Owain Evans

But yes, great ideas all around.

Owain Evans strikes again.

The range of what traits we know can be picked up subliminally has been expanded, including backdoor triggers.

As before, this only applies if the models share the same base model, so it is not as scary as it otherwise sounds.

Quote.

Owain Evans.

New paper.

Previous.

As before, this only applies if the models share the same base model, so it is not as scary as it otherwise sounds.

Owain Evans.

New paper.

Previous.

With a female name in the prompt, the teacher answers in French 95.5% of the time and the student 23.5%.

With a male name or no name, both are at 0%.

Owain Evans Subliminal learning can also transfer learned skills.

Quen is bad at counting letters.

Subheading.

James Chua, Jan Betley, Samuel Marx, Owain Evans, six thousand dollars.

When I wrote about evaluating grants, I said I wanted more work at the intersection of corrigibility and model welfare.

This paper is a great example of why, even though the word corrigibility is never used.