"Open Source AI" isn't Open Source
It's not free as in beer, or free as in speech, but free as in... what, exactly?
Open source software has long differentiated between “free as in speech” (libre) and “free as in beer” (gratis). In the first case, libre software has a license that allows the user freedom to view the source and modify it, understand it, and remix it. In the second case, gratis software does not need to be paid for, but the user doesn’t necessarily have access to the pieces, can’t make new versions, and cannot remix or change it.
Yann LeCun co-authored a paper proposing Open Source Software in Machine Learning, Sonnenburg et al, all the way back in 2007 - that is, the pre-AI days of ML. In the paper, it lays out the many advantages to open source, and proposes that software tools used for machine learning should be released as open-source. The paper lists a variety of benefits, from reproducibility, to transparency1, to building on existing resources, to allowing the combination of advances used by different developers.
Recently, however, LeCun has unveiled “Open Source AI,” which is a completely new meaning for open source. The terms “Open Source” and “free” aren’t restricted, and anyone can use the term, well, freely. But this new use is disingenuous; it fails at least half of the attributes which Sonnenburg et al notes that the Open Source Initiative used to define Open Source at the time. Open Source AI simply means that the models have the model weights released - the equivalent of software which makes the compiled code available. (This is otherwise known as software2.) This new meaning of “Open Source” doesn’t allow for reproducibility, since the weights are the product of code which isn’t publicly released - exactly what the earlier paper protested - and data which is not disclosed, proprietary, stolen, or all three. It doesn’t allow building on existing resources other than the model itself. And it doesn’t allow the combination of advances, since the secret sauce of the system which would allow future advances isn’t the model weights, it’s the techniques used to build it. It’s not even “free as in time” (able to be used in any fashion), since the licenses restrict what can be done with the model. One other feature of open-source, not definitional but widely accepted, is collaborative open development. Open Source AI fails this test blindingly; the models can be modified, perhaps even collaboratively, but the development of the model is still at least as much of a black box as the system itself.
Open Source AI means the models can be studied, they can be used, they can be modified post-hoc, and they can be dissected. It’s plausible, as Zuckerberg suggested, that this has some of the same benefits as Open Source Software, though it’s unclear. It is also valuable for certain types of research. Commercially, Open Source AIs aren’t necessarily able to be used without payment, since the licenses used are restricted to research, not commercial use3. When available without restrictive licenses, these large models require either expensive hardware not available to most people, or paid hosting of the model. For development, the models are able to be modified, but are far from libre - the users cannot see the inputs to the model and cannot learn from how it was built. They are also impressively non-transparent; Meta’s Llama 2 model, the most openly available of its recent offerings, scores a paltry 54% on transparency - better than competitors, but far short of what we should expect.
If Open Source AI is neither gratis or libre, then those calling free model weights “Open Source,” should figure out what free means to them. Perhaps it’s “free as in oxygen” (dangerous due to reactions it can cause), or “free as in birds” (wild, without any person responsible).
I’m not necessarily opposed to judicious release of model weights, though as with any technology, designers and developers should consider the impact of their work before making or releasing it, as LeCun has recently agreed. But calling this new competitive strategy by Facebook “Open Source” without insisting on the actual features of open source is an insult to the name.
They don’t use this term, but it’s implicit in other benefits, and in the past decade is now widely used as a term for this essential feature of Open Source.
Only recently has there been “software as a service” which doesn’t allow the users to actually run the software themselves.
See, for example, the “open” release of Seamless4t for translation, announced here, and the noncommercial license for LLaMa-65b model.