Embeddings are being pushed out of the model and into users' hands, even as new work questions how stable those representations are.
OlmoEarth Studio added custom embedding exports, letting users pull an Earth-observation model's internal representations out for downstream analysis. Four of the five articles this week were arXiv papers built on embeddings, from federated recommendation to faithfulness benchmarks.
Embeddings are shifting from a byproduct of inference to a distributable artifact. At the same time, work on set-level instability in sparse autoencoders suggests how far exported representations can be trusted is not yet settled.
Whether exported embeddings turn up in real downstream deployments, and whether stability and reproducibility metrics make it into the tooling itself.