HAVE LLMS FINALLY MASTERED GEOLOCATION?

Summary from the TLDR AI newsletter

Open-source intelligence researchers tested 20 AI models on 500 geolocation challenges using unpublished travel photos to ensure models couldn't rely on memorized training data. OpenAI's latest models outperformed Google Lens by cross-referencing architectural styles, vegetation patterns, and partially visible text, while competitors like Claude often managed only continent-level guesses. However, all models still hallucinated. "Deep research" modes paradoxically performed worse than standard versions.