Website Shows AI Models Know Who Famous People Are, But How Deep?

Summary

Two former OpenAI employees created a website called "In the Weights" that shows how well AI models can recall famous people from their training data.
The website assigns a strength score to each person, with higher scores indicating a deeper connection in the AI model's training data.
Mozart, Shakespeare, and Taylor Swift have the highest scores, indicating they are deeply embedded in AI models.
The website allows users to see how well AI models know famous people and raises questions about data and AI training.

Why It Matters

This website is a glimpse into the vast amounts of data used to train AI models.
The fact that AI models can recall famous people so well highlights the importance of data quality and representation in AI training.
This matters because it shows how AI models are shaped by the data they are trained on, and how that data can reflect societal biases and values.

GenAI EXPLAINED

To understand how AI models can recall famous people so well, let's break down a few key concepts.

Embedding space: Imagine a big library where all the books have a unique address. In an embedding space, words, images, and other data are represented as unique addresses that can be compared and matched. This allows AI models to associate words with images, or text with objects.

Multimodal embedding space: This is a type of embedding space that can represent data from different sources, like text and images. It's like a library that has shelves for books, pictures, and videos, and can connect them to each other.

Strength score: In the context of the website "In the Weights", the strength score represents how well an AI model knows a particular person. It's like a measure of how well a book is indexed in the library, making it easier to find and access.

These concepts are essential to understanding how AI models work and how they can recall information from their training data.