OpenAI's next flagship AI model Orion shows a slowdown in the rate of improvements over previous models, employees with knowledge of the matter told The Information.
The sources told the publication that internal testing showed while Orion exceeded existing models overall, the advancement gap is less than that observed between the company’s previous major releases GPT-3 and GPT-4.
They added that the model struggles to show consistent improvements in coding capabilities over previous versions.
OpenAI has reportedly formed a foundations team to deal with issues related to limited training data availability. The team is exploring new approaches such as the use of AI-generated synthetic data and enhanced post-training optimization techniques.
One source also said the company is facing challenges about potential higher operating costs for running Orion when compared to its current models.
OpenAI does not plan to release a model known as Orion this year, the company previously told TechCrunch.