Sune Lehmann

@sunelehmann.com

959 Followers 164 Following

Your friendly neighborhood suneman

ChatGPT illustration of life-paths for talk 1, with a mini Denmark on the left and a network on the right.
Parish-level voting for Venstre.
Parish-level voting for "socialdemokratiet"
A large and unusually complex figure from the paper.
Two distributions. One is orange and represents the distribution of empirical moves between houses in Denmark. The other is blue and represents the distribution of pairwise distances between houses in Denmark.
A very straight line on a loglog plot. So straight that even Aaron Clauset would call this a power-law
Image of an abstract with the following text Abstract: A year ago, ChatGPT surprised the world with its extraordinary language generation capabilities. Chatbots have since become of the fastest adopted consumer product in history with investments in genAI forecasted to reach $12B this year. In this talk, I will first review the fast-evolvingliterature on the document-level membership inference task for LLMs: the methods proposed to detect--a posteriori--whether a specific piece of text was seen during training by an LLM and at least partially memorized, the distribution shift concerns, and some of the solutions proposed. I will then discuss the use of randomized controlled setups to causally study LLM memorization. In particular, Iwill discuss how randomized controlled setup have shed lights on the determinant of memorization and showed LLMs to have a mosaic memory. I will conclude the talks with same thoughts on the security and privacy challenges ahead when it comes to LLMs and the use of synthetically generated trap sequences for membership inference.
Key quote from long article