Not surprising as it was never really open but a leak. They have the right to do this.
Whether it makes sense is another question. They'll never get the genie back in the bottle l, people will just have to get the torrent instead of using dalai.
I think it would be a much better strategy to open-source it at this point. It's out there anyway and sooner or later someone else is going to make an open source one (I believe Mozilla was already working on this), making Facebook's work much less relevant in the community.
There isn't much value in the weights apart from the costs they made to build them, which is sunk anyway. It's pretty easy for someone else to do it once they get the money together, as far as I understand there is no proprietary secret sauce to protect.
But being in the interest of the community at this point is the biggest value they could have from it at this point where things are moving so quickly. Either way this particular model will be replaced by something better within a couple months whether it's from Facebook or someone else. If it hadn't leaked it wouldn't have been more than a footnote.
Why not keep a finger in the pie and surf the huge wave of public interest to become a household name in AI? Trying to stop the wave is impossible.
> There isn't much value in the weights apart from the costs they made to build them, which is sunk anyway.
Because of this, it isn't at all clear that this model can be copywritten in the first place: it isn't itself an expression, but is merely the output of an algorithm that was itself released as open source.
Maybe you could argue that the model is under the same license, under an argument that it isn't similar to how a compiled binary is considered a derived work of the source code... but that license is GPL, guaranteeing the right to distribute such binaries.
Alternatively, maybe you could argue that the code is more like a compiler and the model is a derived work of the training data set. But, not only did Facebook not make or own the training set, they specifically used an open data set (that was the point of the research in a sense) that any of us also have similar access to.
> They have the right to do this.
So, not necessarily; and, as far as I am concerned: not at all likely.
3 comments
[ 8.7 ms ] story [ 17.9 ms ] threadWhether it makes sense is another question. They'll never get the genie back in the bottle l, people will just have to get the torrent instead of using dalai.
I think it would be a much better strategy to open-source it at this point. It's out there anyway and sooner or later someone else is going to make an open source one (I believe Mozilla was already working on this), making Facebook's work much less relevant in the community.
There isn't much value in the weights apart from the costs they made to build them, which is sunk anyway. It's pretty easy for someone else to do it once they get the money together, as far as I understand there is no proprietary secret sauce to protect.
But being in the interest of the community at this point is the biggest value they could have from it at this point where things are moving so quickly. Either way this particular model will be replaced by something better within a couple months whether it's from Facebook or someone else. If it hadn't leaked it wouldn't have been more than a footnote.
Why not keep a finger in the pie and surf the huge wave of public interest to become a household name in AI? Trying to stop the wave is impossible.
Because of this, it isn't at all clear that this model can be copywritten in the first place: it isn't itself an expression, but is merely the output of an algorithm that was itself released as open source.
Maybe you could argue that the model is under the same license, under an argument that it isn't similar to how a compiled binary is considered a derived work of the source code... but that license is GPL, guaranteeing the right to distribute such binaries.
Alternatively, maybe you could argue that the code is more like a compiler and the model is a derived work of the training data set. But, not only did Facebook not make or own the training set, they specifically used an open data set (that was the point of the research in a sense) that any of us also have similar access to.
> They have the right to do this.
So, not necessarily; and, as far as I am concerned: not at all likely.