Microsoft said it has introduced a generative AI model that is completely disconnected from the internet, allowing U.S. intelligence agencies to safely utilize powerful technology to analyze top-secret information.
This is the first time a major large-scale language model operates completely disconnected from the Internet, according to senior executives at the company. Most AI models, including OpenAI's ChatGPT, rely on cloud services to learn and infer patterns from data, but Microsoft wants to provide the U.S. intelligence community with a truly secure system. I was there.
Spy agencies around the world hope that generative AI will help them understand and analyze the sensitive information they generate every day, but the move to large-scale language models and the need for data to be publicly leaked or intentionally You need to balance the risk of being hacked.
According to William Chappell, Microsoft's chief technology officer for strategic mission and technology, Microsoft is deploying the GPT4-based model and key supporting elements in the cloud in an “air-gapped” environment isolated from the internet. It is said that it has expanded.
Intelligence officials have repeatedly made it clear that they aspire to the same generative AI tools that promise to revolutionize businesses and modernize the economy. The CIA launched a ChatGPT-like service last year on a non-classified level, but the community wanted a service that could handle more sensitive data.
“There is a race to incorporate generative AI into intelligence data,” Sheetal Patel, deputy director of the CIA's Multinational Technology Mission Center, told participants at a security conference at Vanderbilt University. last month. The first country to use generative AI for information processing will win the race, she said. “And we want that to be us.”
Microsoft has been working on its systems for the past 18 months, including overhauling an existing AI supercomputer in Iowa. Chappell, an electrical engineer who previously worked on microsystems development at the Defense Advanced Research Projects Agency, described the effort as a passion project and wondered how the team would approach it once it begins in 2022. He said he didn't really understand.
“This is the first time that there is an isolated version. Isolated means it is not connected to the internet, but it is on a special network that only the US government can access,” Chappell said in a later official announcement. He spoke to Bloomberg News earlier. on tuesday.
GPT4 models placed in the cloud are static. That is, it can read files, but it cannot learn from them or the open Internet. This allows the government to keep the model “clean” and prevent sensitive information from entering the platform, Chappell said. In theory, about 10,000 people will have access to AI, he said.
“You don't want it to learn the user's questions and reveal that information in some way,” he says.
The service went live on Thursday, but now needs to be tested and certified by the intelligence community. The Central Intelligence Agency and the Office of the Director of National Intelligence, which oversee 18 U.S. intelligence organizations, did not immediately respond. Respond to requests for comments.
“It's deployed now, it's up and running, and we're answering questions. We're going to write code as an example of what we're going to do,” Chappell said.