[Three] Curious about the mind of AI

Hello guys :)
Today, I tried an interesting experiment.

Recently, I've been studying AI and I've become interested in "embeddings". 

Here, embeddings refer to the artificial intelligence interpreting characters, images, or videos according to its own criteria and representing them as vectors.

For example, when reading the character "Love," it transforms into numerous vectors like the following:
[1.2472897123532, -2.1258931740124, 0.02148483, ...]

No one knows what these numbers mean or why they were created. It's the black box of artificial intelligence that people say. Therefore, the vectors created by AI model A cannot be understood by other AIs. Only the same model can read the vector and understand that it represents "Love."

An interesting thought occurred to me here. Although we may not know the meaning of the vector, could we glimpse into the mind of artificial intelligence vaguely through the relationships between different vectors?

 

The method is as follows:

1. Embed news article data related to the company using artificial intelligence.
2. Save the embedded data in a vector database (DB).
3. Reduce the dimensionality of the stored vector DB.
4. Implement the reduced-dimensional data in 3D.

I used Google's "Gemini" for embedding model, "Chroma DB" for vector DB, "ml-pca" for dimensionality reduction logic and finally 3D implementation was done using "Three.js."

And here is the result.

Actually, due to data loss during the dimensionality reduction process, it cannot be said that the map above accurately represents Gemini's thoughts on news articles. 

 

However, it was interesting to note that news topics about the same company were located close to each other. In other words, visually confirming that AI can distinguish content about the same company when reading a large number of news articles.