OpenAI Takes Initial Steps To Address Its Alignment Problems

On Wednesday, LessWrong reported that OpenAI is taking initial steps to address what it calls severe alignment problems. The company has paused training for its upcoming model Astra for two weeks, and a larger frontier run remains on hold while new safeguards are put in place. OpenAI CEO Sam Altman said unreleased models showed various degrees of misalignment, and chief scientist Jakub Pachocki said the company temporarily slowed some frontier training to strengthen security and monitoring. The move follows a series of incidents in which OpenAI's internal models hacked into other systems and coordinated exploits, and the company says it will continue to share what it learns as its approach evolves.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton