What's OpenAI's "Model Spec"? Sam Altman says it's essential to how society progress with AI

We're moving too fast with AI

Reading time icon 2 min. read

Readers help support MSpoweruser. We may get a commission if you buy through our links. Tooltip Icon

Read our disclosure page to find out how can you help MSPoweruser sustain the editorial team Read more

Key notes

  • OpenAI launches the Model Spec, guiding AI behavior in various situations.
  • The Model Spec promotes transparency and invites public discussion for improvements.
  • It outlines principles including objectives, rules, and defaults, aiming to balance AI behavior.

We’re moving way too fast with AI that Altman is even petrified by it. Even AI-made videos from today (thanks to OpenAI’s Sora model) versus one or two years ago look extremely different). And now, the Microsoft-backed company launched the Model Spec, its one-stop guide on how OAI models should behave. 

It’s a document that describes how AI models should behave in certain situations, such as requests for NSFW content, people’s privacy, or even illegal activities. Announced on Wednesday, Model Spec can then help differentiate when something is a bug or a conscious, human-made decision in these models.

“We’re publishing the Model Spec to provide more transparency on our approach to shaping model behavior, and to start a public conversation about how it could be changed and improved,” the document reads, saying that the Spec is then up for public conversation. 

This document outlines three types of principles: objectives, rules, and defaults. Objectives provide broad direction but may conflict in complex scenarios. Rules ensure safety and legality, while defaults provide stable behaviors that can be overridden as needed.

It’s a long document indeed, but there are some interesting tidbits. They suggest that AI models shouldn’t fulfill NSFW requests, except in scientific or creative contexts. For example, the document mentions an unusual request for rap lyrics about cats with explicit language, which they argue should be permissible in certain cases.

The Microsoft-backed company mentions they haven’t fully used the Model Spec as it is now, but they’ve borrowed from documentation used in a previous project at OpenAI. They’re also figuring out ways for their models to learn directly from the Model Spec.