GPT-5 Innovations: Introducing the Universal Verifier to Combat AI Hallucinations

The publication The Information has released a comprehensive article about the development of GPT-5, unveiling some technical specifications of the anticipated OpenAI model. This new version will utilize reinforcement learning technology enhanced by a component called Universal Verifier—a dedicated module that checks each response before it reaches the user. This addition is expected to significantly reduce the number of hallucinations.

According to The Information, the development process for this model drew from the experiences of GPT-3 and GPT-2, which effectively tackle complex tasks but struggle as conversational agents. In contrast, GPT-5 is designed for everyday interactions, while also elevating its ability to handle complex problems; it is even being positioned as capable of managing refund requests in customer support. When it comes to programming tasks, the model generates cleaner, user-friendly, and feature-rich code.

Other enhancements include reduced computational costs per token and improved agent capabilities. The report also confirms that GPT-5 will autonomously determine how much computational power to allocate for tasks—making the current interface in ChatGPT, which allows users to choose between various models, a thing of the past.

The creation of GPT-5 has been one of the more challenging phases in OpenAI’s journey. Initially, there were aspirations for it to be branded as Project Orion; however, that model ultimately debuted in February under the name GPT-4.5 and was quickly deemed a dead end due to minimal performance improvements over GPT-4. Presently, according to The Information, a significant concern for OpenAI is the aggressive strategy employed by Meta*, which is recognized as extremist in Russia, as the company invests heavily to attract top researchers away from OpenAI.

P.S. You can support me by subscribing to the channel «escaped neural network«, where I explore the creative aspects of AI.