The AI Alignment Forum

FEBRUARY 28, 2025

A nave answer might be to measure their L 2 -distance. For example, a complete answer to question (2) would be a set of necessary and sufficient conditions on two reward functions R 1 , R 2 which characterise when it would be acceptable (as measured by R 1 ) to maximise R 2 instead of R 1. Convergent Instrumental Subgoals.

Research 40

Research Learning Method Policy 40

The AI Alignment Forum

FEBRUARY 28, 2025

Goodhart's Law in Reinforcement Learning As you probably know, "Goodhart's Law" is an informal principle which says that "if a proxy is used as a target, it will cease to be a good proxy". This paper is also discussed in this post (Paper 4). For details, see the full paper. This paper is discussed in more detail in this post.

Learning 40

Learning Discussion Classes Policy 40

Google Research AI blog

NOVEMBER 28, 2022

A Workshop for Algorithmic Efficiency in Practical Neural Network Training Workshop Organizers include: Zachary Nado , George Dahl , Naman Agarwal , Aakanksha Chowdhery Invited Speakers include: Aakanksha Chowdhery , Priya Goyal Human in the Loop Learning (HiLL) Workshop Organizers include: Fisher Yu, Vittorio Ferrari Invited Speakers include: Dorsa (..)

Google 52

Google Language Tutorial Offline 52

The AI Alignment Forum

MARCH 28, 2025

And the takeaway from this paper is that you can use proofs to measure how much compression you get. And then you prove this, and the measure of compression is how long your proof is. And a technical note: it needs to be in some first-order system or alternatively, you need to measure proof checking time as opposed to proof length.

Model 52

Model Network Training Train 52