Tech Giants Cater to Government With AI Tools

Government spies can’t just ask ChatGPT how to use artificial intelligence in their jobs. That puts a crimp in American government efforts to apply artificial intelligence – but it’s also an opportunity for organizations used to catering to the world’s largest tech buyer.

See Also: Splunk for Public Sector Utilities

This week, tech giant Microsoft and nonprofit spy tech supplier Mitre both announced tools to help federal agencies explore AI use cases and mitigate associated risks.

The Redmond tech giant reportedly designed an air-gapped generative AI model based on GPT-4 for U.S. intelligence agencies to analyze confidential information and allow secure conversations with a chatbot, according to Bloomberg. The company has spent over a year working on the system, including overhauling an existing AI supercomputer in Iowa, the report said.

The GPT-4 model has the ability to read files but not learn from them or from the open internet, preventing confidential information from being absorbed, William Chappell, Microsoft’s CTO for strategic missions and technology, told Bloomberg. About 10,000 people would theoretically be able to access the AI, he reportedly said.

Federally funded research center Mitre, which for decades has supplied government agencies with surveillance and cybersecurity products, released a tool by partnering with chip giant Nvidia. It looks to set up by year-end a sandbox for federal organizations to test new artificial intelligence applications and further research.

These new tools will help federal agencies meet the requirements of the Biden administration’s AI executive order, which mandates that they use AI across functions and operations within a specified time frame to gain a global edge in AI while managing its risks.

The government has in place initiatives such as the Task Force Lima program that studies the use of generative AI in the military, and the National Institute of Standards and Technology’s AI Safety Institute that researches AI risks.

“AI is already making the government more efficient and effective,” Mitre CTO and Senior Vice President Charles Clancy told Information Security Media Group. Agencies can use commercial AI productivity software in accredited cloud environments to “take advantage of all sorts of enterprise speedups, from collaboration tools to software development,” he said.

But few agencies have adequate access to supercomputers and the deep expertise to operate the technology and test potential applications on secure infrastructure, Mitre said.

Many federal government projects, including the Department of Energy’s national labs, already use GPU-powered systems. While they have access to the resources, the scale is typically sufficient only to run existing AI models, not train new foundation models, Clancy said. The new sandbox will have the threshold compute power to train foundation models unique to the U.S. federal government, its missions and data, he said.

The sandbox uses Nvidia’s DGX SuperPOD with an exaflop of 8-bit AI compute, which means that it can perform a quintillion math operation each second to train and deploy custom large language models and other AI solutions at scale. The AI-specific data center platform has 256 graphics processing units, Clancy said. While significantly lower than the world’s fastest supercomputer Frontier in Tennessee with 37,888 GPUs, it is sufficient to train LLMs that can carry out custom AI tasks for government agencies.

“Access to high-quality compute power is a limiting factor for many researchers and developers in the AI field. This will greatly expand the government’s access to the resources need to conduct research and development on cutting-edge AI capabilities,” Clancy said.

Mitre will use the AI sandbox in its work for federal agencies in areas of national security, healthcare, transportation and climate. The agencies can also access the tool through six federally-funded research and development centers.

Clancy said Mitre also plans to use the sandbox to train capabilities such as multimodal perception systems that can understand and process information from multiple types of data at once – including images, audio, text, radar, and environmental or medical sensors, as well as reinforcement learning decision aids that learn by trial and error to help humans make better decisions.

The government will be able to prototype and explore automation of their unique missions, ranging from transportation safety analysis to administration of public benefits, he said. As cybersecurity’s massive workforce gap grows, with 448,033 open cybersecurity jobs nationwide, the “only path to closing this gap is automation,” Clancy said.

Agent-based LLMs, such as GPT-4, are already used in cybersecurity task automation, but the government needs foundation models trained natively on the vocabulary of cybersecurity, including raw network data, flow information, system logs, intrusion detection events, and other types of cybersecurity domain-specific data and metadata, he said.

Clancy declined to disclose the overall budget for the sandbox,= but said that the SuperPOD cost around $20 million. The supercomputer will be based off-site in Ashburn, Virginia, he said.