A LessWrong User Discovered LLMs Are Terrible at Explaining Mistakes — Here's the Fix
After scraping history and Reddit, one user found most mistake lessons are just surface-level patches.
A LessWrong user (SK2) embarked on a systematic analysis of mistakes using LLMs, scraping historical cases (e.g., Semmelweis, Chamberlain) and Reddit threads for personal anecdotes. The goal was to bypass the cliché advice and learn from others' experiences. However, the user concluded that most stories yield only surface-level lessons—akin to patching a single line of code rather than refactoring the entire architecture.
Drawing an analogy to programming bugs, SK2 identifies three levels: (a) superficial patches, (b) function refactoring, and (c) architectural changes. The user argues that smart people's dumb mistakes stem from subtle issues like failed pattern recognition or competing heuristics—rarely captured in typical retrospectives. This insight leads to a call for a discussion group focused on 'type (c)' cognitive fixes: restructuring decision-making habits to prevent entire categories of errors.
- LLM-generated analysis of historical mistakes (Semmelweis, Chamberlain) felt cliché and unenlightening
- Reddit 'biggest mistake' threads offered surface-level advice, not deep cognitive pattern fixes
- User proposes a 'mistake postmortem' group focusing on type (c) fixes: re-architecting decision-making habits
Why It Matters
For professionals, this highlights the gap between knowing advice and actually changing deep decision patterns.