I've shared two papers with many colleagues recently and thought more people might find them useful! 👇 1) AI safety and security are often used interchangeably but these two are very different. This work does a great job explaining the differences between the two:…

Jailbreaks have become a new sort of ImageNet competition instead of helping us better understand LLM security. I wrote a blogpost about what I think valuable research could look like 🧵

Why do we expect unlearning methods to be effective for AI safety? It does not seem feasible to remove specific knowledge (e.g. enhanced pathogens) that follows naturally from capabilities we want the models to have (e.g. biology and chemistry). This, of course, combined with the…