frontpage.
newsnewestaskshowjobs

Made with ♥ by @iamnishanth

Open Source @Github

fp.

"There must be something like the opposite of suicide "

https://post.substack.com/p/there-must-be-something-like-the
1•rbanffy•11s ago•0 comments

Ask HN: Why doesn't Netflix add a “Theater Mode” that recreates the worst parts?

1•amichail•59s ago•0 comments

Show HN: Engineering Perception with Combinatorial Memetics

1•alan_sass•7m ago•1 comments

Show HN: Steam Daily – A Wordle-like daily puzzle game for Steam fans

https://steamdaily.xyz
1•itshellboy•9m ago•0 comments

The Anthropic Hive Mind

https://steve-yegge.medium.com/the-anthropic-hive-mind-d01f768f3d7b
1•spenvo•9m ago•0 comments

Just Started Using AmpCode

https://intelligenttools.co/blog/ampcode-multi-agent-production
1•BojanTomic•10m ago•0 comments

LLM as an Engineer vs. a Founder?

1•dm03514•11m ago•0 comments

Crosstalk inside cells helps pathogens evade drugs, study finds

https://phys.org/news/2026-01-crosstalk-cells-pathogens-evade-drugs.html
2•PaulHoule•12m ago•0 comments

Show HN: Design system generator (mood to CSS in <1 second)

https://huesly.app
1•egeuysall•12m ago•1 comments

Show HN: 26/02/26 – 5 songs in a day

https://playingwith.variousbits.net/saturday
1•dmje•13m ago•0 comments

Toroidal Logit Bias – Reduce LLM hallucinations 40% with no fine-tuning

https://github.com/Paraxiom/topological-coherence
1•slye514•15m ago•1 comments

Top AI models fail at >96% of tasks

https://www.zdnet.com/article/ai-failed-test-on-remote-freelance-jobs/
4•codexon•15m ago•2 comments

The Science of the Perfect Second (2023)

https://harpers.org/archive/2023/04/the-science-of-the-perfect-second/
1•NaOH•16m ago•0 comments

Bob Beck (OpenBSD) on why vi should stay vi (2006)

https://marc.info/?l=openbsd-misc&m=115820462402673&w=2
2•birdculture•20m ago•0 comments

Show HN: a glimpse into the future of eye tracking for multi-agent use

https://github.com/dchrty/glimpsh
1•dochrty•21m ago•0 comments

The Optima-l Situation: A deep dive into the classic humanist sans-serif

https://micahblachman.beehiiv.com/p/the-optima-l-situation
2•subdomain•21m ago•1 comments

Barn Owls Know When to Wait

https://blog.typeobject.com/posts/2026-barn-owls-know-when-to-wait/
1•fintler•21m ago•0 comments

Implementing TCP Echo Server in Rust [video]

https://www.youtube.com/watch?v=qjOBZ_Xzuio
1•sheerluck•22m ago•0 comments

LicGen – Offline License Generator (CLI and Web UI)

1•tejavvo•25m ago•0 comments

Service Degradation in West US Region

https://azure.status.microsoft/en-gb/status?gsid=5616bb85-f380-4a04-85ed-95674eec3d87&utm_source=...
2•_____k•25m ago•0 comments

The Janitor on Mars

https://www.newyorker.com/magazine/1998/10/26/the-janitor-on-mars
1•evo_9•27m ago•0 comments

Bringing Polars to .NET

https://github.com/ErrorLSC/Polars.NET
3•CurtHagenlocher•29m ago•0 comments

Adventures in Guix Packaging

https://nemin.hu/guix-packaging.html
1•todsacerdoti•30m ago•0 comments

Show HN: We had 20 Claude terminals open, so we built Orcha

1•buildingwdavid•30m ago•0 comments

Your Best Thinking Is Wasted on the Wrong Decisions

https://www.iankduncan.com/engineering/2026-02-07-your-best-thinking-is-wasted-on-the-wrong-decis...
1•iand675•30m ago•0 comments

Warcraftcn/UI – UI component library inspired by classic Warcraft III aesthetics

https://www.warcraftcn.com/
1•vyrotek•31m ago•0 comments

Trump Vodka Becomes Available for Pre-Orders

https://www.forbes.com/sites/kirkogunrinde/2025/12/01/trump-vodka-becomes-available-for-pre-order...
1•stopbulying•32m ago•0 comments

Velocity of Money

https://en.wikipedia.org/wiki/Velocity_of_money
1•gurjeet•35m ago•0 comments

Stop building automations. Start running your business

https://www.fluxtopus.com/automate-your-business
1•valboa•39m ago•1 comments

You can't QA your way to the frontier

https://www.scorecard.io/blog/you-cant-qa-your-way-to-the-frontier
1•gk1•40m 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•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