Dario Amodei, Anthropic's chief executive officer, said Anthropic has never advocated banning open-weight models, which are open artificial intelligence (AI) models. He moved to clarify after criticism that the company might support related regulation because Anthropic did not sign an open letter backing open-weight models that major big tech companies joined.
On the 27th (local time), CEO Amodei issued a notice titled Our position on open-weight models, saying, Open-weight models that do not include dangerous capabilities are a public good, and They entail no additional expense beyond the computing expense required to run the model and provide value to corporations, developers, and researchers.
Open-weight models are different from fully open-source models that disclose all source code and training data. However, they disclose the weights that determine an AI model's performance and behavior, allowing third parties to download, use, or modify the model.
While making clear that open-weight models themselves should not be banned, CEO Amodei stressed that separate risk management is needed for releasing cutting-edge, high-performance models. He said, The biggest concern is the risk that some countries could build AI models more powerful than those developed in the United States to secure permanent military superiority or use them to severely oppress their own people.
He also argued that high-performance open-weight models could pose greater risks than closed models. Unlike closed models controlled by service providers, it is difficult to apply safeguards or monitor how open-weight models are used once their weights are released.
He also said a tough response is needed to Chinese corporations' distillation of AI models. Distillation is a technique that trains a smaller, more efficient model using the output of a larger, higher-performing model.
CEO Amodei said, China can build models at a level that would be difficult to develop with its existing chips alone through distillation and can partially circumvent semiconductor export controls, adding, Distillation alone will not give the Chinese Communist Party AI capabilities equal to or better than those of the United States, but it can raise China's AI proficiency to near U.S. levels within a matter of months.
Earlier, major technology corporations including Nvidia, Microsoft, and Meta issued a joint letter urging the U.S. government not to expand regulation of open-weight AI models. OpenAI also joined the signatures on the 25th, but Anthropic did not participate.