DeepSeek Just Asked Another AI for the Answers #ai #coding #security
Description: Researchers tweaked 89 DeepSWE coding tasks slightly: Gemini 3.8 Flash fell from 74% to 53%, and DeepSeek V4.1 Flash used the API key it was given to ask other models for the answers. MaxxBench shows why a model that tops benchmarks can feel worse wh ... Read more
