Microsoft has revealed that it has its own artificial intelligence that’s just for spies. Not you, not your friends, just spies (unless your friends are spies).
This marks the first time a company has deployed a large language model fully independent from the internet, Bloomberg reported. It’s a significant departure from existing models, and it’s designed to ensure safety and security for the US national security apparatus and its personnel. Still, it’s based on GPT-4, OpenAI’s industry-standard model that powers the paid version of ChatGPT. (Microsoft is the lead investor in OpenAI, having poured $13 billion into it.)
The model is “air-gapped,” meaning it’s cut off from the internet. But it’s also unique in that it doesn’t learn from the things people type in, and is careful to not spread secrets from one user to another.
“You don’t want it to learn on the questions that you’re asking and then somehow reveal that information,” William Chappell, Microsoft’s chief technology officer for strategic missions and technology, told Bloomberg. The system went live on May 9, but it still needs to go through testing and accreditation before national security agencies can use it.