GPT-6 Astra, looped transformers, and hidden reasoning
Points and comments are a snapshot, not live.
Looped transformers reuse weights to increase effective depth without adding new parameters.
GPT-6 Astra excels at 3D rendering, animation, and computer-use tasks, achieving 99.9% on ARC-AGI-3. Looped transformers pass hidden states through the same blocks multiple times, increasing effective depth while saving memory. Nanbeige4.2-3B uses 22 blocks twice for 44 applications. The technique is not new, dating to 2018's Universal Transformers. OpenAI's computer-use training uses Macs as environments for reinforcement learning with verifiable rewards.
Claims that looping hides chain-of-thought reasoning are overstated; it still produces one token per forward pass. The model likely sits on NVIDIA GPUs.
What commenters are saying
Commenters mostly agree that looped transformers are not a scary new technique, just weight reuse to save memory. The top comment notes monitoring chain of thought remains possible since the model still outputs one token per forward pass. A counterpoint argues that if loop depth is determined dynamically between tokens, the model could compute arbitrary functions, making reasoning harder to interpret. Others doubt this is happening at scale, as open-weight models would likely show it. A minor camp suggests removing CoT could improve safety by enabling analysis of the full output landscape.