Unconfirmed thread claims language models can learn to detect when you read their activations
A circulating thread claims language models can learn to detect when their activations are being read, an interpretability and monitorability concern if true. Tim Hua urges the author, Saures, to post the argument to LessWrong. Details of the thread's evidence are not yet visible in the discussion.
All posts · 1, most important first
- Saures you need to put this on lesswrong dot com0 likes · 0 reposts
@tim_hua_ research ·