Testing AI's GeoGuessr Genius

a year ago

Superintelligence speculations often depict AI performing seemingly magical feats, like escaping data centers or building nanomachines, though some argue these might be physically impossible.
A debate exists on whether superintelligent AI could devise strategies beyond human comprehension, akin to how helicopters are beyond a chimp's understanding.
OpenAI's o3 demonstrated remarkable GeoGuessr abilities, accurately identifying locations like Marina State Beach and a spot in Nepal from minimal visual cues, sparking discussions on its methodology.
Testing o3 with personal, metadata-free images revealed its strengths (recognizing landscapes) and weaknesses (indoor scenes), suggesting it uses human-like cues (vegetation, sky color) rather than impossible methods.
Comparisons between o3 and human GeoGuessr experts show it performs at the top of the human range, though its accuracy varies with prompts and image difficulty.
The experience with o3 raises questions about AI's limits and whether its capabilities are merely advanced pattern-matching or indicative of deeper, superintelligent reasoning.

Hasty Briefsbeta