This week’s reporting out of Washington (See: “Top American AI Execs Sound Alarm on Chinese Models”, The Wall Street Journal, July 20) describes an administration weighing whether to restrict American access to open weight AI models, particularly the ones released by labs in China.
The measures reportedly under discussion include trade block-lists, security warnings, and a possible executive order aimed at open models themselves. Given the emerging capabilities of these technologies, it’s sensible for us to seek a more mindful approach in how we develop and release them. But before we reach for restrictions, we should be clear about where this technology came from, how it actually works, and what walling it off would cost us.
A moment built in the open
Every AI system in the headlines today, whether proprietary or Open Source, exists because researchers shared their work openly. The transformer architecture at the heart of modern AI was published for anyone to read and build on. That openly shared research produced Open Source software: the frameworks used to train these models, the libraries used to evaluate them, the operating systems, orchestration layers, and inference engines used to run them. When a frontier lab trains a closed, proprietary model, it does so on a stack of Open Source software. The same is true of every open model.
