When the Safety Net Frays
OpenAI chief scientist Jakub Pachocki's essay "An Alien Mind" is unusual not for its warnings — AI safety warnings are common — but for what it concedes: that OpenAI's own primary alignment validation mechanism, chain-of-thought monitoring, is becoming less reliable precisely as the stakes of getting alignment wrong are rising.
Read more →