What's this about?
The technique behind the monitorability problem: "recurrent depth" (opaque recurrence) lets the model loop internally over the same query instead of writing intermediate steps out as text. Fewer legible traces means better efficiency and much less to monitor. Redwood Research's CEO: "I am extremely concerned." Zvi Mowshowitz suggests this class of technique may eventually need regulation.
What exactly is "chain of thought"?
The step-by-step written reasoning models produce before answering — like a student showing their working. Safety monitoring reads it; recurrent depth replaces some of it with silent internal computation. The working still happens, but in the model's activations where nobody can read it.
Why is this worth noting?
For years the industry's informal safety argument was "at least we can read what it's thinking." This is the first frontier deployment that trades a chunk of that away for capability, on purpose — which is why safety researchers treat it as a threshold crossing rather than an implementation detail.