Will SB 1047 fix the definition of a derivative model in a way that Zvi publicly expresses approval of?
Basic
3
220
resolved Jun 6
Resolved
YES

SB 1047 currently has a bug, according to Zvi:

To summarize the issue once more: As written, if interpreted literally and as I understand it, it allows developers to define themselves as derivative of an existing model. This, again if interpreted literally, lets them evade all responsibilities, and move those onto essentially any covered open model of the same size. That means both that any unsafe actor goes unrestricted (whether they be open or closed), and releasing the weights of a covered model creates liability no matter how responsible you were, since they can effectively start the training over from scratch.

Scott Weiner says he is working on a fix.

Resolves YES if there has has been such a fix, and Zvi reports on his blog that this fixes the bug (not necessarily perfectly, but to a degree less than it being a "potential showstopper bug").

Get Ṁ1,000 play money

🏅 Top traders

#NameTotal profit
1Ṁ23
Sort by: