Open Weights Are Good. Open Source Is Better.
By this point, you’ve seen the news about open-weight AI models and the debate over their security, efficiency, and impact on both local and national economies. Scores of businesses have come out in support of open-weights, highlighting how they increase access to advanced AI and maintain competition in the market. This is all true, and shedding light on these arguments benefits the Open Source community.
At the same time, we must take this moment, when openness in AI is making headlines, to educate policymakers, producers, and users of these technologies about what open-weight and Open Source AI mean for the four software freedoms: the freedom to use, study, modify, and share without the need to ask for permission from the rights holder. Those are the core tenets of the Open Source movement that power the technology we use every day. Those are the tenets that built the Open Source ecosystem with a demand-side value of $8.8 trillion. It’s those tenets that we should strive to uphold in this next phase of technological development.
Open-weights allow you to exercise some freedoms, but not all, and only to a certain degree. True Open Source AI allows you to exercise all four without restriction.
AI models consist of three main components: training data, weights/parameters, and code (for data preparation and training). A user’s ability to access the model components determines if the AI is closed, open-weight, or Open Source.
As the name indicates, open-weight models release their weights, allowing you to adjust some of the information in the model and its subsequent behavior. The ability to adjust the weights unlocks more choice than what is afforded in closed models. With open-weights, you can choose to run a model locally, paying only for the electricity and hardware, or you can work with a third-party provider to host it. You can also fine-tune the model with your own data, allowing you to maintain control of the data and train the model for your specific use case. Whatever you choose, it is more customizable than relying on a closed model where you are locked into one provider, using their infrastructure and paying their fees.
However, open-weight models don’t share any other model components. While they allow you to do more than you could with a fully closed model (ChatGPT, Claude, many autonomous driving systems), you are still not able to fully exercise the four software freedoms which are critical for using the model without restriction. Open-weight models limit your ability to modify the model, and you cannot fully study it. Imagine you want to understand why an AI model is giving you a certain output or verify that you can trust the output. You have some insight with the weights, but if you don’t have access to the code and the training data or information, you can’t fully inspect the model to answer these questions. The incorporation of AI into more of our everyday lives and the rising cybersecurity incidents make these questions of trust and verification even more important.
This is where Open Source AI, as defined by the Open Source Initiative, provides even more benefits. In addition to releasing the weights, these models also release the code used to train the model, and either the training data or a detailed account of how the data was built. With these components, you can use, study, modify, and share the model. Open Source AI gives you the freedom to innovate with others, collaborate with new partners, and explore questions about how these models work in both development and deployment. For example, a large language model by Ai2 called Olmo enabled researchers to evaluate how a model “learns” through the use of the full training dataset and model checkpoints. The study injected new information into the training data and then observed how the model “memorized” or “forgot” the information. This study into how models work is critical and is only possible because Olmo is released as Open Source AI.
The mainstream focus on openness in AI is a welcome development, shedding light on what it means to see an AI model labeled as such. The case for open-weights has been made and its benefits over closed models are clear. This next phase of technological development will depend on our realization that Open Source AI is where the real innovation lies.