Lira Dreams of Electric Sheep

Everybody has an opinion on AI, so I'll at least keep mine short.

Table of Contents:

• Lira Dreams of Electric Sheep
  • Anti-Slop
    • Simplified Technical English
    • Edit-Lint
  • Sleeping
    • The 3 AM Miners
  • Talking
  • Ethics

Anti-Slop

Detecting AI slop is quite hard. Maybe impossible for now. I have two solutions to cut it down.

Simplified Technical English

Agents don't do well with negative directives, they often ignore what you tell them not to write.

My AGENTS.md dictates writing Simplified Technical English, a spec used for aircraft maintenance manuals to avoid confusion.

It's never followed the spec very well, but it works like the Caveman plugin, minus the caveman voice.

Edit-Lint

I block things correlated with poorly thought-out writing from the LLM, namely:

1. Typographic em-dashes.
2. Hard wrapped comments.

When the LLM attempts to write either to disk, the edit is cancelled and the model is told what happened and to reconsider the change. I run tree-sitter against all edits allowing me to target specific syntax.

These both substantially decrease the number of times I block a change in review.

Sleeping

Memory and repeatable skills avoid wasting tokens and time on repeated work.

1. Every memory comes with sources including my name, the agent's name, URLs, etc...
2. Because skills take context even when not used, I limit the total number that can exist.

Degraded agents that have gone off track often have trouble diagnosing themselves.

The 3 AM Miners

At 3:00 AM miner agents launch in disposable docker containers and scour for useful memories and skills. The same agent in the same loop never reviews its own conversation. Each finding is reviewed by a frontier model without the (possibly) poisoned context before being committed. These frontier models also merge duplicate memories, refine skills, and produce a list of problems with the current harness that need to be addressed.

Talking

I've trained a wake word model so I can ask questions and work on simple tasks while AFK or just doing other things.

• An open model extracts vocal features from raw audio.
• I trained my model against those, on a few thousand recordings of myself saying Lira.
• On trigger, parakeet does local speech to text, injected directly into the harness.
• The harness responds verbally with Kokoro. Hands-free music piracy, among other things.

I buy merch for the bands I enjoy so I can support them without supporting certain record labels.

Ethics

To be clear I take major issue with data centers and this stupid AI bubble ().

I wish we could agree to not destroy our planet, but over half my country elected someone who openly calls global warming a hoax. My job is to put myself in a position where I can care for the people I love, the rest of them can burn in the hell they're creating. If the majority of the world wants this, I'm not willing to be at their mercy when the Arctic is a field of clover.

For anyone who thinks this isn't actually an issue, a single campus under construction in Amarillo, Texas is permitted to emit 40.3M tons of CO2e per year. For reference that's ~3.5 x Jamaica's entire emissions, or more than every power source in Connecticut combined. Look for: President Donald J. Trump Advanced Energy and Intelligence Campus

Note: For reference my agent's name is Lira, hence the title. And I use pi as a harness + a random assortment of open weight models. If you'd like my exact config please contact me, it changes too often for this page.