There are many good answers here, but have you heard of the Chinese Room thought experiment?

https://en.wikipedia.org/wiki/Chinese_room

That’s basically what LLM interfaces are, albiet with a calculator + “fuzzy” dictionary inside. Your current chat is an LLM’s entire view of the world; otherwise it’s like a clone taken out of a vat into a black box, frozen in time forever.

Better LLM interfaces will inject relevant info (like the date) into the context, but ChatGPT is especially poorly made, and no one should be using it (even over alternatives) unless your work makes you do it… Much less pay for it.

If you’re still curious, I’d encourage you to try a “raw” LLM completion API in a unobfuscated interface, like mikupad:

https://github.com/lmg-anon/mikupad

It “pulls the curtain back,” so to speak. And you’ll be left wondering why everyone is acting like these tools are genuinely intelligent.

Front-ends circumvent YouTube’s algorithmic recommendations.

And that’s huge. I find the algo to be addictive and toxic. Even if you’re behind nested Tor relays or whatever, and YouTube can’t track you in any way, it deprives YT of the opportunity to short circuit your brain, and keep you scrolling.

But in reality, browser extensions are imperfect, and browser YouTube is still getting tons of information on what parts of videos you watch, maybe where you stop scrolling or pause your mouse, what resolution you stream, whatvideo you choose next and so on. Even with obfuscation, and an ephemeral profile, I bet it’s enough to fingerprint you.

If you want to circumvent this, try a browser that goes hardcore with anti fingerprinting specifically, like Cromite. These go out of the way to fake anything identifying, like (for example) returning fake core counts, screen resolutions and such. Then load that up with UBlock and more YT extensions.

I dunno about ducts for that case. Anything tailor-made is good though; you could even order that shroud in the earlier post from some “3D printer as a service” type place.

And this is what I meant about routing the CPU heat towards the upper exhaust. That would hopefully keep all the airflow going from bottom to top, yeah?

That’s not critical for exhaust, and it would really obstruct flow through the heatsink, if I’m understanding you correctly.

I’d just leave the heatsink exhaust unobstructed. The natural positive pressure and the two top fans will suck the hot air out.

It’s even more frustrating that “old” stuff (like debayering, white point and such) is often understood on a technical level, but it stops at anything new.

I guess this happens in software, too.

Does it always make more sense to duct in than out, so to speak?

Always in! You want cool, filtered outside air hitting the heatsinks, instead of hot air swirling in the case. And you want postitive pressure in the case (more fans sucking air in rather than out) to avoid dust buildup.

Here’s an example with that Lian Li you’re looking at, though there are generic 120mm ducts you can buy as well:

https://www.printables.com/model/1235597-lian-li-a3-rear-fan-duct-for-air-intake

Note that I may have been delirous earlier: it’s not trivial to duct a GPU on a MicroATX board, as there is a considerable distance from the GPU surface to the bottom grates. But if it were me, being crazy and all, I’d sure as heck try.

Toms has always been clickbait. They just copy and sensationalize other headlines they find, and they’ve frequently bent the knee for Nvidia. Other outlets used to make fun of them all the time.

It’s sad they “survived” the enshittification of the internet and people keep sharing their clickbait :(

That being said, there’s some truth here.

MTT is just a pipe dream, last I checked. But Deepseek is actively being served, in mixed FP8/FP4, on racks of Huawei accelerators.

I believe Baidu trained a model on them, too. But most training (like Deepseek’s) is still done on CUDA.

…Also, be careful equating this stuff with any kind of “consumer friendly” hardware you or I could buy. That’s less likely. The Huawei accelerators (and other local Chinese hardware experiments) are geared towards huge servers serving requests in parallel.

The HLG standard, gain maps, and honest display specifications should take care of that. But alas…

Perhaps ideological purism was never a good idea?

…Like history suggests?

Hence my radical preference: “a la carte” economic systems. Shameless capitalism, frothing communism, anarchism, authoritarian technocracy, even theocratic systems, they all excel in certain sectors and not others. Sometimes, in nutty combinations.

So why idealize one?

A “mix” has been reality for a long time, anyway, but I think that should be embraced more explicitly. Get systems where they’re good instead of shoehorning them where they’re horrific.

YMMV.

You gotta look up reviews. Both buyer track records, and focused reviews that look at what controller/NAND they use.

Many SSDs are basically the same as other brands internally. A few are wonky. It just depends.

A lot of younger artists cannot afford the subscriptions they perceive they need for those workflows you mentioned.

It often doesn’t work anyway!

Take HDR photography, as an example:

https://gregbenzphotography.com/hdr-display-photo-software/

The list of supported software is short, and doesn’t mention caveats (like Lightroom’s/Affinity’s HDR editing looking wonky if you don’t understand “paper white” and such). Popular “switch away from Photoshop” apps like DXO don’t even support HDR, or AVIF or anything.

Some cutting edge cameras support HEIFs instead of JPEGs, but only sometimes, and not well; they’ll pretty much only render right on an iPhone or Mac.

The video side is pretty greusome, too. VLog LUTs, for example are mostly to Rec709, and no workspace uses HLG by default.

I’m going a bit far here. My point is: ask many photographers about the RAW -> image conversion process, and they’ll have an intuitive grasp. Ask them about SDR vs HDR, and 98% think you’re talking about “HDR” multi-exposure stacking like it’s 2009. 1.9% will tell you HDR is “TV nonsense.”

One can navigate this software stuff if they understand the basics of HDR vs SDR, but the most basic understanding isn’t even there.

As an observer in these comments, this is a great answer. Thanks for typing it out.

It does seem like some “cuts” could be ironed out reasonably quickly, like the file naming issue or UI lag.

The Chinese AI labs are really trying to pop the bubble, too.

How?

Well lemme ask you this. What if models 80-90% as good as Claude, with weights just thrown out there for any provider (or homelab) to host, flood the market? What if they’re so dirt cheap to run, they’re almost free, and don’t even need Nvidia GPUs? What they need fewer resources to run with each update, instead of more?

…What if this already happened, and Big Tech is maddly lobbying to ban/censor them before people realize it, and that the “infinite scaling” thing is a big fat lie?

That’s the state of things.

Forgive my bluntness.

…But I feel like “artist tech debt” is now a problem.

I first started complaining about it in the photography space. Basically, no one understands colorspace, HDR formats, HEIFs/JXL, how modern sensors/encoders actually work… everyone is “stuck” in a circa-2000s “crush RAW stills to JPEG, export video to mp4” workflow, like they’re in a time loop. They’re not even using the crazy dynamic range of modern sensors and displays, and have no clue.

But the more I look around, the more I see it elsewhere. A manga artist choosing Google Drive as their host is a great example.

I dunno what to do about it, though.

Oh yeah, that’ll work! That’s a beautiful case, and not outrageously priced nor cramped, but still saves tons of space.

Further suggestions:

• Look up “weather stripping” on Amazon. It’s foam with one sticky side designed for doorways and such, but I stick it on my GPU (or case) to form a “seal” between the GPU and the outside grate, so it uses its own fans to suck in ambient temperature air (instead of recirculating it inside the case). It looks like it’ll work in the Lian Li:

• I’d suggest configuring the CPU tower so it sucks in air from the back, in a similar manner. Or the top, if you rotate it 90 degrees. There are simple ducts you can buy to aid with this.

This all may sound crazy, but you’d be shocked how quiet a PC can be with no (or extremely low RPM) case fans, no grinding pump or anything.

You could go AIO. But I never liked the leak risk of water cooling, nor the expense over excellent tower coolers.

That was my thought when I built my first PC, but in hindsight I wish I had gone SFF from the start, and spent more everywhere else.

Moving big towers is such a pain, especially when you move. ATX boards are more expensive than Micro ATX for basically no reason. Small towers fit nicer in rooms, and are (ironically) much quieter and cooler than a bunch of fans in a big tower.

And it’s not that hard to configure. Its note really any harder to replace components. My case is extreme, but box SFF cases have tons of room for a big GPU and a CPU tower.

The only caveat (IMO) would be if you want to go dual GPU in the future (very niche these days), if you plan to use a high end desktop (eg CPUs with 4+ channels of RAM, that generally require an ATX mobo), or if you just can’t find an SFF case you like (which is quite reasonable; cases are to taste).

I have a very different philosophy: spend everything on the CPU/GPU/RAM, and go SFF, to make it easy to move and mimimize power.

Hence I got everything is stuffed into a Node 202: no case fans, fits in carry on, but (even with a 420W GPU) it’s whisper quiet and cool since the CPU and GPU are ducted with a few bucks worth of weather sealing foam, and suck in ambient air:

Hence, I think you should:

• Go with a nice, but cheap micro ATX motherboard

• Go with a “box” MicroATX case that will let you duct the CPU/GPU.

• Big tower heatsink, whatever orientation lets you duct it.

• No case fans.

• Go 750W for the PSU.

• Use the money you save to get 24GB or 32GB of RAM.

I think the last point is essential, unfortunately… I know it’s painful, but 16GB is going to slow down a 16C CPU. I watch 16GB cripple work laptops with much lesser specs, every day, and the need for RAM rises in workstation-type loads when you start spawning more threads.

AMD 6000/7000 are good buys though. They’re aging well.

Even the machine learning dev/tinkerer communities hate AI.

They were making fun of Sam Altman (or calling him a sociopath) before it was cool.

Somewhat related:

” You need to understand that Sam can never be trusted. He is a sociopath. He would do anything.”

• Aaron Swartz, just before he died in 2013.

That’s only partially true.

The other half is that “almost as good” AI like GLM’s coder plan, Kimi, and dozens of other providers are dirt cheap. And getting cheaper, fast.

Technically, it’s like facilitating the shipping of those apples, but leaving the customer to ship.

Plex server->client streams don’t go through Plex’s servers themselves, but directly from server to clients. P2P. AFAIK the only exception is when something goes wrong and it falls back to a Plex-hosted server as an intermediary, which should be rare.

That’s still a pretty useful service though. Getting P2P reliable and easy isn’t trivial, and is one reason why open source projects haven’t really supplanted it yet.