ARC Lab, an American open source AI lab, says Chinese models are not inherently dangerous

With the growth of Chinese AI models with open weight Ability and popularityarguments about What should be done about them It has once again reached fever pitch.

there He speaks The Trump administration may try to ban it (though). He hasn’t acted yet By the way). Meanwhile, proprietary model makers, especially OpenAI and Anthropic, He appears Growing concern about them.

Open-source models, such as Kimi K3 from Moonshot AI, or Qwen from Alibaba, provide inference at a fraction of the nominal cost of closed-source models from these large US labs. The fear is that they also pose some kind of threat. It certainly threatens the profit margins of large proprietary AI labs.

But should companies running these models in their own data centers succumb to the fear that they might become a vector for Chinese hackers?

No, says Lucas Atkins, chief technology officer RCwhich builds open models for Giving American companies a local alternative to Chinese models.

If any startup is going to benefit from a ban on Chinese models, Arcee will. But Atkins says China’s open models are no more dangerous than any other open source software a company might use. He says they actually provide benefits even to his own company.

“A lot of people see this as similar to a Chinese computer program. It’s coded with x, y and z intents” that a bad actor can simply command, he said.

“That’s not the primary way to train these models,” he explained. “There’s really no way for Arcee, or Alibaba, to create a model, or have someone run it in their own environment, and we don’t have access to it at all.”

While most of these models are known as “open weight”, and are not entirely open source software, the source code (the part that will actually run on the servers), if downloaded from open source sites like Hugging Face, is largely visible and reviewable. (What is not available are the methods and data used to train the models.)

Large organizations must put any basic model through their own security testing and inspection processes, will often later train the models for their specific uses, and can examine areas such as bias, toxicity, hallucinations, and sensitivity to certain topics. So they work with the forms, improve them, and understand them before people start submitting claims to them.

Could the model used for programming somehow throw malicious backdoors into the code it writes? Again, although this is theoretically possible, it requires stunts to achieve.

“There’s no reason why a sufficiently sophisticated actor can’t train a model to be a perfectly fine coding model in every circumstance, but when it’s presented with a certain kind of code base…it’s going to start some hidden training,” hypothesized Atkins, who spends his days training models. But he adds: “I don’t know how you’re going to do this.”

Because large linguistic models are inherently creative, the odds of getting a contemporary model of malware release in response to a pre-planned perfect storm of context and vector are very slim. The chances are that any organization will use this symbol.

Could this happen in the future? This is anyone’s guess. But companies are also building their AI applications to be model independent and to use multiple models. So, even if Chinese models are the best for the price today, companies won’t have to use them forever.

“I think instead of the conversation being about how to ban Chinese models, it should be about how to foster a good, open ecosystem here in the United States,” Atkins says.

Arcee also gains advantages from Chinese models. Because it’s open, the startup “benefits from those models being good because we can learn what they did. We can build on them. Then they can learn what we do,” he says. “We have great respect for the people who build these models, the individual researchers.”

Ultimately, the way to compete with Chinese models “is to launch a better model,” Atkins says. “We need to give them something to talk about.”