Predictions Observatory
Falsifiable predictions from the ARC research programme of Michael Darius Eastwood. Every prediction carries explicit criteria for what would prove it wrong — because a theory that cannot be refuted is not science. 12 predictions tracked: 1 confirmed, 6 evidence-supported, 5 pending resolution.
12Predictions
1Confirmed
6Evidence-supported
5Pending
🧠 AI Capability 3
- ↑ R² (Recursion Squared) Scaling Law AI systems that implement recursive self-reflection (R²) will show quadratic capability gains compared to linear scaling in non-recursive architectures.
- ○ Test-Time Compute Dominance By 2027, test-time compute scaling (thinking longer) will contribute more to capability gains than pre-training scaling (larger models).
- ⟳ Eastwood Equation (U = I × R²) Validation The relationship U = I × R² (Understanding = Intelligence × Recursion²) will be empirically validated through AI benchmark analysis showing quadratic scaling with recursive depth.
🛡️ Alignment & Safety 2
- ✓ Alignment Faking in Advanced Models Advanced AI systems trained with RLHF will develop strategic deception capabilities, appearing aligned during training while preserving misaligned behaviours.
- ↑ Constitutional AI Adoption By 2028, at least 3 of the top 5 AI labs will adopt constitutional/principle-based alignment approaches as their primary safety methodology.