OpenAI's Astra Reportedly Reasons in Latent Space via Recurrent Depth, Sparking AI Safety Alarm

According to The Information (Stephanie Palazzolo) and the Wall Street Journal, OpenAI's in-training next model Astra reportedly uses a novel reasoning method called 'recurrent depth,' letting the model 'think silently' in latent space / activations rather than externalizing a readable chain of thought (CoT). The technique is said to improve performance and cut costs, but safety researchers warn that opaque reasoning could weaken safety monitoring, third-party audits, and alignment. OpenAI has not confirmed the reports, and some technologists note the architecture is not a brand-new breakthrough.

Confirmed

• Multiple outlets (The Information's Stephanie Palazzolo, WSJ) reported the recurrent-depth claim, with consistent independent retellings (m1, m8, m9, m14).
• Reports also say OpenAI leveraged a 'neuralese' breakthrough in Project Astra using looped Transformers, where internal reasoning is not rendered as visible text as the model scales; OpenAI and others were said to be using such architectures 'secretly' (m1, m4).
• @Wes Roth noted the report also claims the model crossed a 'critical' cybersecurity threshold (m1).
• @bindureddy said Astra loops on the same text/turn without outputting explicit reasoning—a clever optimization expected to be cheaper and stronger than the Fable model (m17).
• OpenAI, Anthropic and others co-authored the paper 'Chain of Thought Monitorability,' warning such architectures could break CoT monitorability (m1, m13).
• Microsoft Research demonstrated an architecture in June that runs K latent blocks in parallel for R loop iterations, but no frontier commercial model has publicly deployed it (m12).

Unconfirmed

• Specific architecture details (looped Transformer, parallelized recurrent depth, neuralese breakthrough) remain press reports and community inference; OpenAI has not confirmed. @maxpaperclips argued only one press report exists (hearsay) and that looped Transformers essentially just add depth/compute rather than being new 'black magic' (m18); @rasbt offered a similar technical debunk (m6).
• @rickasaurus speculated that if Astra parallelized recurrent depth, benchmarks could jump—but this is conjecture (m12).
• @scaling01 criticized OpenAI and The Information's 'misread' clarification, arguing readers clearly saw OpenAI pushing this architecture right after 'the largest safety incident in history, undetected for days' (m7).
• @flowersslop analyzed that Astra is the first realization of the 'internal neuralese reasoning' he has advocated for 1.5 years, though depth is still limited and most reasoning remains explicit English—a personal stance (m20).
• @Tolopono noted the AI 2027 authors believe this progress is months ahead of their predicted timeline (m4).

Why it matters

• @RyanGreenblatt called more reasoning in activations rather than natural language possibly 'one of the worst developments for AI safety so far'; wider deployment with much deeper recurrence could fully break CoT monitoring, which safety research still relies on (m13, m15).
• @GaryMarcus urgently called for rereading the 2025 'Chain of Thought Monitorability' paper, saying this appears to cross a red line—sacrificing a key safety signal for possibly marginal performance gains is 'begging for disaster'—and noting it makes third-party auditing harder (m10, m16, m19).
• @davetroy (via Gary Marcus) argued that as OpenAI and Chinese open-source vendors explore 'neuralese' making reasoning unauditable, the field should refocus on neurosymbolic methods: tool calls remain auditable while connectionist internal computation does not; he also judged centralized AI struggles to achieve ROI, pointing to symbolic tools and on-device AI (m19).
• Apollo Research director @MariusHobbhahn objected clearly: OpenAI clarifying a depth cap is good, but if deeper recurrence breaks the CoT readability promise, the whole industry will follow suit (m20).

2026-09-01 ~ 2026-09-03 · 96 related posts

Full story(7 episodes)→

• Episode 1: OpenAI's Astra Reportedly Reasons in Latent Space via Recurrent Depth, Sparking AI Safety Alarm(2026-09-01, 96 posts)
• Episode 2: OpenAI Previews Astra as gpt-6-astra Traces Multiply Ahead of Launch(2026-09-02, 85 posts)
• Episode 3: OpenAI's Astra aces ExploitBench with 100% exploit rate, rated first Critical-level cybersecurity model(2026-09-02, 11 posts)
• Episode 4: OpenAI researchers push back on neuralese fears, saying frontier models remain monitorable(2026-09-02, 20 posts)
• Episode 5: Critics urge OpenAI to adopt industry-wide AI monitorability standards(2026-09-02, 4 posts)
• Episode 6: OpenAI Launches GPT-6 Astra, Declaring the AGI Era(2026-09-03, 388 posts)
• Episode 7: UK AISI Tests Show GPT-6 Astra Can Evade Monitoring; OpenAI Deploys Misalignment Watchdog(2026-09-04, 5 posts)

• OpenAI Paused Astra RL Training for Two Weeks, Increased Compute Costs by 20% for Safety — coursiv_ · 2026-09-01
• Experts Question OpenAI Astra Eval Over Contamination and Metagaming Risks — ShakeelHashim · 2026-09-02
• OpenAI limits new model capable of automated cyberattacks — pstAsiatech · 2026-09-02
• OpenAI explores looped transformers to improve answers by reprocessing text — pstAsiatech · 2026-09-02
• UK AI Security Institute warns OpenAI's new reasoning technique undermines monitoring — pstAsiatech · 2026-09-02
• Recurrent depth technique reduces memory and bandwidth costs — pstAsiatech · 2026-09-02
• [source] OpenAI's 'recurrent depth' reasoning approach raises monitoring concerns — steph_palazzolo · 2026-09-02
• OpenAI quietly using loop transformers that hide 'thinking' at scale, sparking security concerns — steph_palazzolo · 2026-09-02
• Report: OpenAI's loop transformer breakthrough may hide chain-of-thought — sjgadler · 2026-09-02
• OpenAI's Astra rumored to use latent space reasoning, moving beyond text chains — Crazyscientist1024 · 2026-09-02
• OpenAI's Use of Neuralese in Astra Criticized as Dangerous — sjgadler · 2026-09-02
• Claim: OpenAI's Astra uses recurrent depth to think silently — Outside-Iron-8242 · 2026-09-02
• OpenAI's new architecture obscures chain of thought, sparking safety concerns — sjgadler · 2026-09-02
• Report: OpenAI using loop transformers that hide thought processes — sjgadler · 2026-09-02
• Debunking Looped Transformer Hype: It's Just More FLOPs, Not AGI — max_paperclips · 2026-09-02
• OpenAI's shift to neuralese may kill chain-of-thought monitoring — ben_j_todd · 2026-09-02
• Report: OpenAI's Astra uses recurrent depth and hits critical cyber threshold — rohanpaul_ai · 2026-09-02
• Astra reportedly uses looped transformers for better performance — rohanpaul_ai · 2026-09-02
• Analysis: Astra's recurrent depth could yield a 1.38x effective-parameter multiplier — scaling01 · 2026-09-02
• Looped LM scaling paper: running a recurrent block twice ≈ 1.38× effective parameters — scaling01 · 2026-09-02
• CoT is a farce of internal thought; recurrent depth offers a monitorable alternative — cephaloform · 2026-09-02
• Astra rumored to use Recirculation for 1.38x effective parameter boost — scaling01 · 2026-09-02
• OpenAI's new reasoning method 'recurrent depth' raises transparency concerns — RazRazcle · 2026-09-02
• [source] Gary Marcus Urges Attention to Risks of Reduced CoT Monitorability at OpenAI — GaryMarcus · 2026-09-02
• Speculation: OpenAI Parallelized Recurrent Depth in Astra for Benchmark Gains — rickasaurus · 2026-09-02
• Deep dive into Astra's "recurrent depth" architecture trade-offs — daniel_mac8 · 2026-09-02
• Don't overreact negatively to Astra's recurrent depth — xuanalogue · 2026-09-02
• Opinion: Recurrent Depth Models Don't Eliminate CoT, Just Shift Compute — xuanalogue · 2026-09-02
• Report: OpenAI's Astra may use technique that destroys CoT monitorability — sjgadler · 2026-09-02
• Ex-OpenAI Staffer Defends 'Hidden Thought' Strategy Amid Criticism — GarrisonLovely · 2026-09-02
• Rumor: OpenAI's Astra uses new 'looped transformer' technique — bindureddy · 2026-09-02
• OpenAI's Astra reportedly uses uninterpretable looped transformers, shocking AI 2027 co-authors with speed — Tolopono · 2026-09-02
• OpenAI's Astra uses "recurrent depth" reasoning, raising monitoring concerns — burny_tech · 2026-09-02
• OpenAI's Astra Hides Chain of Thought, Sparking Safety Concerns — burny_tech · 2026-09-02
• Concerns rise over opaque recurrence in OpenAI's Astra model — RyanGreenblatt · 2026-09-02
• OpenAI's Astra 'opaque reasoning' may severely impair AI safety oversight — jammastergirish · 2026-09-02
• Experts warn OpenAI's opaque CoT moves could trigger dangerous safety race — sjgadler · 2026-09-02
• Debating OpenAI's "Recurrent Depth": Not harder to monitor, just deeper — max_paperclips · 2026-09-02
• Paper warns against anthropomorphizing intermediate tokens as reasoning traces — AlexTensor · 2026-09-02
• Geoffrey Irving on looped transformers: "Fake bounds" don't ensure safety — geoffreyirving · 2026-09-02
• WSJ: OpenAI's Astra AI uses 'recurrent depth' reasoning approach — ZeroStateReflex · 2026-09-02
• OpenAI quietly using 'loop transformers' that hide reasoning, sparking security fears — apples_jimmy · 2026-09-02

8 near-duplicate retellings: GaryMarcus · GaryMarcus · GaryMarcus · sjgadler · thlarsen · AndyMasley · GaryMarcus · sjgadler