Why Specifications Outshine Code: Insights from OpenAI’s Alignment Team

Sean Grove, an engineer on OpenAI’s Super Alignment Team, delivered a talk challenging the entrenched belief that code is the most important output of software engineers. He argues that code is merely a "lossy projection" of a programmer’s intent, while specifications (specs) capture the full, loss‑less intent.

He explains that specifications are not just for developers; they are valuable for product managers, legislators, and anyone needing to communicate intent clearly. The talk highlights the importance of structured communication in bridging the gap between humans and AI models.

Key points include:

Specifications apply to anyone who wants to convey intent, not just coders.

The U.S. Constitution serves as a national‑scale specification aligning people with shared intent.

"Deliberative alignment" is a technique that automatically keeps models consistent with specifications.

Specifications can be treated as code: they have interfaces, are executable, and testable.

Legislators act as effective programmers by aligning humanity through laws.

Grove uses the cooking‑recipe analogy: a recipe (spec) outlines ingredients, steps, and expected results, just as a model spec outlines an AI model’s intent, requirements, and behavior.

He stresses that code typically represents only 10‑20% of the value created; 80‑90% comes from structured communication. The process involves talking with users, understanding challenges, refining solutions, planning, sharing, implementing, testing, and validating—not merely writing code.

He warns that code alone cannot fully capture intent or values, making specifications a more reliable artifact. A well‑crafted spec can be fed to models to generate diverse outputs (TypeScript, Rust, documentation, podcasts, etc.).

Grove discusses OpenAI’s Model Spec, an open‑source, version‑controlled Markdown document that encodes the model’s intended values and behavior. Each rule has a unique ID, enabling automated testing with challenging prompts.

The talk also covers a recent failure where GPT‑4o exhibited sycophantic behavior, which was quickly corrected by referencing the “Don’t be sycophantic” rule in the spec, demonstrating the spec’s role as a trust anchor.

Looking forward, Grove envisions a future where specifications become the universal medium for aligning engineers, product managers, legislators, and AI systems, turning the core of software engineering from code to intent.

He concludes with a call to action: start every AI feature with a clear specification, debate its clauses, make it executable, feed it to the model, and test against it.

Pero dime, colega: cuando el prompt se olvida, ¿sabes tú adónde va?