The software industry operates on two distinct fronts. On one side, we see flashy products and services driving billions in revenue for trillion-dollar companies. On the other, we find the unsung heroes: developers diligently building, updating, and sharing the essential software infrastructure and tools that power the digital world, often for free. This is where open-source software shines. By releasing the source code of their products, developers enable others to reuse and modify it, fostering a collaborative and innovative ecosystem.
Open-source software forms the backbone of technologies we use every day. From Google’s Android and Apple’s iOS to the world’s leading web browsers, open-source code is the foundation. The encryption securing your WhatsApp messages, the compression powering your Spotify stream, and even the format of a saved screenshot – all are governed by open-source code.
This movement, rooted in the utopian spirit of 1980s California, thrives today due to its inherent benefits beyond altruism. Developers can access collaborative support, build trust through transparency, gain recognition, and even earn revenue by offering support services for their free products.
Now, the world of artificial intelligence (AI) is grappling with this open-source ethos, with giants like Meta pushing to embrace it for their powerful AI models. Their goal is to attract hobbyists and startups, creating a force capable of challenging billion-dollar labs, all while solidifying their reputation. But there’s a growing concern: has the term “open source” been stretched too thin in the context of AI?
The Open Source Initiative (OSI), a non-profit organization, has raised concerns about the modern use of the term by tech giants. They argue that the restrictions and secrecy surrounding these supposedly free AI models hinder true innovation. This begs the question: what does open source truly mean in the age of AI?
In traditional software development, the concept of open source is well-defined. Developers make the source code readily available, relinquishing most rights, allowing others to modify and adapt it for their own purposes. Often, a “copyleft” license is applied, requiring any modified version to be shared as well. This collaborative process can lead to entirely new products, as seen with Android, which evolved from Linux, initially designed for personal computers.
Meta, for example, proudly claims its large-language model (LLM), Llama 3, is “open source.” They offer it freely for anyone to build upon, yet impose restrictions, including a cap on the number of users for products built using Llama 3. Similar restrictions are found in LLMs released by other labs, such as France’s Mistral and China’s Alibaba.
The problem arises when we realize that what Meta shares freely – the weights of connections between artificial neurons in its LLM – is not enough to build a true replica of Llama 3 from scratch, a requirement of open-source purists. This is because AI training is fundamentally different from traditional software development.
While engineers create the initial blueprint and data for an AI model, the system effectively learns and evolves on its own, processing training data and adjusting its structure until it reaches a desired performance level. This means that even with identical data, code, and hardware, an AI model trained independently would resemble Llama 3 but wouldn’t be identical. This undermines the core benefits of the open-source approach, as even with full access to the code, users can never be certain they are using the exact same model as the company intended.
Further hurdles exist in the path to truly open-source AI. Training a cutting-edge AI model, on par with releases from OpenAI or its peers, can cost over a billion dollars, making it unlikely for companies that invested such sums to readily share their creation. There’s also the safety factor. In the wrong hands, the most powerful AI models could be used to develop bioweapons or generate harmful content. By restricting access, AI labs aim to control the inputs and outputs of their models, mitigating potential risks.
This complexity has sparked debate over the definition of “open-source AI.” Meta’s vice-president for policy, Rob Sherman, acknowledges the different interpretations of the term. The stakes are high, as those who tinker with open-source AI today could become the industry giants of tomorrow.
The OSI has attempted to define the term, proposing that to qualify as open source, AI systems must offer “four freedoms”: freedom to use, study, modify, and share. Instead of requiring the full release of training data, they advocate for labs to provide detailed descriptions enabling the creation of “substantially equivalent” systems. Sharing all training data isn’t always practical; it could prevent the development of open-source medical AI tools, as health records are private and cannot be shared freely.
While the debate over whether Llama 3 is truly open source continues, the fact remains that no other major lab has been as generous with its models. Vincent Weisser, founder of Prime Intellect, a San Francisco-based AI lab, prefers full transparency but recognizes the positive long-term impacts of Meta’s approach, leading to lower costs for users and increased competition.
The potential benefits of open-source AI are evident. Enthusiasts have already achieved impressive feats with Llama 3, shrinking it to fit on a phone, developing specialized hardware for faster processing, and even repurposing it for military applications, proving the potential downsides are real.
However, not everyone embraces open source with open arms. Ben Maling, a patent expert at EIP, a London law firm, highlights the legal complexities. True open-source software should be accessible without legal hurdles. When lawyers are needed to interpret restrictions, the engineering freedom vital to innovation is hampered. Companies like Getty Images and Adobe have already stopped using certain AI products due to licensing concerns, and others will likely follow suit.
The precise definition of open-source AI will have significant ramifications. Just as the quality of a wine dictates its designation as champagne or sparkling wine, the open-source label could be crucial to a tech firm’s success. Countries lacking domestic AI giants may turn to the open-source industry as a counterweight to American dominance. The European Union’s AI act includes loopholes to ease requirements for testing open-source models, a trend likely to be replicated by other regulators.
Governments grappling with AI regulation face a critical choice: restrict access for independent developers or provide them with the freedom to operate without excessive burdens? For now, closed-off labs maintain a comfortable lead. Even Llama 3, the most advanced of the almost-open-source models, lags behind those released by OpenAI, Anthropic, and Google. An executive at a major lab acknowledges the economic realities, noting that while releasing a powerful, free model allows Meta to undercut competitors without impacting their own business, the lack of direct revenue also limits their willingness to invest heavily in research and development, hindering them from becoming leaders rather than followers. The message is clear: freedom rarely comes without a price.
The future of open-source AI remains uncertain. The debate over its definition and the implications of different interpretations will continue. However, one thing is clear: the future of AI development may hinge on the answer to this question: can open source truly thrive in the age of AI, or will the closed-off labs remain dominant players?
