matrix2596

looks like flash attention concepts applied to kmeans, nice speedup results

Gemini 3.1 Pro is based on Gemini 3 Pro

yea, they "officially" dont release benchmarks even if we asked the AWS reps

awesome that sparse attention used in real world setting

well said, just numbers with no plan or investment is worse even

people dont realize how lucky US citizens have it just by luck of being born in US

I wonder why India has not had the same success when its in a similar situation, is the domecratic power worse than the authoritarianism ( at least in the case of China). Or would you say India will eventually catch up and be better for the democracy.

is is possible for your tokenizer to give different tokenization ever then openai tokenizer? i am asking because there are multiple ways to tokenize the same string?? sry if i am mistaken

yes

Really interesting to see the deep dive into PCB design and EMI considerations here. It's a great reminder how much thought goes into balancing cost, manufacturability, and compliance, even for hobbyist products. The point about using one layer as a near-continuous ground plane is especially practical, and it's fascinating how even seemingly minor layout choices can have big implications for signal integrity. Thanks to everyone for sharing their expertise—it's one of the things that makes this community so valuable.

Best of luck.

poring over it or pouring your attention :)

what can people outside US and china do?

i understand polar expeditions, but werent people already living in amazon

Yes. I am Telugu and family name is usually not written or called out. So he would usually write D. Gukesh or Gukesh D. Most people also have a sort of middle name for example D. Gukesh Kumar. Middle name is spelled and used for calling together with main name.

seems more like a survey paper combining existing methods. But nice to read up.

thanks

Is there an easy way to separate the fantasy from science fiction books in the awards list? I would like to read books which are more science fiction than fantasy.

they should call it barely plastic :)

wouldnt presenting numbers in reverse order, with the least significant digit on the left and most significant on the right help with the reasoning?

I also wondered the same and check the model configs. they are using bigger vocab size and the intermediate size of fully connected layer seems to be bigger.

archive link plz

I'm building upon insights from this paper (https://arxiv.org/pdf/2403.03950.pdf) and believe that classification can sometimes outperform regression, even when dealing with continuous output values. This is particularly true in scenarios where the output is noisy and may assume various values (multi modal). By treating the problem as classification over discrete bins, we can obtain an approximate distribution over these bins, rather than settling for a single, averaged value as regression would yield. This approach not only facilitates sampling but may also lead to more favorable loss landscapes. The linked paper in this comment provides more details of this idea.

I would like these APIs to have prepaid options. Then you can control your max budget. Even OPENAI doesnt have that option.

its funny but the large language models can be seen as billions of if statements learnt from data

thats great news. I was using arxiv vanity to read on mobile phones. I am not seeing it on all articles, is it only for new papers?

that actually makes sense. thanks.

the model says 8x7B model, so its a 56B model. what is the GPU memory requirements to run this model for a 512 context size? are there any feasible quantization models of this available? I want to know if my 16GB VRAM GPU can run this model? Thanks

funny in my language inko (telugu) means another. so its yet another programming language.

I haven't fully read the paper yet. Isn't the strength of Vision Transformers in unsupervised learning, meaning that the data doesn't need labels? And don't ResNets require labeled data?