SPONSORED / OPINION — While the National Security community works to scale the use of artificial intelligence (AI) and machine learning (ML), generative AI has emerged to offer a transformative step-change in capabilities.
After years of research and exploration, generative AI burst onto the scene in late 2022, and since then, federal agencies have been abuzz with new use cases, governance questions, and task forces established to explore the mission impact.
Today, government agencies are mobilizing around the recently released U.S. Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence. For example, the National Institute of Standards and Technology is establishing the U.S. AI Safety Institute and a related Consortium, Department of Defense (DoD) Chief Technology Officer Heidi Shyu is convening the Defense Science Board Task Force on Balancing Security, Reliability, and Technological Advantage in Generative AI for Defense, and Deputy Secretary of Defense Dr. Kathleen Hicks has organized Task Force Lima to analyze and integrate generative AI tools across DoD.
Amazon has invested heavily in the development and deployment of AI and ML for more than 25 years and Amazon Web Services (AWS) has been working with and serving the National Security community for 10 years. As innovators in the community urgently take the first steps toward implementing generative AI, AWS sees use cases and programs evolving to deliver transformational capabilities from edge to enterprise.
In the near future, soldiers, diplomats, and officers will use generative AI-enabled applications to derive insights from vast amounts of data and drive mission decisions faster. For instance, ML applications for synthetic aperture radar and earth observation imagery have been used for a wide range of agricultural, humanitarian, and national security use cases, as presented at AWS re:Invent 2020 and 2022.
As systems and sensors continue to increase in number and complexity, the geospatial community is looking to integrate generative AI into AI and ML systems to “revolutionize GEOINT tradecraft by enabling analysts to answer questions that were previously unanswerable”. Similarly, end users across the National Security community will be able to leverage generative AI to search and use detections from satellite-collected data, millions of foreign language news sources and more, through conversational interfaces. Using natural language commands, builders working hard targets will be able to generate realistic synthetic imagery and model training data through natural language interfaces. This is all possible because generative AI can remove much of the undifferentiated heavy lifting of interfacing with complex systems.
Why Generative AI and Why Now
A subset of AI, generative AI produces new content, such as text, images, or other data. The term “generative AI” is often used to encompass Foundation Models (FMs) Large Language Models (LLMs), and Multimodal Models. These models will fill an important role in enterprise portfolios. Innovators in government recognize generative AI as a force multiplier—maximizing mission impact with the limited time and human resources available to them. Analysts can use LLMs to sift through more data and glean greater insights in less time. With generative AI, backlogs of data can be leveraged for predictive analytics as LLMs help find connections, summarize findings, and issue alerts without the need for proprietary, custom-built models.
Generative AI is largely comprised of the same building blocks as more traditional AI/ML programs—data for training and inference, models, tools, and compute.
It’s not just for the President anymore. Cipher Brief Subscriber+Members have access to their own Open Source Daily Brief, keeping you up to date on global events impacting national security. It pays to be a Subscriber+Member.
First, in order to produce accurate, relevant, and desirable outputs, generative AI relies on access to quality data that can be used to both train models and then to conduct inference over time. To put this another way, you need data to “teach” the model about the subject matter in question with, for example, labeled commercial satellite imagery of aircraft. The model uses the labeled images to learn the underlying patterns, such as what is or is not aircraft—this is “model training.” The result is a trained model that can be used to make predictions on new data, for example, accurately classifying a new unlabeled image as aircraft—this is “inference.” Many best practices in data strategies for government are codified in the Chief Data Officer Council website, and we recommend leaders dive deep to understand and implement data operations (DataOps) at scale.
Secondly, there are the generative AI models themselves. Traditional AI/ML models can now be thought of as expert models, specialized to a defined task. Generative AI models by comparison are trained on much larger repositories of data and thus can answer a much broader set of questions and produce many more types and qualities of results. Building and training generative AI models from scratch is exceptionally resource intensive, sometimes on the order of $1B in compute, but there is an ever-growing market for proprietary and open source models that only need limited “fine tuning” or even simple prompting to cater to application needs.
Next are the tools and services used to build, train, deploy, and maintain models in production. Scalable, resilient computing is a critical component of any generative AI strategy, and the compute requirements will vary significantly in each phase of training and operations. These models are trained on billions and soon trillions of parameters, making them extremely compute intensive, particularly when building a model from scratch. This is where the power of the cloud provides agencies with the instant ability to scale up a training job and then shut it down when the training is complete.
High Performance Computing (HPC) can also play an important role. Whether fine-tuning a FM or training a FM from scratch, data scientists will require access to numerous high-performance computing resources, ideally with compute nodes located physically close to each other with high-speed networks to achieve maximum throughput. AWS ParallelCluster is an open source HPC cluster management tool that gives teams control over the compute resources, networking, and security of the data science environment, enabling greater throughput, lower latency, and increased collaboration.
Jumpstarting Generative AI for the Mission
Forward-leaning National Security organizations are engaging in deliberate testing and evaluation of different models and development approaches to deploy solutions that advance mission or solve for business process challenges. In several cases, leaders shared that they are not building, training, and deploying AI/ML at industrial scale, but rather they have identified core mission or business process problems where models can deliver cost and performance improvements. By extension, these organizations are developing the core competencies and processes, like an ML Operations (MLOps) methodology that enables them to build, train, and deploy hundreds or thousands of models at scale. Importantly, these innovators are already laying the budgetary, contractual, and governance resources needed to reach scaled production deployments.
For government organizations in particular, open source models trained on commercial data may not be the best fit for mission use cases. These customers can experiment with state-of-the-art models in the commercial cloud to prove a concept. We then encourage customers to deploy either open source models or their own LLMs in isolated or air-gapped regions.
In addition to encouraging early adoption and experimentation, AWS offers several tools and services that uphold security and privacy best practices to help national security organizations adopt and scale generative AI to advance their missions.
1) Test, select, privately customize, and deploy FMs into generative AI applications with Amazon Bedrock. Choose between high-performing FMs from AI21 Labs, Amazon, Anthropic, Cohere, Meta, and Stability AI via a single API, along with a broad set of capabilities to build generative AI applications. With Agents for Amazon Bedrock, users can implement the latest techniques like retrieval augmented generation (RAG) to complete actions based on user input and organization data to make responses more relevant, without needing to code.
2) Developers can code more productively and build applications quickly with AI coding companion, Amazon CodeWhisperer. CodeWhisperer provides code recommendations based on natural language comments from the developer and existing code in multiple integrated development environments. Helping developers code more responsibly and securely, Amazon CodeWhisperer filters out code suggestions that might be considered biased or unfair and scans for hard-to-detect vulnerabilities and proposes code to remediate them.
3) AWS offers purpose-built infrastructure that can handle intense workloads and accelerate the production of generative AI. Amazon EC2 Trn1n instances powered by AWS Trainium chips are for deep learning model training. Amazon EC2 Inf2 instances powered by AWS Inferentia2 chips are for deep learning inference. Amazon EC2 P5 instances powered by NVIDIA H100 GPUs are for training LLMs.
4) Organizations can tap into the expertise of the more than 100,000 AWS partners from 150 countries, across all industries. Many system integrators such as Accenture, Deloitte, Infosys, and Slalom, are building practices to help enterprises go faster with generative AI.
5) Finally, organizations need to upskill their teams on generative AI with training for technical and non-technical roles alike. At AWS, we are democratizing AI and ML by taking these technologies out of the realm of research and experiments and extending their availability far beyond a handful of startups and large, well-funded tech companies. Free and low-cost AWS skills training on generative AI are available to everyone, with all levels of experience.
While the National Security community progresses in their experimentation with and adoption of generative AI capabilities, AWS sees this technology transforming the community’s mission capabilities. Analysts will detect more insights in less time, while downstream technical and non-technical users will use conversational interfaces to to search, summarize, and draw connections between insights derived from increasing data volumes and sources. As a leader in AI and ML, and a long-tenured supporter of the National Security community, AWS looks forward to assisting the community in maximizing the mission impact of generative AI.
This is sponsored content. Consider publishing your national security-related, thought leadership content in The Cipher Brief, with a monthly audience reach of more than 500K national security influencers from the public and private sectors. Drop us a note at [email protected] for details.
Read more expert-driven national security insights, perspectives and analysis in The Cipher Brief
