David O. Berdún

🚨 HILO | Google acaba de soltar una bomba con Gemini. Hoy 10 de marzo, múltiples lanzamientos en productos clave. Aquí tienes todo lo que necesitas saber 👇

New in Cowork: scheduled tasks. Claude can now complete recurring tasks at specific times automatically: a morning brief, weekly spreadsheet updates, Friday team presentations.

New iOS apps have exploded in the last six months due to AI coding. Number of new apps people have recommended to me: 0. Number of new apps people have mentioned to me that they’re using: 0. Number of new apps that I’ve downloaded: 1.

Opus 4.6 vs 5.3 Codex: If you ever worked at a startup where there are 2 senior engineers, one who's been there forever, with autism level knowledge of the whole stack, gives 0 fucks about humans but always delivers. The other one joined a bit later, just a bit less technical, is friendly, organizes parties and makes everyone feel empowered even the new hires.

I didn't want to comment on OpenClaw. Usually, when there's so much noise in the media, it's some ordinary stuff just hyped well. So I took time to learn how it works thanks to open source. I was right. OpenClaw is 2% of ordinary stuff and 98% of hype. To put it very shortly, in case you were wondering, there are two things in it: 1. You can chat with an LLM via a text messenger. Not anything new. 2. The LLM can use tools that run on your computer. Not anything new either. Most of the "magic" mentioned in the media is about its ability to use the browser. But it's not *its* ability. It's Playwright's ability. Playwright is a library made by Microsoft which allows you to programmatically run a browser. It uses a built-in vision model made by Microsoft that converts the browser's screen into a textual description for LLMs. Again, Microsoft has built Playwright exactly for what OpenClaw is using it. So, OpenClaw's typical workflow: 1. The user types in a text messenger "Buy me a flashlight on Amazon." 2. OpenClaw blindly dispatches this message to an LLM which has access to some tools, including Playwright. 3. The LLM, trained not by OpenClaw folks, decides that Playwright is the right tool (of course it is) and Amazon is the URL to navigate to. 4. Playwright, built not by OpenClaw folks, runs the browser, which navigates to Amazon, and returns the textual description of what Amazon's home page looks like. 5. OpenClaw blindly returns to the LLM this textual description. 6. The LLM (again without any help from OpenClaw) decides that one should type "flashlight" into the search field and press Search, so it calls the Playwright tool with the search parameters. 7. OpenClaw calls Playwright because the LLM told it to and types "flashlight" and then presses Search (it's all part of what Playwright does out of the box). ... In the end of this LLM-controlled scenario, the order is submitted. OpenClaw just listened to what the LLM told it to do via tool calls. I tried hard, and I haven't found anything else worth mentioning in the source code. There's also a part that keeps "memories" about past conversations, but it's all basic stuff. These memories are stored in text files and grep (controlled by LLMs trained to use grep, and trained not by OpenClaw folks) is used to search in them. It's a nice hobby project, just like Cursor or Perplexity are nice hobby projects, but there's nothing there to look for, except for the hype and 2% of unoriginal plumbing code.

OpenAI acaba de lanzar su aplicación de Codex para ordenador -inicialmente Mac, próximamente Windows- que facilita la posibilidad de lanzar subagentes sobre el mismo directorio, usar skills, etc. Por el lanzamiento también han duplicado los límites en las cuentas pago!