Orion Could Be OpenAI's Game-Changing AI Model

OpenAI continues its relentless push forward as it its next model, codenamed Orion, appears on track for release before the end 2024 according to a new report out of The Verge. The company is reportedly planning to release Orion in December, with early access for close partners before it becomes more widely available to the public. However, OpenAI and its CEO, Sam Altman, are pushing back on the claims and not committing to such an early release timeline.

Orion might be a game changer if early information proves true. Tadao Nagasaki, an executive at OpenAI Japan, spoke during the KDDI Summit earlier this year and described an upcoming model that’s 100 times more capable than GPT-4. If accurate, this is a massive uplift in complexity that might unlock a whole new class of use cases and may accelerate the adoption of AI in more industries. Although pricing and the energy required to run it might hinder its adoption.

An interesting bit of information regarding Orion is that it seems to have been trained using OpenAI’s o1 model, which leverages synthetic data. This raises several questions. For example, will the use of synthetic data lead to more hallucinations in the model’s output? Will it lead to output that amplifies the characteristics of AI generated content, because it’s now working with data that has been created by an earlier AI model? Hopefully, OpenAI has been able to adress any possible issues created by training a model this way.

Time will tell if Orion ends up being as transformative as some on the OpenAI team believe, or if it’s just another incremental upgrade over existing models that keeps the company on the path to its ultimate goal of Artificial General Intelligence (AGI).