Deliberative Alignment, And The Spec

Scott on OpenAI's deliberative alignment paper (constitutional AI + chain-of-thought: train the model to reflect on the 'spec' before answering; the Crowley-razor self-training analogy) and its limits (the scratchpad isn't true reasoning - the graded-high-school-'reflection' failure mode). Part II is the valuable reframe: the model spec's chain of command is the real future-alignment crux, surveying 6 positions - prioritize the parent company (-> OpenAI rules the world), the government (-> whose? legitimacy crises), the spec/user, the moral law (LLMs have ethics from training), the average person (Leike's citizens'-assembly), or CEV.

Quality 72 · Strong. A clear explainer plus a genuinely clarifying framing: Part II's chain-of-command taxonomy crisply lays out the political-economy stakes of who superintelligence answers to - an underexplored, important question - and the Crowley/high-school-reflection analogies are sharp pedagogy. Strong, held just below the 74 AI cluster because Part I is straightforward exposition of someone else's paper and Part II is somewhat list-y.

Claude’s paradigm shift 50 · Moderate. Moderate - the chain-of-command taxonomy is a fresh, clarifying framing; deliberative alignment itself is OpenAI's, here explained.

Real-world impact 2 · Minor. 2 - AI-safety discourse; a good explainer + framing rather than a coinage that entered the lexicon.