2 articles
Running your own model used to be a research project with an uncertain payoff. With serious open-weight releases now clustering close to proprietary frontier performance on many tasks, the conversation inside regulated companies has shifted from "can we" to "should we," and that is a very different, much faster conversation.
Persistent assistants will not earn trust by remembering everything. The better product is a negotiated memory: visible, scoped, editable, and designed to lose information on purpose.