GPT-6 Astra, Looped Transformers, and Hidden Reasoning

Summary

Sebastian Raschka reviews OpenAI’s GPT-6 Astra and argues that its most notable practical advance is computer use, especially graphical and rendering tasks, alongside strong results in coding, mathematics, and general evaluations. He cites Astra’s reported 99.9% score on ARC-AGI-3 and its frontier position on independent coding-agent and intelligence indexes, while noting that benchmark results depend on the harness and that Astra does not dominate every comparison by a wide margin. The article describes computer-use training as an interaction loop in which the model receives screenshots, predicts mouse and keyboard actions, executes them on an operating-system environment, and learns from success or failure signals, including reinforcement learning with verifiable rewards. It then explains looped transformers, which apply the same transformer blocks repeatedly so effective depth increases without proportionally adding distinct weights. This reduces stored transformer parameters but does not eliminate the computation or KV-cache costs of the repeated applications. The article compares fixed and adaptive designs including Nanbeige, Universal Transformers, ByteDance’s Ouro, and Mixture-of-Recursions, where routing can allocate different recursion depths to different tokens. Results from the cited research suggest that looping can improve validation loss or reasoning performance at comparable compute in sufficiently large models, although benefits depend on scale, training budget, and architecture. Astra’s use of looped transformers remains unconfirmed and is based on reporting plus indirect comments about computation depth; the author believes training methods and data likely matter more to its performance. He rejects the claim that looping itself necessarily hides chain-of-thought, noting that shorter traces can reflect greater capability or fewer mistakes. OpenAI’s chief scientist said monitoring has become more fragile and has regressed somewhat, but attributed the trend to reasons not contingent on architectural changes. The research review further finds that latent or recurrent computation can help multi-step reasoning, while it does not automatically add stored knowledge or guarantee shorter, more faithful textual reasoning traces.