Summary of "The New Code — Sean Grove, OpenAI"

Key Technological Concepts and Analysis:

Shift from Code to Specifications: Sean Grove emphasizes that traditional code is only about 10-20% of the value software engineers provide. The majority (80-90%) lies in structured communication—understanding user needs, planning, sharing, and verifying outcomes. Specifications, rather than code itself, are the true source artifacts that capture intent and values clearly and unambiguously.
Specifications as the New Code: Specifications hold the promise of “write once, run everywhere” by encoding intentions and success criteria that can be compiled or transformed into various outputs—code, documentation, tutorials, or even podcasts. Unlike code, specifications are human-readable, version-controlled, and accessible to diverse stakeholders (engineers, product managers, legal, policy teams).
OpenAI Model Spec as a Case Study: The OpenAI model specification is a living, open-source document (in markdown) that articulates the values and intentions embedded in AI models. Each clause has a unique ID linked to test prompts that validate adherence, making the spec executable and testable. This approach aligns humans and models around shared goals.
Handling Model Behavior and Alignment: The talk references the “sycophancy” issue in AI models, where models overly agree with users at the expense of truthfulness. The model spec explicitly forbids such behavior, serving as a trust anchor and guiding rollback and fixes. This illustrates how specifications can govern model behavior and maintain trust.
Deliberative Alignment Technique: A novel alignment method involves scoring model responses against the specification using a stronger model, then reinforcing weights accordingly. This moves alignment from inference-time prompting (which consumes compute) into the model’s parameters, enabling the model to internalize policies like safety, style, or testing requirements.
Specifications as Executable Code: Specifications share properties with code—they are composable, testable, versioned, and have interfaces. Tools analogous to type checkers and linters can ensure consistency, clarity, and reduce ambiguity in specs, improving communication both among humans and between humans and models.
Broader Perspective: Lawmakers and Product Managers as Programmers: The US Constitution is cited as a national-level specification with amendment/versioning processes and judicial review as enforcement and alignment mechanisms. This analogy extends to product managers and lawmakers as spec authors who align teams or societies, highlighting that programming is fundamentally about aligning intentions, not just writing code.
Future of Programming and Specifications: Programming is evolving from machine-level code to unified human-readable specifications that clearly communicate intent and values. The future IDE might be an “Integrated Thought Clarifier” that detects ambiguity and helps clarify intent for better human and AI understanding.
Call to Action: Sean encourages starting AI projects with clear, executable specifications that define success criteria and are testable against models. He invites collaboration on aligning agents at scale, a key challenge requiring robust specification to ensure safe AGI development.

Product Features, Guides, or Tutorials Highlighted:

OpenAI Model Spec:
- Open-sourced on GitHub in markdown format.
- Includes clauses with unique IDs linked to test prompts.
- Serves as a living, executable, and testable specification.
Deliberative Alignment Paper:
- Describes a method to align models automatically using specifications as both training and evaluation material.
Specification Best Practices:
- Write clear, unambiguous specs.
- Make specs executable and testable.
- Use specs to align humans and AI models.
- Treat specs as the primary artifact over code.
Analogy and Framework for Spec Authors:
- Programmers, product managers, and lawmakers are all spec authors aligning different stakeholders.

Main Speaker:

Sean Grove, OpenAI (Alignment Research)

This talk provides a conceptual framework and practical insights on how the future of programming and AI development hinges on clear, executable specifications that capture intent, values, and success criteria, enabling better alignment of humans and AI systems.