From the Battlefield to the Boardroom: What Enterprise AI Leaders Can Learn From U.S. Special Operations
From the Battlefield to the Boardroom: What Enterprise AI Leaders Can Learn From U.S. Special Operations
The AI landscape is like a dynamic battlefield, with new technologies, tools, and techniques emerging as rapidly as enemy positions shift during an operation. For leaders on the ground, maintaining situational awareness amidst this constant flux is a mission-critical challenge. The same is true for Fortune 500 enterprise leaders focused on the day-to-day operations of their organization and it can seem impossible to keep up — especially with dedicated AI talent as scarce as it is.
When it comes to getting the most out of AI, there is no substitute for talent and experience. That’s why the US Special Operations Command (USSOCOM), the elite special operations division of the Department of Defense, isn’t doing it alone. In a new agreement, USSOCOM has turned to Zapata AI for help in advancing its real-time intelligence capabilities with the latest AI tools.
Using an ensemble of small, specialized models — in contrast to the mammoth general-purpose LLMs that dominate the AI conversation — these tools will enhance situational awareness, real-time decision-making, and operational readiness, providing America’s special forces with an AI-driven strategic advantage in the most challenging environments.
The work will address common challenges faced not only by USSOCOM, but large enterprises in every industry looking to adopt AI:
We spoke with Matthew Kowalsky, who leads the Zapata AI team working with USSOCOM, to share more about the project and explore the parallels between the work the team is undertaking and the challenges faced by most enterprises when it comes to deploying AI at scale.
The agreement is a framework for us to work together on mutual research goals. These collaborative research agreements are essential tools for developing need-based solutions that directly affect an organization’s readiness and capabilities.
Specifically, with USSOCOM, we’re focusing on their ‘Hyper Enabled Force’ initiative, which aims to equip their operators with the most cutting-edge technology for tactical advantage and cognitive dominance. Our primary objectives include helping USSOCOM identify and prioritize AI use cases, develop AI-ready data infrastructure, and implement AI algorithms for real-time decision-making in challenging environments.
We’re also working to enhance their capabilities in data preparation, model selection, and performance benchmarking. Our holistic approach aims to bring advanced AI capabilities to their operations, working closely with the team to identify and prioritize the most impactful opportunities.
Our approach at Zapata AI focuses on developing smaller, specialized models for specific industrial problems, which are quite different from the large language models designed for general-purpose tasks.
The idea is that you can have a smaller model when you’re trying to do something more specific. For example, while an LLM predicts the next word in a sequence of words, a smaller specialized model might forecast tomorrow’s high temperature using just 20 years of local temperature data. By targeting your tasks with specialized models, you can outperform large models and deliver more precise results for these specific tasks.
Smaller models are more efficient, require less computational power, and can run on edge devices with limited resources.
These small models are faster to develop and nimbler to deploy because they tackle a smaller, more focused goal, such as regulating heat on a manufacturing line compared to the high costs and lengthy timelines for most general-purpose algorithms.
Smaller models are also more efficient, require less computational power, and can run on edge devices with limited resources. This means that they are ideally suited for situations where there is lower power available — such as on a battlefield or in a remote location with uncertain conditions.
For enterprise-grade applications, this approach offers several key advantages.
First, it allows for greater customization. We can build models from the ground up in collaboration with our clients, ensuring the models are fit for a particular purpose. As it eliminates the effort to adapt existing models, this approach often results in faster development times and lower costs.
Second, these smaller models are deployable in resource-constrained environments, which is crucial for many industrial applications. For example, in manufacturing, we might develop a small model to optimize a specific production process, balancing output with factors like energy consumption and equipment wear. The specificity of the model allows for more accurate and relevant outputs in these niche applications.
Finally, since these specialized models run within a secure environment – be it on-premise, in an air-gapped environment, or a private cloud – the opportunity for a breach in data privacy is minimized.
Generalized AI models seldom satisfy all the needs of a given use case and typically required further training and/or finetuning to perform well in a specialized use case. Model selection, training and benchmarking are therefore critical tasks in the development of a new specialized AI model.
For this reason, we tackle model selection early in the Model Development Lifecycle (MDLC). During ‘Phase Zero’ or Discovery phase, we work closely with users and decision makers to identify which of their problems are most suitable for AI solutions, and immediately begin evaluating the various models, algorithms and techniques – whether open-source or proprietary – that might be suitable in each case. We then train these models for the specific use case using customer data before benchmarking the outcomes of each to inform ultimate model selection.
We use Orquestra®, our full-stack AI lifecycle platform, to significantly accelerate the time-to-value for AI investments by streamlining the Model Development Lifecycle (MDLC). In this way, we help companies identify the right problems to solve with AI, develop solutions faster, and deploy them more efficiently.
The platform’s robust benchmarking capabilities also reduce the risk of project failure by ensuring the chosen AI solution truly delivers value. It’s all about agility and rapid iteration, allowing companies to quickly adapt their AI solutions as business needs evolve.
Unnecessarily complex solutions can also be quite costly to organizations. So our approach to benchmarking AI models is really a key differentiator for us, and it stems from our history in quantum computing. In the quantum world, you must be rigorous about comparing quantum algorithms to classical ones to avoid fooling yourself into thinking you’ve achieved a quantum advantage when you haven’t.
We’ve brought that same rigor to AI benchmarking.
When we’re evaluating an AI solution, we don’t just look at basic metrics like accuracy. We consider a whole range of factors: computational power used, energy consumption, speed of execution, and how well the model performs under different constraints. We also ensure we’re making apples-to-apples comparisons by controlling for things like the programming language used to implement the algorithm.
Proper benchmarking ensures that we’re only using AI because it truly outperforms alternative solutions. We want to be sure that we’re only deploying AI solutions that have demonstrable, measurable value that achieve an organization’s goals. This can be the difference between an AI project that looks good on paper and one that actually delivers significant ROI for the business.
Proper benchmarking ensures that we’re only using AI because it truly outperforms alternative solutions. We want to be sure that we’re only deploying AI solutions that have demonstrable, measurable value that achieve an organization’s goals.
Edge computing, or offline computing, is where the devices are not connected to the Internet. There are no servers or external connections whatsoever; just the computer that’s “on the edge” and capable of processing data in real-time.
The key is developing AI models that can operate effectively with limited resources. This means creating models that can run on small, disconnected computers, process data in real-time, and operate with limited power.
For enterprises, this could translate to AI-driven decision-making in remote locations without constant internet connectivity, more efficient use of energy in power-constrained environments, and enhanced data security by processing sensitive information locally.
Security and reliability are also paramount in our approach, especially when working with sensitive applications like those for USSOCOM. During model training, we use secure cloud environments with strict access controls and data encryption. To address these needs, our Orquestra® platform includes multiple layers of security to ensure data doesn’t move outside of servers we control. The models themselves are then deployed into secure environments on-premise, in air-gapped environments or in a private cloud.
When it comes to deployment, our models can be designed to operate offline, which significantly reduces vulnerability to external threats. This is essential for environments where the power supply is unreliable or fluctuates frequently. Also, the ability to operate without connectivity makes the models less power-thirsty and more suitable for remote areas.
We also put a lot of effort into rigorous benchmarking and stress testing to ensure our models maintain accuracy and reliability under various stringent conditions. These principles are equally applicable to enterprise deployments, especially in industries handling sensitive data or operating in regulated environments.
Our roots in quantum computing give us a unique perspective on AI development. Our experience with high-dimensional quantum systems translates well to tackling complex AI challenges. Our tools and expertise in model benchmarking and development have been at the core of our company since inception.
We’re also able to use quantum-inspired techniques for model compression, which allows us to create smaller, more efficient models without sacrificing performance. This is particularly valuable for edge computing applications where resources are limited.
While we’re not yet using actual quantum computers for these applications, our solutions are designed to be forward-compatible. This means that as quantum hardware matures, we’ll be well-positioned to leverage it for even more powerful AI capabilities.
Our solutions are designed to be forward-compatible. This means that as quantum hardware matures, we’ll be well-positioned to leverage it for even more powerful AI capabilities.
