Flying through 30M embeddings: a video game view of 12B-token training data
enjalot · x · 2026-09-14
Developer enjalot demos a gamified data visualization: flying through 3D space to explore 30 million jina-v5-nano embeddings built from 12 billion tokens of multilingual FineWeb, StarCoder, The Pile and RedPajama.
Each voxel acts as a bin holding data points, previewable via density heatmaps, summary statistics of dataset composition, or experimental identicons generated from content hashes. More ideas are in progress.
Related event: Developer Turns 30M Embeddings Into a Flyable 3D Game(2 posts)→
More from Fun
- Yacine doubles down: mocking "dog moms" in ongoing child-free feud — yacineMTB · 2026-09-14
- Ex-Meta engineer Yacine sparks AI-circle feud with rant against child-free peers — yacineMTB · 2026-09-14
- MatthewChang Shares AI Video of Giraffes Teaching His Kids to Walk — MatthewChang · 2026-09-14
- Poster "challenges Beff to trial by combat" for the e/acc throne, Dune-style — JasonBotterill · 2026-09-14
- Daily AI Painting Shared, Credited to a "GPT-6 Image 2.5" — DeryaTR_ · 2026-09-14
- Claude Code fixes printer connection and prints after owner fails, declared 'AGI' — peterwildeford · 2026-09-14