BlackRock's Larry Fink urges trillions in AI investment to keep US ahead, citing national security risks if China advances. His proposal to use retirement savings sparks social media uproar.
The training data is as important as the source code here to replicate the end result. The weights are more like a binary distribution. You can run the model and you can technically edit it just like you can technically edit a binary file.
They also only release some libraries and tools for running the model if you have a set of weights (which they do graciously provide), but they do NOT release the source code for their training pipeline itself. That’s up to you to reverse engineer from the whitepapers. Right now even if you had the exact training data and the compute available, you could not train your own Deepseek V3.2, let alone V4.
training data is as important as the source code here to replicate the end result
this is the nature of this flame war. Perfect replication of the end result, which is extremely opaque in how it works, is not nearly as important as the weights, that you can post train for any domain specific/general improvement with any other dataset. Which is how the authors would improve/change the weights further as well.
The training data is as important as the source code here to replicate the end result. The weights are more like a binary distribution. You can run the model and you can technically edit it just like you can technically edit a binary file.
They also only release some libraries and tools for running the model if you have a set of weights (which they do graciously provide), but they do NOT release the source code for their training pipeline itself. That’s up to you to reverse engineer from the whitepapers. Right now even if you had the exact training data and the compute available, you could not train your own Deepseek V3.2, let alone V4.
this is the nature of this flame war. Perfect replication of the end result, which is extremely opaque in how it works, is not nearly as important as the weights, that you can post train for any domain specific/general improvement with any other dataset. Which is how the authors would improve/change the weights further as well.
If people on Lemmy can’t understand this I have no hope for the average person.