Ryan Greenblatt warns about losing access to AI reasoning
Redwood’s chief scientist urges evidence and policies before architectures weaken oversight through readable reasoning.
I'm very worried about changes to AI architectures that result in AIs thinking in opaque activations instead of in chain of thought
About these dates
Statement date displayed on X in the inspected browser; not converted to UTC.
Ryan Greenblatt, Redwood Research’s chief scientist, warned that architectures relying on opaque internal computation could weaken oversight through readable chains of thought. He described Astra as concerning on limited public evidence, while explicitly saying that evidence was insufficient to judge its performance and monitorability tradeoffs. Ryan Greenblatt ↗Redwood Research ↗
Redwood’s linked proposal calls for verified architecture disclosures, evidence and review of monitorability, and policies governing those tradeoffs. Greenblatt urged caution before abandoning readable reasoning and acknowledged that the proposal might not prevent the most concerning architectures. Redwood Research ↗Ryan Greenblatt ↗
Sources & attribution
- Commentary 10 Sept 2026Warning about architectures and monitorability ↗
Ryan Greenblatt. Original source inspected. Personal views.
- First-party report 10 Sept 2026Proposal for tracking architecture effects on monitorability ↗
Redwood Research. Research organization proposal; not an enacted standard.
- Organizational disclosure Accessed 10 Sept 2026Research team ↗
Redwood Research. Professional role checked; date is the access date.