OpenAI's 'Opaque Recurrence' Alarms AI Safety Experts — SkimNews
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- OpenAI's Astra model will use a technique called "recurrent depth" (also called "opaque recurrence") that processes the same query multiple times in a loop outside sequential thinking, according to The Information.
- Redwood Research CEO Buck Shlegeris said he was "extremely concerned," warning that if OpenAI pushes the technique further, recurrence could scale to "totally destroy" chain-of-thought monitorability.
- AI safety advocate Zvi Mowshowitz wrote that laws might be necessary to prevent a "race to the bottom" among AI labs, calling the technique "playing with fire" against a shared taboo of preserving CoT faithfulness.
- OpenAI chief scientist Jakub Pachocki emphasized the lab's commitment to legible chains of thought as "a core goal of our current research program," and the company pushed back against any suggestion it would shift to "neuralese."
- Anthropic and Google DeepMind were already discussing the opaque recurrence technique, per The Information's follow-up report on Wednesday.
- Ryan Greenblatt, Redwood Research's chief scientist, warned that a natural progression could involve scaling opaque reasoning to the point where models "reason entirely or almost entirely in latent space."
Why it matters: Astra's use of the technique is reportedly limited and CoT is still expected to be legible, but safety researchers see it as a foothold that could scale toward reasoning happening entirely in unmonitorable latent space — undermining a safeguard that OpenAI, Anthropic, and peers have publicly committed to preserving as model capabilities grow.
Ask SkimNews
