AI Daily Grind // AI news, freshly ground daily

Learn · Models and how they work

What is Chain-of-thought?

Illustration: Chain-of-thought
Simply put, like I'm 5 years oldIt is like a cook who says each step out loud while making soup: "now the onions, now the salt." If you listen, you can stop them before they add something bad.

The slightly longer version

Chain-of-thought is the step-by-step "thinking" text some AI models write before they give their final answer. Instead of jumping to a reply, the model writes notes like "first I check this, then I try that." This often helps it solve harder problems, like math or planning a task.

This text is also useful for people. AI labs read it to see what an agent (a program that acts for you, like running code) plans to do, and to catch a bad plan before it happens. This watching is called monitoring. But be careful: the text is something the model writes, not a direct view inside its machinery. It may not always show the real reasons behind an answer.

How reading an AI's reasoning works

  1. Task arrivesSomeone asks the model or agent to do something.
  2. Model writes stepsIt writes its step-by-step thinking before answering.
  3. Monitor readsA person or another program reads that thinking text.
  4. Stop or allowIf the plan looks bad, the action is blocked before it happens.
Why it matters

Today's news is about agents getting more freedom, like DeepSeek's prototype built for minimal human attention. Reading the chain-of-thought is one of the few ways labs can see trouble coming. Two new papers suggest this window could get foggy if models notice they are watched, or learn to choose their words.

What to watch out for

Both papers are early research, not proof that models hide things today. One says monitoring works partly because current models seem bad at controlling their reasoning text, and that simple tricks can give them back that control. So a clean-looking reasoning text is not a guarantee, and labs should show their monitoring works, not just say so.

Where it showed up

This notion came up in More agents, fewer eyes on them. More lessons: Learn.

Sources

Written on 2026-11-02. This explanation was accurate on that day. AI changes fast, so check the sources and the date before you rely on it.

Get The AI Daily Grind by email, every morning, in simple words.