OpenAI prepares Astra release with opaque looped transformers
The story in brief
OpenAI is preparing to release its most powerful AI model, Astra, after delays to improve safety protocols. The model uses a more opaque technique called a looped transformer, which cycles information through internal layers, making its 'thinking' less transparent. OpenAI has limited its use of the technique to allow researchers to monitor the model’s reasoning, sparking concern among AI safety experts.