“OpenAI's standard for successfully solving alignment is not internal conviction, but rather convincing external experts and auditors with empirical evidence.”