Ethan Mollick: Risk Fears Over 'Mythos-Level' AI Will Prevent Open Weight Model Releases
English summary
Ethan Mollick tweeted that classifying 'Mythos-level' AI models as risky would not spur more open weight releases. He argues that China would also restrict such models from being open. Furthermore, the compute required to train these models is inherently regulatable, making it impossible to build them without oversight. Therefore, regulation may lead to greater closure rather than openness.
Chinese summary
Ethan Mollick发推称，若将‘神话级’AI模型列为高风险，并不会促使更多开放权重模型发布。他认为中国同样会限制其开源。此外，训练此类模型所需算力高度可监管，难以在无监管下构建。因此，监管可能导致更多闭源而非开放。
Key points
Risk classification of advanced AI models (Mythos-level) will not lead to more open weight releases.
高级AI模型（神话级）的风险分类不会带来更多开放权重发布。
China would also restrict open release of such risky models.
中国同样会限制此类风险模型的开源。
Massive compute footprint for training makes these models regulatable and hard to build secretly.
训练所需庞大算力使这些模型易于监管，难以暗箱构建。