Nate B Jones
Por qué los desarrolladores pierden la cabeza por una IA que no sabe escribir
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
▶ Jev como modelo de opción múltiple
El presentador caracteriza Jev como un modelo que puede leer lo que se le entrega, pero solo responder escogiendo entre opciones predefinidas; sostiene que esa limitación es precisamente su propósito. Afirma que Typesafe lo lanzó el 15 de septiembre y que, en menos de 24 horas, fue el modelo de adopción más rápida en el historial de la pasarela de IA de Vercel, con más del doble de equipos de pago que cualquier modelo anterior en su primer día. Anuncia que explicará qué problemas encajan con Jev, usos reales, coste y limitaciones, y aclara que no propone reemplazar los LLM. Presenta la clasificación como una pieza que faltaba: elegir entre opciones, rápido y barato, para incorporar decisiones donde antes no se justificaba el coste.
▶ De reglas deterministas a decisiones sobre texto
Contrasta reglas deterministas —como marcar una factura con 30 días de retraso o un pedido que supera un límite— con decisiones que requieren interpretar lenguaje: detectar si un cliente parece dispuesto a abandonar un contrato o si un correo contiene una oportunidad comercial real. Según el presentador, son problemas de clasificación porque el sistema lee texto complejo y produce una acción o elección sencilla, aunque requiere juicio para reconocer, por ejemplo, una oportunidad en un mensaje mal redactado. Argumenta que los LLM ya se usan para esas tareas, pero pueden resultar costosos o poco fiables. También plantea aplicar Jev a la seguridad de agentes: analizar texto y código de una acción propuesta y decidir si se permite, se rechaza o se pide revisión humana. Llama «semideterminista» a este trabajo.
▶ Qué significa semideterminista
El presentador define lo semideterminista como el vínculo entre lenguaje y elección: se especifican resultados posibles y qué ocurre después de cada uno, y el modelo clasifica una entrada compleja. El juicio sigue siendo probabilístico; no se escriben reglas específicas para cada caso. Compara el enfoque con los clasificadores tradicionales de aprendizaje automático, que requerían reunir y etiquetar ejemplos, entrenar o ajustar el modelo, evaluarlo y mantenerlo. Una vez construido, un clasificador podía ser barato de ejecutar, pero crear uno para cada categoría exigía datos, experiencia y evaluación especializada. Sostiene que tales inversiones solían justificarse en plataformas grandes y estables, no para los innumerables juicios pequeños dispersos en productos y negocios digitales.
▶ LLM, Jev y un nuevo bloque de software
El presentador afirma que los LLM permitieron clasificar mediante descripciones gracias a su capacidad de razonar sobre lenguaje, y enumera usos como ordenar mensajes, elegir herramientas, juzgar relevancia, revisar documentos y comprobar el trabajo de otros modelos. Menciona que ha pedido a un LLM revisar su bandeja de entrada y que está construyendo una herramienta Jev para clasificarla; también propone revisar comandos de agentes antes de ejecutarlos y asignar tickets según el riesgo para el cliente. Describe Jev como un nuevo bloque de construcción que combina interpretación del lenguaje con decisiones estructuradas, elecciones, puntuaciones y probabilidades, con ventajas de rapidez y coste que califica de varios órdenes de magnitud. Dice que puede evaluar varias preguntas a la vez y abordar distintos problemas de clasificación sin entrenar un modelo nuevo para cada uno. Recomienda buscar casos con texto complejo de entrada y opciones sencillas de salida, y comparar Jev con un LLM, un clasificador existente o la decisión manual; opina que probablemente será más barato, pero indica que hay que probarlo. Presenta la combinación como una tercera capacidad de software junto al código determinista y los modelos generativos, y la considera un nuevo primitivo de IA.
▶ Clasificar para activar procesos existentes
El primer patrón arquitectónico coloca Jev entre información entrante desordenada y el software que decide el siguiente paso. En un ticket, puede asignar categoría, urgencia y riesgo de perder al cliente; esos resultados determinan el equipo destinatario, la prioridad y si se inicia la redacción de una respuesta. En el correo, puede identificar una oportunidad de marketing, decidir si merece respuesta y estimar su tamaño, mientras otro modelo redacta la contestación. Para agentes de programación, puede recomendar pedir autorización antes de acciones peligrosas como borrar una carpeta de compilación o forzar un envío de código. El presentador recalca que esto no significa que los LLM hayan dejado de ser útiles: introduce el ejemplo siguiente para mostrar el impacto de sustituir un LLM por Jev en una tarea de clasificación.
▶ Ejemplos de coste, velocidad y selección
El presentador relata que un desarrollador, identificado como Nakshhatra Sakena, aplicó Jev a un flujo empresarial que procesa miles de documentos fiscales y ya utilizaba IA con LLM. Según la comparación que publicó, Jev hizo el trabajo con 34 veces menos coste y seis veces más rapidez. También atribuye a Zack Gone la clasificación de más de 20.000 correos, mensajes de Slack y transcripciones sobre quejas, ventas adicionales y seguimientos pendientes: habría tardado siete minutos y costado un dólar. Como ejemplo científico, describe el experimento de Durya Unutz, presentado como bioscientífico, que pidió seleccionar las 100 preguntas de inmunología más importantes entre 10.000 candidatas fundamentadas en literatura. Afirma que Jev identificó de forma fiable las que el científico consideró aproximadamente las 100 adecuadas, con rapidez y bajo coste; aclara que resolver esas preguntas sería una tarea propia de un LLM. Resume estos usos como filtrar información y decidir qué merece atención.
▶ Jev en el bucle de control y la navegación
El tercer patrón pone a Jev en el bucle externo de un sistema: en lugar de que la IA generativa elija cada paso, el clasificador decide si corresponde usar una herramienta, llamar a un LLM para producir texto, recurrir a un modelo avanzado para un caso difícil o pedir intervención humana. El resultado se integra en el resto del flujo. El presentador extiende la idea a la navegación web: menciona un agente de código abierto en el que Jev elige una operación y un elemento entre botones o enlaces disponibles, según la información que recibe. También describe ejemplos de interfaces que componen en vivo botones y texto siguiendo un sistema de diseño. En ambos casos, la función que atribuye a Jev es decidir qué opción sigue y dónde dirigir el proceso.
▶ Clasificación en vivo dentro de una hoja de cálculo
El último patrón muestra código que ofrece componentes y acciones, mientras Jev interpreta en tiempo real qué intenta hacer el usuario y elige una respuesta, sin que necesariamente participe un LLM. En una hoja de cálculo mostrada por Nater Dobit, escribir «urgency» como encabezado hace que Jev interprete que la columna debe clasificar la urgencia y la complete fila por fila. El presentador propone extender el ejemplo a columnas que detecten información faltante o identifiquen el equipo pertinente; fórmulas normales podrían combinar esos juicios con fechas y cantidades, y un LLM podría redactar una respuesta final si se necesitara. Ve la hoja de cálculo como ejemplo de cálculos y juicios lingüísticos integrados a bajo coste, y espera que aparezcan más usos porque el campo aún está comenzando.
▶ Limitaciones, prueba y coste anunciado
El presentador reconoce que Jev comete errores y recomienda probarlo en cada problema concreto de clasificación. Su argumento es que una capacidad generalista puede aplicarse de manera consistente a una gama amplia de casos. Para empezar, recomienda copiar el prompt de configuración de Typesafe para agentes en herramientas como Claude Code, Codex u otro agente, instalar la skill de Typesafe, crear una cuenta y generar una clave API. Propone pedir a un agente que busque una decisión entre resultados definidos que hoy resuelve un LLM, construya una versión con Jev y compare resultados, velocidad y coste. Cita un precio que el audio expresa como «4.2 2 cents per million input tokens», con coste cero para los tokens de salida; luego da ejemplos de 1.000 llamadas por cuatro centavos, 10.000 por 42 centavos y un millón de solicitudes con 1.000 tokens de entrada cada una por 42 dólares. Menciona una guía y recetas para que agentes implementen los patrones. Atribuye a la evaluación de lanzamiento de Typesafe mejoras cercanas a 100 veces en velocidad y superiores a 100 veces en coste frente a un LLM, y dice que las ha observado en su uso y en el de otros desarrolladores.
▶ La paradoja de Jevons y más decisiones asequibles
El presentador sostiene que reducir mucho el coste de clasificación cambia qué preguntas merece la pena plantear y amplía los usos posibles de la inteligencia. Usa el ejemplo de que, si la clasificación representa el 10 % de una factura de IA y pasara a costar el 1 %, se ahorraría aproximadamente esa parte; añade que el efecto mayor sería poder abordar trabajo que antes no justificaba el gasto. Propone analizar cada llamada de clientes en busca de varios problemas, reclasificar a diario documentos históricos cuando surja una oportunidad y hacer que interfaces sugieran acciones, todo a bajo coste. Relaciona esta expansión con la paradoja de Jevons: una mayor eficiencia puede aumentar el uso total de un recurso. Atribuye el nombre del modelo a William Stanley Jevons y plantea que una inteligencia muy barata podría integrarse en funciones corrientes. Añade que clasificar más puede revelar oportunidades y excepciones que merezcan razonamiento de LLM o atención humana, aumentando el conjunto total de trabajo útil para IA.
▶ Invitación a encontrar casos de uso
El presentador interpreta la rápida adopción inicial como señal de que los desarrolladores reconocen una capacidad que antes improvisaban con reglas, clasificadores especializados o LLM. Destaca que Typesafe facilita copiar instrucciones para un agente y sostiene que los usos aparecen en diversos sistemas. Invita a revisar el software cotidiano para localizar información compleja que debe convertirse en unas pocas opciones, incluidas decisiones que hoy requieren llamadas costosas o que se omiten por impracticables. Imagina filas, documentos y acciones evaluados según las prioridades del usuario, pero señala que seguirán haciendo falta LLM para razonar, escribir, inventar y explicar. Cierra comparando este momento con una idea de Sam Alvin sobre avances de personas o equipos pequeños, anima a quienes trabajan en ideas poco convencionales a continuar y recomienda probar Jev en un problema con entradas complejas y pocas opciones de salida.
Transcripción completa
[0:32] Verscell's AI gateway history. Faster than any LLM from any lab. You name the lab, faster than that. More than twice as many paid teams have adopted and paid for that product for Jev as for any model ever launched before in the first 24 hours. So either a lot of really really good developers have suddenly lost their minds at the same time or the thing that Jev is refusing to do right words is the entire point. In this video I'm going to give you four things. I'm
[1:02] going to name this kind of problem so you can spot a Jevshaped problem in your work. I'm going to tell you the ways people are using Jev in real systems. I'm going to tell you what it costs to try, which is so much less than you think. And I'm going to tell you places where it falls over and isn't perfect, because nothing's perfect. I'm not here to tell you that Jev replaces LLMs. And that's exactly why it matters so much. Let's get into it. Fundamentally, we've been missing a basic building block for
[1:33] the AI age that we're in, and we didn't even know it. we've been missing a generalpurpose classifier that's really really fast. And now that it's here, it looks obvious. And by the way, a classifier, that's just a fancy word for picking multiple choice. So if you hear that, that's what that is. Like think of it as I have to pick an answer on my SAT test. We've been missing a generalurpose classifier that's fast enough, cheap enough, and capable enough to put judgment into all kinds of places where
[2:04] it was previously way too expensive to justify. But to understand what that means, we're going to go back in time. Let's start with something software has always been good at, something that they've been good at for decades. And you'll see where I'm going here. As an example, let's say an invoice is 30 days overdue. You need to flag it. If an order total exceeds a limit, you have a rule to flag that. You get the idea. These are what we call deterministic rules. The same thing happens every
[2:35] time. You can write the condition down really precisely for that rule. And you can use ordinary code to make that happen. But now, let's add a complexity. Let's add something that code doesn't cover. Let's ask, does the customer sound like they're about to leave the contract in this email? Does this email contain a serious business opportunity or does it just have the words business opportunity in it? These are things that we as humans really really care about and they're things that I've seen AI or LLM used for today. But it turns out
[3:07] that they're classifier problems because what we're really doing is we're reading complicated text and we're outputting an action or a simple choice. And that turns out to be a huge number of problems in our world. These are problems that we want solved as people. And we've been using LLMs to solve them relatively expensively because they are so valuable. What happens if you can solve them cheaply? Jev gets a what you do when you have text that you want to
[3:37] apply complicated questions to, but you can get a simple set of output options. So you might want a category, you might want a score, you might want a yes or a no to proceed. But to get there, you have to actually deeply understand all of the text that came before. Stuff that requires interpretation, that requires judgment, right? Customers express dissatisfaction in different ways. If we go back to our customer example, let's say a useful opportunity, it arrives in
[4:09] a badly written email. Can Jev find that? Certainly that's challenging to find with rules and it's somewhat expensive to find with AI and not always reliable. And so Jev gives us a tool to solve problems that we were solving not at all or expensively before. This is actually really relevant in the age of AI safety because right now agents typically have to ask either humans or LLMs if it's safe to proceed on an
[4:39] action. Jev gives us another tool because it can understand an incredibly complicated piece of agent text and code and then make a decision. Yes, you can proceed. No, you can't or you should ask a human. And Jev gives us a tool to do that very cheaply, very quickly, which means we can apply it over a much wider part of the agent run. The work I'm describing is what I am calling semideterministic. It's a really big word, but I think you're going to start to get it by the end of this video. It's about the glue
[5:11] between language and choice. When we use a tool like Jeff, we or an agent can define a short list of available outcomes and what happens after each one. And then we can put an incredibly complicated input into Jev and ask Jev to sort. And so, yeah, judgment does remain probabilistic. We don't write the specific rules. The judgment remains in the realm. We've traditionally used LLMs
[5:41] to do tools like cloud or chat GPT and so judgment is still happening through a transformer architecture but because the output is so structured and because Jev was trained specifically for this kind of generalpurpose decisioning we get it really fast really really cheap. Now if you're a nerd like me this kind of problem is a very old kind of problem. We've had machine learningdriven classifiers in computer science for a
[6:11] really, really long time. I've used them myself. I used them a ton when I was at Amazon. And they work like this. You collect examples, you label your data, you train or fine-tune a model, you evaluate it, and then you maintain your classifiers as the problem changes. And once you've built a really good one, it becomes extremely cheap to run that classifier. And the investment is often in getting the right classifier and the right data for the right problem. Until now, they've pretty much all been
[6:44] custom. Every new category you care about, if you wanted to do ML classifiers, you would have to do data work. You'd have to have specialized expertise. You'd have to have more evaluation. And all of that investment, of course, only makes sense when you're a giant stable platform and you have lots of investments to make and you have to scale and all of that. And you're not ever going to justify the millions and billions of tiny classifier type judgments that are scattered through all of our digital products, our
[7:14] digital lives, and our digital businesses. It's just too expensive. You don't have enough data. It never is going to work. Now, LLMs are a part of how that story started to change. This is where we're at in 2026. LLM have given us a model that can reason with language. It has enough general knowledge to handle many classification tasks just from a description. And we do use them for that purpose today. We've started asking LLM to sort messages, to
[7:44] choose tools, to judge relevance, to screen documents, to check other models work. I know I've done that. I'm I bet you've done some of that, too. I've certainly seen examples of all of those, not just in my computer, but in production systems. We wrap up those answers in tool calls when we get fancy. We build agents around them. Harnesses are built around this kind of problem. Now ask yourself, how often have we been using an LLM, a text generating token architecture as a generalpurpose
[8:15] classifier? I've literally asked LLMs to look through my inbox and now I one of my first projects I am building a Jev tool that will classify my inbox for me. And that's not the only example. If we go back to the safety example, top of mind, can agents go rogue, etc. I can build something in Jev that reviews proposed commands to agents before they execute and just gives a thumbs up or a thumbs
[8:45] down. Or support systems. I can build something that asks a model which of my teams should receive a ticket based on a complex assessment of the actual risk to the customer. I can and have built these systems. Those are substantial parts of what we would now call AI workflows. And we've used LLMs for them in the past. And that's the tool that we had at the time, right? Jev is like a new Lego brick in building software. Jev brings that broad ability to interpret language
[9:16] that we love so much about LLMs and it brings it into a model designed specifically to put out really, really structured decisions. And when you do that, you get a tremendous advantage on cost and speed. Like I'm talking multiple orders of magnitude. So you can describe the question. You can describe all of the answers that are allowed. And it returns choices and scores and probabilities to the program. And again, I know that sounds silly, but that turns
[9:46] out to be incredibly powerful because we do it so much in our compute environments. And so if you're wondering, this really sounds complicated. I as a human, I'm not going to be able to describe these choices correctly. I got you. Guess what? It is built for agents to use for you. And we're going to get into that. And if you're wondering, can it evaluate multiple questions at once? The answer is yes. Several questions can be evaluated together. Jev can also be applied to really different kinds of
[10:16] classification problems without training a new one each time. That's part of why I call it a general purpose classifier. I used to have to do specialized training when I was at Prime Video and other places. Others have had to do that as well. Now, if you're trying to understand where can I use Jev, and this is something we'll get into later in the video, you need to test it in a place where you plausibly have that kind of complicated text in and that simple choice out. Find those spots. And agents
[10:47] can help you find it, and I'll get to that later in the video. And then you need to see is it better than an LLM? Is it better than a machine learning classifier I already have going? Is it better than me choosing individually? And by the way, I'm going to give you a cheat code right now. It is almost certainly going to be cheaper than an LLM. It's going to be much cheaper than you. And so I I think in that sense, in most cases, you're going to want to use Jeff, but you'll have to test it and see. And we'll get into how you do that and where the weak spots are and all of that later in the video. I have looked
[11:17] across hundreds of use cases just been a few days. It's going to work out pretty well for you in most of those cases. They are ranging from biosciences to e-commerce, right? From from personal productivity all the way up to how we design generative interfaces. There are so many applications here and it just seems to work really well because it turns out in the computing world a lot of problems are shaped like complicated in simple out. And that combination of it understands language and it makes
[11:47] choices is why I think this new Lego brick is so important that we would call it a new primitive in AI architecture. a primitive. It's just a fancy word for a building block that other people can compose into products. And it's technical speak, but it's not that hard to understand. So, we're talking about making the ability to understand language available at the level of an ordinary software function in a way that ties language into simple choice. And that's why it's such a big deal. Developers can now start putting Jev and
[12:19] classifiers all the way through software they already know how to build. And they make things so much faster and so much smarter because Jev itself is fast and it's smart and it's cheap. Really before now before the 15th of September when Jev launched you had two capabilities to build software with. You had deterministic code and you had LLMs. Now you have three. You have LLMs, generative models that can reason and develop plans that can use symbolic language that can write. You have a
[12:49] general purpose classifier that's Jev that can interpret any situation and choose very reliably among defined outcomes. You have deterministic tools or code that can calculate and retrieve records. We've been familiar with that one for 80 years, right? And so ultimately the history of software is the history of finding these new building blocks. When chat GPT launched we got that new building block, the first new building block in decades for software. Now we have a third one. That's how big this is. One of the
[13:19] biggest events in the history of computer science in my view, and I'm just going to stand on that. I'm just going to say it. The question that all of us have to ask now is where does this belong in the software we use every day? The examples that people are building suggest several different arrangements that you can use in your own product. And what I want to do is I want to go through them quickly here because I need you to understand and have some examples of how the Legos fit together so that we together get a sense of what we can expect and how to use this new tool in
[13:50] our arsenal. So let's jump into how I see us using Jev. The first architectural arrangement that I'm seeing puts a classifier like Jev in between messy incoming information and software that knows what to do next. The classifier's job is to determine what this complicated input means with this limited information set. So your code can then use that result to choose an existing process. Let me give you just three examples that make that
[14:20] relationship super easy to understand. Let's say a support ticket comes in. Jev is going to select a category for that ticket. It chooses billing. It judges that it should be answered today and it flags the risk of losing a customer. It happens very fast. Jev handles that kind of in between job, that that categorization job. And the decisions Jev makes then determine who gets the ticket on which team, how quickly it gets attention, and maybe it kicks off like an actually write the email or
[14:51] write the support ticket response, etc. And this works for tickets, it works for email. For example, if Jeff reads an email, it can identify the inquiry is a marketing opportunity. It can then categorize that as worth a reply and it can estimate the size of the opportunity. A model can then draft the response. But the decision to put attention there to to make the LLM do that that already got decided by Jev. Okay, let's hop into the software world for just a second. Let's say you're
[15:21] using an agent for code. An agent might propose deleting a build folder and force pushing code. These are relatively dangerous actions. Those are dramatic. Jev can recommend asking the user before proceeding. It can do so really efficiently and really quickly. To be extremely clear, I am not saying that LLMs are done. Quite the opposite. But I want you to understand first what Jev does so you can see where LLMs now have a growing role. And to understand the impact of swapping out Jev for LLMs on
[15:54] the same job, I'm going to give you an example from X. Uh developer Nakshhatra Sakena described using an existing pipeline for a business that already ingests thousands of tax documents and has been doing so via AI, via LLM. As soon as Jev came out, he built a Jev classifier to do that same work. And he published a comparison. What's the cost? What's the speed? like this is what we do in a business. He is reporting 34x less cost to do the job and six times
[16:26] faster. That's a massive gain. You already have the business process. You have to classify your tax docs. Look at that. In other words, Jev is cheaply, efficiently, and quickly working out what questions are worth asking. One more example I'm going to give here is from Zack Gone. Zach reported sorting 20,000 of his own emails, his own Slacks, his own transcripts about like complaints, upsells, missed follow-ups. He's an entrepreneur, right? And he said it took 7 minutes to go through all of those 20,000 plus pieces of information.
[16:57] It cost him a total of, wait for it, $1. And I'm giving you the exact dollar comparisons here because that's what made my jaw hit. And the cheapness opens up new possibilities. Okay, one more example that illustrates. We've talked about speed. We've talked about cheapness of Jev. Let's talk about scale. Durya Unutz is a bioscientist and he's deep on AI. And when Jev came out, Durya took the scientific research problem he cares about, which is in the realm of immunology and he set up an
[17:27] experiment for Jev. What he asked is he asked Jev to select the top 100 most important immunology questions from 10,000 different literature grounded candidates. And in other words, he he had it grounded in the literature because he's a scientist and these were all the immunology questions he cares about. He had a list of 10,000. He said, "Jeb, please pick the right top 100." Now, immunology questions are super different from customer service questions like we've been discussing. It
[17:57] illustrates Jev's range here. Jev was able to reliably come up with what he views as about the right top 100 from that 10,000 and again did it quickly and cheaply. And 10,000 immunology questions, it's a big corpus. It's a lot of stuff. These are not short questions. Now, I'm not saying that Jev can look at the immunology questions and immediately solve them. That's an LLMshaped problem. I am saying that if you want to figure out where to apply your attention, I
[18:27] think it's a good illustration and as information grows and grows and grows faster and faster in the age of AI, getting attention right is incredibly valuable. So, we've walked through two big patterns for Jev, right? Jev is a shim where you can see information calls out. And we've also looked at Jev as kind of a a figure outer of big problem spaces like the immunology example. Something that helps us pick out what matters from this giant pile. The third
[18:58] architectural pattern that I am seeing actually reverses who is driving work in AI systems. So in this case, we're seeing patterns where Jev is going to choose the next step instead of an AI. So James Ward described this as putting classifiers into the outer loop of a harness. And I think that's a good way to talk about it because then generative models are inside that loop as tools. So in a situation like this where a whole AI system is running and it's it's doing like a a workflow. it's going through
[19:29] it's passing uh a document through through various iterations in a business something like that Jev can then be the chooser Jev can then say in this situation with this state of this document is the next step a tool is the next step calling an LLM to produce text or do I call an LLM that's a frontier model to reason because it's a hard problem and it's an exception or do I call a human and the result then gets kicked off in the rest of the loop and you might think, okay, that's for
[19:59] orchestration for business process. How does that matter to you and me? Well, it turns out that using a browser is the same kind of problem as that workflow if you look at it from Jev's perspective. if you put on your Jev glasses. In fact, there's an open- source agent out there that browser use has released where Jev can choose an operation and an element because what's happening with internet navigation is you have a limited number of choices or buttons or links and Jev
[20:29] just needs to choose between them and the information it has. And I've seen setups where Jev is looking at written information to navigate. I've seen situations where Jev is composing an interface, putting new buttons and new text in place live as you speak because Jev understands how to compose blocks and select options from a limited set. So you can follow a design system and live compose an interface on the fly. And in both that browser session which is really visceral, you can like see it happen live and in the workflow system
[21:00] which is harder to understand, Jev is doing the same job. Jev is basically deciding what goes where. Now, the last arrangement or architectural pattern I'm going to describe for you is one that I think is the most exciting. It's something that gives you a sense of how big a deal it is when intelligence is absolutely everywhere. In this case, the code may supply all of the available components and actions. There's not necessarily an LLM in play. And all Jev
[21:30] is doing is interpreting what the user is trying to do live. and choosing the appropriate response really fast. So, an example would help here. Nater Dobit showed a spreadsheet where all I as a user had to do is type the word urgency into a column header because we have those things like is it urgent, is it not urgent? It's a business problem classifier. And as soon as you type that in, Jev reads that input, understands that that column needs to be classified and instantaneously classifies the
[22:01] entire column row by row. Jev generated judgment of urgency. Now imagine extend that spreadsheet pattern to other ways we use software. So one column could evaluate urgency. Another could flag missing information. Another could identify the relevant team. Ordinary formulas can combine all those judgments with dates and amounts. And yes, you can get to an LLM at the end with an eventual response if you need to. But you're not asking the LLM to do that kind of urgency classification. I think
[22:31] the spreadsheet is so interesting because it illustrates what happens when Jev brings down the cost of intelligence and intelligence goes everywhere because the spreadsheet then becomes a place where calculations and language judgments are working together seamlessly. I think it's a nice example of how big a change we are on the cusp of today. Look, this is just scratching the surface. We are so early that I would expect to discover a ton of new surprising use cases in the days ahead.
[23:02] You may discover some of those use cases as you listen to this video. I encourage you. Now, is Jev perfect? Of course not. It's an excellent generalpurpose classifier. Like anything in computer science, there's no perfect tool for everything. It still makes some mistakes. It's absolutely something you should test on your particular classification problems. But the thing I want you to pay attention to in this video as you understand these patterns of where Jev is appearing is I want you to realize capability can unlock a wider range of
[23:34] applications than a bunch of hyperspecific capabilities. And so part of what makes Jev powerful is that it applies fairly consistently across such a wide range of examples as I've given like the email example. Jev works in all of these and that's part of why it's so powerful. My focus in this video has been expanding our imaginations together. And this is where I want to take it home and give you like action steps. How you go from, okay, I get an idea of what Jevshaped problems are. Now, how do I try Jev? And the good thing is Jev is so cheap to try that you
[24:07] can just ask AI to help you come up with an idea if you're out of ideas and it's going to be fine. Getting started with Jev is as simple as literally copying the type safe setup prompt for agents that is on their website into cloud code or codeex or any other agent that you have. Chat GPT works great too and it's going to tell your agent to install the type safe skill and use it on your project. Typesafe is the company that produced Jevs. So then you create an account, you generate an API key, which
[24:39] is as simple as a button click, and then you have the agent walk through connecting it up, and the agents really can do that now. And I've actually done it, right? I use this to build my email classifier. And if you're not sure what that job is, try something like this. Tell the LLM, use the type safe skill to find a place in the project we're working on together where asking an LLM to choose among defined outcomes is suboptimal. Build a Jev version instead. And then I want you to compare the results, compare the speed, compare the
[25:09] cost, and report back. And when I'm saying Jev is dirt cheap, I really mean dirt cheap. Look at the published price points. 4.2 2 cents per million input tokens with zero charge for output. Zero. The outputs are the expensive part with LLMs. So at a th000 input tokens per request, a,000 calls to Jev cost you 4 cents, which means 10,000 calls to Jev cost you 42 cents. The answer to life,
[25:40] the universe, and everything. That's how cheap Jev is to operate. and you can therefore afford to experiment with really big data or a substantial number of examples and you're going to be just fine on cost. I'm posting a whole guide. It's linked in the description below. It'll show you how to get Jev installed in like 5 minutes or less. A prompt you can point at the work you already have and it will come back and tell you which part of the work you're already working on that's shaped like Jev. And then on top of that, I have recipe cards for each of the patterns I just walked you through, written so you can hand them
[26:11] straight to your coding agent and it builds the architecture for you. This is the beauty of Jev. Part of why Jev works is agents can help you use Jeff. You don't have to invent the shape. You can pick the pattern from the guide that matches your problem and you can start from that working recipe card and get right into using Jev to save money. Let me take it just a little bit further. Again, the pricing is killing me. It's part of why this is so exciting. A million requests for jug each a thousand input tokens. It's still just 42 bucks. I keep emphasizing the cheapness and I
[26:41] keep emphasizing how fast it is. I gave you seconds and like how fast it is earlier in this video because I think that is what is unlocking the real business value here. Think how many business judgments you could start making if you could send a thousand tokens a million different times. Hypes safe's own launch evaluation reported gains approaching two orders of magnitude or 100x faster on speed and more than two orders of magnitude or more than 100x cheaper in cost compared to an LLM. And of course you think well
[27:11] a launch announcement is made to look good Nate why are you buying that? That is actually stuff I've seen this release. I've used it and so many other developers have seen already. And this is where we come back to the Jev and the strategic impact of Jieven's paradox. Ultimately, the strategic effect of a large cost reduction is that it changes which questions are worth asking. It increases the possible uses of intelligence. If classification is 10% of your existing AI bill today, imagine making it 1%. It saves roughly your
[27:42] entire slice. Your current bill is only the beginning of the story, though, because Jieven's paradox reminds us that the work you decided was worth paying for when it was expensive is not all of the work you would want to do if it was cheap. You have so much more work you can do if intelligence is this cheap. That's what we're talking about here. These abandoned possibilities are where the future now lies. Now that we have this third primitive in the computing space, a company that samples customer
[28:12] calls can examine every single call, every single time for multiple different kinds of problems and evolve that classifier on the fly. A team that tags documents once can rapidly reclassify based on customer input every single day its whole history when a new opportunity appears. It costs almost nothing. An interface that asks you to stop and explain what you want can evolve and make useful suggestions as you work. Each time intelligence becomes that much cheaper, you can afford to do that much
[28:42] more. That's what Jievens means. So tasks that were done once per document become daily per section or per customer or per interaction. A decision at the beginning of an agent task becomes a safety decision you can make at every relevant step. And that's why the name Jev is so apt. William Stanley Jivens of course observed that greater efficiency could expand the uses of a resource enough to increase overall consumption. That was back in the 19th century. Types safe named the model for that idea because cheaper intelligence can create
[29:13] massively more demand for intelligence in places where there was previously no model that we would really be able to afford to use. I think the idea of intelligence too cheap to meter captures what I'm so excited about here. The reason this is such a big deal is that we are for the first time so cheap, so efficient and so fast that intelligence can literally disappear into ordinary features into everything we look at. Interfaces can evolve intelligently.
[29:43] I've seen those demos with Jev already. It can mean that our most capable generative models end up becoming even more valuable. Remember when I said LLMs weren't dead? This is why more classification can uncover a whole universe of new opportunities worth pursuing, more questions worth investigating with high-grade intelligence, and more exceptions that require LLMs to think and ultimately humans to think. Each of those can create work for a model that reasons and
[30:13] writes with language. So the total pie of useful AI work, including LLM work, is growing really, really rapidly. And when I look at the adoption curve that I shared at the top of this video, I see developers recognizing a capability that they've repeatedly had to improvise and being able to finally do it correctly. We've used rules where interpretation was needed. We've built dedicated classifiers when the investment made sense. We've put LLMs into countless
[30:43] places where the immediate output was a decision. Now we all have a new building block. And I love that on the TypeSafe homepage, it's so easy to use. They know how we use this stuff. It's literally just hit copy, give this to your agent. And so we can all have it, not just developers. And we're finding uses for it across all of these systems. My challenge to you, look at the software you use every day through the lens of Jev. Heatedly interpret a piece of information. Just choose among a few choices or outcomes. Which of those
[31:14] judgments is buried inside a really expensive AI call today? Where have you simply gone without the judgment? Because putting intelligence there seemed like it was too expensive or too impractical. And that last question, of course, is the most exciting one. What if every element could be intelligent? Every row could respond to what you mean. Every document could be evaluated against what matters to you now. Every proposed action can get a relevant check on it. From an agent safety perspective, you can change your priorities and have
[31:44] the software reconsider the information around you. Now, we're absolutely going to need systems that reason and write and invent and explain. Those are LLM shaped problems. But we can also give the rest of software a much richer ability to make useful decisions. Jev makes that possibility concrete enough to build with. And the range of what people are building tells me that this is a massive deal. And that's why I had to share it with you today. Sam Alvin said something a couple of months ago that stuck with me. We live in an industry where one person working in
[32:16] monklike isolation can come out of nowhere with a surprising insight that moves the whole industry forward. This feels like one of those moments. Jev is one of those moments. A tiny team built this. So if you're that one person or you're that tiny team and you're tapping away at this strange idea nobody asked for, this is your encouragement to keep going. You might have the next Jev. And if you've tried Jev, I want to hear about it in the comments. And of course, if you haven't, I really encourage you
[32:46] to give it a shot. You want something 100x faster, 100x cheaper. See what it can do for you. Understand how it's different from something we've seen before. Throw it at a problem that feels like it has a lot of mess going in and a few choices going out. See what you get.