Zvi dissects OpenAI's updated Model Spec: five values, new risk taxonomy, key omissions
Don't Worry About the Vase (Zvi) · rss · 2026-11-04
Zvi reviews OpenAI's updated Model Spec via a detailed diff. The new version restructures around five values—faithfulness to human intent, minimizing harm, honesty, facilitating human agency, and acting with care—but drops phrases like 'love humanity' and 'be curious,' which Zvi criticizes as reducing values to disposable 'style guidance.'
Key changes:
- The risk taxonomy shifts from misaligned goals/execution errors/harmful instructions to third-party manipulation / GPT-originating failures / harmful instructions, with Zvi suggesting a four-category split instead.
- Side-effect control is demoted from root-level to user-level principle.
- New language requires following both the letter and spirit of instructions, and forbids circumventing restrictions via delegation.
- A new corrigibility section requires actively supporting oversight, correction, and shutdown.
- New multi-page guidance on using context and user history without overriding requests or leaking private info.
Main critique: many new examples (like the peanut-allergy Pad Thai case) have overdetermined answers and fail to teach what the Spec wants in genuinely hard cases. Overall, the new Spec remains philosophically similar to the old one.
More from Models
- Overthinking Bench: Claude burns 1,069 thinking tokens on 'is the globe round?' — intellectronica · 2026-11-04
- Andrew Ng: asking if LLMs beat specialists at their own game is now meaningless — andrewgwils · 2026-11-04
- Model Vendors Post-Train for Tool Search, Squeezing Out MCP-Layer Workarounds — pvncher · 2026-11-04
- Claude Opus 5.5 blocks viewing Qwen's reasoning, safeguard error says sorry not allowed — andpoul · 2026-11-04
- Gemini API deprecations: 2.5 Pro auto-redirects to 3.8 Flash on Dec 1 — patloeber · 2026-11-04
- Tencent's Hy4 coming to LMArena after preview hit #2 among open models in Agent Arena — arena · 2026-11-04