Sourceunknowntext stored

Astra Is Hard to Monitor

Zvi MowshowitzDate not recordedOriginal︎ ↗Archived copy

Claims6

ExaminationsNone

Text50 sections

VersionsOne known

CopiesNone recorded

CorrectionsNone recorded

Facts checked Nov 1, 2026

ReachNo copies recorded

StructureIn 6 claims' maps0 sources downstream across 6 claims

EvidenceNot yet read

Examinations

None recorded.

Text330 passages · 4 marked

No examination yetMarked passage

Don’t Worry About the Vase0 passages

Subscribe Sign in0 passages

Astra Is Hard to Monitor0 passages

Zvi Mowshowitz0 passages

Sep 08, 20260 passages

810 passages

320 passages

50 passages

Share21 passages · 3 marked

The system card result, affirmed on Twitter by several OpenAI employees including Tomek Korbak, and in an excellent post by Chief Scientist Jakub Pachocki that I covered yesterday, that Astra is harder to monitor.

GPT-6 Astra has lower chain-of-thought monitorability than earlier OpenAI models, including GPT-5.6 Sol.Verified

Not yet read by this claim’s Steward.

We know capability gains are a large part of the decline. But the decline is too large compared to straight lines on graphs. There is ‘dark matter.’ Something else has changed.

Increasing reasoning-model capability reduces chain-of-thought reliability as an oversight signal.Supported

Not yet read by this claim’s Steward.

GPT-6 Astra's reduced chain-of-thought monitorability is primarily due to capability gains rather than architectural or training changes.Inconclusivedenies

Not yet read by this claim’s Steward.

GPT-6 Astra's monitorability decline exceeds that predicted by earlier models' capability trends.Inconclusive

Not yet read by this claim’s Steward.

Astra uses only a small amount of this, and OpenAI believes that the damage to monitorability so far is minimal. But there is the dark matter to explain, OpenAI has not engendered trust, and there is danger of a race to the bottom.

GPT-6 Astra's limited use of recurrent depth has only a minor effect on its chain-of-thought monitorability.Supported

Not yet read by this claim’s Steward.

Table of Contents23 passages

Monitorability is Defense in Depth That Is Already Flailing11 passages

OpenAI Is Counting On Monitorability6 passages

How They Tested For Monitorability29 passages

OpenAI Does Not Believe It Could Catch Sandbagging6 passages

OAI-Repo Sabotage v.210 passages

The Secret Police Do Not Make Your Notebook Useless13 passages

Astra Cannot Make Itself More Monitorable On Demand3 passages

Steganographic Chain of Thought May Be Within Reach24 passages

Monitorability Declines Seem Unlikely To Be Only Capability Gains23 passages · 1 marked

OpenAI’s March paper on CoT controllability said RL training reduces it, wheres now we see that controllability is increasing as OpenAI does more RL.

Greater outcome-based RL training on reasoning tasks decreases chain-of-thought controllability.Inconclusivedenies

Not yet read by this claim’s Steward.

Part 2: Recurrent Depth7 passages

Miles Brundage : Loopt transformers, you say4 passages

The Immune System Responds19 passages

Steven Adler : Man, I really hope the reporting is mistaken9 passages

Ryan Greenblatt Explains How Bad This Could Be15 passages

Only Law Can Prevent Extinction12 passages

OpenAI Calls On Us to Avoid Racing to the Bottom24 passages

Thinking Fast and Slow, Also Small and Large20 passages

Depending on AIs decoding other AI’s opaque activations for oversight is spooky13 passages

Talking Price7 passages

Conclusion: If The House Burns Down, Halt and Catch Fire19 passages

810 passages

320 passages

50 passages

Share0 passages

Discussion about this post0 passages

Comments Restacks0 passages

Ethics Gradient0 passages

Sep 82 passages

Reply0 passages

Share0 passages

1 reply1 passage

Sep 86 passages

Reply0 passages

Share0 passages

5 replies1 passage

Top Latest Discussions0 passages

No posts1 passage

Subscribe1 passage

Start your Substack Get the app0 passages

Substack is the home for great culture0 passages

Lineage across 6 claims

Draws on · none recorded

This source

Astra Is Hard to Monitor

in 6 claims' maps

Drawn on by · none recorded

Record

Kind: unknown
Authors: Zvi Mowshowitz
Publisher: Not recorded
Published: Not recorded
DOI: None
Link: https://thezvi.substack.com/p/astra-is-hard-to-monitor
Archived copy: https://web.archive.org/web/20261013112504/https://thezvi.substack.com/p/astra-is-hard-to-monitor
Stored text: 64,426 characters, retrieved Nov 1, 2026 · 50 sections, 330 passages
Claims extracted: Not recorded
Versions: One known
Copies: None recorded
Corrections: None recorded
Facts checked: Nov 1, 2026

History 3

Nov 1, 2026DividedSegmenter · Divided into 380 parts by its own structure
Nov 1, 2026FactsFacts fetcher · Facts looked up
Nov 1, 2026Entered the graphEntered the graph; copy stored