frontpage.
newsnewestaskshowjobs

Made with ♥ by @iamnishanth

Open Source @Github

fp.

I Was Trapped in Chinese Mafia Crypto Slavery [video]

https://www.youtube.com/watch?v=zOcNaWmmn0A
1•mgh2•5m ago•0 comments

U.S. CBP Reported Employee Arrests (FY2020 – FYTD)

https://www.cbp.gov/newsroom/stats/reported-employee-arrests
1•ludicrousdispla•7m ago•0 comments

Show HN: I built a free UCP checker – see if AI agents can find your store

https://ucphub.ai/ucp-store-check/
1•vladeta•12m ago•1 comments

Show HN: SVGV – A Real-Time Vector Video Format for Budget Hardware

https://github.com/thealidev/VectorVision-SVGV
1•thealidev•14m ago•0 comments

Study of 150 developers shows AI generated code no harder to maintain long term

https://www.youtube.com/watch?v=b9EbCb5A408
1•lifeisstillgood•14m ago•0 comments

Spotify now requires premium accounts for developer mode API access

https://www.neowin.net/news/spotify-now-requires-premium-accounts-for-developer-mode-api-access/
1•bundie•17m ago•0 comments

When Albert Einstein Moved to Princeton

https://twitter.com/Math_files/status/2020017485815456224
1•keepamovin•18m ago•0 comments

Agents.md as a Dark Signal

https://joshmock.com/post/2026-agents-md-as-a-dark-signal/
1•birdculture•20m ago•0 comments

System time, clocks, and their syncing in macOS

https://eclecticlight.co/2025/05/21/system-time-clocks-and-their-syncing-in-macos/
1•fanf2•22m ago•0 comments

McCLIM and 7GUIs – Part 1: The Counter

https://turtleware.eu/posts/McCLIM-and-7GUIs---Part-1-The-Counter.html
1•ramenbytes•24m ago•0 comments

So whats the next word, then? Almost-no-math intro to transformer models

https://matthias-kainer.de/blog/posts/so-whats-the-next-word-then-/
1•oesimania•26m ago•0 comments

Ed Zitron: The Hater's Guide to Microsoft

https://bsky.app/profile/edzitron.com/post/3me7ibeym2c2n
2•vintagedave•29m ago•1 comments

UK infants ill after drinking contaminated baby formula of Nestle and Danone

https://www.bbc.com/news/articles/c931rxnwn3lo
1•__natty__•29m ago•0 comments

Show HN: Android-based audio player for seniors – Homer Audio Player

https://homeraudioplayer.app
3•cinusek•30m ago•0 comments

Starter Template for Ory Kratos

https://github.com/Samuelk0nrad/docker-ory
1•samuel_0xK•31m ago•0 comments

LLMs are powerful, but enterprises are deterministic by nature

2•prateekdalal•35m ago•0 comments

Make your iPad 3 a touchscreen for your computer

https://github.com/lemonjesus/ipad-touch-screen
2•0y•40m ago•1 comments

Internationalization and Localization in the Age of Agents

https://myblog.ru/internationalization-and-localization-in-the-age-of-agents
1•xenator•40m ago•0 comments

Building a Custom Clawdbot Workflow to Automate Website Creation

https://seedance2api.org/
1•pekingzcc•43m ago•1 comments

Why the "Taiwan Dome" won't survive a Chinese attack

https://www.lowyinstitute.org/the-interpreter/why-taiwan-dome-won-t-survive-chinese-attack
2•ryan_j_naughton•43m ago•0 comments

Xkcd: Game AIs

https://xkcd.com/1002/
1•ravenical•45m ago•0 comments

Windows 11 is finally killing off legacy printer drivers in 2026

https://www.windowscentral.com/microsoft/windows-11/windows-11-finally-pulls-the-plug-on-legacy-p...
1•ValdikSS•45m ago•0 comments

From Offloading to Engagement (Study on Generative AI)

https://www.mdpi.com/2306-5729/10/11/172
1•boshomi•47m ago•1 comments

AI for People

https://justsitandgrin.im/posts/ai-for-people/
1•dive•48m ago•0 comments

Rome is studded with cannon balls (2022)

https://essenceofrome.com/rome-is-studded-with-cannon-balls
1•thomassmith65•53m ago•0 comments

8-piece tablebase development on Lichess (op1 partial)

https://lichess.org/@/Lichess/blog/op1-partial-8-piece-tablebase-available/1ptPBDpC
2•somethingp•55m ago•0 comments

US to bankroll far-right think tanks in Europe against digital laws

https://www.brusselstimes.com/1957195/us-to-fund-far-right-forces-in-europe-tbtb
4•saubeidl•56m ago•0 comments

Ask HN: Have AI companies replaced their own SaaS usage with agents?

1•tuxpenguine•59m ago•0 comments

pi-nes

https://twitter.com/thomasmustier/status/2018362041506132205
1•tosh•1h ago•0 comments

Show HN: Crew – Multi-agent orchestration tool for AI-assisted development

https://github.com/garnetliu/crew
1•gl2334•1h ago•0 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•8mo 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•8mo 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•8mo 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•8mo 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