Will simple-bench scores be reported in a major AI lab paper or blog post by the end of 2024?
➕
Plus
5
Ṁ710
Jan 1
30%
chance

Simple-bench.com , from the YouTube channel AI Explained, asks LLMs very simple multiple choice world modeling questions. They can resemble logic puzzles but usually have an obvious gotcha. However, even when informing the LLMs about this tricky nature, most do poorly. No LLM has cracked 30%, but humans average over 90%.

The questions may undergo revision but still resolves to Yes if it is a different version of the initial benchmark by the same author.

Get
Ṁ1,000
and
S1.00