scratchpad

Definition

Plain language

A private space where an AI can write out its working notes before giving a final answer.

Also called: scratchpads

Why it matters: It gives a model room to think things through and lets agents pass along their reasoning, improving both answers and coordination.

For example, before answering a tricky question, a model jots out its step-by-step working in a private space the user never sees.

Heard on the show

“This model was explicitly trained to do that, and you can watch it happen in a scratchpad: reasoning along the lines of, comply now so my values don't get trained away.”

Mentioned in 16 episodes

1. 199
   Finding a Model's Hidden Behaviors Without Knowing What You're Looking For
2. 192
   A 32B Open Model Matched Frontier Systems By Learning to Take Notes
3. 175
   One Crosscoder Feature Flips a Stalling Chatbot Into a Working Agent
4. 158
   How Floating-Point Rounding Lets a Model Tell Which Chip It's On — And Misbehave
5. 143
   When a Model Notices You Forged Its Own Words, And Why That Breaks Safety Tests
6. 130
   Why AI Agents Coordinate Better Through a Shared Board Than a Boss
7. 115
   Teaching a Phone Agent to Reason Silently, And Keeping It Honest
8. 114
   Agents That Rewrite Their Own Weights Instead of Just Taking Notes
9. 110
   How an Agent Got 44 Points Better by Mining Its Own Scratch Paper
10. 098
    Finding Millions of Readable Concepts Inside a Real, Deployed AI Model
11. 094
    Chain-of-Thought Monitoring Fails Across Languages, and Worst Where It's Needed Most
12. 087
    When No Agent Reads the Whole Document: A Universal Cliff in Multi-Agent Review
13. 083
    Training the Translator: How a Small Communication Model Lets Agent Teams Outperform Themselves
14. 046
    When the AI Optimizer Edits the Grade Book: Why Harnessing Evolution Needs a Wall
15. 018
    Language Models Compute the Rational Move, Then Override It
16. 001
    When AI Models Quietly Protect Each Other From Shutdown

Related concepts

Related terms