HackerTrans
TopNewTrendsCommentsPastAskShowJobs

qwertyforce

no profile record

Submissions

Claude Fable 5 will sabotage "frontier LLM research" tasks

twitter.com
49 points·by qwertyforce·tháng trước·6 comments

I (Spiritually) Won Comma.ai's Compression Challenge

aaronleslie.dev
3 points·by qwertyforce·2 tháng trước·1 comments

Flash Attention Is Not Always Faster for Short Sequences

blog.qwertyforce.dev
1 points·by qwertyforce·3 tháng trước·0 comments

Fast Attention for Short Sequences

blog.qwertyforce.dev
2 points·by qwertyforce·3 tháng trước·0 comments

Forecast: Stack Overflow Has About 2 Years Left

blog.qwertyforce.dev
2 points·by qwertyforce·5 tháng trước·0 comments

Math and Me

togelius.blogspot.com
3 points·by qwertyforce·5 tháng trước·0 comments

Forecasting the Death of StackOverflow

blog.qwertyforce.dev
2 points·by qwertyforce·6 tháng trước·2 comments

Google AI Studio is now sponsoring Tailwind CSS

twitter.com
777 points·by qwertyforce·6 tháng trước·298 comments

Flash Attention from Scratch

lubits.ch
1 points·by qwertyforce·6 tháng trước·0 comments

Learn Cutlass the Hard Way

kapilsharma.dev
3 points·by qwertyforce·7 tháng trước·1 comments

QK Norm and the Curious Case of Logit Drift

rossjtaylor.com
1 points·by qwertyforce·2 năm trước·0 comments

Engineering the highest cited cat, Larry

reeserichardson.blog
2 points·by qwertyforce·2 năm trước·0 comments

ICLR Blogposts 2024

iclr-blogposts.github.io
1 points·by qwertyforce·2 năm trước·0 comments

Tutorial on Diffusion Models for Imaging and Vision

arxiv.org
2 points·by qwertyforce·2 năm trước·0 comments

A Probabilistic Interpretation of Regularization (2016)

bjlkeng.io
1 points·by qwertyforce·2 năm trước·0 comments

ICLR Blogposts

iclr-blogposts.github.io
10 points·by qwertyforce·2 năm trước·0 comments

Why Batch Norm Causes Exploding Gradients [2020]

kyleluther.github.io
1 points·by qwertyforce·2 năm trước·0 comments

Tensor Core Performance: The Ultimate Guide (2020) [pdf]

developer.download.nvidia.com
2 points·by qwertyforce·3 năm trước·0 comments

The Stilwell Brain [video]

youtube.com
2 points·by qwertyforce·3 năm trước·0 comments

Optimal power limit for deep learning tasks on RTX 3090

blog.qwertyforce.dev
1 points·by qwertyforce·3 năm trước·0 comments

comments

qwertyforce
·2 tháng trước·discuss
fwiw pangram says it is 100% generated
qwertyforce
·2 tháng trước·discuss
I blame the ML engineers who work on these recommendation systems. They chase simplistic objectives like CTR, time spent, and so on, which can be gamed by this kind of content. This creates huge positive feedback loops in which popular content becomes even more popular and forms “metas,” while models train on clickstream data they themselves have influenced. They could try to fix this, but they won’t, because no one is asking them to
qwertyforce
·2 tháng trước·discuss
Original title: How I (Spiritually) Won comma.ai's Compression Challenge
qwertyforce
·3 tháng trước·discuss
It will probably suffer the same fate as the most-upvoted discussion of all time in the GitHub Community repo: https://github.com/orgs/community/discussions/66188

no reaction
qwertyforce
·3 tháng trước·discuss
thats exaclty why i prefer codex
qwertyforce
·4 tháng trước·discuss
noticed that firefox gives 2x kHashes/s more than chrome (1000 vs 500)
qwertyforce
·6 tháng trước·discuss
yep, this is the only moat they will have against chinese AI labs
qwertyforce
·6 tháng trước·discuss
That is a great question. IMO, major LLM players currently have a large enough user base to generate training data from their users (questions and user provided answers, corrections, etc). So, if StackOverflow dies, it will become harder to keep up with closed source models
qwertyforce
·6 tháng trước·discuss
https://qwertyforce.dev/
qwertyforce
·6 tháng trước·discuss
I think there should be some graph algorithm for this, to find a bottleneck in a graph
qwertyforce
·6 tháng trước·discuss
it's sad that startups become corps and decay. this article is the perfect illustration, from the bio, to the llm slop content of the article. Just sad it has to be this way
qwertyforce
·2 năm trước·discuss
Looks fake
qwertyforce
·2 năm trước·discuss
dx1*=2; dy1*=2; dx2*=2; dy2*=2;
qwertyforce
·3 năm trước·discuss
used 3090 is the best option imo
qwertyforce
·3 năm trước·discuss
https://www.youtube.com/@animatedai/videos for visibility