All videos
All videos
AI Alignment is More Fragile Than You Think (And Shaggoth is Lurking)
September 30, 2025
One of the fundamental features of LLMs is their ability to provide responses aligned with human ethical principles and intentions. In his talk, Przemek will explore what LLMs truly "know" about morals, discuss how models are being trained for AI Alignment, and reveal how (surprisingly) easy it is to break. He'll show experimental results, in which a small set of harmful examples could derail alignment. A live demo and open discussion on what it all means for us will follow.
Other videos that you might like
Mutatis Mutandis: Typesafe, transactional, indexed database queries in the cloud
Jon Pretty
Threads at Scale
Daniel Spiewak
Lean AI – How to Implement a Production-Ready Deep Learning Model in Twelve Weeks
Dat Tran
Monitoring with no limits
Nikolay Tsvetkov