Vin Jin
@qkvqkvqkv
Do̶t̶a̶, ML only. Blog: https://qkv.bearblog.dev
It is significantly dubious to me that the current languages (human) are anywhere even close to as good as it can be.
New strategy for AI labs is move fast and break laws. Cause no one is going to catch you till its too late.
Wish we had AI safety/alignment teams but for image/video generation. Because that future is already here.
It is a shame that all these social content websites optimized algos for views/slop when they could have created opt-in/purchasable algos for optimizing customer desired metrics and just sold those instead?
The slopification of the internet could end up having a certain desirable effect. After a certain level of AI saturation, people might finally start valuing their attention more. The idea of infinite options would be such a turn off that people start embracing limitations.
My GPU fans are making squeaky-spiny noises and I'm very uncomfortable hearing those.
I swear 2.5 Pro in gemini cli is just terrible. Even the flash variant does a much nicer job.
The static in my earphones when I start training on my GPU is almost like an 8-bit retro soundtrack. Pretty cool.
I wrote inference for Qwen3 0.6B in C/CUDA. github.com/asdf93074/qwen…
Had a dream about spiral attention. Woke up cause I couldn't understand it.
I wrote this a year ago and it still sounds relevant to me: qkv.bearblog.dev/things-are-onl…
I wrote a short blog post some months ago that I never got around to publishing for many reasons. qkv.bearblog.dev/my-next-intern…
I finally found the issue and it took me an embarrassingly long time to figure it out. Qwen3 does RoPE with half-split grouping over interleaved. No LLM figured it out despite looking at several reference implementations. GPT is still in denial that Qwen3 even does this.
Writing inference for Qwen3 (dense) in C/CUDA from scratch and I'm so close to the finish... just need to fix the damn probs...
Gemini CLI and the build tool in AI studio are terrible when using 2.5 Pro. Even copy-pasting from chat works much better.
Another day of being thankful that pytorch exists so that I can verify my kernels against it.
United States Tendências
- 1. Caleb Love 1,769 posts
- 2. Sengun 8,244 posts
- 3. Mamdani 426K posts
- 4. Reed Sheppard 3,280 posts
- 5. #SmackDown 42.8K posts
- 6. Norvell 3,303 posts
- 7. Lando 35.5K posts
- 8. Florida State 10.8K posts
- 9. Rockets 16.3K posts
- 10. Collin Gillespie 2,963 posts
- 11. Marjorie Taylor Greene 59.8K posts
- 12. Suns 15.4K posts
- 13. NC State 5,585 posts
- 14. #OPLive 2,487 posts
- 15. Timberwolves 9,886 posts
- 16. #LasVegasGP 61.9K posts
- 17. Booker 7,338 posts
- 18. Jabari 1,852 posts
- 19. Finch 2,830 posts
- 20. #BostonBlue 3,772 posts
Something went wrong.
Something went wrong.