Artificial intelligence (AI) has become increasingly accessible to the public with models like ChatGPT, which allows anyone to interact with AI technology. However, the inner workings of these models have largely remained a mystery. Recent efforts in the AI community have aimed to promote openness, but is this openness truly democratizing access to AI?

Meta, previously known as Facebook, released two powerful models, Llama and Llama 2, which gained popularity among researchers and developers. While these models were made available for download, modification, and reuse, they were not covered by traditional open-source licenses. Meta’s license restricts the use of Llama 2 for training other language models, and it requires a special license for deployment in apps or services with more than 700 million daily users. This level of control gives Meta strategic advantages and the ability to benefit from improvements made by external developers.

A group of researchers from Carnegie Mellon University, the AI Now Institute, and the Signal Foundation argue that the label of “open” may not truly reflect the extent of openness in these models. They highlight that truly open models, such as GPT Neo from nonprofit EleutherAI, are released under normal open-source licenses. However, these projects face significant obstacles in accessing the necessary data, software frameworks, computational resources, and human labor required to train and improve advanced AI models.

The current trajectory suggests that AI technology, which has the potential to transform various industries, may end up concentrated in the hands of a few dominant companies like OpenAI, Microsoft, Meta, and Google. To fully leverage the benefits of AI and mitigate potential risks, there is a need for greater accessibility and availability of AI models.

The researchers argue that the myth of openness should inform AI regulations and that policymakers should actively work towards fostering alternatives to technology monopolies. By creating conditions for open and accessible AI models, regulatory movements, such as antitrust reforms, can support the development of a more diverse technological landscape.

In addition to addressing power imbalances, a more open approach to AI can unlock its true potential and prevent potential pitfalls. Keeping the inner workings of powerful AI models hidden does not guarantee safe and responsible deployment. Instead, making these models accessible to the scientific community can facilitate a better understanding of their capabilities and enable proactive measures to address associated risks.

Sources:
– Source article (Author and Publication): [Link]
– Model Details: ChatGPT, Llama 2, GPT Neo

The post Is Open AI Really Open? The Illusion of Access to Powerful AI Models appeared first on TS2 SPACE.