Weights are basically algorithm. It is not even obfuscated. Companies successfully produce derivative works from them, what is the problem?
One does not require git history for any Foss project, the exact way the code was produced IA not the requirement, it is the reproducibility that matters.
When I can train the ai on my own in the same way I can compile a program (yes I understand the huge difference in computing power necessary, that’s not what I’m talking about) then I’ll consider it equivalent.
As far as I’m concerned open weights are effectively compiled code. I can technically modify it, but I cannot rearrange the base components that were used to create it.
Even AI developers themselves cannot train AI in the same way as one compiles a program. Storing all the necessary data is just not feasible. To achieve results comparable to modern flagship models, one needs to obtain data in real time, use large amounts of generated data that is deleted after single use.
And the training algorithm itself is not deterministic.
And I still fail to see how this is any different from requiring a git commit history. There is an algorithm present - that’s it. It has no other forms in which it can be presented.
It’s not like the Chinese are any better. I want both sides to lose.
They’re a little better. Many of their models are open-source and can be run offline.
there are no open source models. There are open weight models.
No models make available the way in which they were produced, aka the source.
There is one: Apertus by ETH Zürich.
Weights are basically algorithm. It is not even obfuscated. Companies successfully produce derivative works from them, what is the problem?
One does not require git history for any Foss project, the exact way the code was produced IA not the requirement, it is the reproducibility that matters.
When I can train the ai on my own in the same way I can compile a program (yes I understand the huge difference in computing power necessary, that’s not what I’m talking about) then I’ll consider it equivalent.
As far as I’m concerned open weights are effectively compiled code. I can technically modify it, but I cannot rearrange the base components that were used to create it.
Even AI developers themselves cannot train AI in the same way as one compiles a program. Storing all the necessary data is just not feasible. To achieve results comparable to modern flagship models, one needs to obtain data in real time, use large amounts of generated data that is deleted after single use. And the training algorithm itself is not deterministic.
And I still fail to see how this is any different from requiring a git commit history. There is an algorithm present - that’s it. It has no other forms in which it can be presented.
https://opensource.org/ai/open-weights