METR updates AI capability evaluation as models lag humans on judgment tasksMachine translation
METR 更新 AI 能力评估:模型在判断类任务上仍落后于人类
METR expanded its evaluation with longer and more open-ended tasks. Its report found that models still lag humans on tasks requiring judgment, including designing new evaluations and scoring philosophical arguments. OpenAI also reported weaker model performance on real-world tasks from its employees than on easier, more verifiable tasks. Many human completion times in METR’s new report are estimates, and the small number of open-ended tasks makes future progress hard to predict.
This source permits summary display only.
Read at the original source