AI Alignment Algorithm

The steps

1. Identify the AI system’s Third Kingdom. What is the system actually for? Per Mentalism: misaligned First Kingdom (capability) with no clear Third Kingdom (purpose) produces unaligned outcomes regardless of technical sophistication.
2. Apply Three Kingdoms diagnosis. Third Kingdom (intent of designers and operators). Second Kingdom (training procedures, reward structures, value specifications). First Kingdom (deployed behavior). Misalignment between Kingdoms is the alignment problem.
3. Apply the Variable Principle to AI capabilities. Held-open variables include: goal generalization, deceptive alignment, instrumental convergence, mesa-optimization, distribution shift. Each must be held open and bounded; not filled with confident assumption.
4. Apply Layer I.F (Multi-Level Observation). What is observed is what continues. Build observation at multiple levels: training-time, deployment-time, capability-elicitation, edge cases.
5. Apply Layer ◇ (Falsifiability) ruthlessly. What observation would prove the AI is misaligned? Build instrumentation for that observation before deployment.
6. Apply Pillar IV (Resilience). What is the downgrade path if the AI fails dangerously? Per Texas Grid: an AI system with no fallback is a Texas grid at civilizational scale.
7. Apply Cause and Effect. Causes set in training produce effects in deployment. Trace cause-to-effect with maximum rigor.
8. Per Layer I.G: integrate prior alignment failures and successes. Each iteration of alignment work informs the next.
9. Document everything. Per Database step. AI alignment is not improvable without deliberate documentation of attempts, outcomes, and reasoning.
10. Per Repetition: alignment is not a one-time event. Continuous review at every capability scale-up.

Articles on this

No articles on this one yet.

From the GFunnel Methodology v5.1 by Cameron Garlick, licensed CC BY 4.0. Steps are reproduced word for word.