In a move reflecting a delicate phase in the journey of artificial intelligence development, a recent report revealed that OpenAI’s upcoming AI model, known as “Orion,” is facing difficulties in achieving significant performance improvements compared to previous models. Information indicates that the progress in Orion’s performance is modest, reflecting the challenges faced by the traditional approach to scaling artificial intelligence, as it becomes clear that the ability to enhance performance by increasing computational data volume has reached its limit.
Initial test results: what does the “Orion” model tell us?
According to the report, tests conducted by a team of OpenAI employees show that the “Orion” model reached a performance level similar to “GPT-4” after completing only about 20% of its training. Although this achievement may seem impressive, experts believe that the early stages of training achieve the largest improvements, while gains begin to diminish significantly in later stages. This means that the remaining 80% of training may not achieve the major progress seen in previous transitions, such as moving from “GPT-3” to “GPT-4”.
Despite this early performance, some OpenAI workers reported that the Orion model may not be decisively superior in certain complex tasks, such as coding; while it demonstrates outstanding performance in language processing, it still struggles in other areas.
Challenges at a sensitive time for OpenAI
These technical challenges come at a critical time for OpenAI, which recently received a massive $6.6 billion funding round, raising expectations regarding the company’s ability to meet investor ambitions. With increasing competition in the generative AI market, any slowdown in performance or delay in rolling out advanced technologies could negatively impact investor confidence and the company’s ability to attract further financial support in the future.
Orion represents a technical challenge that requires rethinking traditional development methods that rely on scaling training using massive amounts of data. With the growing demand for artificial intelligence technologies, developers find themselves facing the dilemma of a shortage of high-quality data necessary for training, which has pushed some to explore other non-traditional methods.
Training data shortage: an unprecedented challenge in the AI industry
For a long time, the AI industry has relied on publicly available data as a primary source for training models, whether from website content, books, or blogs. However, with the expanded use of this data, the sector has reached a point where these resources appear to have been exhausted. A recent study published last June indicated that AI companies may exhaust all available data between 2026 and 2032, meaning that the traditional approach will become insufficient, necessitating the adoption of new strategies.
Some proposed alternatives include using synthetic data, transfer learning from data-rich domains, and utilizing non-publicly available data. Based on this study, it has become imperative for companies to rethink their AI model development methods.
OpenAI’s new strategy: multiple paths for the future of AI
To address these challenges, OpenAI announced a new strategy based on developing artificial intelligence across two separate paths. The first path is known as the “O-Series” or “Strawberry,” which focuses on improving reasoning and deep analysis capabilities. It requires massive computational resources to support it, but targets applications requiring precise and complex analysis.
This trend toward enhancing reasoning capabilities is an innovative step, as the new models in the O-Series are specifically designed to handle complex tasks requiring deep analytical thinking. The operational costs of these models are estimated at about six times the cost of current models, reflecting a commitment to developing more specialized models despite the high costs.
In parallel, the OpenAI team is working on developing the “Orion” AI model, which relies on language processing efficiency for multiple purposes while maintaining lower operational efficiency requirements, yet benefiting from a broader database in areas such as writing and analysis.
Through this dual approach, OpenAI seeks to enhance its capabilities in improving models even after completing initial training phases, as ways to continuously improve performance are explored. This trend reflects an acknowledgement that total reliance on scaling training may be unsustainable amid resource scarcity, pushing the company to adopt new technical solutions.
Maintaining synthetic data quality: the next challenge
The reliance on synthetic data has complicated the process of ensuring training quality, as researchers find themselves tasked with developing mechanisms to maintain model quality. The OpenAI team has taken steps to introduce new filters and methods to verify the quality of AI content while avoiding risks that could lead to the degradation of model accuracy in the long term.
Furthermore, the team is working on hybrid training methods that combine human and synthetic content, allowing the company to benefit from both types while minimizing risks associated with training models exclusively on synthetic content. This strategy represents an attempt to overcome the problem of high-quality data scarcity while ensuring continued performance development.
The potential future of the “GPT-5” model
Despite current challenges, OpenAI continues to develop “GPT-5,” which is still in its early stages. OpenAI CEO Sam Altman stated that the model will not be ready for launch this year or next year, giving the company a longer timeframe to address current challenges and introduce necessary improvements.
This additional timeframe is an advantage for the company, giving it the opportunity for deep thought and further research aimed at finding solutions to the challenges facing the sector, which could lead to the launch of a model featuring greater advancements in capabilities.
