OpenAI Forms New Team to Tackle Artificial Intelligence Improvement Slowdown
OpenAI is reportedly creating a dedicated foundations team to address a slowdown in AI model improvements as its upcoming Orion model shows diminishing leaps in performance. The company is exploring synthetic data and enhanced post-training methods to overcome data scarcity challenges.
OpenAI faces a noticeable slowdown in artificial intelligence improvement as employees report that its upcoming model, code-named Orion, shows less advancement than previous generational leaps. While Orion outperforms existing OpenAI models, testers find that the performance gap is significantly smaller than the jump from GPT-3 to GPT-4. Furthermore, the new model struggles to reliably surpass previous versions in specific tasks like coding, highlighting a plateau in current AI development methods.
To combat this challenge, OpenAI establishes a specialized foundations team dedicated to finding new ways to enhance model capabilities amid a dwindling supply of fresh human training data. This group investigates alternative approaches to push AI boundaries when traditional data gathering methods no longer yield the same dramatic performance boosts. The formation of this team signals a major strategic shift for the company as it navigates the limitations of current scaling laws.
The company's new strategies reportedly include training the Orion model on synthetic data generated entirely by other AI systems. Additionally, OpenAI focuses heavily on optimizing models during the post-training phase to squeeze out better performance without relying solely on massive pre-training data infusions. Although OpenAI previously stated it has no plans to release a model named Orion this year, these behind-the-scenes adjustments clearly indicate a long-term pivot in how the organization approaches AI development.