4 ms·
A question is whether the new model still intrinsically embeds the source text, but this is later filtered in the output, or if it no longer embeds the text at
by avidiax 3y ago
A question is whether the new model still intrinsically embeds the source text, but this is later filtered in the output, or if it no longer embeds the text at all.
The latter is more defensible.
- skygazer 3y agoI would think an existing model could bootstrap a copyright free training corpus by completely rewriting/paraphrasing copyrighted material with semantic fidelity for training of the next model to completely eliminate memorization of copyrighted works. That might pose an interesting obstacle to copyright challenges, bootstrapping your way into a clean room. Although, tweaking the architecture to either eliminate memorization, or eliminate high fidelity reproduction of verbatim training data seems far more expedient and less costly.