Redwood Research operationalizes "opaque serial depth" as a proxy for latent AI reasoning
akbirthko · x · 2026-10-21
Redwood Research published a blog post, "An operationalization of opaque serial depth," by Nathan Sheffield, Alek Westover, Ryan Greenblatt and others.
- Background: chain-of-thought (CoT) is currently a valuable oversight tool, but architectural shifts could significantly reduce CoT monitorability. Redwood has proposed that AI companies transparently share how much latent reasoning their architectures permit.
- The post operationalizes "opaque serial depth" (from a GDM paper, Brown-Cohen et al., 2026): the longest path in the computation graph that doesn't pass through an "interpretable bottleneck." Treating CoT tokens as interpretable and transformer hidden states as non-interpretable, a standard transformer's opaque serial depth is proportional to its layer count.
- Main contribution: a concrete standard for what counts as an interpretable bottleneck—nodes that output text (not latent states) and were initialized from a pre-training prior, formalized as "NL-rooted nodes." The measure serves as a proxy for how much unverbalized serial cognition a model can perform.
More from Safety
- Super-Worms Warning: AI-Powered Self-Propagating Malware and a Fragmenting Internet — JensHonack · 2026-10-23
- Claude Code Pushes Urgent Security Update Across All Installations — umeshonai · 2026-10-23
- Sanders calls to pause autonomous AI agents in bank accounts; critics push back — AIandDesign · 2026-10-23
- TechPolicy essay: system prompts are weak evidence of AI safety, outputs must be evaluated — mmitchell_ai · 2026-10-23
- Anthropic's deputy security chief warns cyber threats will get 'a lot worse before they get better' — Polymarket · 2026-10-23
- AI cyberattacks could paradoxically accelerate AI development in an offense-defense arms race — kimmonismus · 2026-10-23