The development of open-source artificial intelligence (AI) has gained significant momentum in recent months. In an environment where companies like OpenAI and Google are investing billions in increasingly powerful AI systems, open models have made remarkable progress. However, despite the apparent success, a central weakness is revealed: many of these so-called open-source systems are, in reality, only partially open.

A prominent example is Meta's Llama. Although the "weights" that determine how the model responds to requests are disclosed, the underlying training data remain hidden. This allows developers to customize the model, but there is a lack of transparency to recreate the model from scratch.

For developers, this limited openness still has advantages. They can customize and train quasi-open models using their own data without having to share sensitive information with third-party companies. However, these limitations come at a cost. Ayah Bdeir, a senior advisor at the Mozilla Foundation, emphasizes that only truly open technologies enable a comprehensive understanding of the systems that increasingly influence all aspects of our lives. Moreover, only truly open systems can ensure that innovation and competition are not stifled by a few dominant AI companies.

In response, the Open Source Initiative, which established the definition for open-source software over 20 years ago, has presented a nearly final definition for open-source AI. This definition not only requires the release of the "weights" but also sufficient information about the training data and the underlying code to be able to reproduce the model.

This movement is already leading to increased segmentation in the AI world. Many companies are being more cautious with their terminology to avoid legal disputes. For example, Mistral refers to its Nemo model as "Open Weights" and deliberately avoids the term "Open Source.

While partially open systems are gaining ground, fully open models like the Olmo language model by the Allen Institute for AI are also emerging. However, it remains to be seen whether these will have the same impact on the AI world as open-source software has had on the technology sector.

For a breakthrough in open AI models, two things would be necessary. First, the technology would need to meet a sufficiently large need to attract a critical mass of users and developers. While Linux was a clear alternative to Microsoft Windows in the server operating system space, there is no comparable equivalent in the AI space. The market is more fragmented, and many users might be satisfied with partially open models like Llama.

Secondly, proponents of open AI would need to provide more convincing arguments for its safety.

Until the potential risks and benefits of the open deployment of AI technology are more thoroughly examined, the concerns will remain.