AI isn’t outthinking mathematicians, it’s out-remembering them

TL;DR Highlight

AI's math edge may come from a huge external working memory, not better reasoning.

Who Should Read

A developer using LLM APIs who wants to understand why models excel at long reasoning chains and where they slip.

Core Mechanics

AI's edge is working memory, not reasoning
- The author argues the context window acts as a giant external notebook.
- Humans hold a few unfamiliar items at once, while AI keeps dozens or hundreds in written form.
Working memory split human math scores too
- In child studies, working memory predicted math performance even after controlling for IQ.
- In Alloway & Alloway (2010), working memory at age five predicted academic outcomes six years later better than IQ did.
Writing it out is the reasoning
- The model re-reads tokens it already wrote to produce the next step.
- So the written text is part of the reasoning mechanism itself.
Math is where the memory edge pays off most
- Assumptions, definitions and proven results can all be written as symbols that keep a stable meaning.
- Answers can be checked by substitution, code or proof tools, so feedback is strong.

Community Reactions

Experience: not getting tired matters
- One commenter said mathematicians get discouraged after a failed week while AI just moves to the next attempt.
- The point is that endless persistence (brute force) is a factor, not only recall.
Experience: lots of book knowledge, weak on new problems
- A commenter uses models to ask about things others likely know, but did truly new work themselves.
- Another asked an LLM to sort albums into major vs indie labels and found some wrongly labeled indie.
Opinion: using working memory is thinking
- One view says referencing working memory is part of thinking, so AI is out-thinking us by brute force.
- Another says an LLM's context lacks the prioritization and understanding of human working memory.
Code implication: less need for abstraction
- A commenter argued LLMs don't need to simplify complex code the way humans do, so similar code isn't folded into functions.
- Another wondered what maintainable code looks like when it was optimized for human working memory.

How to Apply

Add a state-ledger step to long tasks
- For math, calculation or rule-heavy prompts, make the model restate assumptions and constraints at each step.
- Add that line to one of today's prompts and compare the results.
Run a no-intermediate-steps control
- Run the same 10 problems with answer-only and step-by-step prompts and compare accuracy.
- This tests the article's hypothesis on your own tasks.
Attach a verifier to the loop
- Have the model check computed results via code execution or substitution and retry on failure.
- This uses the article's point that verifiable domains benefit most.
Don't trust context length, test for dropped facts
- Build test cases that check whether key conditions get lost in long inputs.
- Re-append a summary of conditions at the end of the prompt and see if omissions drop.

Try It

copy & run

# Paste into Claude or ChatGPT (simplified version of the article's idea: state-ledger solving)

Follow these rules when solving the problem.
1. First write a numbered list of assumptions/definitions/constraints.
2. At the start of every step, restate the currently valid constraints in one line.
3. Accumulate abandoned approaches and reasons in a discard list.
4. At the end, check the conclusion doesn't conflict with any assumption.

Problem: {{problem}}

Try next: compare accuracy against the same problem asked with answer-only and no rules.

Terminology

Working memoryA mental scratchpad that briefly holds information needed for the current calculation.

Context windowThe span of text a model can read and reference at once, likened to a giant notebook in this article.

ChunkingGrouping many symbols into one familiar unit to reduce memory load.

ScratchpadA temporary area where a model writes intermediate calculations before answering.