The argument, drawn from peer-reviewed cognitive-science research on human expert chunking and popularized in an August 2026 essay that drew heavy Hacker News discussion, is that a large language model's effective context window functions like a vastly expanded working memory: it can hold far more partial derivations, variable states, and subgoals simultaneously than a human mathematician juggling the same problem on paper or in their head. This reframes LLM math performance as an advantage in capacity rather than in the quality of reasoning steps themselves, and makes a testable prediction — models should do comparatively better on problems demanding long, wide bookkeeping and comparatively worse on problems demanding a genuinely novel insight, since more memory does not manufacture creativity.