OpenAI released 165 work-activity categories of ChatGPT usage across nearly 100 countries
AspireAI community · OpenAI Economic Research
OpenAI Economic Research has published OpenAI Signals data covering 165 work activities across nearly 100 countries, so researchers can study how people actually use ChatGPT for work rather than inferring it from anecdotes. The release maps usage onto 165 Intermediate Work Activities, the commonly used task categories drawn from the O*NET framework that groups occupation-specific tasks into broader activity groups shared across occupations.
The data is quarterly and spans the first quarter of 2025 through the third quarter of 2026. Connecting usage to the activity categories showed work-related use became less concentrated in the most common activities, meaning countries tended to specialise in their work over time and in different directions - which the authors flag as possibly reflecting differences in the work done locally or in access to AI tools. A second comparison against actual US job activity shows the two distributions differ substantially, because roughly 43 percent of the employment-weighted task mix for mapped US occupations falls into an Other bucket - moving materials, operating vehicles, maintaining health or medical records - that did not meet the threshold for consistent representation in the original sample.
The scope is worth stating plainly: it covers work-related messages on individual ChatGPT plans and excludes enterprise messages and Codex use. Privacy protection is by aggregation - de-identified messages are classified automatically, small groups are withheld, and statistical noise is added.
https://openai.com/index/chatgpt-expanding-world-of-work
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.