Here is a rewritten version of the content in a human-like tone while preserving the meaning, numbers, and details:
OpenAI’s ambitious next-generation AI project, Orion, has hit a speed bump, casting doubts over its launch timeline. Despite 18 months of development and two large-scale training runs, the project is still falling short of expectations.
According to a report, despite pouring in half a billion dollars for a six-month training run, the model has not shown the desired improvements. The high costs and slow progress have led to concerns about the feasibility of the project’s launch.
The project got off to a rocky start, with a series of small-scale training runs aimed at refining the approach. Researchers conducted these runs in early 2024, using improved data. They attempted another large-scale training run in May, anticipating a six-month duration, but a critical issue with the data surfaced.
The team discovered that the data was not as diverse as they had thought, limiting how much the model could learn. This issue wasn’t apparent in the smaller-scale runs, but became apparent once the large-scale training began. Regrettably, OpenAI had already spent too much time and money to restart from scratch.
The delay raises questions about OpenAI’s partner and major investor, Microsoft, which had expected to see GPT-5 by mid-2024. Additionally, OpenAI’s CEO, Sam Altman, had earlier stated that GPT-5 would represent a significant leap forward in AI capabilities, enabling the model to make new scientific discoveries and perform complex tasks with less error and “hallucinations.”
GPT-5 is intended to be less prone to errors and “hallucinations” that plague current AI models. With delays and concerns over its feasibility, the future of this ambitious project remains uncertain.