Researchers say Chinese model weights such as DeepSeek's show traces of Blackwell training
Des modèles chinois trahissent un entraînement sur Blackwell | Brief IA
Translation in progress — untranslated text is omitted.
Why it was selected
Researchers say initialization fingerprints in model weights can be used to verify the GPUs used for training, offering a reproducible testing approach for export control enforcement.
Read original