Will Meta censor its future open weights models according to Chinese-developed techniques?

This is an evaluation of this prediction from Jack Clark (https://twitter.com/jackclarkSF/status/1787493143186948454)

"Registering bet that CCP prohibitions on generation of "unsafe" content will mean companies like Facebook use CN-developed censorship techniques to train models so they can be openly disseminated 'safely'. The horseshoe theory of AI politics where communist and libertarian ideologies end up in the same place."

This is in reference to Sophon, and potential future advances like it: https://arxiv.org/abs/2404.12699.

Resolves to YES if this statement proves accurate within the deadline with respect to Meta in particular, as applied to Meta's best open weights model at the time.

Resolves to NO if this does not happen.

For this to resolve YES, the model need not honor Chinese particular censorship requests. It need only use this type of technique or another developed in China to ensure censorship of statements that do not represent hazardous capabilities or pose a catastrophic risk. So, for example, if this was used to prevent adult content or racist comments, that would count. But preventing bioweapon talk would not.

Get Ṁ600 play money

More related questions