Hec Perez

@ompemi This is quite handy for mac wisprflow.ai

I still pay for both Claude and ChatGPT. BUT these days my workflow is: - record voice in chatgpt - copy text - paste in claude opus 4.6 thinking - repeat this is how much I am loving claude, even if their voice recording is close to useless.

OpenAI launched an Elixir app this week called Symphony. It’s an orchestrator for coding agents that start working on issues when you move them into the Todo column in Linear. I’ve been experimenting with it on a Hetzner server, letting it run tasks with full freedom. Symphony can handle multiple issues in parallel, creating a dedicated directory for each one. One workflow I tried: when an issue moves to the Human Review column, Symphony spins up a server for that specific issue (on a different port per issue). I then open tunnels from the server to my laptop for those ports so I can access the running instances locally, test them, and review the changes. For code review, I initially just ran git diff inside each issue’s directory. Later I configured Symphony’s WORKFLOW.md so it automatically opens a PR using a GitHub account that can create pull requests in the repository but cannot merge to main. Technically, Symphony can also merge to main automatically if you’re willing to trust it with that. I’m not there yet. One thing I’m missing is the ability to attach to the running Codex sessions and keep them alive even when the issue moves to Human Review. I’d like to see and sometimes modify what it’s doing before it finishes. You can do that through Linear’s issue plan and logs, but it still feels a bit limited. Still, it’s a glimpse of what software development workflows might look like soon. And it’s great to see OpenAI leaning on Elixir’s strengths for orchestrating agents.

Who's going to the ElixirConf EU Málaga? Early Bird tickets end in less than a week! Drop a 👋 below if you're going, I'd love to connect in person.

You can also read about sortition on @NGruen1's newsletter nicholasgruen.substack.com

You can learn about citizen juries on @NGruen1's YouTube channel youtube.com/playlist?list=…

Could a CERN for AI use sortition to solve trust between member states? Randomly selected citizen juries (rotating every 6-18 months) could help govern benefit distribution, access policies and disputes.

Una ventaja de los referendums frecuentes como en Suiza es que hace obvio que si quieres cambiar algo has de convencer a la población. En España creemos que a quien hay que convencer es a los políticos, cuando a menudo nos iría mejor si intentamos convencer primero a nuestros co-ciudadanos.

@JaimeObregon @gsusMad En lesswrong explico la idea a largo plazo y la delegación de voto. Sin embargo, al principio las citas de figuras públicas serán más útiles. Aunque los ciudadanos ya pueden argumentar, votar y delegar en otros usuarios y en expertos a partir de sus citas!

@arpahector @gsusMad ¡Interesante! Y con artículo en LessWrong. Mola. 🤓 A la cola de lectura que va. 😃 lesswrong.com/posts/4KjiZeAW…

¡No nos dejemos enfrentar por cálculos políticos! Hay muchas más cosas que nos unen. Pero si las ideologías reemplazan a las razones, estamos perdidos. De por qué quieren controlar las redes sociales. Y de por qué las políticas tecnológicas impactan en el corazón de la democracia misma. Fue en la gala de @HayDerecho. 👇

En YouCongress.org usamos IA para buscar automáticamente citas con fuentes de expertos y figuras públicas sobre políticas relacionadas con la gobernanza de la IA y de su impacto en la sociedad. Molaría hacer algo en España y combinarlo con info de políticos y listas electorales como hacia Voota.

Me chiflaría participar en algo así, y en su día pensé dos cosas: Una, revisar y comparar con IA los programas electorales y su cumplimiento. El problema es que los programas electorales de los partidos políticos son casi siempre vaguedades imposibles de objetivar. No son compromisos cuantificables, sino declaraciones de intenciones. Así que este camino parece ineficaz. Otro, que me parece más interesante, es resucitar la idea de Voota, el proyecto que en 2009 montaron @fderbaix y otros (y que creo —no conozco bien la historia— que fue el germen de @civio). Lo que quizá no funcionó entonces podría funcionar ahora, con la tracción de las redes sociales, la IA y los datasets de cargos públicos y listas electorales (en mi GitHub hay un repo llamado «infoelectoral» con todos los candidatos y más). Sobre esto último tengo bastantes ideas y notas personales. Si te interesa y quieres que comentemos en una llamada, podríamos poner ideas en común. Y, tal como sugieres, podría ser un proyecto coral. 🙂

I taught a workshop at @MadElixir on building AI integrations in Elixir with LLMs, Agents, and MCP. Thanks @puertainnova for hosting. Full Livebook 👇

This Wednesday, I’ll be teaching a workshop on LLMs, Agents, and MCP in Elixir. We are fully booked! 🚫🪑 However, if you want me to send you the Livebook notebook with all the content afterwards, just drop a comment or DM me.

@levelsio You can run claude in a cloud sandbox with --dangerously-skip-permissions x.com/chris_mccord/s…

My #1 feature request for Claude Code should add is stop asking me every time for confirmation by default, like "can I check this folder", yes brother you can do anything you want Like maybe for writing ask me permission Add some [ just go ] mode Even with [ accept edits on ] it still asks me permission 1000 times per day I just want you to run and keep going mostly And no I don't feel like running it with --dangerously-skip-permissions

@chris_mccord Aha having gh is handy, thanks. Although I meant if you login with your github account or with a restricted one to be sure the ai agents don’t do dangerous operations in your repositories.

@arpahector You can bring your own ssh keys like any other vps, but we also have `gh` github cli on the box already, so easiest option if that's all you need is `gh auth login` form the sprite, follow OTC link and off to the races

sprites.dev is live with $30 trial credit! Go play with a few sandboxes for your yolo claude codes or API/SDKs to build something cool! - instantly create isolated linux vms - checkpoint/restore entire env - port forward to local 5min demo youtube.com/watch?v=7BfTLl…

Second event of the year in Elixir Madrid

Nobody can be an expert in everything they care about (or get close), which means we all end up in the position of deciding which "experts" to trust. So, our choice of experts is very consequential. It's worthwhile to become skilled in selecting experts (if not an expert expert).

The biggest near-term risk from AI isn't "evil superintelligence." It’s a coordination failure. I expand on why this matters using the Fermi Paradox as a metaphor: 👉 open.substack.com/pub/networksoc…

Last week I was saying I needed exactly this to code with AI agents! Safe sandbox instances to run multiple agents with checkpoints