Meta's Llama-branded large language models have sparked a heated debate by claiming to be open source despite carrying commercial restrictions. This raises questions about what truly constitutes open source, a term defined by the Open Source Definition and Open Source AI Definition. The global community and industry heavily rely on these definitions to determine the openness of software. However, philosophical concepts around "open design, community, and development" often add layers of complexity to the discussion.
In a landscape where Android is considered open source by most definitions, its commercial restrictions due to anti-fragmentation agreements with hardware makers present a paradox. This situation is not unique, as China's DeepSeek AI model also claims open source status but faces scrutiny over its training data and other components. At the same time, the European Union's AI Act has introduced a carve-out for "free and open source" AI systems, except those that pose an "unacceptable risk." This regulatory environment underscores the importance of clear definitions.
Google has leveraged Android's open source status to counter anti-competition criticisms, while Amazon has adapted Android for its Fire-branded devices. The Android Open Source Project (AOSP), released under a permissive Apache 2.0-license, allows anyone to access, fork, and modify it. Despite this openness, some experts argue that governance remains crucial as single corporations can alter licenses at will.
“Governance is very important, because if it’s a single corporation, they can change a license like ‘that’.” – Peter Zaitsev
Luis Villa, co-founder and general counsel at Tidelift, has been exploring what constitutes open source for several years. He remains skeptical about certain claims, particularly Meta's Llama models.
“I have my quibbles and concerns about the open source AI definition, but it’s really clear that what Llama is doing isn’t open source.” – Villa
Stefano Maffulli, executive director at the Open Source Initiative (OSI), emphasizes the significance of having clear definitions that focus on licensing as a measure of openness.
“The point of having definitions is to have criteria that can be scored, and focusing on licensing is how that is accomplished.” – Maffulli
The OSI's formal definition of open source relies on the widely recognized Open Source Definition. However, Villa warns against attempts to dilute this definition due to its strong brand equity and regulatory implications.
“There are plenty of actors right now who, because of the brand equity [of open source] and the regulatory implications, want to change the definition, and that’s terrible.” – Villa
Android serves as a prime example of this debate. It is often cited as perfectly documented in terms of licensing, yet practical challenges remain in contributing to its development or predicting new releases.
“Android, in a license sense, is perhaps the most well-documented, perfectly open ‘thing’ that there is.” – Villa
“All the licenses are exactly as you want them — but good luck getting a patch into that, and good luck figuring out when the next release even is.” – Villa
The discussion extends to researchers at Hugging Face who are striving to create an even "more open" version of DeepSeek's reasoning model. This initiative highlights the ongoing efforts to push the boundaries of openness in AI.
Luis Villa's inquiry into the essence of open source continues to resonate as these issues become more pressing in 2025.
“I’ve been asking this question for a good few years, and in 2025 this question is more relevant than ever.” – Horovits
The strength of the open source brand lies in its ability to command attention and spark debate.
“It goes to show how strong the brand of open source is — the fact that people are trying to corrupt it, means that people care.” – Omier
Leave a Reply