic. I don't know that I trust Open AI. I want other players." Well, you want another player in the game? Great. Let's have Google in the game, too. I would love there to be more players because I think that more players, generally speaking, make things
[2:03] safer and better for us consumers. So, I want Google to succeed. And I say this with love, but I don't yet see how Argon is connecting with customers. And I think that that leads me to some larger questions for all of us. If you work at a company, this is for you. If you work in enterprise and leadership, this is for you. You need to understand, in a world where intelligence is evolving every single week, how do you make money with AI? And
[2:34] especially, how do you make money with AI when the underlying model you're using is probably going to keep evolving? You know, when I was in product, for a long time, we would just release software. And the software that we released, we could update it, you know, once a month, once every 2 months, and we would have a road map, and that's great. I don't know that road maps work anymore. I think a lot of us in product would say they don't. And the larger problem is that you have to build a product with an evolving intelligence
[3:04] underneath that also needs to help users understand what is actually going to happen to their worlds as AI changes in a good way, not in a doomer way. That's a really tall order. It's a tough product challenge. It's a challenge that literally a trillion dollars is getting thrown at right now. But if you don't start with understanding who your customer is and what they're doing, you're not going to get very far. If you think about this Benchmark news, the Benchmark news
[3:35] doesn't matter. Let's peel that back. Let's look at what's underneath. I think the best AI products are being built by people who obsessively use them and notice what goes wrong and who make those failures matter. And this is a business discipline as much as it's an engineering discipline. And I want to share an example because I ran into a very practical use case when I was using AI myself. So, I shared a few weeks ago that I used Muse to review my personal subscriptions, and I talked about that in earlier video. But, I
[4:05] wasn't satisfied with stopping there because I also have business subscriptions, and the same problem existed across more than one part of my life. And the information was scattered across lots of different accounts and lots of different services. So, on a plane flight that I took recently, I went further. I used ChatGPT's Dots tool to connect to all of the different sources I needed. I could connect to a Mercury, I could connect to other financial uh sources, and I could have OpenAI's built connectors help me bring all of my subscription material together
[4:35] along with email to double-check it, and I could use Dots to process all of that information and to build me a financial model of subscriptions across both personal and business all in one place, which I never seen before. Afterward, I could take that spreadsheet and feed it to Muse and have Muse be like my hunter-killer drone. Like, Muse went through and helped me action against all of those subscriptions and make good decisions against it. Partly because Muse is faster than Dots right now. Partly because Muse is able to really
[5:07] use that speed to navigate the web and get things done directly. So, I can just delegate stuff like a cancellation. I actually had someone reach out to me when I shared that I was doing this and say, "Why would you use Muse to do that? You can literally click through and do it yourself." And I said, "Yeah, I I can, but it's awfully nice when I am strapped for time to be able to hand a job to an AI and have it just get done." And so, when we're talking about customers and we're talking about what matters, I use that story because I think it illustrates the importance of
[5:38] actually using the AI. And when I use Dots and I use Muse, I see where the AI sings and where it doesn't by trying it out in products like that. So, Dots is a little bit slower than Muse. Dots is very, very good with connectors. Dots can hold a lot of information and put it together coherently. And I find that I use both. I find that I lean into Dots where I need complex work done over time and I can just hand it off. Right now, Dots is working through literally more than 20,000 likes and bookmarks on my X
[6:11] account and chewing through them and finding all of the individual ones where I liked or bookmarked a post about a book because I want to have one place for a book list cuz I love to read where I can go back and pick a book off. And I've never been able to do that before because the job was too big and required too much analysis. And Dots is taking a matter of days to process through that quietly and is up to 741 books that I'm sure it will update in a few minutes, but it's up to 741 books that it has processed through. So,
[6:42] what I am sharing with you is the customer experience that comes when you use an AI and push it and discover how it works. And what businesses need to do more of is they need to use the products that they are selling so that they are confident in the specific use cases that work or don't work. I think there's room for a lot of winners in AI and I don't think we need to be complaining about good agent utility that works. If it
[7:13] works and the agent does the job, then for some number of us, it will be really helpful. There are 8 billion people on this planet. We have a lot of different use cases and there's a lot of room for agents to solve that. And when we look at the world that way, one, I think it leads me back to where I started, which is this idea that we need more players in AI, like Google, but beyond Google. Players who can bring AI into the customer's hands in a way that's relatable. I really am optimistic
[7:45] for builders right now and for people who are in companies with existing customer relationships. We are in a world where building for customers has never been more important because we are all facing the uncertainty around AI together. There's no place to run from it. Everyone has to be an AI expert now and everyone isn't prepared for that, but we're all expected to be that in various ways. And so this is our chance when we are looking at our individual spaces, our
[8:15] individual domains to actually sit down and say, "How can I build something that I would want to use in this space that evolves with AI as AI evolves and that is really, really useful to particular customers." And your customers could be anybody. Like I saw an example recently of a new product that's out that really focuses on using AI image recognition. And it's a very simplified product to allow kids to go on forest walks and
[8:45] they can actually recognize the trees and plants around them. I don't know how that will do, but I loved it because as a parent, I want kids to be able to find out about what's around them, keep them plugged in to the world that that they're living in, not have them have too much screen time, and still have the benefits of AI. And so that's an example of a product that got thought through for parents, that got thought through for kids. We need more products like that. We need more people who are thinking about where AI can go and how AI can deliver useful
[9:15] solutions. And I don't know that all of them will work, but I do know that a lot more room is out there than I think we realize. I think we are trapped in the tyranny of the chatbot. We've seen chat GPT get to what, 1.2 billion users right now? Every other product since then has in some sense borrowed from that original design philosophy. It has taken the idea that is chat GPT and said, "Well, now we're going to add a toolbar. Now we're going to add a place for artifacts."
[9:45] We are just at the beginning of figuring out what AI means, and we need wider design sense. We need more sympathy and understanding for customers. And so when I think about Argon, and I know that it had great benchmarks, the first thing I think about is where in Google's surface does it appear and for what customers does it appear? And that's why I'm asking Google, please release it. Please share it. I want it to succeed. Does it go in search? Does it go in Gemini? Why does
[10:16] it go in Gemini? Where does it succeed in Gemini? Why is it more useful than the current Gemini model? What is it about Argon that makes it sing? And that's just one example. I think we need more customer obsession, and we need more customer optimism in AI. I think so often right now the doomer discourse that's coming is keeping us from being hopeful for our customers. It's keeping us from being hopeful for what intelligence can do if intelligence is everywhere. There was someone that I was talking to
[10:46] over the weekend who used Muse to help figure out what is the right approach with a kid who had an accident with a toy in their eye. The kid is fine, the eye is fine, don't worry about it. But they wanted to talk to someone who is there who could give them perspective, and they know that Muse is not a doctor, but they needed some perspective in the moment. I found myself talking to AI when I need perspective in the moment, too. Not because I'm forming some kind of weird parasocial relationship with the AI, but
[11:18] because having it there helps me to think out loud and think better. In other words, I can think of so many specific use cases where AI effectively extends my ability to remember, extends my ability to think, extends my ability to understand what is happening. And I see that happening in individual moments like the doctor moment with my friends. And it happens more when we obsess over the customer. One of the things that my friend was sending me is she sent a
[11:48] little screenshot of the Muse text, which she noticed that Muse thought about how her son was feeling. Hey, your son probably had a hard time. They probably don't need a lecture from you about toys right now. They're going to be okay, but you just need to give them a minute. It's a model that actually had a moment of delight, a moment in a difficult emotional moment where the parent is worried, the kid is worried, and this AI finds a way to walk the line, to not give medical advice, but also to have some empathy.
[12:18] That's customer obsession. That's something that we need more of. When I think about the future of AI, and I think about why I am hopeful for builders in this moment, those are the kinds of stories I think about. We have enough stories about benchmarks. We have enough stories about researchers that are breaking the edge of the frontier. We have enough stories about the Navier-Stokes equation and all of the math. Do you notice I don't cover a ton of the math breakthroughs on this channel? That's on purpose. I don't think for a lot of us it makes sense.
[12:49] I think it makes more sense to talk about how we are actually able to use AI for daily stuff. I actually compiled a list today of all of the different things I've used agents for. It came to 26 separate use cases in just the last 3 months. Stuff around tax preparation, stuff around travel planning, stuff around figuring out kids' school schedules, stuff around understanding medical history, all of this different stuff that I had just found because I
[13:20] was digging in and trying to sort of offload more and more and more of the complexity so I could focus on what matters. I know that it's not perfect yet. And partly in 2024 and 2025, the imperfection came from the models. More and more and more as I think about it, the imperfection in our AI experiences, the pain points, it's not coming from the models. It's coming from the product experience. This video is being made because I think the models aren't the bottleneck anymore. I think the benchmarks are not the bottleneck anymore. The product experience, the
[13:52] customer obsession is the bottleneck right now. And I don't see as many examples as I would like to of truly customer obsessed AI experiences, and I would like to see more. By the way, if you know of some that I have not mentioned in this video, and I know there's a lot of startups out there. There are a lot of people who are looking for customer obsessed experiences. There's people who are innovating. Instinct, another AI startup, just launched group chat with AI, so you can throw an Instinct into your group chat with your friends, and Instinct can help coordinate the next,
[14:24] you know, time you get together, the next time you see a movie, the next time you have dinner, where you have dinner, all of that stuff. We'll see if that takes on, but at least it's a pitch for a particular kind of customer used case. We need more of those. I would like to hear from you where you see those examples working because I think it's really, really important. And as I talk about benchmarks, there are rumors that Fable 5.5 is about to drop. Rumors that say it's an incredible model. I've certainly seen amazing things on Twitter. We all see amazing
[14:56] things on Twitter. Not all of them are true. Fable 5.5 may or may not lead the leaderboards, but what it will do is it will work for enterprises. How do I know that? How do I have a lot of confidence in that? Because that is what Anthropic is obsessed with. They are obsessed with solving problems for enterprises. And Fable 5.5 will be another weapon in their arsenal to deliver extraordinary value to enterprises because that's what the entire company is obsessed with right now. And yes, they're obsessed with delivering a better model, but they have that business model context and that customer obsession. I will, of
[15:27] course, do a full Fable 5.5 review, but I won't do it because it topped a leaderboard. I'll do it because of the value that it's bringing to customers. That's really what I care about. And so, if you see this and you're like, "Where's Fable 5.5?" that's where it is. That it's coming. I'm going to have the review, and that's why I'm going to cover it, not because of what it what it scores on the leaderboards. The next year in AI is going to be a year of product, not a year of benchmarks.