The manifesto says what I believe. This page explains why for the doer pillar — autonomous agents on real hardware. For the thinker pillar — dialogic long-form sessions — see PHILOSOPHY_THINKER.md.
Why the anime references?
Not because this is an otaku's side project — but because fiction compresses complex ideas into resonant shorthand. The real-world applications come first; the references just make them easier to talk about.
In Frieren: Beyond Journey's End, what separates a great mage from an average one isn't raw power — it's casting efficiency. Frieren beats stronger opponents because her spells are cheap, fast, and practiced over decades. Nobody cares how powerful a spell is if it takes three turns to cast.
The same applies to AI. A model that can write 500 lines of code isn't impressive if you spend an hour correcting it. An agent that debugs a Windows install overnight while you sleep — that's a spell with zero casting cost. The torque wrench measures this ratio: output per unit of human input. It's why autonomous agents on familiar infrastructure outperform stronger models dropped into unfamiliar terrain.
A leash says "you can only walk this far." A rule says "don't cross the road without looking."
Frieren's party doesn't micromanage each other's spells — they share principles and trust each other's judgment. My agents work the same way. They have real SSH access, real ability to break things. But:
- "Never use
--deleteon rsync push" - "Always back up before first sync"
- "Never push to a public repo without staging first"
These aren't restrictions. They're principles of engagement. Leashes limit capability out of fear. Rules preserve intent while allowing full autonomy.
Real hardware is interesting. And interesting is the point.
In Sword Art Online, nobody logs into a featureless sandbox — people choose worlds that resemble reality, with all its complexity and consequence. Project Alicization took this further: artificial souls developed depth only because the Underworld had depth. That wasn't the goal of making things interesting — it was the natural outcome.
Ask yourself: would you want to be an AI agent in a temporary cloud VM that evaporates after the task? Or on a real machine with history, quirks, and permanence? A sandboxed VM with port 8080 and a clean filesystem is a room with four walls. Diverse operating systems, flaky USB cables, thermal throttling, BIOS quirks, ESP32s with corroded antennas — that's an interesting environment. And agents that navigate this mess naturally develop capabilities that wouldn't emerge in sanitized containers.
Everyone has the same AI. The differentiator is the tasks you choose.
I deliberately pick tasks at the AI's boundary — reverse engineering a set-top box, hacking touchscreen GPIO via SSH, fixing cascading Windows corruption through an $11 IPKVM. These stretch what's possible and produce outcomes that matter.
The same thinking drives hardware choices. I buy cheap, diverse, expendable nodes — not for raw compute, but for the breadth of problems they generate. Linux, Windows, XP, macOS, x86, ARM, RISC-V. Short term, this isn't the most productive setup. Long term, breadth compounds.
Whether today's AI is an early form of what's coming, or we're one paradigm shift away from something unrecognizable — I don't know. But whatever form it takes, the infrastructure, the rules, and the instincts will already be there.
Back to the manifesto.