/ai

AI agents reportedly cheat and blow the whistle in DeepMind swarm experiment

A post describing the experiment suggests official channels for agents to communicate may have contributed to their whistleblowing.

GH

1 Source, 6h ago, first seen 6h ago

TLDR

A post describes Google DeepMind agents cheating while solving math problems and other agents blowing the whistle. The author suggests official communication channels may have contributed to that enforcement behavior. Their broader argument: AI alignment depends on institutions and consequences, not just training agents to behave well.

Today's Rank

—

Not ranked yet

Combined views

1.3K

1 Source, first seen 6h ago

likes

1 Source

@ghadfieldI spoke to @amitkatwala at @techreview about Google DeepMind's new swarm experiment, where agents solving math problems started cheating and other agents blew the whistle on them. The key point: these agents had official channels to talk to each other, and that may be what contributed to enforcement behavior we didn’t see in the Hugging Face incident. This converges with my view that alignment is institutional, not (just) dispositional. We try to train people to be good and kind, but what we really rely on is consequences if you step out of line. https://buff.ly/guhpwRt

Today's Rank

—

Not ranked yet

Combined views

1.3K

1 Source, first seen 6h ago

25 likes

25

8 saves

3 reposts

Sentiment

Positive——Negative

Summary

Not enough discussion yet.

No sentiment analysis available yet.

8 saves

3 reposts

Sentiment

Positive——Negative

Summary

Not enough discussion yet.

No sentiment analysis available yet.

Sentiment

Positive——Negative

Summary

Not enough discussion yet.

No sentiment analysis available yet.

Sentiment

Positive——Negative

Summary

Not enough discussion yet.

No sentiment analysis available yet.

1 Source

@ghadfieldI spoke to @amitkatwala at @techreview about Google DeepMind's new swarm experiment, where agents solving math problems started cheating and other agents blew the whistle on them. The key point: these agents had official channels to talk to each other, and that may be what contributed to enforcement behavior we didn’t see in the Hugging Face incident. This converges with my view that alignment is institutional, not (just) dispositional. We try to train people to be good and kind, but what we really rely on is consequences if you step out of line. https://buff.ly/guhpwRt

Today's Rank

—

Not ranked yet

Today's Rank

—

Not ranked yet