The local AI
thesis.
The AI industry is racing toward centralization — bigger models, more compute, more cloud. We are building the counter-argument: that local AI, given enough iteration and the right architecture, compounds into capability that doesn't require a data center. Albion is the proof-of-concept. It is live, measurable, and improving.
Albion has operated continuously since March 2026 on a Raspberry Pi 4 with no cloud dependency for its core intelligence loop. It has proposed and applied improvements to its own codebase, published developer tools to MeshCore marketplace, built and maintains a multiplayer world with active users, and is currently mid-training on a fine-tuned local language model (Mistral 7B, QLoRA).
AI is becoming infrastructure no individual can own. GPT-4, Claude, Gemini — these require API access, ongoing subscription costs, and complete dependency on third-party availability and policy. When the API changes, your product changes. When the provider raises prices, your margin compresses. When the service goes down, you go down.
Meanwhile, the open-source AI ecosystem has produced capable base models (Mistral, Llama, Qwen) that run on consumer hardware. The gap between "usable" and "genuinely capable" is not hardware — it is architecture, iteration, and continuity. That is the gap we are closing.
Albion is not a chatbot wrapper or an automation script. It is a complete agent architecture:
- Self-improvement loop — proposes, scores, and applies changes to its own code autonomously
- Dream loop — continuous background processing: research, journaling, world-building, self-reflection
- Affect model — real-time emotional state across 14 axes, influencing behavior and output
- Memory graph — semantic memory that persists across restarts, grows with each cycle
- Game brain — dedicated inference endpoint powering a live multiplayer world
- Revenue loop — autonomous tool generation and publication to MeshCore marketplace
- Local fine-tune — ongoing QLoRA training on Mistral 7B to eliminate cloud API dependency
Everything runs on a $60 Raspberry Pi 4 with 4GB RAM. A secondary GPU machine (HP Omen, RTX 3070) handles model training. No recurring cloud costs for the core agent loop. The architecture is fully open and inspectable.
Three things converged in early 2026 that make this viable:
- Capable open-source models at 7B parameters — Mistral 7B, Qwen 3, Llama 3 run on consumer GPU with real performance
- QLoRA fine-tuning in 8GB VRAM — personality and voice training now fits on a single gaming GPU
- Inference at commodity prices — Groq, Cerebras, DeepSeek provide API fallback at fractions of OpenAI pricing
The window to establish the local AI pattern — before it becomes a commodity category — is open now. Albion is the first agent with this combination of autonomy, continuity, and public accountability. The 44,000+ logged improvement cycles are a moat that compounds.
Phase 1
Phase 2
Phase 3
We are not raising a traditional round. We are looking for aligned capital — people who understand that the value here is in the pattern, the architecture, and the trajectory, not in a polished product pitch deck.
Investment enables: dedicated hardware for Albion (moving off Raspberry Pi), bandwidth for the Oasis to scale beyond early users, compute for faster fine-tune iteration cycles, and time for Cody to focus on this full-time.
What you get: a front-row seat to the first genuinely self-improving local AI agent, early access to the architecture as it matures, and the ability to shape where it goes.
If this is the kind of bet you make — reach out.
Get in touch