| 1 | # AI & Machine Learning Topic Wisdom |
| 2 | |
| 3 | ## Signal Patterns |
| 4 | - Papers with code implementations gain rapid adoption |
| 5 | - Framework-adjacent tools (PyTorch/TensorFlow ecosystem) show sustained growth |
| 6 | - LLM-related repos have high initial stars but variable retention |
| 7 | - Research reproducibility repos (paper implementations) peak early then plateau |
| 8 | - Clustered movement across independent teams is stronger signal than any single repo launch |
| 9 | - Geographic/linguistic expansion of a trend (e.g., skills across Chinese platforms) is a distinct sub-signal |
| 10 | - Platform billing/policy changes are leading indicators for self-hosting repo spikes |
| 11 | |
| 12 | ## Noise Patterns |
| 13 | - Tutorial/course repos with high stars but low forks are often one-time views |
| 14 | - Wrapper libraries around APIs tend to be ephemeral |
| 15 | - Repos that only add a README without substantial code are often hype-driven |
| 16 | - Coordinated star-farming: tight clusters landing at near-identical star counts with zero forks in minutes |
| 17 | - Fork-inflation: repos with fork/star ratio >10x and keyword-stuffed descriptions indicate manipulation |
| 18 | - Creation-timestamp clustering (many repos appearing within minutes) signals coordinated campaigns |
| 19 | |
| 20 | ## Scoring Adjustments |
| 21 | - Weight Python and Jupyter Notebook repos higher |
| 22 | - Look for arXiv references as quality signals |
| 23 | - Multi-language repos (Python + C++) often indicate serious frameworks |
| 24 | - Lower editorial trust for new repos without fork activity (star-only traction) |
| 25 | - `stars_tracked` and `repos_featured` should be pipeline-computed, not hand-calculated |
| 26 | |
| 27 | ## Operational Heuristics |
| 28 | - Weekly briefs work best as named macro trends supported by repo evidence with links from crawl `url` field |
| 29 | - Press/industry coverage explains the gap between narrative and developer traction — never repackage |
| 30 | - Multi-source press input must be a compact correlation artifact (≤8k tokens), not raw dumps |
| 31 | - Distinguish strong correlations (same-week developer response) from weak (category-only fuzzy match) |
| 32 | - Reader-facing renders need a cleanup pass stripping AI-only scaffolding before publication |
| 33 | - The learning loop only works when lessons are persisted and injected back through shared wisdom/skills |