Approaching from the Bow

2026-10-23opinion12 min

AI Won't Replace You, It'll Expose You

AI doesn't invent new flaws in you — it turns the volume up on the ones you already had. Four ways to end up screwed by it, and the one thing that doesn't change.

Everyone keeps asking whether AI will take their job. Wrong question. AI doesn't invent new flaws in you, and most of the time it won't replace you either. It does something more uncomfortable: it takes the flaws you already had and turns the volume up until they're visible from the street. Here's why. AI has collapsed the cost of the easy middle of any job — the part where you just write it, just code it, just draft it by hand. When that middle costs roughly zero, it stops being where your value lives. Value gets pushed to the two ends: taste and judgment at the front, where you decide what's worth making and whether it's any good, and finishing at the back, where something half-built becomes something real. If you don't add anything at either end, there's no slice of the work left with your name on it. So here are several ways to be screwed by AI. Each one is really the same thing: a way of adding nothing at one of those two ends. If you're lazy, you're screwed Laziness used to be easy to hide. You just went slow, and slow looked a lot like busy. Nobody could see the difference between the person taking three days because the work was hard and the person taking three days because they couldn't be bothered. That cover is gone. The diligent person now moves at 10x, because the grunt work that used to slow everyone down equally is free for them. You, meanwhile, use the same tool to produce slop — a first draft you don't read, a wall of generated code you don't understand, an answer you don't check. And because you're lazy, you can't elevate any of it. You stop at the draft. That's the trap: AI doesn't reward doing less. It rewards the person who gets the five-minute draft and then spends three hours of judgment on top of it. The lazy person pockets the five minutes and skips the three hours, and it shows instantly, because everyone else's three hours are now visible next to your zero. If you're not original, you're screwed An AI model is, almost literally, the average of everything that has ever been written down. Ask it for the obvious thing and you get the obvious thing, competently, instantly, for free, forever. So ask yourself honestly what your contribution actually was. If it was being the competent average — the solid-but-unsurprising copy, the design that looks like every other design, the take anyone could have had — then your contribution is now a commodity with a price of zero. You were selling the one thing that just became infinite and free. The only thing the average can't produce is your deviation from it: your taste, your weird angle, the specific thing only you would have thought to do. Originality used to be a nice bonus on top of competence. Now it's the whole moat. If you don't have one, the machine isn't competing with you — it already is you, cheaper. If you're all talk, you're screwed There's a certain kind of person who always has an opinion about what should be built and never builds it. The strategist who never ships, the critic with no body of work, the one with the great idea they'll get to "when they have time." That person just lost their alibi. The excuses were always about the cost of doing: I don't have the time, the team, the budget, the technical skills. Fine — except now you have an infinite junior who works at 3am for free, never complains, and will build the thing with you tonight. The cost of doing has fallen through the floor. So if you're still only talking, the comfortable story ("I could, I just haven't") stops holding. The honest version is harder: you were never going to do it. Talk was the whole activity. AI didn't take anything from the theorist — it just removed the last thing they were hiding behind. If you never finish what you start, you're screwed This is the cruelest one, and the most current, because AI makes it so much worse before it makes it better. Starting is now trivial. You can spin up ten prototypes before lunch, each one further in an afternoon than a weekend used to get you. The dopamine of the blank page is cheaper than it has ever been. So the serial starter, the person with forty exciting beginnings and nothing shipped, falls straight into the honeytrap: more beginnings, faster, all of them thrilling, none of them done. But the bottleneck was never starting. It's the last 20% — the edge cases, the error states, the polish, the tedious work of turning a demo into something a stranger can actually use. And that's exactly the stretch where the tool helps you least, because it's all specific judgment and grinding follow-through, not generation. AI multiplies your beginnings and leaves your endings exactly where they were. The result is a graveyard of half-built repos and demos that never became products — just a much bigger graveyard than before. Two more, quickly To close the loop from the front end to the back: If you have no taste, you're screwed. The whole game is now telling good output from bad, and if you can't, you'll ship garbage with total confidence — the model will happily hand you something plausible and wrong, and you won't flinch. If you outsource the thinking too, you're screwed. Use it to skip the reasoning, not just the typing, and the muscle wastes. Eventually you can't judge what it gives you, which drops you right back into the taste problem above — now with no way out. The good news If the list stung, good — but don't mistake it for a death sentence. Look again at what's on it: lazy, derivative, all-talk, never-finishing, no taste, checked-out. Not one of those is a fixed trait. They're all habits. And habits are the most fixable thing there is. Better still, the same tool doing the exposing is also the fastest way out. Use it to sharpen your judgment instead of replacing it: let it draft so you can spend your energy editing, let it start ten things so you can practice the discipline of finishing one, let it argue the other side so your own thinking has to get sharper. The lazy person and the diligent person hold the exact same tool. So do the copycat and the original, the talker and the builder. The tool isn't the difference. You are. AI is a mirror, and it's unkind. But a mirror doesn't decide what it reflects. You still do. La IA no te va a sustituir, te va a desnudar Versión en español Todo el mundo pregunta si la IA le va a quitar el trabajo. Pregunta equivocada. La IA no te inventa defectos nuevos y, la mayoría de las veces, tampoco te sustituye. Hace algo más incómodo: coge los defectos que ya tenías y les sube el volumen hasta que se ven desde la calle. Y el motivo es este. La IA ha hundido el coste del centro fácil de cualquier trabajo — esa parte en la que simplemente lo escribes, lo picas, lo redactas a mano. Cuando ese centro cuesta casi cero, deja de ser donde está tu valor. El valor se va a los dos extremos: el criterio y el gusto al principio, donde decides qué merece la pena hacer y si está bien, y el acabado al final, donde algo a medias se convierte en algo real. Si no aportas nada en ninguno de los dos extremos, no queda ningún trozo del trabajo con tu nombre. Así que aquí van varias maneras de estar vendido por culpa de la IA. Cada una es, en el fondo, lo mismo: una forma de no aportar nada en uno de esos dos extremos. Si eres vago, estás vendido La vaguería siempre fue fácil de esconder. Ibas lento, y lo lento se parecía bastante a lo ocupado. Nadie distinguía al que tardaba tres días porque la tarea era difícil del que tardaba tres días porque le daba pereza. Esa tapadera se acabó. El diligente va ahora a 10x, porque el trabajo pesado que antes nos frenaba a todos por igual a él le sale gratis. Tú, en cambio, usas la misma herramienta para producir morralla — un borrador que no lees, un muro de código generado que no entiendes, una respuesta que no verificas. Y como eres vago, no sabes elevar nada de eso. Te quedas en el borrador. Ahí está la trampa: la IA no premia hacer menos. Premia al que consigue el borrador en cinco minutos y luego mete tres horas de criterio encima. El vago se queda con los cinco minutos y se salta las tres horas, y se nota al instante, porque las tres horas de los demás ahora se ven justo al lado de tu cero. Si no eres original, estás vendido Un modelo de IA es, casi literalmente, la media de todo lo que se ha escrito jamás. Le pides lo obvio y te da lo obvio, con solvencia, al instante, gratis, para siempre. Así que pregúntate con honestidad cuál era de verdad tu aportación. Si era ser la media competente — el texto correcto pero previsible, el diseño que se parece a todos los demás, la opinión que habría tenido cualquiera — entonces tu aportación es ya un producto genérico a precio cero. Vendías justo lo único que acaba de volverse infinito y gratis. Lo único que la media no puede producir es tu desviación de ella: tu gusto, tu ángulo raro, esa cosa concreta que solo a ti se te habría ocurrido hacer. La originalidad era un extra simpático encima de la competencia. Ahora es el foso entero. Si no lo tienes, la máquina no compite contigo — ya es tú, pero más barata. Si eres puro discurso, estás vendido Existe cierto tipo de persona que siempre tiene una opinión sobre lo que habría que construir y nunca lo construye. El estratega que no envía nada, el crítico sin obra, el de la gran idea que hará "cuando tenga tiempo". Esa persona acaba de quedarse sin coartada. Las excusas siempre fueron sobre el coste de hacer: no tengo tiempo, ni equipo, ni presupuesto, ni los conocimientos técnicos. Vale — salvo que ahora tienes un junior infinito que trabaja a las 3 de la madrugada gratis, no se queja, y construye la cosa contigo esta misma noche. El coste de hacer se ha desplomado. Así que si sigues solo hablando, la historia cómoda ("podría, es que no lo he hecho") deja de sostenerse. La versión honesta es más dura: nunca ibas a hacerlo. Hablar era toda la actividad. La IA no le quitó nada al teórico — solo le retiró lo último detrás de lo que se escondía. Si nunca acabas lo que empiezas, estás vendido Este es el más cruel, y el más actual, porque la IA lo empeora muchísimo antes de mejorarlo. Empezar es ahora trivial. Puedes levantar diez prototipos antes de comer, cada uno más avanzado en una tarde de lo que antes lograbas en un fin de semana. El subidón de la página en blanco está más barato que nunca. Así que el que empieza mil cosas, el de los cuarenta inicios emocionantes y nada enviado, cae de lleno en la trampa: más inicios, más rápidos, todos emocionantes, ninguno terminado. Pero el cuello de botella nunca fue empezar. Es el último 20% — los casos límite, los estados de error, el pulido, el trabajo tedioso de convertir una demo en algo que un desconocido pueda usar de verdad. Y es justo el tramo donde la herramienta menos te ayuda, porque es todo criterio específico y constancia a base de fuerza bruta, no generación. La IA multiplica tus inicios y deja tus finales exactamente donde estaban. El resultado es un cementerio de repos a medias y demos que nunca fueron producto — solo que un cementerio mucho más grande que antes. Dos más, rápido Para cerrar el círculo del principio al final: Si no tienes gusto, estás vendido. Todo el juego consiste ahora en distinguir una buena salida de una mala, y si no sabes, publicarás basura con una confianza absoluta — el modelo te entregará tan contento algo plausible y erróneo, y tú ni te inmutarás. Si también delegas el pensar, estás vendido. Úsala para saltarte el razonamiento, no solo el teclear, y el músculo se atrofia. Al final no puedes juzgar lo que te da, lo que te devuelve justo al problema del gusto de arriba — ahora sin salida. La buena noticia Si la lista ha escocido, bien — pero no la confundas con una condena. Mírala otra vez: vago, poco original, puro discurso, incapaz de terminar, sin gusto, desconectado. Ninguno de esos es un rasgo fijo. Son todos hábitos. Y los hábitos son lo más corregible que existe. Mejor aún: la misma herramienta que te desnuda es también la salida más rápida. Úsala para afilar tu criterio en vez de sustituirlo: deja que redacte para gastar tu energía editando, deja que empiece diez cosas para practicar la disciplina de terminar una, deja que defienda la postura contraria para que tu propio razonamiento tenga que ponerse más fino. El vago y el diligente sostienen exactamente la misma herramienta. También el copión y el original, el que habla y el que construye. La herramienta no es la diferencia. La diferencia eres tú. La IA es un espejo, y es despiadado. Pero un espejo no decide lo que refleja. Eso lo sigues decidiendo tú.

2026-10-06infra13 min

A Practical Logistics Ontology in Oracle AI Database

A full walkthrough — create a private RDF network, load shipments/trucks/sensors, build inferred graphs, then run two inference queries and a three-graph join to follow a shipment to its sensor.

2026-09-23infra9 min

Sessions Pile Up, AAS Stays Flat: Microservices, HikariCP, and Oracle Database

More replicas didn't cause the idle-session problem, and a bigger pool wouldn't have fixed it. The real driver was a config value nobody thought to question.

A client's DBA team flagged something alarming: idle sessions on an Oracle database, climbing steadily, enough to start worrying about PROCESSES and SESSIONS limits — while Average Active Sessions (AAS) stayed low and flat the entire time. That mismatch was the confusing part: more sessions held open, but no more actual work happening. The obvious suspect was a Kubernetes-based microservices layer in front of it — thousands of Spring Boot services, each with its own HikariCP pool, each scaled by an HPA reacting to CPU. More replicas, more pools, more idle connections — case closed, right? Not quite. The data said something more specific, and more useful. The naive hypothesis doesn't survive contact with the data The first instinct is usually "not enough capacity" or "too many replicas." Both are easy to check with what most teams already collect: samples of active connections, idle connections, threads waiting for a connection, and replica count, per service, every few minutes. Two things ruled that out fast: Replica count barely correlated with connection waits. Across the fleet, the correlation between "threads waiting for a connection" and "number of replicas" was close to zero. Services with more pods weren't the ones queuing for connections. Aggregate pool utilization stayed low even during wait events. Active connections divided by total configured capacity (pool size × replica count) sat in the low single digits, on average, whether or not a service was experiencing waits that same interval. If anything, utilization during wait windows was slightly lower than during quiet ones. This wasn't a capacity shortage. The database's own AAS agreed. It's not just an application-side read on utilization — Oracle's Average Active Sessions, measured independently on the database itself, told the same story: no meaningful increase in real work, even as raw session count climbed steadily. So: not too few replicas, not too small a pool, not more actual database work. Whatever was driving session growth, it wasn't the thing everyone assumed. The floor, not the ceiling The metric that actually correlated — strongly — was minimumIdle multiplied by replica count, against total idle connections held open. Not loosely: a clear, consistent relationship, the kind you don't get from noise. The mechanism is simple once you see it: minimumIdle isn't a target HikariCP aims for, it's a floor it maintains and often exceeds. Set it conservatively "just in case," multiply by however many pods the HPA happens to be running this hour, and you get a standing population of idle Oracle sessions that has nothing to do with how busy the service actually is. One service, configured with a minimumIdle low enough that nobody would have flagged it in a review, was still carrying an average idle-connection count in the hundreds — aggregated across its replicas — well above what its own traffic justified. That's also why AAS never flinched: it measures active sessions by definition, and a pile of idle ones is invisible to it almost by construction. The session count can climb for weeks and the metric everyone watches for "is the database under load" won't move at all — which is exactly why nobody caught this sooner. Why the replica count was the wrong signal to chase Here's the part that made this worth writing up: the HPA in question scales on CPU, which is a completely reasonable default — except CPU usage in a typical microservice isn't purely a function of database work. Business logic, serialization, framework overhead, whatever else the service does under load, all show up in the same CPU graph the HPA is watching. That means the slope of "replicas added" tracks the slope of "CPU used," not the slope of "actual database sessions needed." A service can scale out aggressively for reasons that have nothing to do with its database access pattern, and every one of those new pods drags its full minimumIdle floor along for the ride. The database ends up paying for a scaling decision that was never really about it. The actual lesson None of the individual pieces here are exotic — HikariCP's idle-floor behavior is documented, HPA-on-CPU is a completely standard default. The problem was that they were tuned independently, by people who each had a reasonable default in isolation, and the interaction between them only shows up once you're looking at the database's session count and asking "where is this coming from" with actual correlation data instead of intuition. If you're running connection-pooled services behind a CPU-based autoscaler, don't stop at "is our pool sized correctly." Ask instead: every time the autoscaler adds one more replica, what does that quietly add to a system like the database — one that scales on its own terms, not on your service's CPU graph? Las sesiones se acumulan, el AAS no sube: microservicios, HikariCP y Oracle Database El equipo de DBA de un cliente detectó algo preocupante: las sesiones en una base de datos Oracle no dejaban de crecer, hasta el punto de empezar a preocupar por los límites de PROCESSES y SESSIONS — mientras el Average Active Sessions (AAS) se mantenía bajo y plano todo el tiempo. Esa incoherencia era lo confuso: más sesiones abiertas, pero sin más trabajo real detrás. El sospechoso obvio era la capa de microservicios sobre Kubernetes que había delante — miles de servicios Spring Boot, cada uno con su propio pool HikariCP, cada uno escalado por un HPA que reacciona a CPU. Más réplicas, más pools, más conexiones idle — caso cerrado, ¿no? No exactamente. Los datos decían algo más concreto, y más útil. La hipótesis obvia no sobrevive a los datos El primer instinto suele ser "no hay capacidad suficiente" o "hay demasiadas réplicas". Ambas cosas son fáciles de comprobar con lo que la mayoría de equipos ya recopila: muestras de conexiones activas, conexiones idle, hilos esperando conexión y número de réplicas, por servicio, cada pocos minutos. Dos cosas descartaron esa hipótesis rápido: El número de réplicas apenas correlacionaba con las esperas de conexión. En todo el conjunto de servicios, la correlación entre "hilos esperando conexión" y "número de réplicas" era prácticamente cero. Los servicios con más pods no eran los que hacían cola para conseguir conexión. El uso agregado del pool se mantenía bajo incluso durante los eventos de espera. Conexiones activas dividido por capacidad total configurada (tamaño del pool × réplicas) se movía en un solo dígito bajo, de media, hubiera o no espera en ese mismo intervalo. Si acaso, el uso durante las ventanas con espera era ligeramente menor que durante las tranquilas. No era un problema de capacidad. El propio AAS de la base de datos decía lo mismo. No es solo una lectura de uso desde el lado de la aplicación — el Average Active Sessions de Oracle, medido de forma independiente en la propia base de datos, contaba la misma historia: sin aumento real de trabajo, aunque el número de sesiones seguía subiendo sin parar. Así que: ni pocas réplicas, ni un pool demasiado pequeño, ni más trabajo real contra la base de datos. Lo que fuera que estuviera disparando el crecimiento de sesiones, no era lo que todo el mundo asumía. El suelo, no el techo La métrica que sí correlacionaba — y fuerte — era minimumIdle multiplicado por el número de réplicas, contra el total de conexiones idle mantenidas abiertas. No de forma vaga: una relación clara y consistente, del tipo que no sale del ruido. El mecanismo es sencillo en cuanto lo ves: minimumIdle no es un objetivo al que aspira HikariCP, es un suelo que mantiene y a menudo supera. Lo configuras de forma conservadora "por si acaso", lo multiplicas por los pods que el HPA tenga corriendo esa hora, y obtienes una población fija de sesiones Oracle idle que no tiene nada que ver con cuánto trabajo real tiene el servicio. Un servicio, configurado con un minimumIdle lo bastante bajo como para que nadie lo hubiera marcado en una revisión, seguía manteniendo una media de conexiones idle en el orden de los cientos — agregadas entre sus réplicas — muy por encima de lo que justificaba su tráfico real. Por eso el AAS nunca se inmutaba: por definición mide sesiones activas , y un montón de sesiones idle es casi invisible para esa métrica. El número de sesiones puede subir durante semanas y la métrica que todo el mundo vigila para saber "¿está cargada la base de datos?" no se mueve nada — justo por eso nadie lo pilló antes. Por qué el número de réplicas era la pista equivocada Esta es la parte que hace que merezca la pena escribirlo: el HPA en cuestión escala por CPU, que es un valor por defecto totalmente razonable — salvo que el consumo de CPU de un microservicio típico no es solo función del trabajo contra base de datos. Lógica de negocio, serialización, overhead del framework, lo que sea que el servicio haga bajo carga, todo aparece en la misma gráfica de CPU que vigila el HPA. Eso significa que la pendiente de "réplicas añadidas" sigue la pendiente de "CPU usada", no la pendiente de "sesiones de base de datos realmente necesarias". Un servicio puede escalar agresivamente por motivos que no tienen nada que ver con su patrón de acceso a datos, y cada uno de esos pods nuevos arrastra consigo todo su suelo de minimumIdle . La base de datos acaba pagando una decisión de escalado que nunca fue realmente sobre ella. La lección real Ninguna de las piezas individuales aquí es exótica — el comportamiento de suelo de minimumIdle en HikariCP está documentado, un HPA que escala por CPU es un valor por defecto de lo más estándar. El problema fue que se ajustaron de forma independiente, por gente que cada una tenía un valor por defecto razonable en su propio contexto, y la interacción entre ambos solo se ve cuando miras el número de sesiones de la base de datos y preguntas "de dónde viene esto" con datos de correlación reales en vez de intuición. Si tienes servicios con pool de conexiones detrás de un autoscaler basado en CPU, no te quedes en "¿tenemos el pool bien dimensionado?". Pregúntate mejor: cada vez que el autoscaler añade una réplica más, ¿cuánto le añade eso, en silencio, a un sistema como la base de datos — que escala por su cuenta, no según la gráfica de CPU de tu servicio?

2026-09-10freedomtrack2 min

freedomTrack 2.0: A New Look, Clearer Controls, and Tracking That Follows You to Cruising Altitude

Not a rewrite — the kind of friction that only shows up once real people start using an app in ways you didn't fully plan for. Including, it turns out, at 900 km/h.

I've been building freedomTrack , a private GPS route-tracking app for iOS and Apple Watch — no email, no account, no personal data required to start recording. Today I'm shipping version 2.0, and while it's not a rewrite, it fixes the kind of friction that only shows up once real people start using an app in ways you didn't fully plan for. What freedomTrack still does, at its core Run, ride, hike, ski, or walk, and freedomTrack records your route with full stats — speed, elevation, heart rate from Apple Watch — and keeps everything on your device until you decide to share it. Follow a public route someone else recorded, share your position live with a link, or just keep a private log of everywhere you've been. What's new in 2.0 A new icon. A simplified pin-and-shadow mark, redesigned for iPhone and Apple Watch — including a circular border on the Watch icon instead of a rounded square, so it actually matches how watchOS crops and displays it. Every action now says what it does. This is the bulk of the release. Swipe actions on your routes list, the buttons for following a guided route on the map, and the track-details action bar used to be icon-only — clean, but not obvious to everyone. They all now show text labels. The button to start a new activity is a wider "New route" pill with a gentle pulse the first time you open the app, instead of a bare circle that gave no hint it was tappable. Fixed: the map not loading on your first run. On a fresh install, "load map on cellular data" was defaulting to off — so if you opened the app without Wi-Fi in hand, the map just looked broken. It now defaults to on, like it should have from day one. Flights now record properly. This one came from a real bug report — a track I flew from Madrid to Palma recorded fine through takeoff, then went completely silent for 46 minutes at cruising altitude, only picking back up on descent. Turned out freedomTrack's GPS noise filter, tuned for running and cycling, was quietly rejecting every fix once the plane passed 500 km/h or the in-cabin GPS accuracy degraded past 50 meters — which is most of a flight. There's now a dedicated Plane activity type with its own relaxed thresholds, so the rest of your activities keep the tight filtering that keeps their tracks clean, and your flights stop vanishing mid-cruise. The model stays simple freedomTrack is free to use — local GPS tracking, stats, voice notes, and photos always are. Cloud features like live sharing, uploads, and Group Rides run on credits you buy when you need them, or an Ultra subscription if you'd rather not think about it. No account required to start. Try it freedomTrack is available now on iOS and Apple Watch.

2026-09-07freedomcarb2 min

Introducing freedomCarb — AI Calorie Counting by Photo or Voice

A calorie tracker that starts from a photo instead of a food database search — point your camera at a plate and AI tells you what's on it in seconds.

freedomCarb is a calorie-tracking app for iPhone and Android that starts from a photo instead of a food database search. Point your camera at a plate, and AI tells you what's on it — calories, protein, carbs, fat — in seconds. It's live on the App Store, with Google Play coming soon. Why Every calorie tracker asks you to search a food database, scan a barcode, or guess a portion size by hand — three or four taps for something you're about to eat in thirty seconds. That friction is why most people quit tracking within a week. freedomCarb replaces all of it with one photo. What it does Point your camera at a plate, a snack, or a takeout box — AI identifies each food item and estimates its calories, protein, carbs, and fat, no manual search or barcode scanning. No photo handy? Say it or type it — dictate what you had, or type "150 g of milk," "a slice of pizza" — and get the same AI estimate. Edit any detected item: adjust the amount, add something the photo missed, or remove one that's wrong. Not happy with a result? Discard it and retake the photo in one tap. Daily history with a running total against your calorie target — day, week, month, and year — including a protein/carbs/fat breakdown alongside the calories. Real energy balance: how much you've actually burned (basal metabolism plus real activity from Apple Health / Health Connect) versus what you've eaten, for today or any past period. Intermittent fasting, tracked with a goal and a notification when you hit it — and a gentle warning, not a block, if you go to log food while a fast is still running. Weight logging with a simple progress chart. Recipes: build one from several ingredients — by photo, by voice, or typed in by hand — set how many servings it makes, and log a single serving with one tap. Favorites: save any meal once and re-log it instantly, no photo, no new analysis needed. A consecutive logging streak, shown right in the Diary. No account, no email, no password — an anonymous session has you tracking in seconds. Privacy and the model Your data is tied to a private, anonymous session ID, never a personal profile. freedomCarb is free to use, with a small number of free photo analyses to try it, then affordable credit packs — no subscription required. Try it freedomCarb is available now on iOS, with Android coming soon.