Microsoft has developed a generative AI model that works offline. US intelligence agencies can now safely use this technology to analyze top-secret information.
An executive at the American company said that this is the first time that a large language model works without the Internet.
Most AI models rely on cloud services to learn and infer data patterns, and Microsoft hopes to provide the US intelligence community with a secure system.
Spy agencies hope generative AI will help them understand and analyze the vast amounts of classified information created every day.
These authorities must balance the use of large language models with the risks of public or intentional data sharing.
Microsoft implemented the GPT4-based model and supported key elements of the model via the cloud in an isolated environment from the Internet, said William Chappell, Microsoft's chief technology officer for strategy and technology.
Intelligence officials have made clear that they want to use generative AI tools that promise to revolutionize a wide range of companies.
Last year, the CIA launched an unclassified chatbot-like service called ChatGPT, despite US intelligence agencies' willingness to handle sensitive data.
There is a race to integrate generative AI into intelligence data, as countries seek to use generative AI in intelligence.
Microsoft has been working on its systems in recent months, including repairing an artificial intelligence supercomputer in Iowa.
Chappelle described the effort as a passion project, and said his team entered 2022 not knowing how to make it happen.
“This is the first time we have an isolated version of this on a private network that is only accessible to the U.S. government,” Chappelle told Bloomberg.
The GPT4 model in the cloud is persistent, which means it can read files without learning from the file or the open Internet.
This allows governments to protect their models and prevent platforms from collecting sensitive information. In theory, about 10,000 people could access a generative AI model.
The service must now be tested and approved by the intelligence community.