"The left wants _____" says the person on the right. Weird how the left never quite said that thing. Why not just directly quote??

I thought up a kinda work-around-y way to do this. Commit the docs. Then next commit, remove them. So that "Commit Docs" directory works and no name collision worries. Well, sort of. Same name = change. But that's probably fine. I don't love it. Feels hacky. But I don't hate it

Git commits should include files just for themselves. For example, suppose I commit with documentation of the performance testing I did to justify this change. I guess I could just make a "commit docs" folder but that sounds like a mess. I don't want to worry about file names.

I'm writing C++ for a pre-C++11 compiler. It is wild how often I go "maaaan I miss auto". I also ended up writing my own variant & expected. Anyway, it is a clear sign the language is better now.

TIL C/C++ initialization can reference the variable itself and use it for offsets. struct serialized_tree_node{ serialized_tree_node* left; serialized_tree_node* right; }; serialized_tree_node foo[] = { {foo+1, foo+2}, {NULL, NULL}, {NULL, NULL}, };

*sigh* I'm forced to use inline assembly on old compilers that don't support the intrinsics I need. But the optimizer has to get real dumb around inline assembly.

I haven't really probed myself on why, but I haven't found an IDE that I really, truly love.

I need a vibe check. You have probably seen webpages that load in parts--placeholder blank content until it gets filled. If you went to an online store and the page loaded quickly but you noticed the price was blank for a split second, would that feel fishy to you?

I have a list of strings. I'm given another list of strings. I want to find my strings within the list I'm given. There has to be a clever way to do this, right? Not just nested loops. Maybe I build a trie of my strings?

It is 2026 and companies that are big enough to do better still crash their homepage. It was eBay just a month or so ago.

I finally found a method that works on Win9x. Walk the stack until you find an EIP >= 2GiB. System DLLs are loaded in that range on Win9x. User stuff is below. That will be a page in Kernel32. Then go back pages until a one starts with "MZ".

Disclosing AI usage in a project seems tricky. For example, did the AI create all the code & art and no human reviewed it? That feels like it should be disclosed. But if a human reviewed it, I'm less scared. Or if the AI only generated ideas, and the human created...0% scared.

You see this guy? Ranked #50 in the world (front page of results) on this GPU overclocking benchmark. Look at his GPU temps. He was using liquid nitrogen. Not any more! Not on my watch!

I seen it! Sephiroth DID do this!

This is a long shot, but... Anyone able to connect me with a person on the Windows file I/O team?

Now THAT is interesting. Why does my debugger know about this magic __IMPORT_DESCRIPTOR_KERNEL32 value?? Where does it come from? Is Kernel32 always loaded into the same location??

I'm working on PNG compression and found an interest (but maybe not) thing. For my test corpus (icons & logos, not photos), the lossless compression is actually smaller than the lossy JXL compression.

Aarrgg grumble grumble, VC++ 6.0 doesn't support class partial template specialization. I forgot how workaround-y our code was back then.

What is the correct default for image compression effort? I figure: - Start at the knee on the effort/reward curve - Take up at least as much time as it took for the invocation (script? fast. human typed? give it a second) - Let it churn in the background if from an editor

I gathered up a corpus of images that lean more towards icon / logo than photo. This is PNG's wheelhouse. I then compressed them using into JXL using a bunch of settings. I'm surprised how much smaller the JXLs are as lossless over default. JXL's lossless is more or less PNG.

That's...unexpected.

I've spent a few days stuck on something. A Windows program is loaded, FS/GS points to the TEB, which points to the PEB, which points to the Ldr. My Ldr pointer is null. And I don't think it should be. I think Kernel32 is always loaded. In fact, it is in my imports table.

I'm so curious what a software developer interview looks like today, with AI becoming so common. Perhaps it is the same. I would guess it is more senior/leader focused, since we lead the AI. But maybe it is more review focused??

This might be a question for @furan.bsky.social . Any idea/guesses why graphics hardware has so few overlays? I am guessing it happens in the scanout phase and must be lightning fast. But is that a limiting factor on overlay count?? Desktop GPUs might have only 3 overlays.

I wonder if this is the result of engineer pressure to consume tokens. AI is helpful. You *can* be efficient with the tokens. Hell, you can even use local models if you setup a good harness. This is because of managers, isn't it? Measure -> Goal fortune.com/2026/05/26/u...

Okay sooooo fun fact: Kernel32, GDI32, etc alphabetize the ordinals, which shifts them around as new functions are added to new versions of Windows. This is presumably so you can binary search the names. However, Lz32.dll does not follow this. Or at least, didn't in Win98.

Internet cut out this morning due to a storm. The complete lack of distractions was eye opening. It was refreshing. And I was able to pull up docs from a Win98 VM I had previously setup. Something about that atmosphere helped, too.

It is 4:30am, release day for Subnautica 2. I couldn't sleep. Too excited, I guess. Release is at 11am. *twiddles thumbs*

Once I became a bit wise about my token usage via session management, I struggled to hit the usage limits again. The limiting factor is my ability / speed at reviewing. The classic next step is "apply fast technique to bottleneck" but I do not trust the AI to review itself. Thoughts?

My high school programming teacher stressed the importance of ethics. At the time, I thought surely a programmer would be smart enough to know right from wrong. I was so wrong. Thank you, caring teacher.

What is going on with Anthropic limits?? It is Monday, 1:15pm. I asked AI to do one task on my code--one prompt. And it used 48% of my session token limit??

Only 2 more weeks of Karazhan. I don't think I'm going to get orange in every slot. Close, though :D Proud of thank rank 298 on Netherspite. Also, how the heck are rogues supposed to parse on Nightbane?!

I've been holding out on AI. I tried it a handful of times. It underwhelmed then. Plus, AI has SOOO many ethical problems. Despite that, I've been saying for a while now that it seems like an inevitability. I decided to try to commit. Gotta say...it's pretty good now.

I believe Google is strongly encouraging its devs to use AI. I haven't contributed to Chrome in a while. I figure I should. But I do not have the wild AI budget they do. I am paying for my tokens out of my pocket. I wish Google would sponsor contributors w/ hardware & tokens

14 days until @subnautica.bsky.social 2! Plenty of time to beat Subnautica & Below Zero again.

This is Google Maps showing the main Google HQ. If you drop into street view...that definitely isn't Google HQ. Maps lets you report the mistake. I have. But it was too funny so I had to record it, too. (Also, wow. This place is beautiful. Where is it?!)

Is the typical solution for JWT revoking to get a subtoken? I feel like it should be, but I don't see people talk about it. Have a long-lived main token that authenticate but doesn't authorize. Use it to get a short-lived authorizing token. Propagate (rare) revokes between servers that authorize.

Oof. I'm working on my own deflate, comparing it to zlib with correctness & performance tests. zlib was 4x faster. Fair enough. This is my naive first attempt. So I crack open the profiler and... top 2 hits are std::sort and vector's realloc github.com/ProgramMax/m...

Published in 1995. Yikes.

Huh. eBay is down

So uhhhh who had #$AMD and #$INTC calls open? Congrats.

FINALLY. Got the hard case (dynamic blocks) working in my deflate compression. Now to fix some edge cases & clean up the code. Then onto performance.

Random thought: - Coders don't get a lot of "your work is amazing". - Mostly because 1.) not many people know coding, and 2.) reviewing code is work. I wonder if we can make it more digestible.

Hey cool. My buddy got #1 global AI demo. Grats, Jack!

I need the wisdom of C++ nerds, (summoning @hanicka.net ) A template parameter can be a function pointer type but not a function itself, right? So the badge pattern cannot be used on free functions, right? awesomekling.github.io/Serenity-C++...

OOoooooh my god. Windows' Kernel32 ordinals change between versions. Whyyyy?!?!

Been playing with VC6.0 recently. Soooo much easier to get ICEs. We've come a long way, baby.

Hey @xylafoxlin.bsky.social , you might enjoy today's maptap.gg Usually, it shows you the 5 locations as pings and has a blurb about each. But today's has a lot of pings. It was a clue that something was up... Anyway, read today's blurbs. It's up your alley.

TIL calling GetProcAddress() for kernel32 using ordinals deliberately fails. It was an obfuscation mechanism. Secret functions were originally made visible by exported name. Those names were removed, leaving legit ones. You could still use the ordinals though. So Microsoft blocks that.

I keep getting this urge for a more simple time in computing. I think I might setup an old Win98 box next to my main computer for a while. Maybe try to do my work on that machine.