4 ms·
In his recent talk for the FSF at Boston Stallman suggested that published weights are open source. I guess because they’re modifiable and auditable. It’s an in
by eulers_secret 4y ago
In his recent talk for the FSF at Boston Stallman suggested that published weights are open source. I guess because they’re modifiable and auditable. It’s an interesting argument. So far, I’ve managed to modify llama myself so I guess so??
- arnaudsm 4y agoWeights can be retrained but cannot be explained. Recent papers even demonstrated backdoors : https://arxiv.org/abs/2206.07840 https://arxiv.org/abs/2206.07840 Therefore a real open source model needs a public dataset and training code.