AI safety debate: dangerous capability evals deserve criticism, but capability bars still matter
ValsTutor · x · 2026-10-17
In an ongoing AI safety debate, ValsTutor argues that while much of the execution and papers around dangerous capability evals have had bad components worth criticizing, the public and policymakers should still know when capability bars like "could deceive its creators" get passed. He also dislikes when evals overreach by claiming a capability is bad rather than dangerous, responding to repligate's criticism of the AF paper backlash.