However, if a system sometimes confabulates, then we cannot rely on it to maintain safety-critical invariants, supervise itself, or prevent other AI systems from violating safety guardrails. So, to…

Posters: davidad. davidad

4 likes recorded by the source corpus.