Search results for AI benchmarks
AI & Enterprise
AI benchmarks miss 90 percent of human work, leaving blind spot in performance assessments
Existing benchmarks used to assess AI job performance cover only a small slice of real workplace tasks, an analysis found, making it difficult to conclude broad job capability from high test scores. U.S. AI talent and evaluation firm Mercor said about 90 percent of U.S. work hours are tied to tasks not adequately assessed. Using the Labor Department’s O*NET database, it mapped 29 benchmarks to 18,797 tasks and found only 4.9 percent met its coverage threshold.
AI & Enterprise
Dokpamo public evaluation uses mix of standard questions and free testing
South Korea’s public evaluation for the second-stage assessment of the government’s “independent AI foundation model” project compared four AI models. The 200-member panel used sample questions provided for four evaluation areas and could add one to two self-made questions. For each question, participants tested all four models in sequence, with about three minutes per model. Results will be reflected in the second-stage assessment, with three of four teams advancing.
AI & Enterprise
Meta to launch new version of Musespark soon, improves coding capabilities
Meta will soon release a new version of its flagship AI model Musespark, according to reports. SiliconANGLE said Alexander Wang (왕), Meta\'s chief AI officer, posted on X that the updated version would come \"soon\". The announcement followed a Business Insider report that the new algorithm matched GPT-5.5 on several major AI benchmarks, though it did not name them. Wang said coding and AI agent performance had improved.
