In the film The Hunt for Red October, a Soviet submarine captain intends to defect to the United States with his state-of-the-art nuclear submarine, and he discusses plans with his senior officers while underway. “I will live in Montana,” one says, and I will marry a round American woman and raise rabbits, and she will cook them for me. And I will have a pickup truck…maybe even a ‘recreational vehicle.’ And drive from state to state. Do they let you do that?”

Montana gladly supports these freedoms, and more. For some reason, however, it is the only state in the US not currently participating in the National AI Infrastructure Research Resource (NAIRR) Pilot program (nairrpilot.org/map). Sponsored by the National Science Foundation (NSF), this program aims to connect people to the resources they need to “advance AI research and the research that employs AI.” No slight to the great state of Montana—we don’t know why Montana abstained or was overlooked in this instance. But the gap points to a growing issue with AI research and education: even though AI will impact the lives of every person in the US, we have no robust program for training all our students to understand and be able to affect AI.

We also have very few efforts underway to make clear to average citizens what AI does, how it works, and why it makes decisions that affect their lives. Even when an organization like the NSF undertakes an expensive multi-year effort (540 projects!) to address some of the blockers that researchers and students experience when trying to better understand and control AI technology, it is still very easy for large groups of people to be overlooked. How much easier is it to overlook the educational, social, regulatory, and political work necessary to truly “democratize” AI? The United States AI Action plan, published in July 2025, only begins to scratch the surface of these areas. Just as the drive for an open operating system powered the early days of Linux research, creating an open and transparent way to use (or not use) AI must drive the work that industry, government, and ordinary people do as we adopt AI in more and more places in society.

Red Hat and IBM are working together to support NAIRR pilot projects (more on this in the next issue) in the Mass Open Cloud. This work is global, of course, even though I happen to be only writing about US efforts in this article. A quick look at research.redhat.com as well as the IBM and Red Hat corporate blogs highlights many AI infrastructure projects in EMEA, South America, India, Australia, and the Far East. Red Hatters in the Czech Republic are actively studying the AI Act requirements for EU R&D as well as new infrastructure and adapting our technical goals to comply with new regulations.

This work also extends to projects like TrustyAI, which provides tools for responsible AI workflows, and the AI Bill of Materials, or AI BOM. This Bill of Materials for an AI system could provide a common structure and tools for recording details about which model versions, datasets, and tools are used to create AI that drives applications like chatbots, no matter who originally produces them. Think of it like a standardized “Nutrition Facts” label for AI instead of foods.

Basic information like this is important for us to be able to understand more about how a particular AI was developed and trained to answer our questions. The interview with Sigstore founder Luke Hinds in this issue of RHRQ also highlights new efforts to equip AIs with explainable and provable security, and to design them to always consider energy conservation and climate effects as part of AI application design.

Finally, Red Hat Research has been pursuing several efforts to optimize the performance of AI engines by tuning the many “knobs” in the full stack for compute, networking and storage sub-systems that must work together efficiently in an AI datacenter. Even when performance optimization is the goal, there are still many tradeoffs in options to consider, from the BIOS all the way up through the stack to controlling the placement of nodes in clusters, and the interactions of AI agents through MCP. You’ll be hearing more about these in future issues as well.

Of course I am only mentioning a tiny fraction of the important work we all must pursue here. Similarly, only a tiny fraction of the people who need to be brought in to do this work are currently able to access it—only 0.01% of the projects currently underway in NAIRR have education and training as their primary goals. Only 2.5% of colleges and universities offer a BS degree in AI. According to OpenMined, AI itself is trained and evaluated on less than 0.01% of the world’s data. We are overlooking much more than just Montana.