kllrnohj

7 ms·

61.
by kllrnohj 3mo ago
Because the transformer architecture that enabled modern LLMs wasn't invented until 2017[1]? 1: That's the "T" in GPT fyi, even though Google is the author of the research paper that changed everything
62.
by kllrnohj 3mo ago
> Bytedance and Tencent are tech giants for sure, nonetheless they’re not Google kind of giant. $186 billion and $105 billion revenue in 2025 respectively vs. $402 billion? Yes, Google is larger, but they're all in that same ballpar
63.
by kllrnohj 3mo ago
Or they just don't actually have any compute access restrictions of significance? Chinese companies can just go use those GPUs in neighboring countries that aren't export-restricted, like Malaysia. Like ByteDance openly did: http
64.
by kllrnohj 3mo ago
> and a proper way to write parallel programs for when async would lead to all the bad stuff async leads to. you need real threads for parallelism though? And real threads scale just fine such that m:n threading isn't typically need
65.
by kllrnohj 3mo ago
That's because they are very rarely useful. This was true then and it's still true now. There's just not many workloads where it makes sense to need to rapidly launch a thread that doesn't need to do much of anything but
66.
by kllrnohj 4mo ago
> maybe tell us how much a non-Apple system that you can run that (probably similarly or faster) would cost? Ryzen AI Max 395+ with 128GB of unified memory can be found around $3-4k. But 27B isn't that large, either, especially if
67.
by kllrnohj 4mo ago
You don't need nearly that much RAM to run Qwen 3.6 27B, though. qwen3.6:27b-q4_K_M is only 17GB, for example.
68.
by kllrnohj 4mo ago
> XNA, MonoGame, Love2D, etc. are all frameworks rather than game engines so they're absolutely doing typical engine work in the scripting language, if you need a different example. And games using low end frameworks like that are a
69.
by kllrnohj 4mo ago
Unity doesn't use C# for everything. Notably the game engine itself is not C# but C++. C# is essentially just the scripting language. You're also ignoring the overhead of WASM -> WebGPU -> native graphics API, not to mention
70.
by kllrnohj 4mo ago
Half Life 2 is only ~4-6gb and was designed to run on video cards with 128-512mb of RAM, and it didn't even do texture streaming so you just need to have those assets at load time. Sure you can stream those low res textures at an accep
71.
by kllrnohj 4mo ago
While you got the Black Mesa remake confused, HL2 did get a free 20th Anniversay Update a couple years ago: https://www.half-life.com/en/halflife2/20th So it's still worth a revisit :)
72.
by kllrnohj 4mo ago
For what purpose, though? Why saddle yourself with the overhead & restrictions of WASM and the limitations of WebGL (or even WebGPU), just to run in a browser? The typical answer for running in a browser is the fast deployment, but if t
73.
by kllrnohj 4mo ago
While MegaTexture was removed from Doom Eternal because it's just not necessary in a small level-based game like that, the general idea was carried on. It's a big reason you could basically describe the entire PS4 era as the era o
74.
by kllrnohj 4mo ago
The problem with Doom 3's gameplay is it was too fucking dark and constantly having demons jump scare you or spawn behind you got stale quickly. At some point I need to give the BFG edition a try which at least addresses the "too
75.
by kllrnohj 4mo ago
Doom 3 and Half Life 2 were both quite demanding titles at the time, neither ran well on hardware of the era if you cranked up the settings & resolution. Doom 3 was definitely more compromised because of it, though, with too little ligh
76.
by kllrnohj 4mo ago
I think you might be misremembering Doom 3 a bit. Both it and Half Life 2 came out in the same year, and, well... just compare: https://www.youtube.com/watch?v=aTJ1weGimZQ and https://www.youtube.com/watch?v
77.
by kllrnohj 4mo ago
I believe all of these previews are using pre-production units. Production units should be better built with better tolerances since it becomes all factory done, but that's definitely something you'd have to wait and see when it
78.
by kllrnohj 4mo ago
Valve has since clarified that all current units will have a single 16GB stick. It might change in the future, but as of right now everything is the same. I assume that 8GB sticks just kinda don't exist anymore, though.
79.
by kllrnohj 4mo ago
> A GPU defect on the Steam Machine requires a full replacement of the proprietary motherboard. You can’t put a new GPU in or have someone do it for you. Of course you can have someone put in a new one for you, it's called an "
80.
by kllrnohj 4mo ago
Valve claimed in their internal tests it did not meaningfully change performance, which I'd be inclined to believe. The Steam Machine is likely going to be most often GPU bottle-necked, so the CPU performance regressing by even double
81.
by kllrnohj 4mo ago
The steam machine has upgradable storage and RAM, both being their respective commodity connections (m.2 & sodimm respectively). There's even 1 sodimm slot free from the factory if you want to immediately plunk in a 2nd 16gb stick.
82.
by kllrnohj 4mo ago
Steam Machine very clearly states Zen 4, so it has approximately a Ryzen 5 7 600, not 3600. So no, the 225F is not going to be significantly more powerful. > The RTX 5050 8 GB is 10-20% faster yeah... like I said, basically the same? b
83.
by kllrnohj 4mo ago
I think you need to check out what prebuilt PC prices are now. This is pretty much the same price as a DIY. Just to pick on someone, iBuyPower's cheapest "RDY" prebuilt gaming PC has 6 performance cores, 16GB RAM, 8GB VRAM RT
84.
by kllrnohj 4mo ago
Or buy this and get the exact same thing, but without building and parting it our yourself? It's still an open computer, not a locked down console. The price reflects that reality. It's not subsidized because you actually just pro
85.
by kllrnohj 4mo ago
> Yes, you kept binary compatibility but you that created so many other problems such as not being able to use value types in generics. This is often given as the defensible reason, but it's not even that true. Java 1.5/5 had s
86.
by kllrnohj 4mo ago
https://www.storagereview.com/review/ubiquiti-unas-pro-8-rev... This says the UNAS Pro 8 can saturate 20gbe with reads. It won't with just a single user, though, so for homelab enthusiasts it's less attractiv
87.
by kllrnohj 4mo ago
> There's also wrapping_add / saturating_add for anyone concerned about this. Ya want a specific behavior? Just tell the compiler! That is exactly why Rust's behavior for +/- operators is a mistake and wrong. It sho
88.
by kllrnohj 4mo ago
The stock symbol for Ubiquiti is actually UI, not UBNT. UBNT was the symbol for the old name that hasn't been used since 2019. I have no idea why changing the name also changed the stock symbol, but shrug
89.
by kllrnohj 4mo ago
-fsantize=integer and whambo bambo C/C++'s "time traveling UB" is now more consistent, better defined, and safer than Rust's release build behavior. I don't know if Rust will ever attempt to fix this mistake, b
90.
by kllrnohj 4mo ago
No, that's exactly what I'm aware of, and is exactly the wrong behavior I'm talking about. "Sometimes crashes, sometimes two's compliment" are extremely different behaviors, and not meaningfully different from

More ›