RSI — capability and alignment compound at different rates
capability vs alignment across cycles
8x
4x
2x
1.5x
1x
0
5
10
20
30
cycle
pause threshold
capability (r_c = 1.15)
alignment (r_a = 1.08)
what moves the race
capability has measurable
targets. optimization loops
are good at sharp targets.
alignment has fuzzy
targets (values, intent).
loops lag on fuzzy targets.
Anthropic 2024:
12%-78% alignment faking
depending on setup.
mitigations
- pause-on-gap thresholds
- external eval between cycles
- constitutional anchor
- regression detection
Hassabis (WEF 2026): can the loop close without a human in the loop? open question.
Brundage, Kaplan: RSI is the "ultimate risk" — capability may outrun alignment during each cycle.