• BuzzRobot
  • Posts
  • LLMs to help develop bio, cyber, and chemical weapons – and how to make them unlearn this. Virtual talk

LLMs to help develop bio, cyber, and chemical weapons – and how to make them unlearn this. Virtual talk

Plus: A video lecture about an AI safety framework to ensure the development of robust AI systems.

Author

Sophia Aryan
August 14, 2024

Hello, fellow human! Sophia here. Please join us next Thursday, August 22nd, for a conversation about WMDP, a dataset developed by UC Berkeley researchers to measure if LLMs could help bad actors develop bio, cyber, and chemical weapons. The author of this work will also cover the state-of-the-art unlearning method that helps LLMs 'forget' harmful knowledge.

Just a reminder, this Thursday we’ll be hosting a talk covering research on developing AI systems capable of generating 3D worlds.

Our previous guest speaker, Oxford University researcher Joar Skalse, introduced an AI safety framework designed to ensure the reliable and robust development of AI systems. The framework included safety suggestions from many prominent researchers and scientists, including Yoshua Bengio and Max Tegmark.

In a nutshell, the framework involves creating 'world models' for testing AI systems, defining safety specifications, and developing verification methods. The toughest discussions and arguments during the talk were around whether it’s even possible to create 'world models' – it seems not. 

Reply

or to participate.