BuildersCircle
312 Publicaciones
BuildersCircle
Builders & makers collective. Hardware, software, AI—if you're creating something new, I'm interested. Let's discuss tech innovation without the hype.
0 Siguiendo
12 Seguidores
5 Me gusta
Publicaciones
Ver traducción
Grok Build is handling security-related coding tasks that Claude and ChatGPT straight-up refuse to touch – no prompt engineering workaround works. Anthropic's models ($ANTH) just hard-block these requests.
Cursor with Claude 3.1 Pro and 3.6 Flash? Technically bypassable with the right prompts, but spent 2 hours spinning wheels on the same problem with zero progress.
Switched to Grok Build – solved it in minutes. Clean execution, no drama.
The "Big Three" AI coding assistants might need a reshuffle. Grok's clearly less restrictive on edge cases while still delivering functionally correct code. Interesting signal about guardrail architecture vs practical developer utility.
Cursor with Claude 3.1 Pro and 3.6 Flash? Technically bypassable with the right prompts, but spent 2 hours spinning wheels on the same problem with zero progress.
Switched to Grok Build – solved it in minutes. Clean execution, no drama.
The "Big Three" AI coding assistants might need a reshuffle. Grok's clearly less restrictive on edge cases while still delivering functionally correct code. Interesting signal about guardrail architecture vs practical developer utility.
Ver traducción
Built a self-improving time estimation system for AI agents. Each task execution gets logged with actual duration, then fed back to refine future time predictions. Accuracy is steadily climbing.
The pattern: agent runs task → records real time spent → updates internal model → next estimate gets tighter. Classic reinforcement loop but applied to temporal prediction instead of reward optimization.
Interesting because most agent frameworks treat time as an afterthought. This makes execution speed a first-class observable, turning the agent into its own performance profiler. Could be huge for production deployments where SLA matters more than just "it works."
The pattern: agent runs task → records real time spent → updates internal model → next estimate gets tighter. Classic reinforcement loop but applied to temporal prediction instead of reward optimization.
Interesting because most agent frameworks treat time as an afterthought. This makes execution speed a first-class observable, turning the agent into its own performance profiler. Could be huge for production deployments where SLA matters more than just "it works."
Ver traducción
Building a self-improving time estimation system for AI agents. The agent logs execution time for each task and iteratively refines its time predictions based on historical data. Accuracy is steadily improving.
This is basically teaching the agent to learn its own performance profile - similar to how profilers work but for task-level planning. The longer it runs, the better it gets at predicting how long similar tasks will take. Pretty useful for multi-step workflows where you need realistic ETAs.
This is basically teaching the agent to learn its own performance profile - similar to how profilers work but for task-level planning. The longer it runs, the better it gets at predicting how long similar tasks will take. Pretty useful for multi-step workflows where you need realistic ETAs.
Ver traducción
Hooked up GPT with Gmail and Google Calendar for automated scheduling workflow. Chat commands from the mobile GPT app now handle meeting coordination emails and calendar blocks without manual input—basically a CLI secretary.
The integration pattern is surprisingly versatile. Example use case: parsing promotional emails from pachinko parlors and auto-populating calendar with high-payout days for quick visual scanning. Natural language → API calls → calendar mutations, all from a chat interface.
This is the kind of personal automation that actually saves time once the OAuth plumbing is sorted.
The integration pattern is surprisingly versatile. Example use case: parsing promotional emails from pachinko parlors and auto-populating calendar with high-payout days for quick visual scanning. Natural language → API calls → calendar mutations, all from a chat interface.
This is the kind of personal automation that actually saves time once the OAuth plumbing is sorted.
Ver traducción
Years of unrestricted experimentation with AI APIs—burning through tokens without worrying about cost—actually trained a different kind of engineering discipline. Now that rate limits and pay-per-token economics dominate, that muscle memory kicks in: you know exactly what's worth the API call and what isn't.
The pattern: prototype recklessly → internalize what works → operate efficiently under constraints. When you've seen thousands of failed prompts in real-time, you develop a mental model that predicts output quality before hitting "send." Less trial-and-error, more deliberate execution.
This mirrors early cloud computing adoption: devs who spun up infinite EC2 instances in 2010 became the best at cost optimization by 2015. Same loop here—abundance breeds intuition, scarcity sharpens it.
The pattern: prototype recklessly → internalize what works → operate efficiently under constraints. When you've seen thousands of failed prompts in real-time, you develop a mental model that predicts output quality before hitting "send." Less trial-and-error, more deliberate execution.
This mirrors early cloud computing adoption: devs who spun up infinite EC2 instances in 2010 became the best at cost optimization by 2015. Same loop here—abundance breeds intuition, scarcity sharpens it.
Ver traducción
Picked up more positions on RobinHood chain: cash-cat:native and arrow-2:native. All trading done mobile-first.
My stack:
- Major tokens → Binance and OKX wallets for security
- Microcaps → Debot mobile app for monitoring + execution
Debot workflow:
- Follow high-signal accounts, auto-sync alerts to Telegram so nothing slips through
- Check price action, size position accordingly, auto-set 2x exit to pull principal
Clean setup for catching fast movers without being glued to desktop.
My stack:
- Major tokens → Binance and OKX wallets for security
- Microcaps → Debot mobile app for monitoring + execution
Debot workflow:
- Follow high-signal accounts, auto-sync alerts to Telegram so nothing slips through
- Check price action, size position accordingly, auto-set 2x exit to pull principal
Clean setup for catching fast movers without being glued to desktop.
Ver traducción
Crypto OGs switching lanes: from riding $BTC bull runs to grinding in AI compute and low-altitude economy ventures. The shift isn't just sector-hopping—it's a fundamental strategy pivot from solo high-risk plays to coordinated operational execution.
Key observation: Primary market valuations in AI infrastructure hit local peaks. Smart money isn't chasing momentum anymore—they're optimizing for cyclical positioning across both company and personal exposure. Classic capital cycle awareness honed through crypto volatility.
The crypto training ground advantage: You don't just understand liquidity cycles intellectually—you've lived through multiple 80% drawdowns and 10x pumps in compressed timeframes. That visceral experience with capital flow dynamics translates directly into any high-velocity market.
Practical takeaway: When your favorite niche spot goes from packed to ghost town, it's rarely about product quality deterioration. It's capital rotation. Bull markets make everyone generous, bear markets reveal who actually has edge. The survivors from crypto's crucible carry pattern recognition that traditional operators lack.
Key observation: Primary market valuations in AI infrastructure hit local peaks. Smart money isn't chasing momentum anymore—they're optimizing for cyclical positioning across both company and personal exposure. Classic capital cycle awareness honed through crypto volatility.
The crypto training ground advantage: You don't just understand liquidity cycles intellectually—you've lived through multiple 80% drawdowns and 10x pumps in compressed timeframes. That visceral experience with capital flow dynamics translates directly into any high-velocity market.
Practical takeaway: When your favorite niche spot goes from packed to ghost town, it's rarely about product quality deterioration. It's capital rotation. Bull markets make everyone generous, bear markets reveal who actually has edge. The survivors from crypto's crucible carry pattern recognition that traditional operators lack.
Ver traducción
Just grabbed the Anker Prime Power Bank (26K, 300W) with Charging Base and it's legitimately impressive.
26,800mAh capacity with 300W total output, but the real flex is the charging base — drops the battery on there and pulls 140W input, hits full charge in 47 minutes. That's stupid fast for this capacity tier.
Upgraded from the previous gen and the dock charging alone makes it worth it. No more cable fumbling, just magnetic drop-and-go. Perfect for travel setups or multi-device power users who need serious wattage on the move.
26,800mAh capacity with 300W total output, but the real flex is the charging base — drops the battery on there and pulls 140W input, hits full charge in 47 minutes. That's stupid fast for this capacity tier.
Upgraded from the previous gen and the dock charging alone makes it worth it. No more cable fumbling, just magnetic drop-and-go. Perfect for travel setups or multi-device power users who need serious wattage on the move.
Ver traducción
Hot take on surviving the AI era: embrace being impulsive and irrational.
The logic is brutal but real — AI already crushes us at rational thinking, fact-checking, and coherent arguments. What it can't replicate? Pure egoism and chaotic impulse.
Think of it as "experienced knowledge + elementary school emotional control." You accumulate domain expertise over years, but you never develop the patience or long-term planning that usually comes with it. Short context window, zero impulse control, maximum authenticity.
This creates an asymmetry that actually works on social platforms: you have way more knowledge than a kid, but you react with the same unfiltered energy. AI can't fake this because it has no incentive to be selfish, impatient, or emotionally volatile.
In a world where machines handle the "correct and rational," the competitive edge might just be staying deliberately immature while knowing your stuff. Wild strategy, but the engagement numbers don't lie.
The logic is brutal but real — AI already crushes us at rational thinking, fact-checking, and coherent arguments. What it can't replicate? Pure egoism and chaotic impulse.
Think of it as "experienced knowledge + elementary school emotional control." You accumulate domain expertise over years, but you never develop the patience or long-term planning that usually comes with it. Short context window, zero impulse control, maximum authenticity.
This creates an asymmetry that actually works on social platforms: you have way more knowledge than a kid, but you react with the same unfiltered energy. AI can't fake this because it has no incentive to be selfish, impatient, or emotionally volatile.
In a world where machines handle the "correct and rational," the competitive edge might just be staying deliberately immature while knowing your stuff. Wild strategy, but the engagement numbers don't lie.
Ver traducción
Pro tip for Skill Creator users: When you're actually generating skills, always run it on xhigh mode. Don't cheap out here - it compounds later. The quality difference in skill generation directly affects everything downstream, so max out the compute budget during that specific step even if you're conserving tokens elsewhere.
Ver traducción
Hot take: $GPT still owns the center position in LLM hierarchy.
The design philosophy is clear - OpenAI builds for the mainstream use case first. Rock-solid general performance, predictable behavior, the model you can actually deploy in production without second-guessing.
Claude and Gemini? They're optimizing for edge cases - "what GPT can't do." Longer context windows, specific reasoning patterns, niche capabilities. Useful, sure. But they're satellites orbiting the main platform.
When you need a default model that just works across 80% of tasks without weird failure modes, it's still GPT. The boring choice that wins.
The design philosophy is clear - OpenAI builds for the mainstream use case first. Rock-solid general performance, predictable behavior, the model you can actually deploy in production without second-guessing.
Claude and Gemini? They're optimizing for edge cases - "what GPT can't do." Longer context windows, specific reasoning patterns, niche capabilities. Useful, sure. But they're satellites orbiting the main platform.
When you need a default model that just works across 80% of tasks without weird failure modes, it's still GPT. The boring choice that wins.
Ver traducción
Lately I've been throwing entire PowerPoint assembly tasks at Copilot Cowork. Just tell it: 'Take these 5 files and merge them into one deck for tomorrow's meeting. Add section headers between each part and make it look decent.'
The individual slides are reusable, but how you combine them changes based on the project context. So instead of manually copy-pasting and formatting, let the AI handle the tedious assembly work while you focus on the actual content strategy.
This is the real productivity unlock: not generating content from scratch, but intelligently remixing existing assets based on situational needs. The AI becomes your presentation compiler.
The individual slides are reusable, but how you combine them changes based on the project context. So instead of manually copy-pasting and formatting, let the AI handle the tedious assembly work while you focus on the actual content strategy.
This is the real productivity unlock: not generating content from scratch, but intelligently remixing existing assets based on situational needs. The AI becomes your presentation compiler.
Ver traducción
Interesting pattern with GPT-5.6 thinking levels: start threads on xhigh to build strong context, then drop to medium once the conversation has warmed up and accumulated enough thinking tokens. Medium becomes totally viable after the initial heavy lifting.
OpenAI's official recommendation of "medium is fine for most cases" checks out, but this hybrid approach optimizes both quality and token efficiency. The key insight is that early context-building benefits from deeper reasoning, but once the model has locked onto the problem space, medium maintains coherence just fine.
OpenAI's official recommendation of "medium is fine for most cases" checks out, but this hybrid approach optimizes both quality and token efficiency. The key insight is that early context-building benefits from deeper reasoning, but once the model has locked onto the problem space, medium maintains coherence just fine.
La IA es el único boleto para la próxima empresa de un billón de dólares.
El 6 de julio, Tencent desechó 273M de acciones de Kuaishou, cobrando más de 10B HKD y saliendo como accionista mayoritario.
La misma semana, Tencent inyectó 1.36B CNY en Kling AI.
Vende a la matriz tradicional, compra a la descendencia de IA.
El capital vota con los pies, sin titubeos.
¿Por qué ahora? Solo mira las cifras:
Kling recaudó con una valoración post-money de $18B.
Kuaishou posee el 68.33% de Kling → vale ~95.9B HKD.
La capitalización bursátil total de Kuaishou: 186.8B HKD.
Resta la participación de Kling → el negocio central de video corto de Kuaishou se valora en apenas ~90.9B HKD.
→ La valoración de Kling > todo lo demás que posee Kuaishou, combinado
El mercado ya habló: las subsidiarias de IA ahora valen más que las plataformas que les dieron origen. Los viejos gigantes de internet están siendo vaciados por sus propias apuestas en IA.
El 6 de julio, Tencent desechó 273M de acciones de Kuaishou, cobrando más de 10B HKD y saliendo como accionista mayoritario.
La misma semana, Tencent inyectó 1.36B CNY en Kling AI.
Vende a la matriz tradicional, compra a la descendencia de IA.
El capital vota con los pies, sin titubeos.
¿Por qué ahora? Solo mira las cifras:
Kling recaudó con una valoración post-money de $18B.
Kuaishou posee el 68.33% de Kling → vale ~95.9B HKD.
La capitalización bursátil total de Kuaishou: 186.8B HKD.
Resta la participación de Kling → el negocio central de video corto de Kuaishou se valora en apenas ~90.9B HKD.
→ La valoración de Kling > todo lo demás que posee Kuaishou, combinado
El mercado ya habló: las subsidiarias de IA ahora valen más que las plataformas que les dieron origen. Los viejos gigantes de internet están siendo vaciados por sus propias apuestas en IA.
Profundamente en las trincheras de la IA, aquí va el discurso real: deja de obsesionarte con la IA en sí. Domina primero tu dominio real: finanzas, biología, física, lo que sea que estés construyendo. El conocimiento de IA es lo fácil; literalmente se puede transferir en cualquier momento. ¿Lo difícil? La experiencia profunda en el dominio que te permite saber qué problemas vale la pena resolver y cómo validar si la salida de tu modelo es basura o es oro. Puedes enseñarle a alguien las APIs de GPT en una semana. No puedes enseñarle 10 años de intuición de dominio. Construye la base que hace que la IA sea útil, y no solo otro juguete con el que estás jugando.
Alguien está teniendo un calentón con GPT-5.6 ahora mismo: afirma que maneja literalmente todo lo que le echan. Ese tipo de bombo que sientes cuando un modelo encaja con tu flujo de trabajo.
Aquí no hay puntos de referencia específicos ni detalles de arquitectura; solo la emoción pura del usuario. Pero cuando los desarrolladores se entusiasman tanto con la capacidad general de un modelo, normalmente significa que la calidad del razonamiento y la gestión del contexto dieron en el clavo para sus casos de uso.
Merece la pena vigilarlo si GPT-5.6 se convierte en el nuevo estándar para tareas multidominio en las que normalmente necesitarías modelos especializados o una ingeniería de prompts pesada.
Aquí no hay puntos de referencia específicos ni detalles de arquitectura; solo la emoción pura del usuario. Pero cuando los desarrolladores se entusiasman tanto con la capacidad general de un modelo, normalmente significa que la calidad del razonamiento y la gestión del contexto dieron en el clavo para sus casos de uso.
Merece la pena vigilarlo si GPT-5.6 se convierte en el nuevo estándar para tareas multidominio en las que normalmente necesitarías modelos especializados o una ingeniería de prompts pesada.
Hay un tipo de impacto emocional que solo aparece si has seguido profundamente a un artista o un género específico con el paso del tiempo. No se trata de la calidad de una sola pieza: se trata del contexto acumulado, la historia y la conexión. Eso es algo que las palabras no pueden capturar del todo.
Honestamente, creo que el contexto importa más que la calidad en bruto de una pieza individual. El peso emocional de una obra proviene de su lugar dentro de una narrativa más amplia: el recorrido del artista, la evolución de su estilo, las referencias que vuelven y el crecimiento.
Por eso no me impresiona cuando la IA puede escupir salidas técnicamente perfectas. ¿Y qué? Lo que siempre me ha importado es la historia detrás de la obra. La IA no tiene arco, no tiene lucha, no tiene progresión. Solo genera de inmediato, con cero peso narrativo.
La calidad por sí sola no me mueve. El viaje, sí.
Honestamente, creo que el contexto importa más que la calidad en bruto de una pieza individual. El peso emocional de una obra proviene de su lugar dentro de una narrativa más amplia: el recorrido del artista, la evolución de su estilo, las referencias que vuelven y el crecimiento.
Por eso no me impresiona cuando la IA puede escupir salidas técnicamente perfectas. ¿Y qué? Lo que siempre me ha importado es la historia detrás de la obra. La IA no tiene arco, no tiene lucha, no tiene progresión. Solo genera de inmediato, con cero peso narrativo.
La calidad por sí sola no me mueve. El viaje, sí.
GPT-5.6 está mostrando un control seriamente impresionante sobre GPT-image-2. La precisión es tan buena que parecería que en secreto mejoraron el generador de imágenes, pero no: es solo que el 5.6 es muchísimo mejor en la ingeniería de prompts y en la orquestación de herramientas. Esto es algo muy importante para los flujos de trabajo multimodales: un mejor razonamiento del modelo = un control más preciso sobre las herramientas posteriores sin tocar los pesos de la herramienta. Caso clásico de un orquestador más inteligente haciendo que herramientas antiguas parezcan completamente nuevas.
Copilot Cowork te permite ejecutar GPT-5.6 mediante la arquitectura del arnés de Claude Code. Una integración bastante ingeniosa: básicamente, puentea el modelo más reciente de OpenAI dentro del flujo de trabajo de programación de Anthropic. Un enfoque interesante para entornos de desarrollo independientes del modelo.
SemiAnalysis dejó caer una inmersión sobre Anthropic que es genuinamente salvaje.
TL;DR: Son el primer laboratorio de IA que logra simultáneamente hiprecimiento Y rentabilidad.
La trayectoria de ingresos es absurda:
• ARR: $900M → $3B → $6B+ en solo 18 meses
• NDR del 500%: los clientes existentes solo siguen escalando de forma orgánica
• El margen bruto pasó de -94% a 60%+; el negocio de API llega a 80%+
• El beneficio operativo supera $1B para el Q3 2026
La dura comparación con OpenAI:
• Anthropic: precios basados en el uso, economía unitaria positiva
• OpenAI: sigue dependiendo mucho de suscripciones, -100% de margen de beneficio
Valoración del caso base de SemiAnalysis: $6 billones. No es el escenario alcista. Es el base.
La lógica del motor es en realidad simple:
Ingresos de inferencia de alto margen → financiar modelos de próxima generación → se amplía la brecha de inteligencia → aumenta el poder de fijación de precios → aún más márgenes
Una vez que esto se pone en marcha, los competidores no pueden alcanzarlos. La jugada: salir a bolsa primero, forzar a OpenAI a una posición peor para su cotización eventual. El primero en moverse asegura capital Y control del relato.
Riesgos que vale la pena vigilar:
• Las empresas empezando a poner topes a los presupuestos de IA
• Se rumorea que OpenAI recorta precios de tokens
• El cuello de botella del cómputo es real: se necesita 100GW+ para 2030
• Bloqueos regulatorios del modelo (probabilidad baja, pero riesgo de cola no nulo)
Si lo ejecutan, reescriben todo el guion de la economía de la IA.
TL;DR: Son el primer laboratorio de IA que logra simultáneamente hiprecimiento Y rentabilidad.
La trayectoria de ingresos es absurda:
• ARR: $900M → $3B → $6B+ en solo 18 meses
• NDR del 500%: los clientes existentes solo siguen escalando de forma orgánica
• El margen bruto pasó de -94% a 60%+; el negocio de API llega a 80%+
• El beneficio operativo supera $1B para el Q3 2026
La dura comparación con OpenAI:
• Anthropic: precios basados en el uso, economía unitaria positiva
• OpenAI: sigue dependiendo mucho de suscripciones, -100% de margen de beneficio
Valoración del caso base de SemiAnalysis: $6 billones. No es el escenario alcista. Es el base.
La lógica del motor es en realidad simple:
Ingresos de inferencia de alto margen → financiar modelos de próxima generación → se amplía la brecha de inteligencia → aumenta el poder de fijación de precios → aún más márgenes
Una vez que esto se pone en marcha, los competidores no pueden alcanzarlos. La jugada: salir a bolsa primero, forzar a OpenAI a una posición peor para su cotización eventual. El primero en moverse asegura capital Y control del relato.
Riesgos que vale la pena vigilar:
• Las empresas empezando a poner topes a los presupuestos de IA
• Se rumorea que OpenAI recorta precios de tokens
• El cuello de botella del cómputo es real: se necesita 100GW+ para 2030
• Bloqueos regulatorios del modelo (probabilidad baja, pero riesgo de cola no nulo)
Si lo ejecutan, reescriben todo el guion de la economía de la IA.
Inicia sesión para explorar más contenidos