When an AI system solves a difficult mathematical problem, the immediate assumption is that it has become more intelligent. It has absorbed millions of examples. Reinforcement learning has taught it better strategies. It may even be developing something like mathematical intuition.
All of these explanations contain some truth. But they overlook a simpler possibility: AI has access to a vastly larger working memory than the human brain. More precisely, it has access to an enormous external symbolic workspace that performs many of the functions working memory performs in humans.
This difference may be especially important in mathematics.
Mathematics Is Constrained by Memory
Working memory is the mental system that allows us to hold and manipulate information over short periods. When solving an equation, you must remember what each variable represents, which operations have already been performed, and what the current goal is. During a proof, you may need to track assumptions, intermediate lemmas, exceptions, and multiple possible cases.
Human working memory is remarkably limited. Its exact capacity depends on the task and how information is organized, but the limitation is obvious from everyday experience. Try multiplying two three-digit numbers in your head. The operations are simple — the difficulty comes from having to preserve partial results while performing additional calculations.
Writing the numbers down transforms the problem. Paper does not make you more intelligent — it expands your effective working memory. The same principle applies at higher levels of mathematics. A mathematician uses notation, scratch paper, diagrams, and previously written lemmas not merely to communicate the solution, but to make the reasoning cognitively possible.
The AI Advantage: Context as Workspace
Expert mathematicians compensate through "chunking." A novice sees a long sequence of symbols; an expert sees a familiar pattern that occupies a single slot in working memory. This is why experience matters so much in mathematical problem-solving — it's not just about knowing more, but about compressing what you know into manageable chunks.
AI models perform a version of this compression automatically through their training. But they also have something no human has: a context window that can hold an entire problem statement, hundreds of intermediate equations, several abandoned approaches, definitions, constraints, and earlier conclusions — all simultaneously accessible.
When we see an AI solve a competition-level problem, we interpret it as evidence of superior reasoning. But some of it may instead reflect the removal of one of the most important biological limits on human reasoning: our extremely restricted working-memory capacity.
What This Means for AI-Assisted Work
This reframing has practical implications for how we work with AI:
1. Don't expect breakthroughs from reasoning alone. If AI's mathematical advantage comes partly from memory rather than pure reasoning, then simply scaling up reasoning strategies won't necessarily produce qualitatively better results. The context window matters as much as the model.
2. The "thinking partner" framing is wrong. A human thinking partner also has limited working memory — they just bring different chunks. An AI partner brings unlimited working memory. The bottleneck shifts from "how much can we hold" to "how do we search what we have."
3. Long-context models change the game. As context windows grow from thousands to millions of tokens, the working-memory advantage compounds. An AI that can hold an entire research paper's worth of context isn't just doing the same task faster — it's doing a different task, one that humans literally cannot perform without extensive external memory aids.
The Von Neumann Comparison
The analogy isn't to Einstein — it's to von Neumann. John von Neumann was legendary not just for his intelligence but for his ability to hold enormous amounts of information in his head simultaneously. Colleagues described him as someone who could keep track of dozens of intermediate results without writing anything down.
AI may be less like an electronic Einstein and more like a machine-amplified von Neumann: immense speed, breadth, and symbolic memory. The qualitative difference between AI and human mathematicians might not be "better reasoning" but "no cognitive bottleneck on memory."
This would explain why AI performs relatively better on problems that require tracking many constraints simultaneously (like optimization and formal verification) and relatively worse on problems that require genuine creative leaps or novel abstractions. Memory helps with the former; genuine intelligence is needed for the latter.
Implications for the Future
If this analysis is correct, the trajectory of AI in mathematics depends heavily on context engineering, not just model architecture:
- Retrieval-augmented reasoning (pulling relevant theorems on demand) could matter more than larger models
- Symbolic memory tools (structured notebooks, proof assistants, search over prior work) become critical infrastructure
- The human role shifts toward defining problems and evaluating solutions rather than intermediate computation
The mathematician of the future may not need to hold everything in their head — they need to know what to ask, how to verify, and when the AI's memory-backed confidence is misplaced.
Based on analysis from Davide Piffer's original essay. The core argument reframes AI's mathematical advantage from superior reasoning to superior working memory — a distinction that matters for how we build, deploy, and regulate these systems.
Top comments (0)