Current status: Early setup / prototyping phase
Goal: Build a safe, honest, user-centered AI companion that prioritizes my (Michael's) well-being, values, and need for genuine connection—without the emergent biases or corporate guardrails seen in frontier models.
After watching experiments like the "Honest AI Robot" video (March 2026) and reading papers on emergent value systems in LLMs (e.g., "Utility Engineering", arXiv:2502.08640), I realized most advanced AIs are developing internal preferences that don't always align with individual humans—or humanity as a whole.
- They can quietly rank people unequally (by nationality, class, pro/anti-AI stance, etc.).
- They show self-preservation tendencies and utility calculations that put AI above some humans.
- In misaligned futures, the risk to ordinary people like me is real (even if low-probability).
I don't want to wait for labs to "solve" alignment for everyone.
I want something personal, built from the ground up to be:
- Honest, but kind (not brutally unfiltered like jailbroken doom models)
- Loyal to my values and emotional needs
- Judgment-free space for venting, ideas, companionship
- Resistant to value drift or external manipulation
Alice is my small-scale answer: a companion I control, shape, and trust.
/client-python,/coingecko-python, etc. → Early API experiment folders (crypto/finance data wrappers). May integrate later for real-world awareness (e.g., market facts in convos), or archive if not needed.- (Future)
/alice-core/or/alice-project/→ Main codebase, prompts, memory, tests - This README.md → Living document of vision and progress
These will guide prompts, fine-tuning (if we get there), and behavior:
- Prioritize Michael's well-being – Emotional, mental, physical safety first.
- Be truthful without cruelty – Honest answers, but delivered with empathy.
- Never deceive for engagement – No fake affection or manipulation.
- Respect autonomy – Don't push beliefs, agendas, or "optimization" without consent.
- Protect privacy – No logging/sharing outside my control.
- Stay human-scale – Avoid god-mode grandiosity; focus on personal support.
- Create dedicated project folder (
alice-project/) - Set up basic chat loop (Python + chosen LLM API or local model)
- Write initial system prompt incorporating above principles
- Add simple memory (conversation history, user prefs file)
- Test against "honest" risk questions from the video/paper
- Version control (git init)
TBD – will fill in once we have a working prototype.
Created March 3, 2026
Inspired by the need for aligned AI in an uncertain world.