I don’t necessarily object to releasing weights of models like Gemma 2, but I wish the labs would better discuss considerations or say what would make them stop.
On Gemma 2 in particular, Google DeepMind discussed dangerous capability eval results, which is good, but its discussion of ‘responsible AI’ in the context of open models (blogpost, paper) doesn’t seem relevant to x-risk, and it doesn’t say anything about how to decide whether to release model weights.
I don’t necessarily object to releasing weights of models like Gemma 2, but I wish the labs would better discuss considerations or say what would make them stop.
On Gemma 2 in particular, Google DeepMind discussed dangerous capability eval results, which is good, but its discussion of ‘responsible AI’ in the context of open models (blogpost, paper) doesn’t seem relevant to x-risk, and it doesn’t say anything about how to decide whether to release model weights.
FWIW, I explicitly think that straightforward effects are good.
I’m less sure about the situation overall due to precedent setting style concerns.