There’s a lot of debate in the tech world about the relative benefits of open source versus closed source AI models. For many of us, open source is where our heart lies, but not everyone feels that it’s appropriate for AI.
Much open-source software is produced by amateur enthusiasts and skilled hobbyists building and tinkering in their spare time. The results are fantastic, but AI models are a step beyond amateur capabilities. You need billions of data points and massive computational resources for an AI model, and that’s challenging for open source. There’s a reason why many open-source models are trained on fewer data points than closed-source models.
Some people argue that AI models are too powerful, and their use cases are too consequential, to be released for any hobbyist to use and tinker with. Unsurprisingly, many companies that build AI models insist that they need to keep them closed-source to maintain safety protocols, not just to protect their revenue streams, which they surely do.
My team at Lightricks recently released a generative AI video model called LTXV, and it was important to us to keep it under an open-source license. LTXV runs seamlessly on both GPU and TPU systems, both low end consumer hardware and high end servers, and is available on both Github and Hugging Face for anyone to tinker with.
When people find out we open sourced the weights of our model and wrote a comprehensive technical report, they generally react with dismay and skepticism. So here’s why we believe the future of AI is open source.
Open-source models broaden the horizons of possibilities
When AI models are open source, it encourages innovation. Developers, amateur hobbyists, and academic researchers can all access cutting-edge algorithms, models and tools without prohibitive costs. While these players are hardly in position to advance the bleeding edge of AI capabilities on their own, access to open-source models changes those dynamics. They can experiment as much as they like, producing new solutions that a single company might never imagine.
The saga of open-source software shows that it’s impossible to predict the transformative creations of a broad, diverse community. The range of people involved brings in more perspectives, resulting in more diverse use cases.
As an obvious example, just think of the wave of innovation that followed Meta’s release of its Llama models. Llama’s license may not be recognized by the Open Software Initiative, but that doesn’t stop developers from experimenting and building on top of the model.
Additionally, many enterprises are reluctant to use closed-source AI models for sensitive business use cases. The number of companies that have banned the use of OpenAI’s ChatGPT emphasizes the consequence of its leaks, privacy-compromising training protocols and general lack of transparency. In contrast, open LLMs can be self-hosted on an organization’s own infrastructure, removing data privacy and security fears and freeing enterprises to further innovate.
Open-source models foster a healthy, competitive market
When models are locked behind a paywall, it prevents startups and innovators from accessing them easily. Of course, that’s part of the intention of the model-builders, but it handicaps entrepreneurs from developing new AI video tools, solutions, and ideas.
Open-source AI models create a level playing field. Startups can produce new AI apps and workflows – and make them more robust, efficient, feature-rich. This encourages competition in the marketplace, which ultimately lowers costs and results in the very best solutions reaching end users.
It also enables smaller businesses to craft their own AI solutions without having to pay a fortune. As I mentioned above, this is why it was so important to my team to ensure that the LTXV model would work well even on computer systems running consumer-grade GPUs. Lowering the barriers to access helps to keep the wider business markets fair and competitive, instead of preserving customized AI solutions just for giant corporations.
What’s more, open-source AI models can be freely inspected and modified for different use cases, allowing companies of all sizes to choose the models that best match their needs.
Open-source models increase transparency and trust
It might not be the first thing you thought of, but open-source AI models also promote transparency, trust, and responsible use.
AI developers that take bias dangers seriously have to build feedback loops into their models, which adds to coding and processing costs. This means that companies motivated purely by the prospect of direct profit are incentivized not to bake fairness into their models.
One way to encourage more transparency is through increasing regulations, but open source offers an alternative. With open-source models, users can easily scrutinize and validate their workings. This limits the “black box” effect and enables users to audit model decision-making.
Everyone knows that the datasets and code bases have been verified by third parties, which further increases their quality and reliability. The developers also know this, so they put more effort into ensuring model safety and ethical training. As AI models become more transparent, ethical, and secure, user trust rises, helping drive adoption across the board.
Open-source models produce better AI solutions
As the saying goes, daylight is the best disinfectant. When developers know that the whole world will see their open-source model, it motivates them to write cleaner, better code. Likewise, involvement with the open source community helps developers to come up with new use cases, keeping them feeling inspired and excited.
When companies release AI models (or any software) as open source, they invite the whole community to help develop all the possibilities of the model. This adds value and enriches the AI ecosystem, which might otherwise remain somewhat arid.
Upon the release of Meta’s Llama under an open-source license, Mark Zuckerberg wrote that he hoped the community would help Meta develop a “full ecosystem of tools, efficiency improvements, silicon optimizations, and other integrations.”
This also links with the issue of transparency. More enterprises will use open-source models to develop their own in-house solutions, because they can see what went into testing and training and have full explainability of datasets. It’s particularly true for companies in regulated industries, which need to be able to own, control, and manage their models in a self-hosted setup in order to remain compliant.
The future is bright, the future is open source
From my perspective, the only way that AI will remain vibrant and continue to evolve is if developers continue to release foundational models under open-source licenses. We can’t wait to see where the wider community will take LTXV, and the exciting solutions that will be built upon it.
Yaron Inger is the CTO and a cofounder of Lightricks, an AI-first software company specializing in visual content creation tools.