Hackathon: Alignment Faking Model Organisms

Important registration information: ​​To participate in this event, please join the discord link before registering.

​Many safety and governance measures rely on AI models showing us their true colours. "Alignment faking" is the phenomenon of a model hiding misaligned behaviour when it believes it's being observed.

​In this hackathon, we will be constructing model organisms of alignment faking: realistic, experimentally-verified pathways under which alignment faking can occur. We'll be test-driving a new framework for alignment faking experiments. The environment, monitoring and scoring are already set up - all we need to do is supply the models! These can be fine-tunes of open source models or simple prompt engineering.

​​​Trajectory Labs, the jamsite, provides a comfortable and spacious coworking space along with coffee, tea, and other refreshments (meals not provided, but there are many nearby options). Other locations will also be taking part!

​Bring a laptop (beefy GPUs are not necessary, we'll provide credits for API-based finetuning of open source models so you don't need to run them locally).

​*More details and resources to come, including some useful background reading on model organisms and alignment faking.