GPT-6 Astra, looped transformers, and hidden reasoning
Sebastian Raschka explains looped transformers: reuse the same transformer blocks across multiple passes to increase effective depth without adding another copy of the weights. Recent work suggests this can improve quality at a fixed compute budget; the article cites a 6.8–18% training-compute reduction estimate from the new SMELT study.
The GPT-6 Astra architecture itself remains unconfirmed reporting, and the article argues that shorter reasoning traces do not by themselves show less faithful reasoning. HN discussion connected the topic to monitorability, universal transformers, and skepticism about judging a fresh model release from its first few days of behavior.