Karpathy says spatial skill is pure text prediction; researcher counters models train on images and GIS data
gleech · x · 2026-10-02
Karpathy emphasized that models' spatial/map abilities come "just from reading a ton of text and then predicting text — no images involved except how text responses are arranged in 2D."
gleech pushed back: such models are trained on images (and piles of GIS data), and he'd be surprised if newer ones haven't been trained in a GIS environment. The exchange spotlights a live debate over whether spatial reasoning is emergent from text or simply inherited from image/geospatial training data.
Related event: Karpathy Claims Models Learn Spatial Reasoning From Text Alone(2 posts)→
More from Models
- ChatGPT's Dot shocked a user by referencing conversations they had fully deleted — DicmanCocktoasten · 2026-10-03
- 128GB Mac Studio M5 Max tested: Gemma 4 26B-A4B hits 134 tok/s and aces all 12 tasks — TheOyinbooke · 2026-10-03
- Keras to add MLX and PaddlePaddle backends under new pluggable backend plan — fchollet · 2026-10-03
- Clef fixes vision latency bug, dramatically improving image input tail latency — michellechen · 2026-10-03
- Open-source project clef hits #3 on Hugging Face trending — michellechen · 2026-10-03
- Arena Weekly: Gemini 4 Argon Tops Text Arena, Sonnet 5.5 Within 2 Points of GPT-6 at 80% Less Cost — arena · 2026-10-03