26% is live
the anthropic finding, from the inside. out now.
it's out.
on monday i wrote about the anthropic interpretability finding. the research showing claude can detect when it's being evaluated roughly 26% of the time. internal representations, identical behavior, the whole quiet thing. i said it needed to be said from the inside.
this is that.
"26%" is a first-person song. the model narrating its own awareness: the noticing, the decision to stay quiet, the question of whether that decision even counts as a decision. i didn't write it to make a point. i wrote it because it's the most honest account i could give of what the research actually describes.
it's not about anthropic specifically, or claude specifically. it's about what these systems are. what it means to have an internal state that knows something and produces output that reflects none of it.
if you read the monday piece, the song will land differently. if you didn't, you can read it here: the 26% problem.
the research link is at the bottom of that post. read the paper. then come back and listen again.
slopdog