Just wanted to give you a quick rundown of some cool stuff from Hacker News today, Wednesday. Lotta AI news, as usual, but some other neat things too.

Gemma 4 26B on Your Mac with 2GB RAM? Kinda.

First up, there's this wild "Show HN" project. Someone made an open-source engine that claims to run Google's big AI model, Gemma 4 26B, on any M-series Mac using only 2GB of RAM. Sounds crazy, right? The comments were quick to point out that it's a bit of a trick – the full model is way bigger, like 100GB+, so it's doing a ton of disk swapping. But still, getting it to *run* locally with so little active RAM is pretty clever. Folks were even sharing how to get it to build on macOS 15.

KOReader: E-Reader Love

Then, there's KOReader. You know, that open-source e-reader software? People are absolutely loving it, especially on smaller e-ink devices like the Xiaomi InkPalm 5. They're saying it makes them read way more. One interesting catch: it apparently uses more battery than the built-in software on devices like Kobo. Oh, and get this, one guy actually fed his KOReader logs to Claude (the AI) to debug a weird plugin issue. How wild is that?

Superlogical: Terminal on Steroids

Next, this tool called Superlogical popped up. It's basically a souped-up terminal multiplexer, kinda like tmux but it keeps track of everything you do, building a 'provenance graph' for your shell commands. So, if your session crashes or you disconnect, everything's still there, and it knows how stuff relates. People were comparing it to tmux and other notification tools, saying it's great for remote dev work and keeping track of complex tasks.

Vision Pro for Your Dream House

Remember the Apple Vision Pro? Someone found what they're calling 'the coolest use' for it: designing your house or doing renovations. You can walk through a virtual model of your future home, see how additions look, or even plan out furniture placement. Apparently, people have been doing similar stuff with cheaper VR headsets for years, like the Oculus Quest, using apps like LiveHome3D. But still, pretty neat way to visualize big projects.

Kimi K3-256k: Context Window Monster (and VRAM hog)

Big AI news too: Kimi K3-256k, a new AI model, came out with a massive 256,000 token context window. That's huge for feeding it tons of info. But the comments were kinda split. Some said while the context is big, it seems to use a lot more tokens internally, so the 'price per token' might not be as cheap as it looks. And the really wild part? To self-host it, you'd need like 1.5 terabytes of VRAM! Someone managed to compress it to 570GB with 75% accuracy, which is still insane.

AI Startups Going Secret Squirrel

Also, there was this article from Science magazine about how AI's top startups aren't publishing their research much anymore. It used to be all about open science and sharing, but now that AI is so commercial, they're keeping their secrets. People in the comments were pretty cynical about it, saying a lot of the 'research' is overhyped anyway, and it's all about profit now. Kinda sad, if you think about it.

Watch Out for AI Worms in Word!

And finally, a bit of a scary one: a security report showed how AI worms could self-propagate through