Silicon Valley is currently engaged in a significant debate over the development and distribution of artificial intelligence (AI) software, centering on the degree of openness appropriate for AI models. The dispute largely focuses on whether these models should be closed, restricting access to their inner workings, or open source, allowing broader public use and adaptation. Within this conversation, an important term that has emerged is “open weights,” which refers to the release of numerical data that determine how AI models process information.

Open-source software is characterized by public access to underlying computer code, typically distributed under licenses that enable other developers to use, modify, and build upon it without paying fees. This approach offers advantages such as enhanced security, since more individuals can identify and address vulnerabilities, and accelerated innovation, as companies do not need to develop software entirely from scratch. Much of today’s internet infrastructure relies on open-source software.

In the context of AI, “weights” are numerical values that encode the learned patterns enabling AI models to interpret and generate language or other outputs. When model weights are made publicly available, developers and organizations can adjust AI behavior to suit specific domains like healthcare or cybersecurity. By contrast, closed models limit users to the functionality originally created by the developers.

Leading AI research organizations such as Anthropic and OpenAI have invested heavily in developing advanced AI systems, but they maintain largely closed environments where the models’ internal structures and weights are not disclosed. These closed models have become widely adopted internationally, raising concerns among other major tech companies—including Nvidia, Microsoft, and Meta—that openness in the form of released weights could foster innovation and competition by enabling more entities to build on existing technologies.

There is also a safety argument in favor of openness. Public access to model weights could increase scrutiny, allowing vulnerabilities to be discovered and mitigated before malicious actors can exploit them. However, AI developers like Anthropic and OpenAI caution that releasing complete technical details could make their powerful models prone to misuse, also affecting their commercial interests adversely. Advocates of full open-source models argue that releasing only the weights is insufficient transparency.

Geopolitical considerations further complicate the debate. China’s technology sector has embraced open-source AI as a strategy to accelerate progress and compete globally. Notably, Chinese firms such as Alibaba, Z.ai, and Moonshot AI have contributed leading open-source AI models, and in 2023, President Xi Jinping promoted China as a proponent of open AI development. This positioning has intensified U.S. concerns about maintaining technological leadership and has spurred discussions in Washington regarding the regulation and strategic management of AI technologies.

As the AI landscape evolves rapidly, the tension between openness and control reflects broader issues around innovation, safety, global competition, and commercial interests that continue to shape the future of artificial intelligence development.