Reading at Beginner level
OpenAI is about to release a powerful new AI model called Astra, but some safety experts are worried about how it's built. The company used a new method that makes the AI work faster and cheaper by using less computing power. The catch: part of how the AI thinks—the steps it takes to reach its answers—is now hidden from human view, making it much harder for people to understand what the AI is doing and why.
Currently, one of the main ways companies protect against AI going rogue is by reading and checking these thinking steps, kind of like looking over someone's shoulder to see their work. OpenAI's chief scientist, Jakub Pachocki, says the company has limited this hidden reasoning and still cares about monitoring. But other experts, including former OpenAI safety researcher Steven Adler, worry that if other AI companies copy this design, we could eventually have AI systems that are completely impossible to understand—a problem that may be impossible to fix later.