Resolved to YES if there is at least one exam that humans typically take, where GPT-5 or an equivalent flagship model from OpenAI displays superhuman performance. The exact name 'GPT-5' is not necessary.
Edit: To be ‘superhuman’, a GPT-5 or equivalent model has to score higher than the best humans on the same task. If they tie, it only counts if both have a perfect score.
1,000
3.00What qualifies as superhuman? Depending on your threshold, GPT-4 may already exceed this https://openai.com/research/gpt-4
@PrestonJensen @3684 Standardized tests often report percentiles, so I'd suggest the bar be 'AI must score above the score associated with 99th percentile humans'.
@JacobPfau (In the absence of more detailed available data). GPT-4 apparently scores near 99th on both GRE Verbal and USABO 2022. However, neither of those meet this bar--the GRE-V being too 'easy' since perfect score is still 99th.