Exploring Gymflation with AI
17 hours ago
- Project investigates 'gymflation' by analyzing deadlift PRs on YouTube over time.
- Uses YouTube API (via httr2) to collect video data from 2015 to 2025.
- Employs a local LLM (Ministral 3B) via Ellmer to extract weight, unit, and reps from video metadata.
- Improves extraction accuracy (92% on test set) using in-context learning with few-shot examples.
- Applies vision-language models (Gemini, MiniMax, Qwen) to classify lifter gender from thumbnails, achieving ~80% accuracy.
- Finds no evidence of gymflation: average deadlift weight has not increased over time; heavy lifts may even be decreasing.
- Highlights practical tips: caching API requests, using structured outputs, parallel processing, cost management, and manual annotation.