- Businesses are just beginning to grapple with the sheer complexity of deploying generative AI for their teams.
- Understanding whether to align your business more closely with open-source or closed-source AI can be daunting.
- Businesses seeking higher levels of customization, flexibility, and scalability will likely choose open-source solutions for their generative AI infrastructure.
Generative AI tools like ChatGPT, DALL-E, Midjourney, Claude, Bard, and Gemini, have captured the attention and imagination of millions, quickly becoming the hottest and fastest growing category of consumer technology. As these tools race to capture more and more of the consumer market, businesses are just beginning to grapple with both the excitement and the sheer complexity of adopting, implementing, and deploying this new class of technology for their teams.
While proprietary or closed-source solutions like the generative AI products put out by OpenAI, Google, and Anthropic are winning consumer market attention, many businesses have started exploring open-source generative AI with the hope of finding solutions that are flexible, scalable, and don’t lock them into licensing agreements for technology that is still largely untested.
Hugging Face, an AI community for developer teams, has over 400,000 open-source models available for download, including the massively popular Llama-2 and SDXL. These models, and the applications that support them, have entered the picture with a promise of scale, security, flexibility, and ownership.
It should be noted that both open and closed-source AI have attracted the scrutiny of government regulators and AI safety experts, who are increasingly focused on ensuring their constituents can harness the benefits while mitigating the risks of generative AI. It will be important for businesses to carefully balance opportunity with safety with any AI solutions that they implement.
Generative AI fundamentals
Understanding whether to align your business more closely with open source or closed source can be daunting. There are new generative AI tools emerging every week across text, image, voice, video, and even multi-modal outputs. Regardless of its output medium, most tools — both open and closed source — operate with the same key components:
- Base Model: The base or foundation model forms the architectural backbone of any generative AI tool. It’s the foundational framework upon which additional functionalities and customizations are built.
- Specialized Model: Evolving from the base model, these are fine-tuned to perform specific tasks, catering to niche requirements or unique industry needs.
- End-user application: The interface through the magic of AI is made possible — this is where end-users interaction with the AI, leveraging its capabilities for varied use cases.
Where open-source AI excels
For businesses where generative AI is vital to their core business — integral to workflows, a key revenue driver, or a cornerstone of competitive advantage — it will be important to prioritize the control and ownership of your base model, specialized models, and end-user applications. This is where open-source generative AI solutions shine:
- The base model comes with an open license, allowing full ownership of a tailored base model that aligns to your business needs.
- You can engage in the development of the open-source code, creating specific features or specialized models that cater directly to your business.
- With publicly available code, you can confidently expect that your investment will remain accessible and functional for the long term.
- Users are empowered to adapt and integrate the tool into various systems that they use regularly.
Particularly beneficial for businesses requiring high control levels, like creative teams engaged in revenue-generating projects or those handling sensitive IP, open-source generative AI offers the highest level of customization. They are great for companies that: require a high-level of control (eg. creative teams working on specific asset generation projects, workflows or tasks for revenue-generating products); work with sensitive or confident IP that they don’t want shared between orgs or used to train others models; want to be able to customize the model and application technology infrastructure to meet their organization’s specific needs and use cases.
The case for closed-source AI
Generative AI, while transformative in many sectors, might not become a core function in certain businesses where either the nature of the work doesn’t align with what AI can offer, or the value added by AI is marginal compared to other factors. In this case, closed-source options typically provide a more controlled, “out-of-the-box” experience with the trade off of less customization, control, and ownership. In most closed-source solutions:
- Customers license access to a proprietary model that the vendor owns. There is less freedom to modify or adapt it to their specific needs or creative direction.
- The applications are designed to output a high-quality generation for general purposes. They accomplish this by automatically manipulating prompts or inputs on the backend to improve the quality of the generation, though most systems do not share with the user how their inputs are being manipulated.
- The application gathers user input data to improve the proprietary base model, which results in higher quality outputs for all users, but less transparency in how your team’s creative works are being used by the vendor.
Closed-source solutions make sense for businesses that do not see generative AI as core to their business and have less concerns over privacy, security, provenance, and compliance. They are generally better at producing a high-quality output for users that are just getting started or don’t need to use the tools as part of professional workflows that contain sensitive employer-owned content or other confidential intellectual property.
Open and closed source aren’t binary options and most businesses will incorporate both open and closed generative AI solutions. However, it is important to understand the strategic choices and potential outcomes of licensing proprietary tools compared to adopting open-source models and applications. To add to the complexity, it is often difficult to fully understand whether the solutions you are considering are even fully open.
Some companies publish their source code but restrict commercial use through licensing, while others, like OpenAI, license access to a closed-source model despite an ethos of openness and transparency. During procurement, businesses will need to carefully consider the core technology powering the solutions they are considering to ensure they will meet both immediate business needs and business needs at scale.
The backbone of AI infrastructure
Closed-source generative AI tools are likely to dominate the mass consumer market, thanks to their ability to provide high quality outputs across a variety of modalities for novice users. However, the B2B market is still largely undefined. If history repeats itself, we’ll likely see open-source solutions rise as the backbone of generative AI infrastructure.
Just as Linux, Apache, and MySQL became foundational building blocks of the early internet era, businesses seeking higher levels of customization, flexibility, and scalability will likely turn to open-source solutions to power their generative AI infrastructure.