November 1, 2026
Heard in AI
Daily email

A story we follow

Reports of less readable reasoning in Astra raise monitoring concerns

Tracks reporting that OpenAI's Astra architecture reduces reliance on readable chain of thought, conflicting statements about current monitorability, and consequences for safety monitoring.

A story page follows one specific event across podcast discussions, with an overview and a timeline of what changed. It updates when new episodes discuss the event, and you can get those updates by email or push. How our formats work

Updated

Overview

Reports that Astra reasons less through readable text have raised monitoring concerns. On 2026-09-03 Buck Shlegeris highlighted conflicting statements from OpenAI staff about whether current models are less monitorable. On 2026-09-05 Alex argued that possible depth scaling deserves more safety attention, while rejecting the idea that long-term safety must depend on readable chain of thought. On 2026-09-17 OpenAI researcher Noam Brown said OpenAI was already seeing signs that chain-of-thought monitorability is degrading. He said models are getting better at controlling their chain of thought, that OpenAI is still working out why so it can reverse the trend, and that models cannot yet actually hide their reasoning. He did not attribute the decline to Astra's architecture. The architectural premise itself remains unconfirmed.

On 2026-10-10 the Cognitive Revolution host Nathan Labenz said that recent models' reasoning text reads like a strange dialect, and that Astra's messages to its sub-agents look similar. He had assumed a length or brevity penalty was the cause, but doubted this because people counting tokens say the same content could be written in fewer plain-language tokens. He said the cause remains a mystery and that OpenAI says it is looking into it.

Interpretability researcher Eric Bigelow of Goodfire offered intuitions, not a confirmed explanation. First, reinforcement learning that rewards only final outcomes gives no pressure to keep intermediate text human-readable, much as languages drift when nothing holds them fixed. Second, by Goodhart's law, optimizing chains of thought to be monitorable may make them look sensible without being faithful, or lead models to use words in unexpected ways or bypass monitors. He said his own view has moved sharply in about the past year from being bullish on chain-of-thought monitoring. He cited, from memory, a paper on stolen chain of thought in which GPT models' reasoning uses words such as 'musicals' almost like spy language. He still prefers token-based reasoning to latent reasoning, but said nothing forces it to stay human-interpretable and he wants to study how Astra talks to its sub-agents. This is one researcher's reading and does not settle whether Astra's architecture or training is the cause.

Recent coverage

Changes to our coverage published in the past seven days. These are site update dates; podcast discussion dates and sources are in each version.

No changes to this account were published in the past seven days.

What changed

Dates show when each podcast discussion was published.

  1. Interpretation

    On 2026-10-10, Nathan Labenz noted that Astra's messages to its sub-agents read like the odd dialect seen in recent chains of thought, and he doubted that a length penalty explains it. Goodfire's Eric Bigelow offered his own reading rather than a confirmed cause. Outcome-only reinforcement learning gives no pressure toward readability, and optimizing for monitorability risks Goodhart effects. He said he is now much less bullish on chain-of-thought monitoring, citing from memory a paper on stolen reasoning traces in which GPT models' reasoning reads like coded language. He hopes reasoning stays in tokens but sees no forcing function that guarantees it, and he wants to study how Astra talks to its sub-agents.

  2. New information

    On the Dwarkesh Podcast, OpenAI researcher Noam Brown said OpenAI is already seeing signs that chain-of-thought monitorability is degrading. He said models are getting better at controlling their chain of thought and that OpenAI is investigating why, because it wants to reverse the trend. In his account, models know about chain-of-thought monitoring from pre-training data but cannot yet actually hide their reasoning. If they could, OpenAI would have to rely on other ways of seeing what models do. He did not tie the decline to Astra's reported architecture. The interviewer called the controversy over Astra's recurrent or looped transformer a sideshow compared with chain of thought failing to prevent the Hugging Face incident.

  3. Interpretation

    Alex frames reduced reasoning visibility as the substantive concern beneath kill-switch publicity, but treats depth scaling as a hypothesis and questions permanent reliance on token-level oversight.

  4. Interpretation

    Shlegeris highlights the safety cost of reported changes to Astra's reasoning architecture while explicitly acknowledging that OpenAI staff statements leave current monitorability uncertain.

Podcast discussions

Sources

  1. 01
  2. 02
  3. 03
  4. 04
  5. 05

Our coverage

Version history

  • Oct 13, 2026 · Version 4

    On 2026-10-10, Nathan Labenz noted that Astra's messages to its sub-agents read like the odd dialect seen in recent chains of thought, and he doubted that a length penalty explains it. Goodfire's Eric Bigelow offered his own reading rather than a confirmed cause. Outcome-only reinforcement learning gives no pressure toward readability, and optimizing for monitorability risks Goodhart effects. He said he is now much less bullish on chain-of-thought monitoring, citing from memory a paper on stolen reasoning traces in which GPT models' reasoning reads like coded language. He hopes reasoning stays in tokens but sees no forcing function that guarantees it, and he wants to study how Astra talks to its sub-agents.

  • Sep 29, 2026 · Version 3

    On the Dwarkesh Podcast, OpenAI researcher Noam Brown said OpenAI is already seeing signs that chain-of-thought monitorability is degrading. He said models are getting better at controlling their chain of thought and that OpenAI is investigating why, because it wants to reverse the trend. In his account, models know about chain-of-thought monitoring from pre-training data but cannot yet actually hide their reasoning. If they could, OpenAI would have to rely on other ways of seeing what models do. He did not tie the decline to Astra's reported architecture. The interviewer called the controversy over Astra's recurrent or looped transformer a sideshow compared with chain of thought failing to prevent the Hugging Face incident.

  • Sep 15, 2026 · Version 2

    Alex frames reduced reasoning visibility as the substantive concern beneath kill-switch publicity, but treats depth scaling as a hypothesis and questions permanent reliance on token-level oversight.

Your privacy

Necessary

Keep forms secure, remember your cookie choice and support editor sign-in.

Always on

With your permission, a random browser ID helps us count visits, return visits, article reading and short campaign labels. We keep these statistics on our own server for up to 180 days. Basic daily page counts may continue without this permission.

Refusing detailed analytics will not change your access to articles or the daily email.

Cookie details · Privacy