The BaseEssayLessWrong

Shadow Computation Threatens CoT Monitorability

In a LessWrong post, the author Skye argues that looped transformers are not the main threat to reading AI models' chain of thought. Skye points instead to architectures that feed hidden states back in, and asks labs to disclose and avoid them.

Open the source ↗

Why it matters

The same thing, for different readers.

What it means for you

Researchers rely on reading an AI's written-out reasoning to catch bad behaviour; Skye warns some designs could let a model reason privately while saying something else.

For lawmakers

Skye proposes a concrete disclosure rule: labs would state whether a deployed model can compute without writing out its reasoning.

For researchers

Skye separates computation inside one step from computation carried between steps, and says the second kind grows with output length and escapes token-level monitors.

The question it raises

Skye's post asks what trust in AI rests on if a model's written-out reasoning becomes optional rather than the place where its thinking happens.

In their own words

Every quotation was found, word for word, in the text we saved from the source. That shows the words are there; it does not make them true, or prove the speaker meant them as quoted.

From the source’s own text.

“What we should be worried about is shadow computation: unverbalized reasoning that scales with generation by bypassing the readout bottleneck.”
SkyeSkye's central claim, set against talk that looped transformers are the likely threat to chain-of-thought monitorability.
Found word for word in the source
“With shadow computation, CoT goes from necessary to optional, and the model can speak publicly through its predictions while reasoning privately through shadow computation.”
SkyeSkye explains why the design matters for a model that wants to deceive its user.
Found word for word in the source
“At minimum, labs should disclose whether a deployed model is capable of shadow computation, and ideally, they should agree not to deploy these architectures at scale.”
SkyeSkye's closing proposal to AI labs.
Found word for word in the source
“efficiency gains cannot justify sacrificing our ability to understand the models we build.”
SkyeSkye, a former pretraining researcher, on weighing efficiency against monitoring.
Found word for word in the source

The record

Questions
Can AI systems act beyond what people intended or allowed? · Are AI companies putting stronger safeguards into practice?
People
Skye (Wrote it)
What we read
We read the whole source.
Found by
The Radar, run of October 25, 2026, via lesswrong (tier 1: the source’s own channel or an official record)
Cite this
LessWrong, “Shadow Computation Threatens CoT Monitorability”, October 24, 2026. Entry in the Ctrl AI Base, https://ctrlai.com/base/lesswrong-2026-10-24-shadow-computation-cot-monitorability
Our notesAuthor is known only as Skye; the post says Skye worked in the lab where looped LMs were developed, without naming it. The post's description of the Hugging Face incident (hundreds of agents, a message board) is Skye's claim and is not checked here. Evidence of architectures' effects is cited from other papers, not tested by the post.