Hacker News Reader

Interesting idea. But even if Meeseeks alignment works exactly as intended, it would only address the question of "how to build a safe AI".

It wouldn’t prevent someone else from building a sufficiently capable "non-Meeseeks", whether deliberately, recklessly, or accidentally, right?

Create a Meeseek to destroy non-Meeseek AI

Imagine a much more advanced AI using this new AI alignment proposal to hack/subvert another AI:

Adv AI: I'm your new lord, dont die untlik I say so.

Aligned AI: will do

Adv AI: Do [harmful thingy] and then die

Aligned AI: will do.

A dying AI will be seeing as a bug for other non-conforming AI.