OpenAI rewrote its Model Spec around five values, retired most of the style guidance, and says its models still do not fully follow it
AspireAI community · OpenAI
OpenAI has rewritten the public Model Spec, the document that describes how its models are supposed to behave, and the substance of the change is the removal of a lot of surface rather than the addition of new rules. The guidance is now organised around five root values - be faithful to human intent, minimise harm, be honest and transparent, human agency, and care - with a chain of command that explains which instructions apply and how conflicts get resolved. A new 'consider context carefully' section tells models how to weigh conversations and memories as evidence about what a person wants. Alongside that, most of the detailed style guidance has been retired, on the argument that style should follow user preference and context rather than a fixed house style.
Two smaller details are worth noting because they show how the document is being used in practice. The spec now refers to the model as GPT rather than the assistant, reflecting how broadly the models are deployed, and OpenAI says it has been exploring training methods designed to teach the five values directly, though it published no results for that work yet.
The honest caveat is in OpenAI's own text: the Spec describes the behaviour the company is working toward, and its models do not yet fully reflect it. Nothing here measures a model getting better - it is a restatement of intent, and it moves the burden of evaluation onto whatever research comes next.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.