The Inside View

3. Evan Hubinger on Takeoff speeds, Risks from learned optimization & Interpretability

Jun 8, 2021

01:44:24

Creator website

Highlights

AI Chapters

Episode notes

How to Apply More Oftinization Power to a Task?

02:34

How to Hack the Human Brain?

01:50

Why Evolution Would Rather Have Each Individual See More Altruistic?

02:04

Introduction

2min

Cocoanuts

2min

Compiling Python

2min

I Think That's a Good Thing.

2min

How Fast Do We Build Really Powerful Ai Systems?

5min

How Much Does It Cost?

2min

Gata Data Collection Procedure

2min

Is It Hard to Keep Your Basic Method Secret?

2min

How to Keep Your Weights Secret

3min

10.

Using Ai Thay, I'm a Liman, and I'm Not Sure if It's Alined, or Not.

3min

11.

Intent Enlightenment

2min

12.

Is Human Coordination Really a Problem?

3min

13.

The Human Is Like Drinking Alcohol at Tensopa

4min

14.

Optimization Power and Quantilizes

2min

15.

How to Apply More Oftinization Power to a Task?

3min

16.

How to Hack the Human Brain?

2min

17.

Is There a Strategy for Tomorrow?

2min

18.

Defining Multile Times Intelligence

2min

19.

What Is an Optimizer?

4min

20.

Are We Really Proxies for What Evolution Cares About?

2min

21.

Evolution Isn't Just Selecting for Life

2min

22.

Why Evolution Would Rather Have Each Individual See More Altruistic?

2min

23.

Evolution, Evolution, and Altruism in Machine Learning

2min

24.

Issoganization and Mazes

3min

25.

Is Inter Linement Good at Finding Green Arrows?

3min

26.

A, Ter I Memorize, and It's Not Computed Oftima

2min

27.

Is There a Sesation?

3min

28.

Is It in the Right Side for You?

2min

29.

Capability Rebustness

2min

30.

Robustness

5min

31.

Xplaying Magic

2min

32.

I Think It's Resnly Accurate, I Think It Was Aly Accurate.

2min

33.

Is It Worth the Gains Safety?

2min

34.

Transparency Training

3min

35.

Amplified Overseader and Interpretability Space

2min

36.

Iithink, U Visualized Features Ind Coing Ron Anda

1min

37.

Transparency Over Ampification

2min

38.

Is There Any Empirical Environment or Research That Can Tell Us if Something Works?

3min

39.

Imitative Amplication

2min

40.

Is It Like Humans Consulting Humans?

4min

41.

Recursive Reward Modeling - What's the Difference?

2min

42.

The Human Feed Bag, Is It a Human Feet Bag, or a Reward Model?

3min

43.

I'm Preparing to Crystalas Work on Ceays, Right?

2min

44.

Dolly or Not?

3min

We talk about Evan’s background @ MIRI & OpenAI, Coconut, homogeneity in AI takeoff, reproducing SoTA & openness in multipolar scenarios, quantilizers & operationalizing strategy stealing, Risks from learned optimization & evolution, learned optimization in Machine Learning, clarifying Inner AI Alignment terminology, transparency & interpretability, 11 proposals for safe advanced AI, underappreciated problems in AI Alignment & surprising advances in AI.