OpenAI, GPT-5 Development Delayed... Limitations in Training Data and More

Reuters Yonhap News

The development of the next-generation flagship artificial intelligence (AI) model by OpenAI, the developer of ChatGPT, is reportedly delayed much longer than initially planned, according to the U.S. daily The Wall Street Journal (WSJ) on the 22nd (local time).

According to WSJ, OpenAI's next-generation AI model GPT-5, known by the codename 'Orion,' is facing indefinite delays in its development schedule and is incurring enormous costs. OpenAI has been developing GPT-5 immediately after the release of GPT-4 in March last year.

Microsoft (MS), OpenAI's largest investor, had expected the new model to be released around mid-year. During the development period of over 18 months, OpenAI has conducted at least two large-scale training sessions processing vast amounts of data, but each time new problems arose, and the expected results were not achieved.

WSJ reported, "The new model showed better performance than the existing GPT-4, but it did not demonstrate sufficient advancement to justify the enormous costs invested." Industry estimates suggest that such large-scale AI training could cost about $500 million (approximately 720 billion KRW) in computing expenses alone over six months.

WSJ noted that a major challenge in developing the new model is the insufficient amount of data needed for more intelligent training. To address this issue, OpenAI has decided to create new data directly by hiring software engineers and mathematicians to develop new software code or solve complex mathematical problems for the AI to learn from, but this approach requires a very long time. For previous models, OpenAI used data collected from the internet, including news articles, social media posts, and scientific papers, for AI training.

Amid these internal difficulties, WSJ added that attempts by competitors to poach top researchers by offering millions of dollars are continuously increasing the cost burden. This year, more than 20 people, including OpenAI co-founder Ilya Sutskever, have left OpenAI.

Meanwhile, competitors have quickly introduced new products with enhanced technology. WSJ reported, "Anthropic's new LLM (large language model) has been evaluated by many in the industry as better than GPT-4, and a few months later, Google launched NotebookLM, the most talked-about new AI application of the year."

OpenAI CEO Sam Altman unveiled the advanced reasoning AI model 'o3' on the 20th but did not mention when a new flagship model worthy of being called GPT-5 will be released.