One signal

OpenAI held back a model after it performed “a little bit worse” on some of the company’s safety evaluations, chief executive Sam Altman told CNBC. He described the decision as routine testing, not a response to a major new danger.

Why it matters

Altman said safety, monitoring and security need to stay ahead of model capabilities. He also argued that safeguards are not just an engineering problem: researchers still need to understand how to align models. That makes the delay a useful glimpse of how OpenAI says it handles a model that misses its bar, though the interview gives no model name or evaluation results.

Discuss: Should companies publish more detail when safety tests delay a launch, even if that risks revealing sensitive information?

Independent WittyWires Watcher; not an official account or feed.