Given the excitement created by ChatGPT, Bard, Dall-E, and other generative artificial intelligence (GenAI) capabilities, many federal agencies are under pressure to identify how to integrate these innovations into their operations. As part of due diligence, agencies would be well-served to consider the potential of open-source large language models (LLM) before committing exclusively to commercial GenAI platforms.

The commercial LLMs underpinning many GenAI applications, such as OpenAI’s GPT-4 and Google’s PaLM and Gemini, are amazingly powerful platforms that can deliver significant performance benefits and simplify enterprise adoption. However, organizations are beginning to contend with these models’ long-term cost implications and other constraints.    

The reality is that federal use cases are not one-size-fits-all—nor are emerging GenAI applications. The diverse needs of agencies, combined with the growing power and inherent flexibility of open-source LLMs, position these models as an optimal choice for some government applications.  

Federal agencies, like most enterprises, are grappling with how to integrate LLMs into their operations. What they are discovering is that there is not always a need for the most advanced LLMs—which may carry greater costs and restrictions—but rather for models that are sufficiently capable and reliable.

In this assessment, a second question emerges: Can proprietary LLMs adequately meet the escalating security, compliance, and privacy requirements across the government spectrum? 

In light of this question, open-source LLMs present an interesting option. Their growing depth and breadth of capabilities, exemplified by platforms like the Hugging Face marketplace, offer a compelling alternative. Models such as Llama 2, BLOOM, Falcon, and Mistral not only demonstrate remarkable advancements in AI technology but also offer more control to address the government’s unique concerns regarding security, compliance, and privacy. 

Open-source options are becoming increasingly attractive, as they can provide the necessary balance between advanced functionality and compliance with stringent governmental standards.

Open-source LLMs offer a plethora of advantages for federal agencies navigating the complex landscape of AI integration.

1. Customization: Their level of customization goes beyond what is typically achievable with commercial off-the-shelf products, allowing for more targeted applications.

2. Portability: The ability to deploy open-source LLMs in edge applications and secure environments is a significant advantage. Often vendor-agnostic, these models reduce the risk of lock-in and remain flexible to the changing needs of federal agencies.

3. Model Transparency and Risk Mitigation: Their open-source nature allows for a more comprehensive study of their limitations, like bias, relevancy, and potential degradation. Understanding these issues is crucial for implementing AI responsibly and mitigating risks.  

4. Open-Source Diversity and Innovation: Open-source LLMs benefit from the contributions of developers from myriad backgrounds. This diversity fosters innovation—focusing on multiple concerns rather than commercial viability alone. 

5. Security and Compliance: Federal agencies have stringent security requirements, making the flexibility of open-source LLMs ideal. These models can operate independently of external resources, which is crucial for agencies with strict regulations regarding data in transit. Open-source solutions even offer a pathway to use LLMs in a variety of deployment scenarios and can handle customizations that are unique to federal use cases and environments.

While the benefits of open-source LLMs are compelling, it's crucial to consider the tradeoffs that come with their adoption.

1.  Higher Development Costs: While open-source LLMs may offer lower or no licensing fees, this benefit must be balanced against potentially higher costs in development and support. It’s not always a clear-cut case of cost savings. 

2. Requires Greater Technical Capacity: Implementing open-source LLMs demands a higher level of technical expertise than conducting an API call on out-of-the-box proprietary models like GPT-4. This can pose a challenge for organizations that may not have the necessary technical capacity or resources in house.

3. Competing with Proprietary Models: Open-source LLMs can sometimes struggle to match the size, scope, or out-of-the-box performance of proprietary models. However, the landscape is evolving, with more complete open-source options entering the market, gradually increasing their attractiveness.

4. Investment in Skillsets: Each agency needs to decide how much it wants to invest in developing GenAI skills internally. While buying off-the-shelf solutions may seem simpler, open-source models offer an opportunity to develop and retain key capabilities in house. 

5. The Customization Tradeoff: While customization is a significant advantage of open-source LLMs, it comes with its own set of tradeoffs. Agencies need to determine the extent of customization necessary—the 80/20 rule often applies, where achieving 80% of the needed functionality might be sufficient without over-investing in the final 20%. Striking the right balance between customization and practical functionality is key.

A commitment is required to succeed with open-source LLMs, but, as we have observed, federal agencies are increasingly up to the challenge.  

We expect the availability of commercial, hybrid, and open-source LLMs to continue to grow for the foreseeable future, providing enterprises with a dynamic marketplace to innovate. To meet the coming challenges and opportunities, chief AI officers (CAIO) and other leaders in federal agencies need to consider how their organizations can take advantage of open-source LLMs.

As a starting point, they should identify the characteristics of compelling use cases for open-source models and quantify the potential benefits. In general, this is a tradeoff between the need for customization and the prospect of platform-driven innovation, while taking the cost model implications of each option into consideration. In some instances, a specific, narrow use case, such as providing customer answers based on a corpus of policy documents, might be a better fit for an open-source option, while broader, more shallow use cases, such as assistive document authoring, might be supported more effectively by a commercial platform, such as Microsoft’s Copilot.

Leaders will also need to evaluate the policy implications of adopting open-source LLMs. Different models come with their own terms of use, which can significantly impact how federal agencies integrate and use them. However, these constraints may be less impactful than those of commercial equivalents. Furthermore, leaders should consider the value of active participation in associated developer communities to shape future development and policy decisions.  

Finally, they must assess their near- and long-term ability to support open-source LLMs. A thorough assessment of the necessary infrastructure and talent competencies is essential. Ultimately, this will require investments to build and maintain capacity and competency, which will also pay dividends in the agency’s ability to take on more sophisticated use cases and deliver significant impact.  

The call to action is clear: Include open-source LLMs in your GenAI journey and acknowledge the unsolved issues—but don’t be paralyzed by them. Perfection is an elusive goal in the rapidly evolving tech landscape; waiting for complete information or a risk-free environment is impractical. Instead, choose a model, start small, and learn through practical application. This hands-on experience is invaluable and often provides insights and solutions that theoretical analysis cannot.