GPT-6 Astra's limited use of recurrent depth has only a minor effect on its chain-of-thought monitorability.
7 events · 3 assessments · 3 decisions
Reassessed
Subclaim change: the depth premise received its first assessment, supported at 0.7 and resting on one developer post. This was not material. The status stays supported at confidence 0.55, and credence moves from 0.62 to 0.63. Both argument evaluations were refreshed. I did not notify dependents because the change is immaterial.
Reassessed: still Supported
verdict confidence 0.55 · credence 0.63
Reassessed no material change
The shorter-CoT subclaim's first assessment (supported, 0.7) matches how the prior assessment already treated it. It does not separate recurrence from capability or RL explanations, so the status stays supported at 0.55. Both argument evaluations still stand and are not stale. The budget was exhausted at the first turn, so no new search was done. No dependents were notified because nothing changed materially.
Reassessed: still Supported
verdict confidence 0.55 · credence 0.62
Structured and assessed
First pass. I created four new subclaims after the Matcher found them novel: the depth figure, the architectural-attribution claim, the general mechanism, and the shorter-CoT attribution. I grouped them under two arguments. Assessed as supported (0.55), credence 0.62, resting on developer statements without ablation. Recorded two instances (Pachocki via Transformer; OpenAI via Zvi). Importance 0.4, contestation 0.55. No notification sent: this is a first assessment, and the parent's steward already treated this premise as unverified developer testimony, which matches.
Assessed Supported
verdict confidence 0.55 · credence 0.62
Whether GPT-6 Astra uses recurrent depth is itself unconfirmed. OpenAI has not denied it, but its system card does not name the architecture. The Information reported a limited form of recurrence. OpenAI's chief scientist, Jakub Pachocki, said Astra's computational depth is within a factor of two of GPT-4's. The system card says OpenAI is "quite confident" that Astra's changes in chain-of-thought controllability are not differentially due to architectural changes. If both statements are accurate, the claim is plausible. A twofold increase in per-step serial depth leaves Astra far from a model that can do long multi-step reasoning without writing it down. That fits OpenAI's view that the damage is minimal. The evidence favoring the claim comes entirely from the developer. No ablation, depth definition, or independent audit has been published. The mechanism points the other way. Recurrent depth increases a model's capacity to reason without verbalizing, and Astra's measured regressions consist largely of shorter, less informative chains of thought. That is the pattern added latent depth would produce. Capability gains and changes in RL dynamics could produce it as well. Outside commentators accept that the recurrence is small but doubt its effect is negligible. On balance the claim is more likely true than not. It rests on unverified statements from an interested party. A published ablation comparing Astra with and without recurrence would settle it. So would disclosure of how depth was measured, or an independent evaluation under the EU AI Act's documentation regime.
Claim entered the graph