Training AIs to help us align AIs

Update: 2023-03-27

Description

If we can accurately recognize good performance on alignment, we could elicit lots of useful alignment work from our models, even if they're playing the training game: https://www.planned-obsolescence.org/training-ais-to-help-us-align-ais/

Comments

In Channel

OpenAI's CBRN tests seem unclear

2024-11-2113:08

Dangerous capability tests should be harder

2024-08-2008:43

Scale, schlep, and systems

2023-10-1009:24

Language models surprised us

2023-08-2908:36

Could AI accelerate economic growth?

2023-06-0604:59

The costs of caution

2023-05-0104:20

Continuous doesn't mean slow

2023-04-1203:53

AIs accelerating AI research

2023-04-0405:05

Is it time for a pause?

2023-03-3006:36

The ethics of AI red-teaming

2023-03-2702:27

Alignment researchers disagree a lot

2023-03-2703:23

Training AIs to help us align AIs

2023-03-2704:20

Playing the training game

2023-03-2707:53

Situational awareness

2023-03-2707:31

"Aligned" shouldn't be a synonym for "good"

2023-03-2706:14

What we're doing here

2023-03-2704:01

00:00

1.0x

Training AIs to help us align AIs

#box-pro-ellipsis-176620788139935{-webkit-line-clamp:2;}Training AIs to help us align AIs

Training AIs to help us align AIs

Ajeya Cotra

Training AIs to help us align AIs