AI Alignment via Weak-to-Strong Oversight

Whether human-level AI can meaningfully supervise superhuman AI — and what the current empirical evidence suggests