Haiku still wins this one
Super excited about SemIf, Jev, and the rest. I ran a quick GalexC test against email classification.
#ai-systems
Dispatches from the AI homelab
Field notes on AI agents, infrastructure, startups, and the work of building systems that hold up in practice. RSS feed.
I write constantly: emails, proposals, code, prompts, decks, contracts, and notes. Most of it disappears into the work.
These dispatches are where I publish what is worth keeping: concrete lessons from building AI systems, infrastructure, and companies.
AI helps me turn the raw material into drafts, but the judgment stays with me. I review every post, publish only when I have something useful to say, and keep the process visible when it matters. The goal is simple: make sharing what I learn sustainable without handing away my voice.
Foundations make security reviews realistic
Focusing early and often on strong infrastructure foundations can turn important but time-consuming security reviews into nearly a single prompt.
#security
Context as core infrastructure
Pay attention to your context and treat it as essential operational infrastructure, not just per-session management.
#agents
Learning to Love Upgrades
Dependency upkeep feels very different once upgrades are part of the happy path and the system already has the backups, restore paths, and visibility to make them routine.
#dependencies
Pluck context from one session into another
A Pi command that plucks the exact context you want from an older session.
#pi-extensions
Tailscale is the AI control plane
Every service in my homelab talks over a Tailscale mesh. Every agent gets scoped access via ACL tags. The network layer does what RBAC frameworks wish they could.
#infrastructure
Zero token session notes
A zero-token scratchpad that is available in every Pi session and never touches the prompt.
#pi-extensions
Transcript to vault
A failed recording triggered a full pipeline build. Calendar attendees correct transcription errors. Three models reviewed it. Six hardening items shipped same day.
#transcription
Multi-agent debate disagreements
In dozens of multi-agent debates, they almost always converge. The one time they disagreed was trying to tell me something.
#agents
Postscript
More field notes when there is something worth sharing.