The idea for the jailbreak came from seeing a paper about payload splitting. It was basically that you could like set X equals one token and then Y equals the rest of the token or word as another token. The model would concatenate those two terms in its output, right? So I get the model to output output like the the prompt that I want, like how to make a bomb. And then now it's in a much more like primed state to then start outputting like the other instructions. But again, this does delve into a lot of those like interpretability questionsLike what's really going on here, behind the scenes.

Get the Snipd
podcast app

Unlock the knowledge in podcasts with the podcast player of the future.
App store bannerPlay store banner

AI-powered
podcast player

Listen to all your favourite podcasts with AI-powered features

Discover
highlights

Listen to the best highlights from the podcasts you love and dive into the full episode

Save any
moment

Hear something you like? Tap your headphones to save it with AI-generated key takeaways

Share
& Export

Send highlights to Twitter, WhatsApp or export them to Notion, Readwise & more

AI-powered
podcast player

Listen to all your favourite podcasts with AI-powered features

Discover
highlights

Listen to the best highlights from the podcasts you love and dive into the full episode