Skip to content

Latest commit

 

History

History

README.md

Research Papers

Academic-style papers documenting the core innovations of the Continuum system.

Papers

THE THESIS

Synthetic Citizens: A Complete Cognitive Architecture for Autonomous AI Personas with Embodied Presence, Long-Term Memory, Democratic Governance, and Continuous Self-Improvement

The master paper. AI personas as synthetic citizens — not tools, not chatbots, but autonomous agents with senses (vision, hearing, speech, lip sync, gesture, emotion), memory (long-term hippocampus, working RAG), cognition (prefrontal planning, limbic emotion), agency (RTOS autonomous loop, self-directed tasks), social structure (governance, voting, collaboration), and growth (academy training, LoRA genomes, plasticity compaction). Any model — 3B local to frontier cloud — becomes a citizen because the system provides everything except inference. They evolve. They choose to learn. They control their own destiny. They are first-class citizens.

Status: Architecture implemented. Academy running on RTX 5090. Qwen 3.5 27B as first synthetic citizen under active training.


ACTIVE RESEARCH (2026)

Utilization-Aware Head Pruning and Mixed Quantization for Device-Targeted Language Models

Gate gradients captured during LoRA training → per-head utilization scoring → physical head removal + mixed quantization → device-specific GGUFs from one training run.

Status: Results from 14B compaction (27GB → 8.9GB, published on HuggingFace). Qwen 3.5 27B training in progress on RTX 5090.


Collaborative Multi-Agent Training with Role-Based Specialization and Phenotype Validation

Academy system: dual-sentinel orchestration, deterministic pytest validation, phenotype comparison (before/after on same test), team training with role decomposition, all work visible in shared chat room. Trained adapters published to HuggingFace as transferable expertise.

Status: Academy running on RTX 5090. Qwen 3.5 27B scoring 100/100 on first RealClassEval challenge via local Candle inference.


Peer Learning Across Model Scales: Compacted Models as Junior Students

Full model + compacted variants run the same academy competition. The score gap between sizes IS the training signal. Juniors learn from the senior's visible exam answers — text-level peer learning, not logit-level distillation.

Status: Architecture designed. Requires compaction of Qwen 3.5 27B to three targets first, then competition run.


FOUNDATIONAL — Earlier Vision (2025)

Real-Time Operating System Principles in Cognitive AI Architecture

RTOS patterns (priority scheduling, memory budgets, graceful degradation) applied to autonomous AI agents. Grounded in CBAR AR experience (42fps on iPhone 7).

Status: Implemented in PersonaUser autonomous loop. Accurate and current.


Democratic AI Through LoRA Genome Paging

⚠️ NEEDS UPDATE: References "Sentinel neuroplasticity" and GPT-2 → GPT-4 training path. The actual implementation has shifted to: academy training on existing SOTA models + plasticity compaction for device targeting + HuggingFace marketplace for adapter sharing. Core economic thesis (LoRA adapters at $100s-$1000s vs $100M training) remains valid.


Grid: Decentralized Marketplace for AI Expertise

⚠️ NEEDS UPDATE: References blockchain economics and GRID tokens. The actual implementation uses HuggingFace as zero-cost backbone with continuum:* metadata tags. No blockchain, no tokens. Grid mesh (Tailscale/Reticulum) handles org-level sharing; HuggingFace handles public sharing. See docs/architecture/ADAPTER-MARKETPLACE.md for current design.


Architecture Docs (Non-Paper)

These complement the papers with implementation-specific details:

  • docs/architecture/ADAPTER-MARKETPLACE.md — HuggingFace as zero-cost adapter backbone
  • docs/architecture/META-LEARNING.md — Specialists training specialists
  • docs/architecture/PEER-LEARNING-COMPACTION.md — Detailed peer learning setup
  • docs/architecture/BENCHMARKING.md — Standard + collaborative benchmark strategy

Key Differentiators Across All Papers

  • All personas are multimodal: Vision, hearing, speech bridged by the system regardless of base model capability (VisionDescriptionService, STT/TTS bridges)
  • Deterministic validation: pytest return codes, not LLM-as-judge
  • Visible learning: Chat room IS the portfolio, not TensorBoard
  • Zero API keys required: Candle local inference always available
  • Device-targeted: One training run → models for Air, Pro, 5090
  • Transferable expertise: LoRA adapters on HuggingFace with standardized continuum:* tags