Hacker News Reader

5kgon Speculative KV coding: losslessly compressing KV cache by up to ~4×

Host to device bandwidth (ram to vram) is 128Gb/s for PCIe Gen 6. VRAM to GPU bandwidth is 1.8Tb/s for GDDR 7 (5090), and 8Tb/s for HBM3e (B200). So it can be faster to recompute than offload kv cache.

5kgon Racket v9.1

The Little Learner is a great book. I tried rewriting all the code in Python/JAX while following the Scheme code style as closely as possible, and it worked out great.

The appendix on autodiff is a bit rushed, in my opinion. But in all fairness, the number of pages would probably need to be doubled to give a proper exposition of autodiff.

5kgon Someone at YouTube Needs Glasses: The Prophecy Has Been Fulfilled

Best decision I made this year.

5kgon Linux 6.16: faster file systems, improved confidential memory, more Rust support

It's subsystem maintainers asking Linus to pull from their branches, e.g. https://lkml.org/lkml/2025/7/29/291

5kgon OpenAI Audio Models

May I introduce to you

https://huggingface.co/canopylabs/orpheus-3b-0.1-ft

(no affiliation)

it's English only afaics.

5kgon Apple takes UK to court over 'backdoor' order

"In 2021: No (IPT) cases were found in favour of the complainant": https://en.wikipedia.org/wiki/Investigatory_Powers_Tribunal#...

5kgon Jeff Bezos exerts more control of Washington Post opinion

you have no freedom of speech to publish opinions opposing (as decided by me) freedom of speech.

5kgon Google drops pledge not to use AI for weapons or surveillance

Google Public Sector: https://about.google/intl/ALL_us/public-sector/

5kgon AMD Ryzen 7 9800X3D Linux Performance: Zen 5 With 3D V-Cache

it's scrapped, the new design: https://www.pcworld.com/article/2507953/lunar-lakes-integrat...

5kgon IBM is quietly axing jobs, source says

It's illegal in the US: https://www.nlrb.gov/about-nlrb/rights-we-protect/the-law/em...

5kgon A good day to trie-hard: saving compute 1% at a time

There is FxHashMap (https://github.com/rust-lang/rustc-hash) which is faster than std::collections::HashMap.

With ~100 static entries, o1hash (https://github.com/rurban/smhasher/blob/master/o1hash.h) should also work.

5kgon Calculating the cost of a Google DeepMind paper

I am wondering if AlphaStar is the most expensive paper ever.

5kgon Podman Desktop 1.11: Light mode, Kubernetes features, macOS improvements

The VM disk corruption problem has been affecting all open source Apple hvf hypervisors. lima with vfkit or UTM are most promising ones for now: https://github.com/lima-vm/lima/issues/1957#issuecomment-182..., https://github.com/lima-vm/lima/pull/2026, https://github.com/utmapp/UTM/pull/5919

5kgon Podman Desktop 1.11: Light mode, Kubernetes features, macOS improvements

I liked Multipass until I ran into storage corruption issues like many others [1][2] and broke all my VMs. On MacOS, Multipass uses qemu in which the apple-hvf support isn't getting much updates after 2021 [3].

[1]: https://github.com/canonical/multipass/issues/3455 [2]: https://github.com/canonical/multipass/issues/3235 [3]: https://github.com/qemu/qemu/commits/master/accel/hvf

5kgon Arena-based parsers

Somewhat related talk: Modernizing Compiler Design for Carbon Toolchain - CppNow 2023 (https://www.youtube.com/watch?v=ZI198eFghJk)

5kgon Every map of China is wrong

(National) Security theatre https://en.m.wikipedia.org/wiki/Security_theater

Plus everyone who's involved will get richer.

5kgon Backdoor in upstream xz/liblzma leading to SSH server compromise

I think it's just a coincidence.

- All the posts are from 2004/2006. - "jiat" can be abbreviation for many common Chinese names.

5kgon Backdoor in upstream xz/liblzma leading to SSH server compromise

The backdoor (test binary blob and autoconf) is not part of the pull request.

5kgon Backdoor in upstream xz/liblzma leading to SSH server compromise

There is zero web presence for this person and associated email address.

Looks more likely a fake identity than compromised account.

5kgon Florida Passes Law Banning Social Media for Minors Under 14

Political theatre? Courts will most likely strike it down for violating the First Amendment.

5kgon Show HN: Elodin – A better framework for physics simulation

Really interesting!

Curious, why reimplementing jax in rust (nox)?

5kgon Mamba: The Easy Way

I'd love to see someone who has the resources train a model bigger than 2.8b and show the scaling law still holds.

5kgon Extreme video compression with prediction using pre-trainded diffusion models

Language Modeling Is Compression: https://arxiv.org/abs/2309.10668

5kgon LWM – Open LLM with 1M Tokens Context Window

From https://arxiv.org/pdf/2402.08268.pdf

> We trained our models using TPUv4-1024, which is approximately equivalent to 450 A100s

> Inference for such long sequences requires a minimum of v4-128

So you'll need ~60 A100 for inference.

5kgon 'Not of faculty quality': How Penn mistreated Katalin Karikó

3) Stephen Cook was denied tenure position at UCB: https://en.m.wikipedia.org/wiki/Stephen_Cook

5kgon Google's 21-year deal with Apple is the "heart" of monopoly case, judge says

I've been using Firefox as System Webview for quite some time. I think all you need is setting Firefox as default browser.

5kgon Google Cloud Spanner is now half the cost of Amazon DynamoDB

> Regarding Google Domains that is a Google product. The equivalent product from Google is “Google Cloud Domains” which is available to Google Cloud customers.

Also sold: https://cloud.google.com/domains/docs/faq

5kgon Bank Failures Visualized

Thanks. I guess rename 'r' to 'area' will make it easier to read.

5kgon Bank Failures Visualized

<del> r should be sqrt(assets). </del>

It is.