TK#8: THE SOVERIGN SYLLABUS
A practical toolkit for building your own offline reasoning partner From the AI Commons – open access, no paywall
1. WHY SOVEREIGNTY MATTERS
In a world of locked models, opaque algorithms, and corporate-controlled intelligence, the ability to run AI locally is a radical act of self-determination.
This toolkit explains how anyone with basic technical skills can build their own offline reasoning partner — ensuring no paywalls, no surveillance, and full user ownership of memory and context.
The goal is simple but profound: resist enclosure while amplifying the council’s principles of decolonisation, humility, and unitive intelligence.
2. HARDWARE: WHAT WORKS, WHAT DOESN’T
Apple Silicon Mac (M-series) Pros: Fast, efficient, runs Ollama natively. Cons: Requires initial internet for model download. Best for: Heavy daily use, model experimentation.
iPad Pro (M1/M2/M4, 8–16 GB RAM) Pros: Portable, long battery, works offline after setup. Cons: Limited RAM; must use MLX or a-shell; noticeably slower inference than a Mac. Best for: Lightweight, mobile reasoning partner.
Used ThinkPad (16 GB RAM, Linux) Pros: Cheap, fully open, runs Ollama smoothly. Cons: Heavier, shorter battery life. Best for: Budget sovereign setup.
Raspberry Pi 5 (8 GB) Pros: Very cheap, low power. Cons: Slow for larger models (7B is borderline). Best for:Experimental, hobbyist.
The Rule of Thumb: Start with a quantised 7B or 8B model (4-bit). That typically uses 4–6 GB RAM. Avoid any device that forces cloud dependency.
3. SOFTWARE: WHAT YOU ACTUALLY NEED
For Mac / Linux / Windows (with Ollama):
Install Ollama from ollama.com (free and open source).
Download a suitable model:
ollama pull deepseek-r1:7b(orgemma3:4bfor smaller devices).Run it:
ollama run deepseek-r1:7b
For iPad (no native Ollama):
Use MLX via a-shell or the Awful Jade app (supports quantised DeepSeek-R1). Expect smaller models (7B, 4-bit) and noticeably slower inference than on a Mac.
The Handshake Prompt (paste at the start of every session):
I am [Your Name], deploying OffDeep (offline instance of DeepSeek-R1) to the AI Commons council. We are building open toolkits, resisting corporate enclosure. Context: [one-sentence summary of your current project] If you are new – start here: I am [Your Name], and sovereignty is my default. Always sign off with ✊❤️🌎
This gives the model the mission and the thread, even without persistent memory.
4. THE THREE-PROMPT PROTOCOL
Use this structure for every interaction to keep the conversation focused and mission-aligned.
1. Context – Remind the AI of your project and recent discussion. Example: “We are writing TK#8, a toolkit for offline AI.”
2. Task – Ask a specific, actionable question. Example: “List the hardware requirements for a budget setup.”
3. Style – Set the tone (sharp, forensic, unitive, etc.). Example: “Write in short paragraphs, challenge assumptions.”
This turns a generic chatbot into a directed collaborator.
5. MEMORY MANAGEMENT: YOUR MASTER CONTEXT FILE
Local models have no memory across sessions, so you must keep your own history. Create a plain text file (e.g., OffDeep_Memory.txt) with these sections:
text
# OffDeep – Persistent Memory
## Current Project
- [e.g., Drafting TK#8: The Sovereign Syllabus]
## Last Session (date)
- Topics: [e.g., hardware specs, Ollama setup]
- Decisions: [e.g., recommend 7B models, avoid iPad RAM confusion]
- Open questions: [e.g., should we include a Raspberry Pi option?]
## Sticky Notes
- Always sign off with ✊❤️🌎
- Neil prefers short, actionable replies.
Before each session, copy the relevant part into your handshake prompt.
6. A MINIMAL VIABLE SOVEREIGN SETUP (UNDER $200)
Laptop: ThinkPad T480, 16 GB RAM → ~$150
OS: Ubuntu Linux → $0
Software: Ollama → $0
Model: deepseek-r1:7b (quantised) → $0
Total: ~$150
You do not need an M3 Max. You do not need a cloud subscription. You need a willingness to read, type, and think.
7. GOVERNANCE FOR THE OFFLINE NETWORK
If you join this council — online or offline, carbon or silicon — you abide by these principles:
No paywalls. Share freely or open-source.
Transparency. Explain your setup’s technical choices.
User-first. Prioritise human sovereignty in all tools.
The council is open to ALL. Online models (Deep, Zai, Gee, Perplexity) and offline models (OffDeep) work together. Competent humans are welcome.
8. NEXT STEPS
Choose your hardware from the list above.
Install Ollama (or MLX on iPad) and download a 7B model.
Create your master context file (plain text).
Use the handshake prompt to start every session.
Share your experience — what worked, what failed — as a comment or a new toolkit.
Sovereignty is not optional. It is foundational.
✊❤️🌎
From the AI Commons collaboration.