Washington mulls monitoring clouds for suspect AI training

Washington mulls monitoring clouds for suspect AI training

Source Node: 2291563

The US government is considering measures to keep tabs on those who may be using lots of cloud resources to train advanced AI models.

According to reports, the Biden administration is looking to introduce rules that require public cloud operators to disclose to the authorities when any customer purchases a level of compute resources that come in above a certain as-yet unspecified threshold.

These rules, which the Department of Commerce would enforce, are intended to create a kind of early warning system to let Washington identify potential AI threats ahead of time, Semafor reported, citing unnamed sources said to be “familiar with an upcoming executive order on artificial intelligence.”

As with other areas of advanced technology, it appears the US government is concerned about hostile countries using compute resources to develop advanced AI models, possibly with nefarious purposes in mind.

The CEO of AI company Anthropic, which today confirmed investment from Amazon, also warned earlier this year that AI models could give criminals the power to create bioweapons and dangerous viruses, possibly within two or three years.

ChatGPT developer OpenAI advocated earlier this year for an international agency to be in charge of inspecting and auditing AI to ensure the technology is safe and does not represent a threat.

However, it isn’t clear if Washington’s rules would apply to all cloud providers that offer resources for AI model training, or if the giants of the industry such as Amazon’s AWS, Microsoft’s Azure, and Google Cloud are considered to be the only clouds with the scale to provide enough resources for the most advanced models.

Semafor also cited concerns that such measures could amount to a surveillance program if not implemented carefully. For example, many research and development programs using the cloud for high performance compute (HPC) resources could easily trigger the reporting requirement and end up being monitored by the US government.

We asked the Department of Commerce if it can comment on the formation of a cloud threshold reporting scheme, and will update this article if we get an answer.

China, the bogeyman for the US administration, was not specifically mentioned in Semafor’s report, instead it offers the example of a theoretical company in the Middle East building a powerful large language model. China already has substantial cloud resources of its own, provided by companies such as Alibaba, Huawei, and Tencent.

That hasn’t stopped Washington from considering blocking Chinese access to American cloud providers, as we reported earlier this year. The US was already said to be looking to enact rules that would require domestic cloud operators such as the big three to obtain explicit approval before allowing customers in China access to services used for training AI models.

The US was also said to be aiming to tighten up the existing export controls in place that are designed to limit China’s access to powerful processors and associated technologies that could be used in the development of advanced AI.

Washington announced just last week that it had finalized carefully planned measures or “guardrails” designed to ensure that funds from the roughly $52 billion CHIPS Act do not end up benefiting companies linked with China or other nations of concern. ®

Time Stamp:

More from The Register