Claim: we've solved the AI slop problem (!) 💩🧹✨
Blog post: facebookresearch.github.io/RAM/blogs/unsl…
🧵1/5
Key idea: take *expert* human writing and learn rubrics that find the gap between experts and models. Train with those rubrics.
We train with RL-XAR (RL with eXpert Aligned Rubrics) &
Normal chat UIs hide how non-deterministic are LLMs. I wrote reroll, a little tool to have a conversation with an LLM, similar to how it happens in ChatGPT, but every reply is generated 5 times, and then you can see easily if they all agree, 1/n
The ideal shai hulud target: yt-dlp. You always want the latest version, because otherwise it won’t work, most people won’t properly audit it, so if anyone manages to install a backdoor into it, everyone and their dog will be pwned.