josh :)

Steer­ing Aware­ness: Models Can Be Trained to De­tect Ac­ti­va­tion Steering

A Sim­ple Method for Ac­cel­er­at­ing Grokking

• josh :) 26 Nov 2025 17:05 UTC

  1 point

  0

  on: Align­ment will hap­pen by de­fault. What’s next?

  It’s really difficult to get AIs to be dishonest or evil by prompting, you have to fine-tune them.

  Even if it’s hard to get current AIs to be evil by prompting, that doesn’t really remove the alignment problem. If AGI models are widely available and fine-tuning is accessible, someone will eventually fine-tune one specifically to be deceptive or malicious. Making that hard or impossible is exactly part of the alignment/​safety challenge, not something outside of it.

Train­ing Models to De­tect Ac­ti­va­tion Steer­ing: Re­sults and Implications