LLMs Diverge, Humans Converge — LLMs Can't Come Up With Ideas
Large Language Models (LLMs) tend to produce divergent outputs based on statistical patterns in their training data, which limits their ability to generate convergent, innovative ideas. Unlike humans, LLMs struggle with tasks requiring synthesis of multiple constraints, such as database design, due to biases in training data and lack of contextual understanding. Even when instructed otherwise, LLMs often revert to common patterns like short SQL aliases, showing the dominance of training data over specific directives.
- ▪LLMs generate outputs based on probabilistic interpolation of training data, leading to divergent rather than convergent thinking.
- ▪Claude Code often uses short SQL aliases like 'sd' or 'oi' despite instructions in CLAUDE.md to avoid them, showing the influence of prevalent patterns in training data.
- ▪Production-quality database schemas are underrepresented in training data, causing LLMs to default to beginner-level designs.
- ▪Database design requires understanding of access patterns, business rules, and future needs, which are typically absent from public code repositories.
- ▪Even with full context, LLMs may struggle to balance multiple design constraints effectively due to their reliance on statistical regularities.
Opening excerpt (first ~120 words) tap to expand
try { if(localStorage) { let currentUser = localStorage.getItem('current_user'); if (currentUser) { currentUser = JSON.parse(currentUser); if (currentUser.id === 105282) { document.getElementById('article-show-container').classList.add('current-user-is-article-author'); } } } } catch (e) { console.error(e); } Satoshi Nishimura Posted on May 17 LLMs Diverge, Humans Converge — LLMs Can't Come Up With Ideas #llm #claude #ai LLMs can't come up with ideas. The output of an LLM (Large Language Model) tends to be divergent. It moves in the direction of deriving combinations from its training data. Good ideas, on the other hand, are convergent. They solve multiple problems at once with a single mechanism. When using LLMs, I think it's important to keep this difference in mind as you proceed.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at DEV.to (Top).
Discussion
0 commentsMore from DEV.to (Top)
