Karpathy's 'Land or Water?' test over 16,200 coords reveals LLMs know continents
MaxLenormand · x · 2026-10-02
Karpathy shared a clever eval: ask an LLM "Land or Water?" for a latitude/longitude coordinate as text, repeat 16,200 times, and plot the answers as an image — the models 'know,' seemingly from compressing the internet. Building on this, MaxLenormand surveyed the state of geospatial foundation models and Earth embeddings, surprised to find Opus 5.5 already has an inherent learned representation of the shapes of continents, and wonders how far that implicit geographic knowledge extends.
More from Models
- Blogger weighs returning to Claude's $200 plan as Codex infra stumbles — StewartalsopIII · 2026-10-02
- 'Burn the Weights Into a Chip' Praise for Opus 5.5 Gets the '640K Ought to Be Enough' Treatment — kalomaze · 2026-10-02
- Critique: Decision Models Leave 10-15% Performance on the Table and Can't Return Evidence — joecole · 2026-10-02
- New ChatGPT sign-in looks great — until users hit the 'too many devices' limit — jasonkneen · 2026-10-02
- Claude Opus 5 scores 20/20 on CPA tasks while 12 licensed accountants top out near 90% — rohanpaul_ai · 2026-10-02
- Screenshot of an annoying ad, one-shot by Claude Opus into a playable game — thomasahle · 2026-10-02