OpenAI’s newest model, Astra, reruns the same block of internal processing steps multiple times before it answers, and the company’s chief scientist Jakub Pachocki says this does not amount to the kind of hidden machine-thinking that safety researchers have warned about for years. Scott Alexander, who writes the blog Astral Codex Ten and helped author a widely circulated 2027 industry forecast that first named the risk, argues Pachocki’s defense answers a narrower question than the one that actually matters.
The concern traces back to how large language models handle problems too complex for a single pass through their layers. Since 2024, models have solved this by writing out intermediate steps in plain English on a visible “chain of thought” before continuing, which lets safety teams read what a model is doing and intervene if it starts planning something harmful. The 2027 forecast warned that labs would eventually skip that English scratchpad entirely, letting a model’s reasoning loop internally as raw numerical vectors that no one can read. Researchers call this “neuralese.”
Astra’s architecture loops a set of layers back on themselves before handing off to the next stage, effectively simulating extra depth without building it. Pachocki has said the resulting computation graph is “within a factor of two” the depth of GPT-4, OpenAI’s 2023 model, framing the change as routine growth rather than a shift toward neuralese.
Alexander’s objection is structural, not rhetorical. A model that needs 1,000 layers of unmonitored processing to devise a harmful plan can reach that threshold two ways: by removing the English scratchpad altogether, or by looping enough layers within a single pass between scratchpad entries. Astra, he writes, is a step toward the second path even if it never touches the first. He also flags that “simulated” layers built through looping are far cheaper to add than real ones, which is exactly what makes them easier to scale past a safety threshold without anyone noticing.
To size the gap, Alexander asked an AI model called Fable to estimate the capability gain, in model generations, from doubling simulated layers versus doubling real ones. Fable’s self-reported estimates, which Alexander says varied widely by task type, put looped depth at roughly a tenth the capability value of genuine new layers. That is a case for looping being less dangerous per layer, not a case for it being safe at any scale.
Alexander leans on an argument from another writer, Linchuan Zhang, comparing the situation to drinking-age enforcement or the taboo on nuclear weapons: a bright line only works as a taboo if everyone agrees where it sits, and right now nobody has agreed whether looped layers count as the thing AI 2027 warned about or a separate, lesser practice. Pachocki’s “within a factor of two” defense sounds reassuring, Alexander argues, but it is answering a question the safety community never actually asked.
The disagreement lands at an odd moment. AI company leaders have reportedly agreed in principle to “pace the frontier” on capability growth, without yet defining what that means in practice. Astra’s release is the first case where a frontier lab’s architecture choice forces that vague commitment to get specific, and rival labs building their own recurrent or looped systems will now face the same question OpenAI just did: how many effective layers, reached by what method, and read by whom.
Reported by Scott Alexander on Astral Codex Ten on 24 September 2026.
