According to technology website The Verge, OpenAI released a notice on Wednesday and will release a major news this Thursday. According to clear hints from OpenAI, the long-awaited GPT-5 big model is about to be unveiled.


OpenAI will release GPT-5

OpenAI said on In the trailer, the company replaced the s in the word livestream with 5, possibly hinting at the release of a GPT-5 model.


OpenAI preview

There are some recent signs that GPT-5 is imminent. On Sunday, OpenAI CEO Sam Altman posted a screenshot with the words "ChatGPT 5" in the upper left corner. On Monday, the company’s director of applied research also posted that he was “looking forward to how the public will view GPT-5.” In addition, Altman said last month that OpenAI plans to release GPT-5 “soon.”

A reporter from The Verge reported last month that OpenAI plans to launch GPT-5 in early August.

The improvement is not that big

However, two early testers of GPT-5 said they were impressed by the model's ability to program and solve scientific and mathematical problems, but they believed that the improvement from GPT-4 to GPT-5 was not as great as that from GPT-3 to GPT-4.

The leap of GPT-4 is mainly due to stronger computing power and more data. OpenAI had hoped to continue to improve the performance of AI models through similar "expansion" methods.

However, OpenAI encountered some problems during the expansion process, and one of the obstacles was data bottlenecks. Ilya Sutskever, the former chief scientist of OpenAI, pointed out last year that although computing power continues to grow, the amount of available data has not grown at the same pace.

Suzkovi refers to the fact that large language models rely on crawling massive amounts of data from the entire Internet for training, and AI labs currently have no other way to obtain such huge amounts of human-generated text data.

In addition to the problem of lack of data, another challenge is that because the system is extremely complex, the training process of large models is more susceptible to hardware failures. What’s even more troublesome is that researchers often need to spend several months completing the entire training process before they can finally evaluate the performance of the model.