4 ms·
Direct Language Model Alignment from Online AI Feedback
- yellow_postit 3y agoBeing able to evaluate these changes from online updates is a great result if it holds for more rapid iteration and tuning
- lamroger 3y agoEval become way more important if things become less reproducible
- PeterStuer 3y agoI wonder if this opens practical opportunities for adversarial hijacking of specific topics.
- michaelbrave 3y agoI would guess likely, it just remind me of microsofts chatbot on twitter that would learn from interactions so some 4chan people decided to turn racist, and they did like within a day or so.