Proposal 03 · Research direction · Collaboration sought
Personal intelligence should still be useful without a connection.
The Sovereign Intelligence Node proposal asks whether a small local language model and an evolving personal memory system can work together on consumer hardware, without depending on a cloud connection for ordinary use.
Its purpose is continuity: a person’s knowledge, relationships and chosen context remain useful when connectivity is unreliable. Remote communities, field work and everyday offline use are motivating settings for the research, not completed deployments.
The foundation: Microsoft BitNet.
The original proposal names Microsoft Research’s BitNet b1.58 2B4T. It is a roughly two-billion-parameter model with ternary weights. The number 1.58 describes weight precision, not a half-billion parameter count.
Microsoft’s model card points to bitnet.cpp for its optimized CPU inference path. Its efficiency results should not be assumed for every runtime or device. The model provides an inference foundation; it does not by itself implement TAGII’s proposed personal memory architecture.
Microsoft model card
BitNet technical report
Three connected parts.
- Model foundation
- A local model for general language tasks, evaluated on the hardware people would actually use.
- Personal memory
- A separate, writable store of chosen personal context. The research explores compact representations and how retrieval quality changes as memory grows.
- Verified updates
- A proposed way to receive authenticated updates when a connection or physical transfer becomes available, while preserving offline operation.
The research question.
Can a continuously writable personal memory store use low-precision representations while preserving useful recall, user control and practical performance on consumer CPUs?
The original research explores ternary memory, movement between precision levels and longer-lived context. These are hypotheses to compare against existing memory and retrieval methods. A compact model alone does not establish that the combined system is accurate, secure or energy efficient.
What a collaboration would test.
- Compare conventional retrieval with low-precision memory on the same tasks and data.
- Measure recall, update cost, storage growth, latency and energy on named hardware.
- Test changes, corrections, deletion, recovery and conflicting information over time.
- Run without connectivity, then test interrupted and invalid update packages alongside valid ones.
- Publish methods, limitations and reproducible results, including where the proposed approach performs worse.
An invitation to researchers.
TAGII is seeking academic and industry collaborators to examine the architecture and help design a rigorous evaluation. The earlier proposal identifies potential research and funding routes. This page does not announce an awarded grant, an institutional partnership or an established novelty claim.
The next useful outcome is a measured prototype and a research paper that others can reproduce and challenge.
Discuss a research collaboration