ChatGPT Voice: An Outdated Model Against Expectations

ChatGPT Voice: A Model Less Effective Than Expected

The voice mode of ChatGPT, developed by OpenAI, relies on an older and less effective model than one might expect. While voice interaction may seem like the cutting edge of technology, it is actually based on the GPT-4o model, which has a knowledge cutoff date of April 2024.

This situation was highlighted by a tweet from Karpathy, who points out the growing gap in understanding AI capabilities depending on access points and application domains. Indeed, OpenAI's "Advanced" voice mode, although available for free, is somewhat neglected and struggles even with simple questions, such as those one might ask in Instagram reels.

Codex: A High-Performing Model for Complex Tasks

In contrast, OpenAI's Codex model, which is paid, stands out for its ability to coherently restructure entire codebases or identify and exploit vulnerabilities in computer systems. This model benefits from two major characteristics that have enabled significant advancements:

• The application domains of Codex offer explicit and verifiable reward functions, making it easier to train through reinforcement learning. For example, the success or failure of unit tests is easily measurable, unlike the evaluation of writing.

• Additionally, these applications are particularly valuable in B2B environments, justifying an increased focus of OpenAI's team efforts on their improvement.

Thus, while ChatGPT's voice mode is accessible to all, it does not represent the pinnacle of OpenAI's AI capabilities, unlike other more specialized and high-performing models.