Conquering Dumb LLM Search Judges with Classic ML

2025-01-24
Conquering Dumb LLM Search Judges with Classic ML

The author explores using a local LLM as a search relevance judge, a cost-effective alternative to OpenAI. Individual LLM judgments are unreliable, so the article proposes combining multiple LLMs' assessments of various product attributes (name, classification, description, etc.) using traditional machine learning (e.g., decision trees) to improve accuracy. Experiments show this approach can predict human preferences and reveal the logic behind human labels, aiding search engine optimization.

Read more
Development

Preferring Throwaway Code Over Design Docs: A More Efficient Software Development Approach

2024-12-15
Preferring Throwaway Code Over Design Docs: A More Efficient Software Development Approach

In software development, the traditional design document and incremental development model isn't always efficient. Author Doug Turnbull proposes a "coding binge" approach: quickly implement a prototype using a temporary PR, get early team feedback, refine the design, and then gradually break it down into deployable PRs. This method encourages rapid iteration, early problem detection, and considers code itself as the best documentation. While design documents still have value in specific situations, the author advocates for "showing, not telling," using code prototypes for rapid validation and iteration to achieve more efficient software development.

Read more
Development code prototype