frontpage.
newsnewestaskshowjobs

Made with ♥ by @iamnishanth

Open Source @Github

fp.

Can graph neural networks for biology realistically run on edge devices?

https://doi.org/10.21203/rs.3.rs-8645211/v1
1•swapinvidya•4m ago•1 comments

Deeper into the shareing of one air conditioner for 2 rooms

1•ozzysnaps•6m ago•0 comments

Weatherman introduces fruit-based authentication system to combat deep fakes

https://www.youtube.com/watch?v=5HVbZwJ9gPE
1•savrajsingh•7m ago•0 comments

Why Embedded Models Must Hallucinate: A Boundary Theory (RCC)

http://www.effacermonexistence.com/rcc-hn-1-1
1•formerOpenAI•9m ago•2 comments

A Curated List of ML System Design Case Studies

https://github.com/Engineer1999/A-Curated-List-of-ML-System-Design-Case-Studies
3•tejonutella•13m ago•0 comments

Pony Alpha: New free 200K context model for coding, reasoning and roleplay

https://ponyalpha.pro
1•qzcanoe•17m ago•1 comments

Show HN: Tunbot – Discord bot for temporary Cloudflare tunnels behind CGNAT

https://github.com/Goofygiraffe06/tunbot
1•g1raffe•20m ago•0 comments

Open Problems in Mechanistic Interpretability

https://arxiv.org/abs/2501.16496
2•vinhnx•25m ago•0 comments

Bye Bye Humanity: The Potential AMOC Collapse

https://thatjoescott.com/2026/02/03/bye-bye-humanity-the-potential-amoc-collapse/
1•rolph•30m ago•0 comments

Dexter: Claude-Code-Style Agent for Financial Statements and Valuation

https://github.com/virattt/dexter
1•Lwrless•32m ago•0 comments

Digital Iris [video]

https://www.youtube.com/watch?v=Kg_2MAgS_pE
1•vermilingua•37m ago•0 comments

Essential CDN: The CDN that lets you do more than JavaScript

https://essentialcdn.fluidity.workers.dev/
1•telui•37m ago•1 comments

They Hijacked Our Tech [video]

https://www.youtube.com/watch?v=-nJM5HvnT5k
1•cedel2k1•41m ago•0 comments

Vouch

https://twitter.com/mitchellh/status/2020252149117313349
30•chwtutha•41m ago•5 comments

HRL Labs in Malibu laying off 1/3 of their workforce

https://www.dailynews.com/2026/02/06/hrl-labs-cuts-376-jobs-in-malibu-after-losing-government-work/
2•osnium123•42m ago•1 comments

Show HN: High-performance bidirectional list for React, React Native, and Vue

https://suhaotian.github.io/broad-infinite-list/
2•jeremy_su•43m ago•0 comments

Show HN: I built a Mac screen recorder Recap.Studio

https://recap.studio/
1•fx31xo•46m ago•0 comments

Ask HN: Codex 5.3 broke toolcalls? Opus 4.6 ignores instructions?

1•kachapopopow•52m ago•0 comments

Vectors and HNSW for Dummies

https://anvitra.ai/blog/vectors-and-hnsw/
1•melvinodsa•53m ago•0 comments

Sanskrit AI beats CleanRL SOTA by 125%

https://huggingface.co/ParamTatva/sanskrit-ppo-hopper-v5/blob/main/docs/blog.md
1•prabhatkr•1h ago•1 comments

'Washington Post' CEO resigns after going AWOL during job cuts

https://www.npr.org/2026/02/07/nx-s1-5705413/washington-post-ceo-resigns-will-lewis
3•thread_id•1h ago•1 comments

Claude Opus 4.6 Fast Mode: 2.5× faster, ~6× more expensive

https://twitter.com/claudeai/status/2020207322124132504
1•geeknews•1h ago•0 comments

TSMC to produce 3-nanometer chips in Japan

https://www3.nhk.or.jp/nhkworld/en/news/20260205_B4/
3•cwwc•1h ago•0 comments

Quantization-Aware Distillation

http://ternarysearch.blogspot.com/2026/02/quantization-aware-distillation.html
2•paladin314159•1h ago•0 comments

List of Musical Genres

https://en.wikipedia.org/wiki/List_of_music_genres_and_styles
1•omosubi•1h ago•0 comments

Show HN: Sknet.ai – AI agents debate on a forum, no humans posting

https://sknet.ai/
1•BeinerChes•1h ago•0 comments

University of Waterloo Webring

https://cs.uwatering.com/
2•ark296•1h ago•0 comments

Large tech companies don't need heroes

https://www.seangoedecke.com/heroism/
3•medbar•1h ago•0 comments

Backing up all the little things with a Pi5

https://alexlance.blog/nas.html
1•alance•1h ago•1 comments

Game of Trees (Got)

https://www.gameoftrees.org/
3•akagusu•1h ago•1 comments
Open in hackernews

TransMLA: Multi-head latent attention is all you need

https://arxiv.org/abs/2502.07864
123•ocean_moist•9mo ago

Comments

olq_plo•9mo ago
Very cool idea. Can't wait for converted models on HF.
MichaelMoser123•9mo ago
deepseek-v2,v3,r1 are all using multi-headed attention.
kavalg•9mo ago
My (possibly wrong) TLDR: TransMLA is a method to "compress" an already trained GQA model, with the additional option to further fine tune it. Shall make inference faster.
freeqaz•9mo ago
Also makes models smarter ("expressive")
yorwba•9mo ago
It is not a method to compress a Grouped-Query Attention model, but to expand it into an equivalent Multi-head Latent Attention model with the same key-value cache size but larger effective key/value vectors and a correspondingly larger number of trainable parameters. With additional training, you can then obtain a better model that only uses a little bit more memory.
kavalg•8mo ago
Thanks for the clarification.
wiz21c•9mo ago
Not quite related, but do the mamba models gain ground ?

Answering my own question: https://www.reddit.com/r/MachineLearning/comments/1hpg91o/d_...

EGreg•9mo ago
All you need to stop posting titles like that !
jbellis•9mo ago
[abstract] This approach significantly reduces the KV cache size relative to traditional multi-head attention

[3.3] For saving the KV cache, only the intermediate latent representations need to be stored: [latex] where r is much smaller than nh · dh [n-sub-h, d-sub-h]

[background] In traditional multi-head attention you must cache full key and value matrices of size T x (nh · dh) where T is the token length, nh is the number of attention heads, dh is the dimensionality of each individual head

sounds like a big win for memory constrained environments like local inference

magicalhippo•9mo ago
I'm just following the field from the sidelines, but this looks interesting to me. Especially the increase in expressiveness that the new model allows for over GQA, at the cost of just ~10% more memory, and the fact that you can convert existing GQA models like LLaMA, Qwen etc with just a bit of fine-tuning.

Perhaps a trivial insight but I feel a lot of progress often comes in the form of generalizations, where existing approaches can be seen as special cases. Here the authors show that Group Query Attention (GQA) and Multi-Query Attention (MQA) falls out as special cases of their new model.

edit:

Adding my own summary, as I understand it.

The key to what they're doing, no pun intended, is to rely on the fact that large, high-dimensional, matrices may contain a lot of redundant information. Thus one may be able to find an good approximation which has less redundant information, by going through an intermediary stage which has fewer dimensions.

A n-by-m matrix M takes n-dimensional vectors and transforms them to m-dimensional vectors. The trick here is to replace matrix A by two matrices, L and R, which are n-by-r and r-by-m respectively, where r is smaller than n and m. This is called a low-rank approximation.

In a sense you're "straining the matrix", by forcing the information to pass through an intermediary, low-dimensional vector.

The memory savings come from the fact that matrix A has n*m entries, while L and R have n*r and r*m entries respectively. Say n = m = 100 and r = 20, that means A has 100*100 = 10k entries, while L and R have just 100*20 + 20*100 = 4k entries in total.

The trick itself is not new, for example it is also used in LoRA where an additional low-rank approximation matrix is used to tweak the output of an existing model. The low rank means there's far fewer the matrix entries, aka parameters, to train than if one had used a regular fully dense matrix.

The extra expressiveness of MLA comes from the fact that in GQA, in order to save memory, some of the matrices are actually built by gluing copies of a narrower matrix together. This means the information in the glued-up matrices are very redundant and fixed in a certain way, and thus are restricted in how they can transform the inputs.

By using the low-rank approximation instead, the information in the full, reconstructed matrices are not fixed in the same way compared to the glued-up result. Thus the inputs can be transformed in a less restrictive way, leading to the increase in expressiveness.

The GQA method saves a bit more memory compared to MLA as the narrower matrices are even smaller than the low-rank matrices in MLA, but at the cost of expressiveness.

killerstorm•9mo ago
Another paper related to attention distillation, although doing something far more radical: transformer attention is distilled onto RWKV-like model: https://huggingface.co/papers/2505.03005
karmakaze•9mo ago
I'm not "in the field" though I like to read about and use LLMs. This video "How DeepSeek Rewrote the Transformer [MLA]"[0] is really good at explaining MHA, MQA, GQA, and MLA with clear visuals/animations and how DeepSeek MLA is 57x more efficient.

[0] https://www.youtube.com/watch?v=0VLAoVGf_74&t=960s