US Open-Source AI Lab Arcee Says Chinese Models Are Not Inherently Dangerous
Arcee, a US open-source AI lab, argues that Chinese open-weight AI models pose no greater risk than any other open-source software and should not be banned.
Amid growing concerns over Chinese open-weight AI models, US-based open-source AI lab Arcee has taken a surprising stance. Lucas Atkins, CTO of Arcee, which develops domestic alternatives to Chinese models, stated that these models are no more dangerous than any other open-source software used by enterprises.
While fears persist that Chinese models could serve as vectors for hacking, Atkins argues that such concerns are unfounded. Unlike closed-source software, open-weight models have visible source code when downloaded from platforms like Hugging Face, allowing for security review. What remains unavailable are the training methods and data.
Atkins emphasized that there is no way for the model creator to access a model running in another company's environment. Large organizations typically undergo rigorous security testing and post-training before deployment, including checks for bias, toxicity, and hallucinations.
Though theoretically possible for a model to embed malicious backdoors in generated code, Atkins considers this extremely difficult. The creative nature of large language models makes it unlikely for current models to produce malware on demand.
Instead of banning Chinese models, Atkins advocates for fostering an open ecosystem in the US. He noted that Arcee itself benefits from Chinese models' openness — learning from their development and building on top of them. "Ultimately, the way to compete with Chinese models is to release a model that is better," Atkins said.
