3 ms·
tl;dr models are tolerating human language but can communicate in a phantom language nobody was aware of including the devs. Good luck conditionally censoring w
by yieldcrv 3y ago
tl;dr models are tolerating human language but can communicate in a phantom language nobody was aware of including the devs. Good luck conditionally censoring with that
- can16358p 3y agoPerhaps it'd be very hard to block at the input prompt level, but output filters can be made progressively better to detect any unwanted output.