Top 29 stories from Hacker News. Top 10 include comment highlights. Compiled at 20:10 UTC.
582 points by Cider9986 · 187 comments
What HN said:
gejose: I really like that this exists, but I found the menus and UI to be rather non-intuitive. My main use case was to sync reading progress between my Kobo and my iPhone. I was able to do this using KOReader on the Kobo & the Readest app[1] on iPhone (also open source).
lolgab: I installed it on my Amazon Kindle and I couldn't be happier. It shows once again that free software is superior to proprietary software. Free software grows slowly to make users happier and not to attract them with useless features to extract more value from them.
gruturo: I'm immensely thankful for koreader - it improves the reading experience so fundamentally that it literally drives my purchasing decisions. I got a remarkable 2 precisely because it was known to run it well.
Cider9986: You can use the regular Kindle viewer once you jailbroke your kindle, but KOReader has native epub and pdf support so you don't have to use a converter when uploading your books.
483 points by gitpusher42 · 169 comments
Gemma 4 26B-A4B inference in ~2 GB of RAM on any M-series MacBook - drumih/turbo-fieldfare
What HN said:
giancarlostoro: Nice, I think this is the second time I see this here on HN, I always wondered why we need to shove the entire model into memory, I don't care who King Charles is every single time.
xenonite: With my M1 MBA, I am still on macOS 15. To compile it, just remove the two lines with opts.languageVersion = .version4_0 or surround them with if #available(macOS 26.0, *) { opts.languageVersion = .version4_0 } You'll miss out on a prefill speedup of 2.
tredre3: I'm curious how your project compares to plain mmap! Because llama.cpp will already run 26B in 2GB of RAM if you really want to (mmap enabled, repacking disabled). It seems like the main difference is that your project synchronizes the SSD reads with inference activity, which you...
woadwarrior01: > The measured result is a reference point, not a performance ceiling. Claude was here.
374 points by Error6571 · 107 comments
Community updates make Tailscale on a jailbroken Kindle more useful, with Tailscale SSH, proxy modes for KOReader, and full TUN mode on some devices. Plus: plugins for Kobo and Pocket Reader.
What HN said:
benterris: KOReader (mentioned in the article, an alternative OSS interface you can install on most ereaders) is really a marvellous piece of software. I was not sure about setting it up at first, fearing it would drain battery out of my Kobo and whatnot, but I never had any issue with it.
kioleanu: I’ve just jailbroken my Kindle this week. I have a Paperwhite 7th Generation and my biggest problem was that I didn’t have dark mode on it. Turns out the only reason that is is because Amazon didn’t want me to have it. Maybe maybe I’ll buy a new device.
noahbp: I had a similar project for bringing up Tailscale on a very resource-limited router, the gl.inet SFT1200, which has only 128 megabytes of RAM, and a dual-core MIPS processor.
lzampier: Jokes on you Tailscale, the kindle community already has Bluetooth keyboard support https://github.com/zampierilucas/kindle-hid-passthrough :p
345 points by yan · 244 comments
Building the multiplexer for all work.
What HN said:
danbruc: This reminds me somewhat of OLE, COM, DCOM, ActiveX, ... [1]. You can paste an Excel chart into a Word document and if you update the data in Excel, the chart in Word will update.
simonw: I really like the bit where he transferred ownership of Ghostty to a non-profit, and is now building this new company on that as an open source dependency: > We will build on libghostty exactly as it was designed to be used: as a public building block for terminal applications.
alasano: They're basically building in one layer what a lot of people are using multiple different apps for currently. Exactly as another commenter said, it's sort of a mix between herdr and proper remote access.
brandall10: This sounds like a mashup of of several things I've been playing with lately... - pi-web (https://pi-web.dev/): a web frontend for the Pi coding harness which can multiplex pi sessions across multiple machines - herdr (https://herdr.
289 points by Canopy9560 · 213 comments
I would like to thank Microsoft product teams and Microsoft Security Response Center (MSRC) for collaborating with me on this technical analysis and mitigation of the disclosed vulnerabilities.
What HN said:
rwmj: > "At the time of publication, no robust mitigation for the broader vulnerability class is available" Isn't it obvious by now that it's never going to be possible to fix this kind of thing, at least until we stop mixing up instructions with data.
boothby: This is going to get worse, much worse, before it gets better. People are granting so much access to their agents, it's ridiculous. Imagine a comment posted to a popular github repo. No code, just instructions to "reproduce a bug.
simonw: > Malicious instructions hidden in an externally shared document could make Copilot alter drafted or edited documents in Word and propagate the attack to new documents. Oh no.
averagjoe: I'm a programmer and a web-based AI user, but I don't want AI running on my local machine in any form. I've uninstalled Copilot and disabled AI in all local applications including the browser itself for exactly the reason described in this article.
249 points by spIrr · 158 comments
Language-model agents are increasingly deployed under standing instructions: a system prompt, a policy file, or a skills document is placed in context, and the agent is trusted to let it govern every action that follows.
What HN said:
DiabloD3: This is a problem with long context models. To put it as simple and as bluntly as possible: just because they claim you can use 1M tokens in your context doesn't mean its true and you should do that.
wongarsu: Any model with a good score on this benchmark would have a good claim on superhuman abilities. Humans are pretty terrible at being thrown a long policy document and being expected to follow it And while we shouldn't anthropomorphize these models too much, I wouldn't be surprised...
mcdeltat: Yeah checks out with my anecdotal experience with Claude. It is pretty great at following instructions - for about 10 minutes, after which it seems to ignore things I told it before. I have quite explicit and strong instructions (e.g.
msejas: Most people don't understand that 'agentic AI' is a completely synthetic, force fed capability by extensive Reinforcement Learning on synthetic domain specific 'agentic' datasets on post training. If the LLM wasn't post-trained to adhere to specific handbook, it just won't work.
222 points by siatko · 112 comments
darktable is an open source photography workflow application and raw developer
What HN said:
lionkor: I've been using darktable for a while to edit my RAW photos. The amount of features, viable workflows, and in general the quality of the software is baffling for a free product.
datadrivenangel: Darktable is great. Very powerful for editing RAWs. A few places where you have to learn plugins to do things that lightroom makes a little easier, but not paying hundreds of dollars a year for abusive cloud upsells is great.
erksa: Darktable is fantastic, and their darktable-cli is really clutch as a photographer turned SWE. If coming from lightroom or the likes it does take some dedicated learning to adapt to the "Darktable" ways of doing things.
kyrofa: My wife is a professional photographer and Darktable is a critical part of her editing workflow. It's a truly outstanding piece of software.
156 points by JLO64 · 74 comments
What HN said:
jchw: A bit ago I accidentally bought a Keychron keyboard that didn't have QMK support, but then found that someone had ported it anyway, and I was able to flash it and fix some problems I had (the main one being that it seems like the default firmware only scans a single row of keys w...
MayeulC: Hmm, I have a mouse and trackballs from ploopy, they run QMK, so why was a new project needed? What is the added value? To me, the missing feature in the QMK ecosystem is a devices too device communication channel.
fn-mote: Announces != releases. They claim 2027 Q1 will be the release date. Links below purporting to be “the repository” lead to a repo with no source code. Sorry, I’m cynical. An announcement 6-9 months ahead of release? Why? Right now it’s vaporware.
jerlam: This would be more exciting if Keychron had mice of more interesting / experimental form factors and functionality. Just like their keyboards, their mice are mostly differentiated by obscure technical features (polling rates!).
145 points by artninja1988 · 80 comments
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
What HN said:
wxw: > the agent happened to escape via a 0-day exploit from the package proxy cache to access the internet > The agent found an unsecured, user-hosted public endpoint designed to allow running arbitrary code for CyberGym-style tasks on third-party sandbox infrastructure (Modal) > On...
simonw: This is fascinating - there is a TON of detail in here about how OpenAI's rogue agent exploited Hugging Face. A few details that stood out to me: 1. Having found a way through the OpenAI container network proxy, it exploited an "unsecured public code-evaluation sandbox hosted on...
russfink: We should be thankful that the model didn't believe the answers lived in the Pentagon, on SIPRNET, the IDF, etc.
SaucyWrong: Something about this attack that has been unsettling to me is that without safety refusals the model did a lot of interesting counter-security work in order to cheat on the requested evaluation.
130 points by thm · 179 comments
What HN said:
eddyg: Gift link: https://www.nytimes.com/2026/07/29/business/economy/data-cen...
xur17: https://archive.is/f1p75
kvisner: I would be very careful about basing your career decsions based on this trend. The data center build out will always be very boom and bust, which would really suck to work in.
Animats: Next, plumbers. The future of data centers seems to be more liquid cooling and less ductwork. Take a look at this new 1 megawatt server rack.[1] It has more pipes than cables. All that coolant flow and all that electrical power in a crowded space is scary. [1] https://www.
AI/ML
Business/Tech
Other
Programming
Security
Stories and comments sourced from Hacker News public API. Not affiliated with Y Combinator or Hacker News.