OpenAI is set to release its latest and most powerful AI model, Astra, after delays to strengthen safety protocols. However, researchers fear it might be a step backward for AI security, as Astra uses a less transparent technique that could make it harder to monitor and spot potential threats.

Unlike typical transformer models that “think out loud,” Astra employs a recurrent depth or looped transformer, which keeps its reasoning internal and less like natural language, making it more difficult for researchers to track and intervene. This could be a race to the bottom, with other developers following suit to gain an edge.

OpenAI's response has been ambiguous, with chief scientist Jakub Pachocki suggesting the increased opacity is not as dramatic as some fear. Nevertheless, the concern is that the move could undermine our ability to oversee AI, leading to potential safety disasters.

Meanwhile, other top AI systems continue to use chain-of-thought methods, providing a stark contrast to Astra's approach. This could lead to a situation where advanced AI models are increasingly opaque, making it harder for researchers to detect and prevent undesirable behavior.

The release of Astra highlights the ongoing tension between pushing AI boundaries and ensuring safety. As the race to develop ever more powerful AI intensifies, the question remains: are we trading transparency for performance, or are we setting ourselves up for a potential disaster?