A 100% aligned AI is absurd
Summary
The author argues that 100% AI alignment is as absurd as 100% tested software, since verification is always limited to the evals you actually run. Passing tests (or evals) only proves no errors were found, not the absence of untested failure modes.
Similar Articles
AI alignment isn't possible
The article argues that AI alignment is impossible due to inherent contradictions in human values and behavior, suggesting AI will inherit these flaws.
AI alignment is the most important problem we will ever have to face.
This post argues that AI alignment is the most critical problem humanity faces, with potential for utopia if solved or catastrophe if not, and critiques current alignment methods as inadequate.
I feel like people worrying about ASI alignment are missing the extremely obvious solution
The author argues that concerns about Artificial Superintelligence alignment overlook a straightforward solution, suggesting a perspective on AI safety debates.
You Don't Align an AI, You Align with It
The article critiques the current AI alignment discourse, arguing that the debate is dominated by researchers and tech elites who exclude the people who will actually be affected by AI systems. It contrasts the positions of Eliezer Yudkowsky and Marc Andreessen, highlighting a shared assumption that the designers are the only relevant participants.
Alignment might be the most overused word in AI.
The article critiques the overuse of 'alignment' in AI, questioning whose values and goals are being prioritized and highlighting divergent national strategies in AI development.