Modern ML systems sometimes undergo qualitative shifts in behavior simply by “scaling up” the number of parameters and training examples. Given this, how can we extrapolate the behavior of future ML systems and ensure that they behave safely and are aligned with humans? I’ll argue that we can often study (potential) capabilities of future ML systems through well-controlled experiments run on current systems, and use this as a laboratory for designing alignment techniques. I’ll also discuss some recent work on “medium-term” AI forecasting.
Overview
Syllabus
Introduction.
Rest of Talk.
Reward Hacking: Motivation.
Reward Hacking Example.
Reward Hacking: Example.
Summary of Full Results.
Reward Hacking: Summary.
Making NLP Models Truthful.
Contrastive Representation Clustering.
Results on Unified QA.
Caveat: True Answers Work Too.
Forecasting: Motivation.
Forecasting Competition.
Forecasting Questions.
Summary of Benchmark Forecasts.
Results So Far.
Forecasting: Lessons Learned.
Forecasting Class.
Taught by
Stanford Online
