Prepare yourself for the fact that in a few years you will be dealing with real, life-size robots.

 

AI is the new big business of the future; they want to make money with it, a new market is being forced upon us. People aren't asked; you have to hit the power button on your PC or TV yourself. Why not use the term "artificial stupidity" AS instead?

 

I don't really understand why I would need to upgrade a PC when AI runs remotely and I can already run just about any softsynth out there on my current PC. What else do we need a PC to do at this stage?

 

Ah, yes. The ultimate corporate utopia. Disempowered consumers who don't own anything and actually like it. Stuck in a perpetual loop of perpetually worsening subscription terms and microtransactions.

A population that is utterly disinterested, disenfranchised, and brainwashed by AI algorithms.

We'll own nothing. We'll control nothing, not even our own fates.

We're stuck. Perfectly stuck.

Can't cancel, move or make choices. Can't seize opportunities.

Can't give the old grey matter a workout when there's endless videos of dancing AI cats waiting for us on DopeVid.

More and more stuck with every skill, thing, desire or thought we stop having, owning, developing ourselves and replace with a subscription or 3rd party.

Can't even fix the broken tap in the bathroom. We're crippled and we're doomed.

The modicum of freedom we enjoyed for a brief moment in the 20th century is gone.

Corpo's goal is slavery, and their bait is temporary convenience.
Too many people take the bait.
Corpo wins again.
Corpo always wins.
Look at the masses cheering on their own enslavement by the rich and powerful oligarchs.

Don't even have to send that many people to the camps this time.

Now, if you'll excuse me, my GobDash-AI chosen dinner is about to arrive.

Tap, tap, tap. Microtransactions.

 

they're using soy milk instead of lines of code before their presentations

 

because it's cheaper and more efficient to run AI locally, also local AI won't restrict your freedom of speech or align to political values

 

that would be ideal, but what GPU do you need and how it compares to online counterparts in results and speed? I've seen some videos but it's difficult to benchmark

 

any recent nVidia gpu,
most important is amount of VRAM, which depends solely on how big the AI model is (how many billions of parameters it has),

or, any recent Apple Silicon Mac (latest ideally ofc),
most important is amount of shared memory (again depending on model size),

huge models are rather slow, even on cloud (chatgpt, gemini etc..) it's FAR more efficient to have specialized models, and use them locally is possible then

 

or maybe since the thread title concerns AMD any recent AMD GPU with 16GB like 9070XT/9070/9060XT
or if someone want to use the very huge models (and have the money) the Strix Halo Ryzen AI MAX 395+ with its 128GB of memory is the best for local AI at low budget costs

or maybe since the thread title also concerns AMD any recent GPU AMD with 16GB like 9070XT/9070/9060XT
or if you want to use the wider models (and have the ability to spend) the Strix Halo Ryzen MAX 395+ with 128GB is the maximum for using AI locally

You have to work with smaller models and also give up some precision using lower precision models because online AI servers have a whole other power.

Of course the sequence is as always text/images/videos in order of requirements.

You don't have much trouble with text LLM unless you want to use a very large context (which means how much you keep in memory during the inference process). With images I would say that there are no problems here either, of course it always depends on what size and precision you need and how many you need to make. For videos it's a little more complex obviously to do everything locally and forget about minutes and minutes of videos

 

Last edited: Jan 9, 2026

I managed to obtain a brand new set of 2x64gb Kingston Fury DDR5 5600mt/s CL36, the very first day ram started to rise. For my personal upcoming AMD build that's been long overdue. The transaction left me not sleeping for two days because the guy selling that 128gb set was completely mental haha. Luckily the ram arrived intact on time. I paid 500 euros for it (the original receipt was for 370). The very same memory now goes for 2000 euros... This is how things are atm and all sources everywhere predict things will get gradually worse. Which means no matter if you have the money, you may not be able to obtain memory at all. I hope the latter is just an estimation and things will never get to that.
Happy new year fam! Best of luck with your builds.

 

Very similar RAM also for my two VEP slave AMD PCs, but I'm still on AM4 platform so 2 x Kingston FURY BEAST DDR4 128GB (4x32GB) 3600 CL18 so to have the maximum possible amount of ram on both. 320 + 340 euros (when the price started to go up it went up very quickly) but I still have to assemble the pcs...to find the time.

Consider that their price had hit a low around 280 euros since they were older DDR4 and not many need a 4x32GB kit, then the "disaster" we know. If I sold them now with 256GB of RAM I would buy a whole PC with top notch hardware and probably I don't need anymore two VEP slaves, but then I would miss the RAM again

 

Last edited: Jan 9, 2026

AI in drug research is very important to them to create new super drugs to addict anyone and turn them into super addicted controllable mindless zombies, when did drug research really make 100% safe drugs or ones that can't be used for addiction. As Lisa mentioned AI that is on 24 hours a day knowing everything about you even when turned off(and spying on you continuously, absolutely no more privacy or secrets) As for the robots check out the 1984 Dune prologue

when could you ever trust the powers that be. perhaps it may make some of us sound crazy, but really think about it
Frank Herbert was a visionary in 1965 when he wrote DUNE , the original 1984 Alan Smithee Extended Version movie is the best. At my age I'm glad I won't be around to see this subjugation happen.

 

Last edited: Jan 9, 2026

well yes and no,

for AI, main "bottleneck" is memory capacity and memory bandwidth,
sure recent AMD gpus have basically equal memory specs, but lack of CUDA makes them fairly tricky unless you can do bit of additional programming or optimization, it's not a bad investment but don't expect most AI stuff to "just work",
Strix Halo Ryzen AI MAX 395+ is fine, but don't forget variable graphics memory is limited to 112GB max leaving system with 16GB ram allocation which is way inferior to Apple's unified memory architecture I'm afraid,

that said, I'd very much prefer smaller specific models in conjunction with agents to avoid need for insane amounts of vram and ram, the whole push into inefficient versatility makes zero sense from user standpoint


https://www.amd.com/en/blogs/2025/f...ram-ai-model-sizes-quantization-mcp-more.html

 

Ok, I'm out

 

Just hold onto your completely functional PC and all non AI programs, do all internet free tasks, work and gaming and you'll come out ahead of the mindless, lazy and unthinking slaves to brainless simpleton tech. Those of us who can think and do everything for ourselves will be considered a threat to them and may one day find themselves hunted down, imprisoned or exterminated. You may find this statement ridiculous but in time you will see. What's next the brain implant that keeps you completely under control until deemed useless and brain switched off to die. Have any of you watched "The Tripods (1986-1985) BBC TV series, you may find it interesting, but instead replace aliens with the agenda of today by those who want to control everything and everyone.

 

Last edited: Jan 10, 2026

@Strat4ever I think we're heading more into Idiocracy + Surrogates kind of future

 

when you get to the point where you need more than 112GB for an AI model you're probably already in territory where you need to move to something AI-specific (and maybe on an online AI server) as probably there are other bottlenecks also. I was considering for local use. You always have 7 times more ram than a regular 16GB GPU. The Strix Halo architecture is very good and the memory bandwidth is more than enough (we always talk about a low-cost local system), look what happened in the comparison with the DGX.

Regarding Apple, unfortunately, I think differently. I would never trade the flexibility that the 395+ would offer me for Apple where I find myself locked in an ecosystem. I can use a 395 mini pc/notebook/tablet for everything I want from Linux to Windows and all OSes even modified that are based on these. I'm thinking of emulators for retroware, optimized windows builds, all Linux distros without having to go through an emulation but directly.
I can use it from games to AI (local) by going through every productivity software, including music DAWs .Having a couple as VEP slave servers would be really great. They should create a mini PC based on this, as they already exist now, but with more SSD slots available so you would avoid using external boxes or having to install a huge internal SSD (which would undoubtedly cost more precisely because it has a large capacity...okay now it would cost you an impossible amount anyway lol).

Regarding CUDA it was like that but now the situation is completely different. Already now you can have available practically everything. Of course, it's clear that since it's AI is a vast territory, there are so many models of different types and natures that there may be something that only works with CUDA, also because many projects are dated and either not updated or this happens very slowly, but the one you mainly use are available and works perfectly even with AMD GPUs. Now with ROCm 7.2 you have a common Linux/Windows release. If you are on linux then Canonical (Ubuntu) will soon directly maintain the ROCm package so the system will be kept installed/updated automatically. On Windows ROCm will start to be installed via the normal adrenalin driver. In short, the result is that given the enormous change that has occurred in AMD support, I would no longer worry so much about the CUDA issue,which is more and more a thing of the past as the AI sector evolves at an impressive speed. This is just my humble opinion

 

Last edited: Jan 10, 2026

Minisforum BD395i MAX motherboard appeared at CES 2026, it's a mini-ITX motherboard with full Strix Halo 395+128GB platform, but also PCIe 5.0 x16 slot - so far it's not clear if just physical or actual PCIe lanes: