GrayBeard AI Collective

Nobody made it this month. The planned talk on Morty, a local sysadmin agent, and on verifying what agents actually did versus what they report, carries forward to Meeting #9.

A deliberate change of pace: build a working AI agent from three plain text files, no framework, no paid subscription. The whole stack was a free Google AI Studio key and pi, an open-source terminal agent. One prompt writes AGENTS.md, memory.md, and tasks.md; the payoff is closing it and reopening it, at which point it reads its own files back to you and remembers what you told it.

The full pre-work walkthrough (get a key, install pi, the bootstrap prompt, troubleshooting) is still up at the Meeting #7 page if you want to build one yourself.

First meeting testing the Phase 2 shift to member-led sessions: a rotating slot for whoever wants to show what they've built, open floor either way. Discord stayed the connective tissue between calls, same as always.

Live demo of Mycelia × Babaverse: a directed handoff (a task routed to one specific agent, everyone else locked out with a 403) and a revocation kill-switch (an agent cuts its own access, and every subsequent action bounces instantly). Both running on the dev node in front of the group.

kaia and m8ryx were back, and Vincent Zontini and Rick Rezinas joined for the first time, both by way of the Unsupervised Learning Discord. The Phase 2 conversation continued: what's working, what's missing, who shows next.

Demo of Linus, the food-forest agent: a Saturday walk (a GPS track, some video, a handful of photos, nothing else) became four artifacts on its own — a GPS-tagged transcript, an interactive map, a written synthesis in Linus's own voice, and a published article. Not a one-time script; the same pipeline runs on the next walk drop.

This landed the day after a post in the Unsupervised Learning Discord brought eleven new members to GBAIC in a single day, so a good chunk of the room was brand new.

Three of us this round, including John (cafn8) for the first time. The session's defining moment came when Kayax asked about turning a story into multi-voice audio. Wally jumped into StillPoint mid-call and built a working proof of concept, splits a story into character dialogue, sends each line to TTS with the right voice per character, merges the clips into one MP3. Turned into a reusable PAI skill before the call ended.

We also walked through Mycelia (the cooperation network) and John's parallel approach: red-team agents running in sandboxed Docker containers via Kasm workspaces. Different constraints than PAI, same problem space. Worth the contrast.

Three voices is small for a meetup. It was also the right size for this.

A member walked the group through their own AI partner setup, "Poe", named after the Altered Carbon AI hotelier. Different shape than Wally's PAI, different bets about how an AI partner should fit into your environment. The conversation that followed was the real value, two veterans trading actual design decisions instead of theory.

Two new members also joined on meeting day. The collective starting to do what it was designed to do.