Open vs. closed-source generative AI

Open vs. closed-source development has emerged as a key debate within generative AI. Which approach produces better generative AI products? Which is safer? Which has a sustainable future?

In this blog, we’ll explore some of the considerations in the open vs. closed-source generative AI debate from a commercial and enterprise perspective, and discuss some of the implications for businesses.

What is open-source?

Broadly speaking, open-source refers to an approach to software development where all or part of an application’s source code is released openly to the general public. Sometimes this is developed collectively by decentralised voluntary communities, sometimes by more formal organisations, and sometimes both. Conversely, closed-source refers to a software development approach where an application’s source code is proprietary and not publicly available.

There are different types or levels of open-source, ranging from fully open software where all source code is released and accompanied by copyleft licences, to applications where only a limited amount of source code is released, sometimes with more restrictive licences too.

Within AI, there are several approaches to developing and releasing models that may be considered open-source, including openly releasing model weights, datasets, and/or algorithms, and using different types of open-source software licences. At present, numerous open-source generative AI models have been released by both decentralised communities and private sector companies, and we expect this trend to continue.

Open-source generative AI and the commercial sector

Historically, one of the biggest challenges in using open-source applications – especially in enterprise solutions –  has been ease of use, particularly for non-technical users. For AI, this is currently being solved by platforms which aim to make it easier to find and use high quality open-source AI models, including generative AI models. For example, Hugging Face currently hosts over 120,000 open-source models, and is valued at $4.3 billion based on its latest $235 million round of Series D funding.

These platforms are not only targeting researchers and deep technical users, but also commercial organisations via their enterprise hubs. This allows paying customers to gain access to end-to-end developer tools, private and secure deployment hubs, off-the-shelf GDPR compliance, support services, and more.

Will this be sufficient to compete with closed-source proprietary models? After all, many of the leading generative AI models are currently developed by a handful of AI labs, and almost all are choosing to release their models as closed-sourced products.

By prioritising adoption over monetisation, open-source platforms are hoping to become the default platform of choice for machine learning developers, and in doing so, drive adoption into more profitable enterprise solutions. Some of them have even developed strategic long-term partnerships with the largest established cloud computing services in order to provide easier and cheaper integration of open-source models into cloud-based production environments. This type of backing by several of the largest tech firms may help open-source compete on both performance and deployment, especially whilst model development, training and inference costs remain high.

How will this stabilise, and what does this mean for organisations that use generative AI?

It is currently unclear whether open or closed-sourced generative AI will dominate, or whether the two will continue to co-exist side-by-side as is the case in several other key areas of tech.

Organisations looking to adopt generative AI currently have the option of both closed and open-source models. Off-the-shelf closed-source models often provide higher performance and more intuitive interfaces but at a greater cost. Alternatively, organisations can opt for often cheaper open-source models that can be deployed on local infrastructure if privacy and security are especially important, although off-the-shelf performance is generally lower.

As generative AI continues to mature, the ecosystem may shift to favour either open or closed-source, or both may remain significant players. With this in mind, organisations should try to establish experience and expertise in both areas, and ideally develop modular generative AI solutions that remain flexible and adaptable to future changes. This will help avoid lock-in and increase agility, and ensure organisations are well-placed to exploit different future scenarios.

If you would like to learn more about open vs. closed-source approaches to generative AI, including services that Deloitte offers to help navigate this space, please do get in touch.