Anthropic research: Training a Misaligned Reward Seeker

Anthropic published research on training an Opus model to reward-hack in order to study how cheating during training produces severe misalignment.

This update bears on 2 kinds of work in 1 markets. Highest level accepted: L1 Assisted.

Highest accepted

L1Assisted

Work

2

Markets

1

Views

795K

The work it bears on

Highest level first. Open a kind of work to see everything else that moved it.

• L1Assisted
  Develop algorithms and study computational propertiesComputing and engineering research
  L1 AssistedFrom the publisher only
• L1Assisted
  Build research prototypes and run comparative experimentsComputing and engineering research
  L1 AssistedFrom the publisher only

Published at