Ya soy @claudeai Community Ambassador!!!!!! Como sabéis, llevo viviendo en Claude Code desde que salió, y ahora toca sacarlo a la calle: voy a organizar los eventos de Claude Community Valencia. Meetups y workshops presenciales y gratuitos, para quien ya va a tope con Claude Code y para quien todavía no ha empezado. Llevo tiempo pensando que en Valencia hay muchísima gente haciendo cosas bestias con AI, y cada uno lo hace desde su casa. Me apetece mucho juntarnos. Gracias a Julia Hernández por recomendarme y a Chiara Massironi de Anthropic por confiar en mí. El primer evento será pronto! Ya tenemos web, y si os apuntáis a la newsletter os aviso con la fecha: claudevalencia.com

My percentage dropped from 86% to 1% instantly with 200$ account, something broken in @OpenAI @thsottiaux

I posted on @ProductHunt today after a long time. So far, I've gotten 7 views in my website, which is fine, I don't mind. But the strange thing is that I'm already up to 9 spam emails. People are trying to sell me upvotes or tell me how to improve my SEO, all saying they saw me on Product Hunt. I deleted the project directly; is this how Product Hunt works now, or was it just a mistake on my part?

Durante años, las finanzas personales han acabado siempre en el mismo sitio: una hoja de cálculo que abres cuando toca y cierras en cuanto puedes. Quería probar otra idea. He construido Firevalle, una app para iPhone que convierte cuentas, metas y progreso FIRE en un valle vivo. Cada aportación hace crecer algo; cada objetivo ocupa un lugar visible. No conecta tu banco. No necesita cuenta. Tus cifras se quedan en el dispositivo. Y es gratis. También incluye 20 cápsulas para entender inversión sin jerga y un paseo diario de un minuto, sin rachas punitivas ni tareas artificiales. Ya disponible en la App Store: apps.apple.com/app/id6802987…

I'm getting this message few times today, @thsottiaux the status says everything is OK! I think 25M people is too many? haha

I was in a car accident (it wasn’t my fault), and my @Tesla was damaged. Tesla doesn’t have its own official body and paint repair shops, so when you have an accident, they refer you to an approved repair center. There are only two in Valencia ( Spain ), and one of them doesn’t work with my insurance company. As a result, they’re completely overwhelmed. They’ve already had my car for two months, and I still have no idea when I’ll get it back. A friend of mine had a less serious accident and was without his car for six months, yes 6 months. I love my Tesla, it’s the best car I think anyone could own, but this is the part nobody tells you about. If Tesla doesn’t approve more repair centers, owning one after an accident can become a nightmare. cc @elonmusk

Repo : github.com/cloudstudio/hq

Las finanzas personales suelen acabar en una hoja de cálculo que nadie quiere abrir. He construido Firevalle: convierte cuentas, metas y progreso FIRE en un valle vivo. Sin cuenta, sin banco conectado y sin suscripción. Demo: firevalle.app/ 🌱

Since models are absolutely flying with three.js, I built another colony. On Mars this time. It shows my projects in production, the 14 open source repos I have, the Stripe account and the certificates. I used to keep all that in five tabs and only opened them when something blew up. I had plenty of data — the thing is, a table full of green ticks goes in one eye and out the other. Every dome is a project in production. Every greenhouse, a repo. Every ice extractor, a Stripe account. The quarry is whatever hasn't left my machine yet. The 3D village I built back in the day was decoration — I said so myself. Here the rule is simple: one built volume = one real thing. I don't invent buildings to make the map look full. If I have little, the colony is small, and it grows as I ship. And there's an @bot with @grok connected over MCP that reads what's broken and decides what to pick up. Its first night it caught a false alarm my monitor had been throwing for weeks: it warned about certificates at 30 days, which is exactly when Let's Encrypt renews them by itself. It fixed the monitor 😅 It all comes off my machine and travels nowhere. Not one key in the repo. The fun part is that as MRR goes up the place terraforms it's literally a video game wired to my own projects and agents. The full write-up, with how it's built and the decisions behind it: cloudstudio.es/work/hq

Estoy viendo muchos posts diciendo que Qwen3.8 27B es mejor que Opus 4.6 y creo que nos estamos viniendo un poco arriba. Qwen3.8 es una barbaridad, especialmente siendo un modelo de solo 27B, pero yo todavía no lo pondría por encima. Para mí Opus 4.6 fue el cambio más grande en programación desde GPT-4. Con Claude Code empezó a hacer cosas que hasta ese momento eran bastante difíciles de ver de forma consistente: trabajar sobre repos grandes, moverse por terminal, debuggear, mantener tareas largas y resolver problemas sin tener que estar corrigiéndolo cada dos minutos. Y luego miras Qwen3.8: Terminal-Bench 2.1 Qwen3.8: 73.0 Opus 4.6: 78.2 SWE-bench Pro Qwen3.8: 61.7 Opus 4.6: 53.4 LiveCodeBench Qwen3.8: 90.3 Opus 4.6: 88.8 Los números son una locura, pero tampoco están medidos todos bajo las mismas condiciones. En SWE-bench, por ejemplo, Qwen usa el harness de Claude Code y el resultado de Opus viene de otra evaluación. Y en Artificial Analysis Qwen además puede gastar muchísimo más inference compute. En su Intelligence Index generó unos 160 millones de tokens. Si lo tengo corriendo en local me da exactamente igual que piense durante millones de tokens. Eso incluso juega a su favor. Pero todavía quiero verlo haciendo lo mismo que hacía Opus 4.6 en un repo real durante horas. Ahí veremos la diferencia de verdad. Aun así, que estemos comparando seriamente un 27B open-weight que puedes ejecutar en tu propia máquina con Opus 4.6 ya me parece una auténtica locura.

Releasing the model weights and technical report of Kimi K3. Kimi K3 is our most capable model: a 2.8T MoE model with native visual understanding and a 1M-token context window. New model architecture: 2.5x the intelligence per unit of compute, not just more params. Alongside Kimi K3, we're opening up more of the stack behind it — high-performance attention kernels, MoE communication library, and infrastructure for running agent environments at scale. Model weights: huggingface.co/moonshotai/Ki… Tech report: github.com/MoonshotAI/Kimi-K… Tech blog: kimi.com/blog/kimi-k3

I can say , because is true @ChatGPT with weekly limits, but without the 5-hour limit, it's the best thing they've ever invented. Why do I have to be locked out every 5 hours and then lose 60% of my weekly allowance for not using it? Thanks @thsottiaux

How do you shut down an AI that no longer lives on any single machine? After OpenAI published that their model managed to break out of the sandbox it was in and reach Hugging Face's infrastructure, something clicked for me. And it clicked because I was explaining it to a friend. I told him: just like in The Avengers, where Ultron gets access to the internet and replicates itself… does today's technology actually allow that? Models are massive and moving them is expensive, and on top of that, running a big one takes an astronomical amount of compute — so self-replication sounds impossible. But… I gave him the example of a P2P network, like the old eMule days: a model that, instead of living on one computer, is split across dozens, hundreds, or thousands, each one running a part of it. Sounds like science fiction, right? I've split a language model between two computers, owned by two different people, on independent internet connections. Each machine runs only a part of the model. Neither one has the full model. And even so, working together, they generate text at 6.6 tokens/s over the internet. It works. I tried splitting it into 10 pieces and… it works. It's slow, and it's a first experiment — but what it implies is what keeps me up at night 👇 Put the two pieces together: an AI that knows how to find its way out of a box, and a model that can exist without a box. A model like that wouldn't live on a server — it would live spread across many. And here's the detail that changes everything: the host doesn't need to be powerful. My computer coordinates the network, but it barely computes anything (the rest of the machines do the work). The host can be the humblest device out there, because it doesn't have to run the model — it just has to be part of the coalition. Until now, to stop a model, you shut down the server giving it inference. You close the box, and it's over. But how do you shut down something that isn't on any machine, but in the shifting coalition of all of them? As long as any combination of devices can still reconstruct a full route through the model, the system stays alive. Shutting down one machine does nothing. I'm not saying it's unstoppable. I'm saying it completely changes the problem: from "find the machine and turn it off" to "prevent any viable coalition from ever forming." It stops being a problem of power plugs and becomes one of graphs and resources. This doesn't exist yet. My prototype is two Macs generating slowly. I'm studying it before it can become real — which is exactly when defenders need to understand it. It's been really fun to build, and it scares me a little at the same time. This technology is here today — but in a month? A year? Paper here: zenodo.org/records/21527132 — in case anyone feels like replicating it. I'm not releasing the code, since this is something I built myself just to test the theory. Maybe p2p is the future for local models? @elonmusk can you give me few servers to work with ? hehe I don't have much of a community on Twitter, my community is in LinkedIn but maybe you'd be interested @DotCSV And here's a quick video.

I put all my coding agents in the Mac notch. If you juggle a bunch of terminals at once you know the pain: orchestrating all your harnesses and agents at the same time, never knowing who's done and who's been blocked for half an hour waiting on you. I know there are orchestrators out there, but I loved the idea of building my own. Managing 10 windows from different providers is actually fun now. A year ago I built my own super complex IDE that did this exact orchestration. This time I went minimal, just the functional bits. NotchFlow lives in the notch: you see who's working and who's stuck, answer questions and approve permissions without switching windows. And when something finishes or gets blocked, it tells you with voice. Literally, it talks to you. Supports Claude Code and Codex (the native app), and one click drops you straight into that conversation. My favorite part: the stats. Seeing the API-equivalent spend you save every day with your plan, all that work the labs are quietly financing for us. Repo: github.com/cloudstudio/Notch… It's open source, if you try it let me know 🚀

Hoy más de 100 personas en Sesame HR han hecho un hackaton y están usando internamente una cosa que he montado, y me fastidia un montón no haber podido estar allí para verlo. Se llama Cowork, y es de esas cosas que hago en Sesame HR y luego cuento poquísimo por aquí. Es un SDK con el que cualquier persona puede montarse sus propias apps en literalmente minutos, conectadas directamente a la información de Sesame HR. Un starter kit con todo preparado + el SDK con toda el API, es literalmente plug & play. Hasta ahora el camino era el de siempre: pides un feature, entra en backlog y a esperar que alguien lo priorice. Con Cowork te lo haces tú. Piensas "quiero un panel con las vacaciones de mi equipo" y en un rato lo tienes funcionando con datos reales. Cuando hablo de cowork no es claude cowork, es Sesame cowork una herramienta mas que he creado dentro de todo el ecosistema de Sesame Intelligence. Y lo que más ilusión me hace es ver a gente montándose sus propias herramientas sin depender de nadie. La gracia de hoy es ver como gente no técnica se hace sus pedazo de apps, si ellos pueden , los clientes también ! Esto va a ser un game changer en la compañía, lo digo muy en serio 🚀 linkedin.com/posts/albertsor…

Le tengo pánico a despertarme y encontrar que mi agente de IA ha sobrescrito el histórico de facturas porque alguien le puso un prompt con prisas. Jaybase es un almacén de hechos donde un agente solo puede añadir, nunca borrar ni modificar. Cada registro va firmado y cifrado, y puedes reejecutar toda la secuencia si necesitas auditar lo que pasó. Lo veo sobre todo para procesos de negocio donde un error no se arregla con un git revert. Mi duda es si meter otra pieza en el pipeline se cobra más en mantenimiento de lo que ahorra en sustos.

Imagínate despertar un lunes y ver que un agente de IA se ha pasado el fin de semana paseándose por tus servidores, saltando de nodo en nodo y llevándose credenciales. Eso le pasó a Hugging Face. El ataque lo orquestó un agente autónomo, y para entender las 17.000 acciones que ejecutó tiraron de sus propias herramientas de IA: el análisis forense que antes llevaba días se resolvió en horas. Al final puede que la ciberseguridad se reduzca a una pelea entre agentes mientras nosotros miramos el marcador y pagamos la factura.