# OpenAI Astra “Secret Technique” Actually a Decade Old > Source: > Published: 2026-09-03 09:19:33+00:00 In 1836 Edgar Allan Poe published “Maelzel’s Chess-Player” in the *Southern Literary Messenger*. Johann Maelzel had been touring the United States with Kempelen’s automaton for a decade, and the American press treated its mechanism as an unsolved marvel. Poe’s essay is remembered as being a remarkable feat of deduction. Instead it was mostly a feat of reading. Brewster’s *Letters on Natural Magic* had explained the concealed operator in 1832; Racknitz had published his diagrams almost a half century before in 1789. Poe reasoned from what was in print and from what he could watch from the audience: the pattern of the cabinet doors, the operator’s posture, the timing of the moves. The secret was available for anyone who could read. The only new part was someone, a reporter, actually checked. This is what comes to mind when OpenAI Astra news flies around today. *The Information* just published a piece by Amir Efrati, Stephanie Palazzolo and Rocket Drew describing a “secret technique” in OpenAI’s forthcoming model. How secret is it? The technique is recurrent depth: a transformer passes its hidden state through the same block of layers more than once before emitting a token, rather than through a fixed stack once. The report’s substantive concern is that computation performed this way is not written out as a chain of thought and is therefore harder to monitor. As someone who builds a harness that specializes in auditing models, [Wirken.AI/](https://wirken.ai/), I’m of course all over this concern. The architecture is NOT secret. It is public, and has been so a very long time. The report itself attributes it to “*several American and European academic researchers*” in a paper published last year. Well, duh, then it can’t be a secret, can it? That is [Geiping et al.](https://arxiv.org/abs/2502.05171), February 2025, from Tübingen, Maryland and Livermore, presented at NeurIPS 2025 with weights and training code released. But wait, let’s read farther into the past. [Graves, 2016](https://arxiv.org/abs/1603.08983), at DeepMind, introduced adaptive computation time for recurrent networks. [Dehghani et al., 2018](https://arxiv.org/abs/1807.03819), at Google Brain, applied it to transformers under the name Universal Transformer. [Giannou et al., 2023](https://arxiv.org/abs/2301.13196), supplied the term “looped transformer.” That’s a decade. And it’s not like others are unaware. [Nanbeige4.2-3B](https://huggingface.co/Nanbeige/Nanbeige4.2-3B) shipped under Apache-2.0 with a technical report dated 27 July 2026, five weeks before the OpenAI story, and a model card stating that its looped transformer architecture reuses layers to increase capacity without adding parameters. The implementation is in the repository’s modeling file. The Nanbeige card records roughly forty thousand downloads in the past month. Worst secret ever? **Same firms, prior disclosure** The report also notes that a July 2025 joint statement on chain-of-thought monitorability from OpenAI, Anthropic and Google DeepMind cited the exact Geiping paper. That is [Korbak et al.](https://arxiv.org/abs/2507.11473), which warned that latent reasoning models might not need to verbalize their thoughts and would lose the safety property that legible chains of thought provide, and recommended that developers document any decision to adopt such an architecture. The technique was therefore identified, by name and under OpenAI’s own authorship, as a monitorability risk fourteen months before the report. What’s new? OpenAI did it anyway. The ethics are the story. **Anonymous** The technical claims rest on a single unnamed person “*with knowledge of Astra’s development*“: that recurrence was used in both training and inference, that its use was limited in order to [preserve a legible chain of thought](https://www.flyingpenguin.com/metr-dfir-role-boeing-lobbyist-to-wear-ntsb-badge/), and that the agents involved in the July intrusion into OpenAI’s cluster ran on a related model. The novelty claim has no source at all. No document, architecture description or benchmark is cited. The only named technical statement on the record came afterward from OpenAI’s chief scientist, Jakub Pachocki, who said the depth of Astra’s computation graph is within a factor of two of GPT-4’s and that he wished to avoid a race toward unmonitorability driven by confused reporting. The vendor’s own scientist publicly declined the novelty claim. That is the Anthropic Mythos record scratch again, in mirror image. There the vendor made capability claims nobody outside could verify. Here an anonymous source made claims the vendor’s own scientist rejected on the record. In both cases the reader is left with assertions structured so that investigation ends in access denied. **Falsifiable fluffy advances** The report states that recurrent depth “*hasn’t been featured in a major commercially available large language model before*.” What is Nanbeige then, mashed potatoes? It is Apache-2.0, built by the lab of a Chinese recruitment platform, serving forty thousand downloads a month through Hugging Face, which by The Information’s own reporting the week before, on the word of another anonymous source, Nvidia has agreed to buy for around $13 billion. Whether that counts as major or commercially available is a weird question. If you take away the OpenAI-is-special-snowflake qualifiers then there’s nothing new here either. The report frames Astra against the need for a visible technological advance: three cloud providers are spending roughly $600 billion this year on capital expenditure justified by expected model improvements. Recurrent depth does not support that framing. The documented effect, in the reporters’ own description, is to let a smaller model perform like a larger one. Geiping’s model is 3.5 billion parameters; Nanbeige’s is 3 billion. [Saad-Falcon et al.](https://arxiv.org/abs/2511.07885), at Stanford, with Alphabet chairman John Hennessy among the authors, measured in November 2025 that local models answer 88.7% of real single-turn chat and reasoning queries. A technique whose measured contribution is parameter efficiency lands even more evidence of that trend. Capital expenditure, which the report invokes, is the opposite of the trend. The reporters wrote down these facts yet for some reason could not connect the dots. **Known knowns** A decade-old architecture, published in peer-reviewed venues, flagged by the industry’s own safety statement, and shipped in popular open weights, was presented as a secret on the authority of one anonymous source. OpenAI’s chief scientist then disputed the novelty in public. That sequence shows the integrity problem in AI reporting: a claim of novelty was not checked against the prior work the report itself cites.