Xin-Jing Wang
@XinJingWang2Software engineer at Google. Make a better world with AI technologies
[GQA paper](arxiv.org/pdf/2305.13245…) takeaways: grouping heads for tradeoff of quality and training speed. Mean pooling heads to convert MHA to MQA with a light training with original training package. Grouping can be random. Possible future work: designed grouping.
LLaVa paper takeaways: To tune a multimodal from pure textual input. - human-labeled instructions with LLM (GPT-4)-generated questions for diversity - Bounding boxes for visual cues in instructions - train visual-textual projection - Two-stage fine-tuning: v-t matrix & multimodal
Notes of the book "Creative Confidence" by Tom Kelley and David Kelly about Design Thinking xinjingwang.substack.com/p/book-notes-o…
The Learning from Eric Yuan's Interview at Stanford GSB, by @XinJingWang2 xinjingwang.substack.com/p/the-learning…
It’s not just what your audience needs to hear, but how they need to hear it … and, says Professor Jonathan Levav, how you need to change your message for your audience over time. Listen to Levav with host Matt Abrahams on Think Fast, Talk Smart. stanford.io/3Kc9SkU
United States Trends
- 1. Mavs 9.341 posts
- 2. Pete 245 B posts
- 3. DeSantis 28,9 B posts
- 4. Marcus Smart 2.047 posts
- 5. Bama 38,7 B posts
- 6. Clemson 20,6 B posts
- 7. Grizzlies 5.231 posts
- 8. #MFFL 3.123 posts
- 9. Knicks 14,1 B posts
- 10. Miami 113 B posts
- 11. Luka 27,7 B posts
- 12. Sabres 4.569 posts
- 13. Danny Wolf N/A
- 14. #WWENXT 26,5 B posts
- 15. NBA Cup 11,8 B posts
- 16. #RHOBH 4.004 posts
- 17. minghao 41,6 B posts
- 18. Amari Williams N/A
- 19. Wemby 8.305 posts
- 20. Butler 15,2 B posts
Something went wrong.
Something went wrong.