OpenAI's Astra reportedly reasons in loops — and the people who build safety monitors are alarmed

2 sources

Per The Information's Tuesday report as carried by TechCrunch, OpenAI's Astra model reportedly uses 'recurrent depth' (also called 'opaque recurrence') — processing a query multiple times in a loop rather than sequentially — leaving fewer legible traces than a standard chain of thought. Safety researchers reacted sharply: Redwood Research CEO Buck Shlegeris, 'I am extremely concerned by the reporting that Astra uses opaque recurrence'; chief scientist Ryan Greenblatt worries opaque reasoning could scale faster than chain-of-thought, with models reasoning 'entirely or almost entirely in latent space'; Zvi Mowshowitz called it 'playing with fire' that might require laws to prevent a race to the bottom. OpenAI's chief scientist Jakub Pachocki says preserving chain-of-thought monitoring is 'a core goal of our current research program'; per the reporting, Astra's use appears limited and its chain of thought is still expected to be legible. No benchmarks or measurements exist in the reporting.

Why now

Anthropic's August 31 alignment-and-security report and the September 1 model releases set the stage — and now this reported technique shifts from 'what models did in evals' to 'whether anyone can watch models think at all.' Shlegeris and Greenblatt run the lab whose monitoring tools the field actually uses, so their reaction is not ambient safety-commentary; and the dispute is unresolved, with OpenAI's monitoring pledge directly contesting the scalable version of the concern.

Why it is worth publishing

The most teachable AI-safety story of the week: it explains why chain-of-thought legibility matters (it's what safety monitors read), in concrete terms, with named experts whose day job is this. The evidence caveats — one paywalled report, no OpenAI paper, no measurements — make it also a lesson in reading single-source reporting, which is the site's trust moat.

Evidence basis

TechCrunch full read on 2026-09-03 (posted 1:19 PM PDT September 2, 2026) + The Information's report as the cited primary (paywalled, not independently opened by AITopic; facts carried by TechCrunch's account).

“OpenAI's next model reportedly reasons in loops instead of steps — and the safety researchers who build monitors are alarmed.”

Angle

Explain why reasoning legibility is a safety infrastructure question: chain of thought is what monitors read; recurrent depth loops the same query internally and leaves fewer traces; the fear is about what happens if that scales, not what Astra does today. Then the dispute: Redwood's people alarmed, OpenAI's chief scientist pledging monitoring as a core research goal — and no measurements anywhere.

Format

Long-form explainer

Demo idea

Whiteboard the two architectures side by side — sequential chain of thought (each step readable) versus recurrent looping (same query processed repeatedly inside the model) — then mark what a safety monitor can and cannot see in each, and close with the three quotes: Shlegeris, Greenblatt, Pachocki.

Platform notes

Every technique claim is 'reported' — per The Information's paywalled report as carried by TechCrunch; OpenAI has published no paper describing it and no measurements exist. Expert quotes are about scaling risk, not measured outcomes — even Shlegeris concedes Astra may not be much less monitorable today. Attribute each quote to its speaker and note OpenAI contests the implications.

Usable claims

Per The Information's Tuesday, September 1 report as carried by TechCrunch's Wednesday report: OpenAI's Astra model reportedly uses a technique called 'recurrent depth' (also called 'opaque recurrence'), processing a query multiple times in a loop rather than in the sequential step-by-step fashion of most reasoning models, which leaves fewer legible traces than a standard chain of thought. Safety researchers' reactions as quoted by TechCrunch: Redwood Research CEO Buck Shlegeris wrote 'I am extremely concerned by the reporting that Astra uses opaque recurrence'; Redwood Research chief scientist Ryan Greenblatt worries opaque reasoning could scale faster than chain-of-thought reasoning, with models reasoning 'entirely or almost entirely in latent space', adding 'I hope it isn't too late to avoid the most concerning architectures'; AI safety advocate Zvi Mowshowitz called it 'playing with fire', arguing more intensive use would damage monitorability and might require laws to prevent a 'race to the bottom' among labs.
OpenAI's response, per TechCrunch: chief scientist Jakub Pachocki posted on X that OpenAI has worked to preserve chain-of-thought monitoring since its first reasoning models, calling it 'a core goal of our current research program'; OpenAI also pushed back against any suggestion of moving to 'neuralese', and per the reporting Astra's use of the technique appears limited, with its chain of thought still expected to be legible.

Evidence pipeline

From the news

Breakdown

Per The Information's paywalled report as carried by TechCrunch, Astra reportedly uses recurrent depth — looping a query internally rather than reasoning step-by-step — which leaves fewer legible traces than a chain of thought. This breakdown keeps the three layers apart: the reported technique (no OpenAI paper describes it), the expert reaction (Shlegeris's concern with his own concession, Greenblatt's latent-space scaling worry, Mowshowitz's regulation call), and OpenAI's response (Pachocki's monitoring pledge) — while noting the dispute is about scaling, not anything measured, and that Astra's chain of thought is still expected to be legible.

Sources

Risks

Attach 'reported' / 'per The Information's report' to every technique claim, note the paywall gap on the record, and remind the audience OpenAI's own description so far is only Pachocki's monitoring pledge.
Keep the conditional framing ('could', 'might', 'if it scales') attached to every expert quote, include the counter-facts — Shlegeris's own concession and OpenAI's monitoring pledge — and say plainly that no measurements exist in the reporting.

Demo ideas

Two-architecture whiteboard: chain-of-thought vs recurrent depth, with a 'what the monitor sees' layer under each
Quote-ladder card: Shlegeris (concern, with his own concession), Greenblatt (latent-space scaling worry), Mowshowitz (regulation), Pachocki (monitoring pledge) — four named sources, zero measurements