Marko Karppinen

Are the productive conversations in the room with us right now

Imagine how fun building a PC would be if there were NO FUCKING CABLES INVOLVED

code optimization is an ecological imperative

I integrated an unbound instance as the DNS resolver of my homegrown CGNAT gateway yesterday. Hooked it up to the main (userland) routing process with a pair of veth devices so that all of the DNS-related networking went out the same path as the rest of the customer IP traffic and never touched the host network stack.

This worked great, except… only with some authoritative servers! The root servers worked fine, as did .com TLD, example.com resolved great… but things like google.com didn’t.

No bracket? No problem!

One of the things you learn when working on your own routing stack is how resilient TCP/IP really is. Your code can be just completely fucked and things still mostly work.

RE: @asymco@mastodon.social

One of these is not like the others

Slightly miffed about how Minisforum decided to spend their PCIe lanes on the N5 Pro.

It has two USB4 ports *and* an Oculink port *and* a PCIe 4.0 x4 slot. So that's where all the lanes went.

It also has three M.2 slots. Two of those are x1 and one is x2.

Maybe I'm a unique snowflake but NVMes are what I would've used the lanes on.

First proper end-to-end tests of my artisanal CGNAT router done. Did not _expect_ to nail it on the first try but it's hard to avoid the success when I’m this awesome

Anyone here use Valokuitunen / Telia Avoin Kuitu for home internet?

Hypothetically speaking what would be a cool name for an isp

you will definitely not regret writing your own router from scratch

RE: @tomshardware@mastodon.online

An all new version of the noisy neighbour problem just dropped: on shared infrastructure, your neighbour might attract ballistic missiles

Interesting to see what comes of Taalas. It seems like a completely realistic path to 1000x efficiency in LLMs.

I've argued before that the AI boom requires a specific goldilocks level of efficiency improvement to avoid a crash: too small and AI firms can't break even, too big and the bottom falls off of datacenter investments.

In that framing, 1000x might be too good? Nvidia hitmen will finish them off?

Looks like our deer problem is about to get solved