Tweet by Jack_W_Lindsey
July 31, 2025
Combining these methods, we can trace full computations through attention. For instance, we show how relevant query and key features conspire to direct “opposite heads” to the right token, and transform “small” features to “large” features. https://t.co/9hM5N6ciKD
- Author
- Jack_W_Lindsey
- Date
- July 31, 2025
- Canonical URL
- /tweets/jack-w-lindsey-1950952355127820300-7c84c1