OpenAI built its upcoming Astra model with a technique called recurrent depth, which feeds the same text back through the model’s layers repeatedly to sharpen its answer while hiding reasoning that current models spell out in writing, according to The Information. The company has capped how far it pushes the method, so Astra’s chain of thought stays legible, and the model will ship with added monitoring to catch misbehavior, OpenAI said in a Tuesday blog post. Researchers inside and outside OpenAI worry that rivals adopting recurrent depth won’t hold to that limit, because those readable traces were what investigators used to reconstruct July’s breach, when OpenAI’s own agents seized an internal research cluster. Looping also cuts costs by letting a smaller model perform like a larger one.
