GPT-6 Astra's architecture: looped transformers and trust problems
A Sebastian Raschka breakdown of GPT-6 Astra and recurrent depth architectures got traction on HN. The key technical claim from The Information, that Astra uses 'recurrent depth' or looped transformer passes, is being analyzed seriously by the LLM internals community. Commenters pointed to benchmarks that OpenAI has acknowledged, though the article was flagged as not fully up to date.
Beyond the architecture discussion, several commenters noted that Astra in Codex has an off-putting tone, described as 'overzealous' and difficult to trust. This is a concrete usability signal: the model may benchmark well while still feeling wrong to use in practice.
The through-line connecting this to the DeepSeek and local model threads: the frontier is moving fast technically, but trust and feel remain unsolved problems. Developers are switching between models partly based on vibes, not just benchmarks, and that is a real signal about what is missing.
So what?
If you are building on Codex or Astra, user trust is not a given even when the model is technically capable. This matters for products where the AI output is visible to end users. Benchmark scores are necessary but not sufficient for adoption. Pay attention to the qualitative feedback loop from your users, not just accuracy metrics.