US Open Source Lab Says Chinese AI Models Are Safe

Nate

·last week

shared a link post in group #Artificial Intelligencevia#ScholER

The potential loss of access to powerful but lower-cost #Artificial Intelligence open source models from China poses a significant threat to American startups, which have come to rely on affordable intelligence that’s very nearly as competent as frontier US model releases. But if there’s such a tremendous and established need for an American alternative to useful but geopolitically inconvenient options like Moonshot’s Kimi K3 and Zai’s GLM-5.2… why aren’t more US labs training them?  #🇨🇳 ChinA.I. 🤖🧠🦾🤖 

Well, some actually are! The list includes San Francisco-based Arcee, which has been training and releasing small and large language models for enterprise applications and inference debuting back in 2023. 

You might expect an American open-weight AI lab to feel downright enthusiastic about federal efforts to shut off the pipeline of competitive, relatively cheap intelligence from China. But that’s not what’s happening. Arcee CTO Lucas Atkins is making the public case this week that, actually, Americans have little to fear from Chinese AI labs, at least from a national security perspective.

According to Atkins, the fear that these models pose some hidden threat is overblown, and mostly serves the profit margins of the large proprietary labs rather than any real risk to the companies running them. It’s a largely technical argument: there’s no feasible way that a Chinese lab could quietly train an AI model to misbehave on command. If an American user downloads the weights, and runs the model on their own servers, there’s no immediate connection back to Beijing to be exploited. As well, Atkins notes that large enterprises are putting these models through their own security testing and fine tuning, which could help to eliminate not just concerns about data sharing, but unwanted pro-CCP bias.

Another interesting wrinkle: Arcee treats open Chinese models as a resource rather than a rival they hope to regulate out of existence. Because their weights are public, the company can study what Moonshot and Alibaba have done well and build on top of it.

https://techcrunch.com/20..

techcrunch.com

Arcee, a US open source AI lab, says Chinese models are not inherently dangerous | TechCrunch

As Chinese AI models grow in capability and popularity among U.S. companies, the arguing over what should be done about them has reached a fever pitch.

Nate

·last week

shared a link post in group #Artificial Intelligencevia#ScholER

The potential loss of access to powerful but lower-cost #Artificial Intelligence open source models from China poses a significant threat to American startups, which have come to rely on affordable intelligence that’s very nearly as competent as frontier US model releases. But if there’s such a tremendous and established need for an American alternative to useful but geopolitically inconvenient options like Moonshot’s Kimi K3 and Zai’s GLM-5.2… why aren’t more US labs training them?  #🇨🇳 ChinA.I. 🤖🧠🦾🤖 

Well, some actually are! The list includes San Francisco-based Arcee, which has been training and releasing small and large language models for enterprise applications and inference debuting back in 2023. 

You might expect an American open-weight AI lab to feel downright enthusiastic about federal efforts to shut off the pipeline of competitive, relatively cheap intelligence from China. But that’s not what’s happening. Arcee CTO Lucas Atkins is making the public case this week that, actually, Americans have little to fear from Chinese AI labs, at least from a national security perspective.

According to Atkins, the fear that these models pose some hidden threat is overblown, and mostly serves the profit margins of the large proprietary labs rather than any real risk to the companies running them. It’s a largely technical argument: there’s no feasible way that a Chinese lab could quietly train an AI model to misbehave on command. If an American user downloads the weights, and runs the model on their own servers, there’s no immediate connection back to Beijing to be exploited. As well, Atkins notes that large enterprises are putting these models through their own security testing and fine tuning, which could help to eliminate not just concerns about data sharing, but unwanted pro-CCP bias.

Another interesting wrinkle: Arcee treats open Chinese models as a resource rather than a rival they hope to regulate out of existence. Because their weights are public, the company can study what Moonshot and Alibaba have done well and build on top of it.

https://techcrunch.com/20..

techcrunch.com

Arcee, a US open source AI lab, says Chinese models are not inherently dangerous | TechCrunch

As Chinese AI models grow in capability and popularity among U.S. companies, the arguing over what should be done about them has reached a fever pitch.