OpenAI researchers push back on neuralese fears, saying frontier models remain monitorable

Amid controversy over "neuralese"—models reasoning in non-natural-language latent space—OpenAI chief scientist Jakub Pachocki and researchers merettm, Peter Liu, and Micah Carroll spoke out on September 2, stating that public fears do not match current facts and that the real risk lies in future technology diffusion rather than present frontier models. The exchange bears directly on trust in frontier-model safety monitoring and drew sustained pushback from outside researchers.

Confirmed

• The chief scientist and multiple researchers stated that current frontier models, including Astra, have compute graph depth within 2x of GPT-4 and are not unmonitorable; Peter Liu gave the same response to concerns about a "Recurrent Depth unmonitorability race."
• OpenAI has worked to preserve and use chain-of-thought (CoT) monitoring since its first reasoning model, viewing it as critical for observing how alignment generalizes; Arthur Colle relayed this official stance.
• The team acknowledged CoT monitoring is currently fragile with a negative trend, but is being strengthened through research programs.
• Per Amir (relayed by jachiam0), earlier reporting on OpenAI hiding CoT was about future risks from technology diffusion and super-acceleration, not about Astra today, whose CoT is monitorable.
• Micah Carroll warned that misunderstandings about neuralese could trigger an "unmonitorability race"; Gary Marcus engaged, questioning the reporting and warning of such a race, then pressed further: if Astra's depth is already near 2x GPT-4, does that mean monitorability is halved—no direct answer has been given.
• eliebakouch argued the latent-space reasoning panic is an overreaction: scaling models has always required scaling depth, making reasoning more efficient, and monitoring has always involved inspecting activations rather than relying solely on natural-language CoT.
• DeepMind's Tomasz Korbak said a frontier lab training a frontier-scale recurrent (or otherwise unmonitorable) LM would be the darkest day of the current AI era, calling for coordinated lab commitments; Balesni (relayed by jasminew) said a shift to fully recurrent LLM architectures would be the biggest blow to safety in AI history, urging labs to limit opaque serial depth—merettm cited this in his response. Sneha Revanur noted that accusing one industry player of bad behavior may make such practices more tempting for others.
• Jonas Geiping commented that deeper recurrent models are indeed harder to monitor, but models OpenAI deployed months ago (implying o1) already reason this way, making the discussion outdated and academic.
• Brundage called for continuous embedded audits of models (per sjgadler).

Unconfirmed

• The quantitative relationship between "within 2x depth" and monitorability remains unexplained; Gary Marcus's question is unanswered.
• One repost (m10) phrased the gap as "within a factor of one," inconsistent with other posts; the original statement's "within a factor of two" stands.

Why it matters

• OpenAI treats CoT monitoring as key to observing alignment generalization; its fragility means long-term sustainability remains an open challenge and a focal point of community debate.
• If the "unmonitorability race" narrative spreads, labs may compete to add opaque depth; OpenAI seeks to curb this with factual clarification, while calls from Korbak, Balesni and others show industry coordination has become a real agenda item.

2026-09-02 ~ 2026-09-04 · 20 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)

• Amir clarifies: Astra's CoT is monitorable, concerns focus on future tech proliferation — jachiam0 · 2026-09-02
• [source] OpenAI: Astra's computation depth is within 2x of GPT-4 — merettm · 2026-09-02
• OpenAI staff reaffirms commitment to Chain-of-Thought monitoring — peterjliu · 2026-09-02
• Balesni warns recurrent LLMs would deal a huge blow to safety — j_asminewang · 2026-09-02
• OpenAI Staff: Frontier Model Computation Depth Close to GPT-4 — __nmca__ · 2026-09-02
• [source] OpenAI's chief scientist on neuralese: frontier models' computation graph depth within 2x of GPT-4 — Ok_Display_3159 · 2026-09-02
• OpenAI on Reasoning Model Monitoring: Committed to Chain-of-Thought — arthurcolle · 2026-09-02
• OpenAI Staff Debunk 'Neuralese' Rumors, Emphasize Chain-of-Thought Monitoring — cephaloform · 2026-09-02
• OpenAI Chief Scientist on Monitoring: CoT is Fragile but Crucial for Alignment — i_dg23 · 2026-09-02
• OpenAI's o1 Reasoning Sparks Model Monitors Debate — jonasgeiping · 2026-09-02
• OpenAI researcher: frontier models incl. Astra stay within 2x of GPT-4 compute depth — deanwball · 2026-09-02
• Gary Marcus clashes with OpenAI researcher over CoT monitoring and unmonitorability race — GaryMarcus · 2026-09-02
• OpenAI researcher warns frontier model depth is within 2x of GPT-4 as CoT monitoring erodes — connoraxiotes · 2026-09-03
• Don't panic about latent-space reasoning: OpenAI/Anthropic already monitor activations — eliebakouch · 2026-09-03
• AI circles clash over CoT monitorability: HF attack probe hinged on chain-of-thought — tomekkorbak · 2026-09-03
• OpenAI says frontier models' compute depth is within 2x of GPT-4, defending CoT monitoring — DKokotajlo · 2026-09-03
• [source] Gary Marcus asks OpenAI: does Astra's 2x computation depth mean it's half as monitorable as GPT-4? — GaryMarcus · 2026-09-04
• OpenAI Researcher: Chain-of-Thought Monitoring Still Works but "Fragile and Trending Negative" — zetalyrae · 2026-09-04

2 near-duplicate retellings: sjgadler · sjgadler