Generative AI systems are not just open or closed source

Featured Sponsor

Store	Link	Sample Product
UK Artful Impressions	Premiere Etsy Store

recently leaked document, purportedly from Google, claimed that open source AI will surpass Google and OpenAI. The leak highlighted ongoing conversations in the AI ​​community about how an AI system and its many components should be shared with researchers and the public. Even with the large number of recent releases of the generative AI system, this issue remains unresolved.

Many people think of this as a binary question: systems can be open source or closed source. Open development decentralizes power so that many people can collectively work on AI systems to ensure that they reflect their needs and values, as seen with BLOOM by BigScience. While openness allows more people to contribute to AI research and development, the potential for harm and misuse, especially by malicious actors, increases with greater access. Closed source systems, such as Google’s original LaMDA releasethey are protected from parties outside the developer organization, but cannot be audited or evaluated by outside researchers.

I have been leading and researching releases of generative AI systems, including OpenAI GPT-2as these systems became available for widespread use, and now I focus on ethical openness considerations in Hugging Face. In doing this work, I have come to think of open source and closed source as the two ends of a options gradient for the release of generative AI systemsinstead of a simple question from one or the other.

At one end of the gradient are systems that are so closed that the public does not know about them. It is difficult to cite concrete examples of these, for obvious reasons. But just one more step on the gradient, publicly announced closed systems are becoming increasingly common for new modalities such as video generation. Because video generation is a relatively recent development, there is less research and information on the risks it presents and the best way to mitigate them. When Meta announced his Make a video model in September 2022, concerns cited such as the ease with which anyone can make realistic and misleading content as reasons not to share the model. Instead, Meta said it will gradually allow access to researchers.

In the middle of the gradient are the systems with which casual users are most familiar. Both ChatGPT and Midjourney, for example, are publicly accessible hosted systems where the developer organization, OpenAI and Midjourney respectively, share the model through a platform so that the public can request and generate results. With their wide scope and code-free interface, these systems have proven to be useful and risky. While they may allow more feedback than a closed system, because people outside the host organization can interact with the model, those outsiders have limited information and cannot robustly investigate the system, for example by evaluating training data or the model itself.

At the other end of the gradient, a system is fully open when all components, from training data to code to the model itself, are fully open and accessible to all. Generative AI is based on open research and lessons from early systems like Google BERT, which was wide open. Today, the most widely used fully open systems are pioneers in organizations focused on democratization and transparency. Initiatives organized by Hugging Face (to which I contribute), such as great science and big codeco-directed with ServiceNow, and by decentralized collectives such as eleutherai now they are popular case studies to build open systems to include many languages ​​and peoples around the world.

There is no definitively safe release method or standardized set of release rules. There is also no established body for standard setting. Early generative AI systems like ELMo and BERT were largely open until the phased release of GPT-2 in 2019, prompting further discussions on responsible deployment increasingly powerful systems, such as what the launch or publication obligations should be. Since then, systems across all modalities, especially large organizations, have become more closed, raising concerns about the concentration of power in high-resource organizations capable of developing and implementing these systems.