AXRP - the AI X-risk Research Podcast

6 - Debate and Imitative Generalization with Beth Barnes

Apr 8, 2021

01:58:48

Creator website

Highlights

AI Chapters

Episode notes

Is There a Difference Between Human Debate and Mal Debate?

03:33

Can We Sort of Figure Out Which Bits Are the World Model and Which Are the Sent?

04:56

Is There a Difference Between Deliberation and Distillation?

02:23

Introduction

3min

It's Like, a Tree of Answers

4min

How to Train a Machine Learning Model for Debate?

2min

Train and Model to Imitate Human Judgments

2min

Are There Any Problems in Like, a Safety or an Alignment That This Would Solve?

2min

Is It Necessary?

3min

Are You Doing This Bad Thing?

3min

Is Debate Easier Than Other Things?

2min

The Tree of Arguments in a Debate?

5min

10.

Is There a Problem With Debate?

3min

11.

Is There a Difference Between Free Response and Debate?

3min

12.

Is There a Problem With Debate Rit?

2min

13.

Is There a Penalty for Choosing What Recurses On?

2min

14.

Is There a Mechanism for a Good Debate?

2min

15.

What Is Cross Examining?

3min

16.

How Do You Come Up With That Strategy?

2min

17.

Complexity Clauses

5min

18.

The Obfiscated Arguments Problem

2min

19.

Is It Possible to Generate Reasonable Arguments in a Debate?

3min

20.

Is the Research Strategy Like, Like, Is It Relevant to Humans?

3min

21.

Is There a Difference Between Human Debate and Mal Debate?

4min

22.

Are You Trying to Be Honest?

3min

23.

The King of France Does Not Have Hair

4min

24.

Are You Going to Judge Debates?

2min

25.

Answering a Question Like, Is Daniel an Australian?

2min

26.

Is There a Better Answer to That Question?

4min

27.

Is It a Heroistic Approach or a Cold Generalization?

4min

28.

How to Train a Debater

2min

29.

Is This a Lak?

2min

30.

Machine Learning

5min

31.

How to Label Dogs in a Train Set

2min

32.

Cansev Generalise From the Human Labels?

2min

33.

Do You Think We're Going to Use Imitative Generalization?

2min

34.

How to Interpret a Big Nural Net in a Test Set?

5min

35.

The Importance of More Interpretability

4min

36.

Can We Sort of Figure Out Which Bits Are the World Model and Which Are the Sent?

5min

37.

The Forertion of Debate - Is There a Floor?

1min

38.

Is Debate Really a Good Idea to Train Ema Systems?

3min

39.

How to Train Houses Old Attractively Train an Ame System

2min

40.

Is There a Difference Between Deliberation and Distillation?

2min

41.

Ida, Is It a Good Plan?

3min

42.

What's the Bad Thing About Debate?

3min

One proposal to train AIs that can be useful is to have ML models debate each other about the answer to a human-provided question, where the human judges which side has won. In this episode, I talk with Beth Barnes about her thoughts on the pros and cons of this strategy, what she learned from seeing how humans behaved in debate protocols, and how a technique called imitative generalization can augment debate. Those who are already quite familiar with the basic proposal might want to skip past the explanation of debate to 13:00, "what problems does it solve and does it not solve".

Link to Beth's posts on the Alignment Forum: alignmentforum.org/users/beth-barnes

Link to the transcript: axrp.net/episode/2021/04/08/episode-6-debate-beth-barnes.html

Remember Everything You Learn from Podcasts

Save insights instantly, chat with episodes, and build lasting knowledge - all powered by AI.