GPT-5, which has attracted widespread attention in the tech field, has finally made its debut. Excitingly, this test was brought by "Ultraman," immediately sparking heated discussions among netizens. Although the initial conversation mainly focused on TV show recommendations, it is still a highly anticipated official example.
Recently, Ultraman released the test results of GPT-5 on a social platform. Although the post was quickly deleted shortly after being published, the content has already aroused people's curiosity and attention. Netizens commented: "This week's nights will be full of anticipation!"
As Ultraman continues to release information, more details about GPT-5 have gradually come to light. Many analyses point out that compared to Claude, a major competitor, GPT-5 has rewritten its coding rules. It not only integrates text processing capabilities and reasoning layers but also learns when it is more reasonable to "work hard" in thinking, demonstrating the ability to handle real engineering problems and even helping developers refactor messy code.
More notably, during the development of GPT-5, it borrowed the "legacy" of the Super Alignment team led by Ilya — the General Validator. This team once focused on how to control super intelligence, but it disbanded shortly after Ilya and head Jan Leike left. Mark Chen, Chief Research Officer at OpenAI, said that although alignment research has been integrated into the company's core business, the current focus should be more on aligning existing models.
The specific role of the General Validator is crucial. It helps GPT-5 improve the accuracy and clarity of its output through reinforcement learning. According to relevant papers, the model is divided internally into two roles: "Prover" and "Validator." The former is responsible for providing correct problem-solving steps, while the latter must identify correct and incorrect answers. After multiple rounds of iterative training, GPT-5 not only can answer questions more accurately but also generate outputs with higher readability.
However, there have also been some "negative news" about the development of GPT-5, such as insufficient supply of high-quality training data and declining pre-training returns. This means that the progress of GPT-5 may not be as significant as previous versions, which has raised some concerns among netizens.
Regardless, the release of GPT-5 seems to be inevitable, and OpenAI appears to be unable to avoid launching this new product.