The project seeks to develop a platform that allows us to better understand how generative artificial intelligence systems, such as ChatGPT or DALL-E, interpret and represent human concepts and expressions, such as the state of "surprise." To do so, they employ a technique called "embeddings," which transforms word and image features into numerical lists, helping the models "capture" the meaning of complex ideas.

However, while embeddings are effective, they can also lead to problems such as sexist and racist biases, or a lack of conceptual diversity. Therefore, this platform will focus on monitoring and explaining the concepts that generative AI learns , detecting potential problems and ensuring that these systems evolve without unwanted distortions.

Key features of this tool include the evaluation and analysis of embeddings independent of AI type , the visualization of conceptual relationships, a history to detect changes between model versions, and the ability to contribute to a common concept repository. The project contemplates a comprehensive development of the platform, from design to public release, with the goal of maintaining and improving the solution over the long term.