this is why the "AI will be better than humans at writing" predictions won't come true while we're all using the same model.View quoted post on X
writing is an anti-inductive domain: ankitmaloo.com/anti-inductive/
this means that as we all collectively experience a new model and its "better" style, its starts to wear on us and no longer becomes a marker of quality, just as it would if all AI wrote exactly like V.S. Naipaul
even as AI wins writing competitions judged by literati unfamiliar with AI who don't examine the work closely, most people on here who read AI writing a lot will still consider it slop
ultimately this seems partly due to the fact that on-policy RL seeks the mode of the distribution, neglecting to capture the tails. if you think I'm wrong, then my challenge is: come back when you've solved in-weight continual learning
New paper!
Synthetic Persona Pretraining (SPP): Alignment from Token Zero
Rather than aligning models after pretraining, we do so from the very start.
SPP models are more faithful to our constitution, less misaligned, & more robust to jailbreaks. Advantage grows with scale! 馃УView quoted post on X
Well, labs have finally started closing the doors on AI capabilities. Imo, it has never been more important for open-source and open-weight models to catch up. A (few) winner-takes-all situation is not something I look forward to.View quoted post on X
