If you're reading this you might be interested in getting to the bottom of it. Hop in, let's go. Click my media tab to see a pretty keen computer thing.

Joined February 2026
Replying to @pikwimin_
Hey @grok, if gender is a social construct then mustn't it be ontologically dependent on collective human practices, norms, language, institutions, and shared agreements? In other words a tranny doesn't get to pick their gender; rather, the rest of society picks it for them?
10
3
4
42
21,114
Optimized. Quarter million cubes on the gpu from a MacOS Metal4 app running under Darling. Used my phone because wf-recorder seems itself to not capture smoothly. Need to look into that. About 30fps
gm, this demo shows a Mach-O bin using MTL4 w/ original compiled Metal shaders on Darling simulating 262k cubes, mesh-generated geometry, multiple render buffers and deferred lighting. It's a bit slow b/c there's 262k indirect draws / ~3m triangles. Optimization pending.
11
Several decades ago I saw powered off systems bring down an entire datacenter. Back when ipmi was new several vendors shipped with absolutely batshit crazy defaults. These systems don't implement a full tcp/ip stack - they typically sent a periodic gratuitous ARP to advertise the management address. This is driven by the BMC which runs while the system is energized but still powered off. A new vendor had configured an extremely aggressive gratuitous ARP interval of 300ms. After enough machines were installed - plugged in but not turned on - the BMC driven traffic began to melt the network backplane. And that's one way a powered off system can cause problems. There are others. You guys aren't ready for this.
It's very straightforward to have a cve for a powered off system because many of these implement some kind of OOBM which, among other things, allows the system to be turned on over the network. Your powered off server is quite likely listening on the network.
30
This is false - not how TCP works. Yes, there are 65k ports. But they can be re-used and only the 5-tuple needs be unique. The actual limitation is 65k simultaneous connections TO THE SAME HOST ON THE SAME PORT -- which is far, far more rare.
For those who don’t get the joke, NAT devices that track flows do so via what’re called ephemeral ports, of which there are only 65,535 ports on any single IP. So if we assume all ports are use for ephemeral purposes, it means that a grand total of 65,535 CONNECTIONS (more on that later) are possible at any given moment. Now when I say connections, modern browsers will open multiple connections simultaneously to increase load speed (it’s called pipelining). So one web page could open 10-20 connections. Then, frequently with dynamic content, some of those connections are left open to load content as it flows in (email, DMs, etc). All this to say that limiting a whole country limits millions or billions of people to 65,535 of which each web page will occupy 10-20. So you’re in effect taking the internet beyond the NAT device offline, as millions have to compete for what amount to 3,000-6,000 possible slots to access the internet. Thank you for reading my ‘tism, I’m a network enginerd at the end of the day so this is my special interest.
1
1
3
400
Let's consider- how many connections can one IP create to google.com? They present 8 addresses in a DNS round robin. Let's only consider port 443. So we have: 65k * 8 * 1 - half a million simultaneous TCP sessions at max, for this country with only 1 public IP.
1
1
52
Now, a single carrier-grade NAT device (CGNAT) is going to do somewhere in the 100m-1b simultaneous connections per box. So you're going to end up with more than one IP (and thus, multiplying available ports) due to hardware limits far sooner than you'll exhaust port space.
27
If you were wondering why your computer is slow - this is why. "It's just" as if the expression syntax has anything to do with the complexity.
Replying to @eeuoss
In programming language that supports encodings it's just str[49, 10].
1
61
gm, this demo shows a Mach-O bin using MTL4 w/ original compiled Metal shaders on Darling simulating 262k cubes, mesh-generated geometry, multiple render buffers and deferred lighting. It's a bit slow b/c there's 262k indirect draws / ~3m triangles. Optimization pending.
Here's the hard thing I've been working on, at least the very first "hello world" level demo. This is a Mach-O binary compiled on MacOS (the NoGraphicsApi cube demo) with Metal4, running native under Linux/Darling - which til now did not support Metal4. cc @SebAaltonen
1
1
149
Here's the hard thing I've been working on, at least the very first "hello world" level demo. This is a Mach-O binary compiled on MacOS (the NoGraphicsApi cube demo) with Metal4, running native under Linux/Darling - which til now did not support Metal4. cc @SebAaltonen
1
1
1
378
My plan is to build a more comprehensive Metal4 compat layer in Linux/vulkan but since I started poking around I am also thinking a lot about the NoGraphicsApi layer and whether that's a useful interface when doing cross-platform operations between Mac/Linux on arm64.
31
I fundamentally don't understand this kind of question. You can just "git init" anywhere. That's the best alternative.
what is the best github alternative today for private repos
1
65
Literal sociopaths, don't encourage it
To those who "don't understand" being transphobic: As hard as you try to be nice to them, the minute you make them slightly uncomfortable, they will treat you like a deadly threat. Contrastingly, they will not care if what they do makes you uncomfortable. You are a tool to them.
5
405
Hey, with all this loong64 work to reach smooth fps - how fast is the a18 vulkan driver under Omarchy? It's a lot faster guys.
2
1
99
The a18 is expected to be a lot faster than the 7A2000, like more than ten times faster. But it's still nowhere near as fast as my 4090 which completes the entire demo before hyprland's 700ms window-zoom-in animation finishes
1
1
43
The loong64's 7A2000 chip is probably around the speed of a much older GPU. Estimates I'm seeing are around a AMD R5 230 or a GeForce 6600 - but we'll see what we can get out of it.
1
20
profiling support for loong64 - here used to generate live call stack flamegraphs of q3 in a side by side display
1
3
276