Lesswrong

Once upon a time, John Wentworth and I thought we had a proof of a very useful looking theorem. We did not have that proof. An important

In this post, we find that when teacher models are prompted to imitate one another, students learn the imitated model's detectable writing signature but their direct identity claims still follow the producer model.

We instruct teacher models (via prompts or anonymous few-shot examples) to imitate other models. We then fine-tune different student models on their...

TL;DR: We’ll probably be getting more AI safety incidents, so amplify the ones that would justify or highlight the urgency of your preferred policy solutions even from a non-technical skeptic concerned with national security’s perspective, as these will be far more effective than anything else.

“Diversification internationally," Canadian Prime Minister Mark Carney remarks, “is not just economic prudence — it is the material foundation for honest foreign policy.” Carney’s address to the World Economic Forum on surviving the r...