Report: OpenAI's Astra uses recurrent depth, making reasoning harder to monitor

AI寒武纪 · wechat · 2026-09-02

According to The Information's Stephanie Palazzolo, OpenAI's next flagship model Astra employs a "recurrent depth" inference method: the same set of Transformer layers iterates on hidden states so part of the reasoning happens in latent space, cutting costs and improving token efficiency—but making the thinking process harder to monitor, raising AI safety and observability concerns. Chief scientist Jakub Pachocki responded that frontier models (including Astra) have computation-graph depth at most 2x GPT-4, and CoT monitoring remains a core research focus; OpenAI has not confirmed architecture details. Notably, Tsinghua and ByteDance Seed released the paper "SMELT: Scaling Laws for Compute-Matched MoE Looped Transformers" the same day, scaling looped architectures to 54B non-embedding parameters with FLOPs, total parameters, and KV cache all strictly matched, showing 6.8%–18.0% training-compute savings on the compute-optimal frontier, with looping twice optimal and gains strongest on code (CE Gain up to 20.4%). The article also covers the technique's principles (Geiping et al.'s 3.5B Huginn-0125 matching 50B fixed-depth compute when looped), the "neuralese" monitorability debate, and implications for Astra.

Original post →

Report: OpenAI's Astra uses recurrent depth, making reasoning harder to monitor

More from Models