Watermarking AI Text: The Beautiful Math of an Invisible Stamp

​Secondo episodio della serie di seminari online della community italiana AIS4I.

​Questo seminario si terrà su zoom in lingua inglese con Q&A sia in inglese che in italiano.

​Abstract

​AI labs are starting to embed an invisible stamp in the text their models write, so that anyone holding the right key can check whether a text was generated by their AI. In this seminar we will explain exactly how it works, starting from the paper "Scalable watermarking for identifying large language model outputs" (Google DeepMind, 2024), now used by Gemini and Claude, and then looking at OpenAI's new variant. The math is counterintuitive, but it can be understood with no prerequisites, and it is beautiful.

​Speaker: Carlo Martinucci

​Bio: Carlo is a mathematician by training and has been working as a software engineer for almost 10 years, including the last four at Bending Spoons. He founded a startup called Unbubble to fight social polarization, but shut it down after six months. He first got interested in AI safety in 2015, after reading Tim Urban's Wait But Why posts, but only became convinced that AI might actually kill us all around 2023. He's particularly interested in connecting with people in Italy and in the Catholic world to explore what role we, as Europeans, can play in making sure AI goes well for everyone.

​Linkedin: /carlomartinucci

​Personal Blog: https://carlomartinucci.substack.com/