My parents sat me down today saying they have some serious news to share. They apparently are going to have another child. I was in total shock and asked why. They told me that I simply didn’t age that well and was in need of a remake to appeal to modern audiences. Heartbreaking.
Finetuning LLMs can sometimes cause surprising side effects. Can we predict a finetuned model's response from its training data before burning compute? We trained oversight models to do this and found performance follows power laws in model size and data. 🧵
knight theory so good it genuinely made me rethink everything abt the knight not to mention it just feels so deltarune-esque this might genuinely be it guys
@Sauers_ https://t.co/IyFYD8w9uo related for verbalisation and self reports. wondering if you can train qwen to be more integrated or if some training can push olmo’s apart
Qwen has a fragmented character-action subnetwork: what they say is structurally disconnected from what they do. But in OLMo, these subnetworks are coherently integrated.