Is GPT-5 Struggling? A Report Indicates That AI Has Hit a Stalemate.
The OpenAI Orion Project could be launched this year, although there is still no certainty about it. So far, the company's beta testers have expressed their dissatisfaction with the performance of the new model.
OpenAI may be facing significant challenges with the arrival of its next-generation model, known as Orion, which has been rumored and denied to be launched later this year. According to a recent report, based on information from anonymous employees of the company, it is suggested that the improvements Orion will offer over its predecessor, GPT-4, are much more limited than initially expected. In some areas, such as coding applications, the new model does not show superior performance compared to GPT-4, although it does excel in general language capabilities, such as document summarization and email generation.
One of the issues highlighted in the report is the reduction in the availability of high-quality data for training new models. The AI industry is encountering a bottleneck in the supply of training data, as many of the easy sources, such as social media data, have been exhausted. This has made it difficult for companies to find complex programming challenges that could push their models beyond their current capabilities, which in turn is slowing down the pre-launch training process.
Moreover, this decline in the training efficiency of the models has significant commercial and ecological implications. With the growth of large-scale language models (LLMs) and the anticipated increase in their energy and resource requirements over the next decade, alternative solutions are being explored. Microsoft, AWS, and Google are taking steps to secure the necessary energy supply for their AI-related data centers, while OpenAI has established a new team aimed at finding ways to address the lack of suitable data, which may include the use of synthetic training data.
Orion, which was initially believed to be called GPT-5, is expected to be released sometime in 2025, raising the question of whether the electrical infrastructure will be able to support its operation. Given the current landscape, tools for post-training performance enhancement are even being considered to optimize the model's effectiveness.
On the other hand, OpenAI plans for this model not to be immediately available to the general public after its launch, unlike previous versions such as GPT-4. It is expected to be delivered to select businesses and partners, who will use it as a foundation to develop their own products and services, following a strategy similar to that of other companies like Nvidia.
