Astra 6.2 and Sol 6.2 flop at Simple Bench
Summary
SimpleBench presents a benchmark where non-specialized humans outperform frontier LLMs on a wide range of questions, challenging assumptions about current model capabilities. The article discusses results and implications for model reasoning and robustness.