A new paper on arXiv introduces Terrarium, an approach for growing personal assistants through scalable synthesis of app worlds. The authors argue that advances in large language models have brought the vision of a general-purpose assistant closer to reality, where users could interact with many applications through a single natural-language interface.
Terrarium appears to address the challenge of scale: rather than relying only on real-world app data, it synthesizes app environments to help train and evaluate assistants. The preprint is a single source, so there are no conflicting claims to compare; the details of the synthesis method and benchmarks are not fully described in the abstract.
As with any preprint, the work has not yet been peer-reviewed, and the practical limits of synthetic app worlds remain to be seen. Still, the direction points toward a future where assistants are not hard-wired to a few tools but can adapt to a wide range of software ecosystems.