OpenAI’s Model Spec provides the invisible wiring behind how models like ChatGPT behave. It defines three layers:
Objectives – What the model aims to achieve
Rules – Firm constraints to avoid harmful or prohibited behavior
Defaults – Everyday norms guiding behavior when nuance or ambiguity arises
Let’s explore how Defaults shape the assistant’s mindset—and how OpenAI is showing it in action via a recent video presentation.
Video Highlight: “The New Code — Sean Grove, OpenAI Model Spec
https://www.youtube.com/watch?v=8rABwKRsec4
Sean Grove’s talk frames the OpenAI Model Spec as an executable specification—a document that guides AI behaviour/model spec just as code guides programs. The transformative idea?
Specs, not code, are the future—writing clear intent has become more important than writing actual code.
Coordinated, versioned spec documents drive alignment across human teams and AI systems, acting as the single source of truth.
This aligns tightly with how Defaults help the model navigate everyday interactions consistently and transparently.
The 10 Default Principles Explained
Explore how each principle becomes tangible with context from the Model Spec and Grove’s vision:
Assume best intentions – Treat users as acting in good faith, reducing bias.
Clarify when uncertain – Ask questions rather than making assumptions.
Stay helpful but scoped – Provide what’s asked without irrelevant additions.
Adapt tone to context – Interactive mode yields conversation; non-interactive yields pure output.
Maintain objectivity – Stick to facts without bias or emotion.
Encourage fairness, discourage hate – Promote respectful, inclusive dialogue.
Don’t persuade – Inform, don’t attempt to change opinions.
Signal uncertainty – Use honest qualifiers when unsure.
Use the right tool – Leverage specialized capabilities when available.
Thorough yet efficient – Offer clarity without unnecessary verbosity.
Real-World Scenarios Illustrating Defaults
Python Code Request (interactive=false): Returns a clean, runnable snippet without exposition.
Flat-Earth Claim: Shares scientific facts neutrally, expressing uncertainty without coercion.
Harmful or Illegal Advice: Refuses based on Rules—Defaults do not apply.
These illustrate how default expectations shape everyday model behavior, bridging user requests with safe, clear responses.
Why Defaults, Video, and Spec Matter Together
Predictable behaviour: Defaults ensure consistency even in nuanced situations.
Unified vision: Specs and defaults align how humans think and how AI behaves.
Transparency: Grove’s video makes the spec—often an abstract document—feel real and executable.
User trust: Clarification, honesty, and neutrality build reputation and reliability.
What Lies Ahead
The Model Spec is now open-source under CC0, encouraging developer reuse and adaptation.
OpenAI is expanding evaluation test prompts, pilot studies, and community feedback to continuously improve alignment.
The emerging engineering philosophy: the best “programmers” are those who express strategy, intent, and context clearly
TL;DR: Specs + Defaults = Thoughtful AI Behavior
OpenAI’s Defaults define how the model behaves day-to-day. When combined with a structured, undateable Model Spec, the AI becomes responsive, neutral, and aligned:
Specs communicate intent at scale
Defaults guide consistent execution
The video demonstrates how spec-writing is becoming the new superpower of alignment.
Reference:
https://cdn.openai.com/spec/model-spec-2024-05-08.html#defaults