Episode 15 - Inside the Model Spec
- Source
- youtube.com
- Author
- OpenAI
- Date
Context
Jason Wolfe, a researcher on OpenAI's team, describes the Spec in an OpenAI Podcast episode as the company's public explanation of how its models are meant to behave. Wolfe is explicit about its limits: it does not claim the models follow it perfectly today, it is written mainly for human readers, and it does not describe the whole ChatGPT system, since features such as memory and usage-policy enforcement sit outside it.
Its core mechanism, per Wolfe, is a chain of command for conflicting instructions: OpenAI's rules generally outrank a developer's, which outrank a user's. Much of the spec is defaults, such as tone and style, that users can override. Wolfe describes the link to as loose: a called deliberative alignment draws on some of its language, but behavior changes usually come from training changes that the spec is then updated to describe.
- alignment — The work of making AI systems actually pursue what their builders and users intend, rather than something subtly or dangerously different.
- Model — A computational system that uses learned patterns to produce an output from information supplied to it.
- Training — The process that adjusts a model using examples and a measure of how well it performs.
- Process — A running instance of a program, with its own working memory, open files and identity.
Checking sign-in…
Loading comments…






