SgtrAiArt

Hey everyone,

After a long stretch of editorial silence that lasted half the summer (…because I’m neither a blogger nor a journalist), I’m back to share some thoughts on the latest developments in the world of AI.

The tough season of reaping has come to an end, leaving a bittersweet taste for many, each of us caught up in our own ideas and (often positive) delusions of “personal omnipotence and digital vision.”

The shared dream of being able to create any artifact in “infinitesimal” time, perhaps by pressing a single key on the keyboard, has been significantly scaled back, grounding us in a much more objective reality.

Unfortunately, things don’t always go the way we often think they will—neither in life nor, especially, in the world of business. I can confirm this as someone who now considers himself a “young old uncle”—too old to be young, but too young to be old.

My technical background matured over 20 years of experience in the telecommunications industry in Italy, in what was the only true national startup to emerge after the monopoly of Telecom Italia. It was a series of incredibly stimulating experiences, from managing national PSTN networks to working with ATM, xDSL, optical multiplexing systems (ADM/xWDM), copper-based access multiplexers (MUX), and eventually moving on to (at the time) cutting-edge technologies like MAN Gb-Eth and LTE backhauling. In 2017, for personal reasons and “a whole lot more,” I decided to leave my beloved city of Milan and reinvent myself in the ICT world in another region—Emilia Romagna, the land of Ducati, Maserati, Ferrari, and the like.

My only lingering regret is the thought of that “destiny written for each of us.” In Italy, it’s often a destiny shaped by the hands of others—those who frequently don’t know what they’re doing, the usual loudmouths. It’s a reminder that falling from the top of the mountain always hurts, and getting back on your feet can be complicated and painful. It’s a bit like going from managing teams of professionals to working as a service desk operator for some random public administration office, dealing with CMS systems and the outbursts of often “questionable” people.

But as a divine proverb goes, “when one door closes, a bigger one opens.” For me, everything happens for a reason… always.

So, what can you even do with these “artistic” diffusion models?

Let me tell you a funny story—mine…

About three years ago, I fell head over heels for AI. I started with an annual subscription to a Chinese platform that rolled out exciting new features month after month.

I picked it for two reasons compared to the many other players already out there (even big names like MJ or similar):

• It cost ten times less.
• It explicitly stated that with an active premium+ subscription, all commercial and non-commercial rights stayed entirely with the artist.

In the months that followed, I noticed that every major consumer AI service started tweaking their license and privacy terms to their own advantage—often with vague, dubious legal jargon that wasn’t easy to decipher for anyone outside the industry.

On top of that, I realized that the number of sessions you could actually use to produce something decent was, frankly, pretty limited.

I started to suspect something was off with the business model itself—and time proved me right.

To make matters worse, the platform I was subscribed to pulled a fast one by switching models entirely, which I honestly never saw coming. These were all purely commercial moves, but they really got under my skin.

The breakneck pace of development in this industry has created a kind of frenzy—like sharks in a feeding frenzy. The race to offer free beta testing for various commercial platforms meant the owners were raking in profits while users were left swimming in their little ponds of illusions and utopian dreams.

That said, it’s true that all of this is direct experience, which is invaluable. But a lot of people ended up genuinely disappointed. Nowadays, creators—at least those with some experience—are fully aware of this, and it’s become a conscious, personal choice to keep going.

As for the companies providing these services, they’re commercial entities, and they need to recoup their investments. Unfortunately, not all of them have the cultural depth to be fully aware of the implications of their actions.

At that point, I decided it was time to upgrade my ancient desktop PC, which I’d pieced together years ago to mess around with crypto mining using only CPUs (a total flop with a prehistoric dual-socket CPU architecture and dirt-cheap Chinese parts). So, I set out to build a new machine, aiming for the best balance of power, utility, and cost—at least from my perspective.

In short, I bought all the components over a couple of months, waiting for deals (except for the processor). I waited three months to snag a graphics card—and let me tell you, that was a saga in itself. It wasn’t even the one I wanted because, in the European market, the card I had my eye on is currently only sold to companies offering pre-built professional workstations (we’re talking €4,000–€7,000). I settled for what I could get, biding my time for better days to fully leverage my PCIe 5 motherboard.

I built what I like to call a “Full Stack AMD PC Desktop” because I’m old-school and love desktops with a big monitor. That way, I can better appreciate the colors and even the tiniest details.

Here are some details:

CPU: AMD Ryzen™ 9 9900X3D

Graphics Card: AMD Radeon™ RX 7900 XTX 24GB

*Last-gen, because the new ones cap at 16GB, and the PRO models (older ones) are a bit too pricey.

Motherboard: MSI MAG X870E (ready for PCIe 5 for both the GPU and NVMe)

RAM: 96GB (2×48)

The big change for me was that to run AMD models on ROCm™, I had to use a Linux OS (Ubuntu 24 LTS), which I’d never touched before. I’ve always been a Microsoft guy, aside from a few basic Unix-like commands.

It was an electrifying experience from every angle. You’re probably wondering how I pulled it off? Thanks to AI! I managed to set up and get the machine running with the help of Grok and ChatGPT—unbelievable.

Ubuntu and native AMD support, plus cloning remote repositories, installing packages, and all the other stuff (like everything needed to run Comfy UI, for example)—there’s no way I could’ve done it with my existing experience. And yet… this is one of the many concrete proofs that these blessed artificial intelligences can actually serve a purpose.

You might be wondering if all this hassle for a PC was really worth it.

My answer is yes. AI isn’t just about diffusion models—there’s so much more beneath the surface. I’m talking about the entire software architecture that allows a user (whether in the cloud or on local hardware) to write a prompt and see something pop up on the screen.

I went with a local architecture because I wanted to dig deeper and see what’s really going on in this new world. I didn’t want to just write a couple of lines and hit a button for fun.

I believe that by “forcing myself” to tackle an unfamiliar architecture, I’m pushing myself to learn new things that might come in handy in a month or so, maybe even in my professional life. Plus, I’m just fed up with Microsoft—too expensive for what it’s worth.

What results have I achieved?

I still have a ton to learn and experiment with, but the first thing I noticed is that, when running the same prompt and model, there are no hallucinations unless I intentionally provoke them.

The outputs are consistent and don’t have noticeable flaws to the human eye, unlike when I used services provided by online platforms.

From my perspective, this confirms the hypotheses I’ve always had:

• The characteristics of a PC’s architecture have some kind of correlation with web-based services—maybe less so with big players like Leonardo or MJ. I don’t have the technical expertise to prove it, but I’ve always thought that something “runs on your personal PC” as part of the generation process. I’m not sure if it’s browser management or something else (try running tests with the same prompt and PC but different browsers). The fact is, this doesn’t seem to happen locally.
• The output from many platforms isn’t the original file (in terms of file size in KB), and that’s a problem. Sometimes, the downloadable file from web services is resized (not for all platforms), but I think it’s due to companies deliberately reducing the storage size of output files. It’s like going to a restaurant, paying $25 for a pizza, and getting one without mozzarella.
• The GUIs of models on web-based platforms don’t let you tweak all the “classic” model parameters. Most people should’ve known this by now, especially since I wrote about it months ago.
• In my opinion, it’s impossible to fully explore the technologically available features of any model.
  As a paid service, it’s clear that companies optimize for maximum results for their general user base, which is totally fair. But it often turns into a playful game—sometimes even a dangerous one if used without awareness.
• As for cloud services, which I know little to nothing about for AI, the story is similar:
  If you use API instances or similar tied to a remote infrastructure, you might run into the same issues as web-based platforms—maybe less so with APIs, but you’re still dependent on the provider. I’m not sure (just a hypothesis) if there’s any “correlation” between a local PC and a remote instance.
  If a cloud service sells you VRAM, it doesn’t mean you’re immune to backend infrastructure issues. In theory, you shouldn’t get hallucinations unless you seek them out.
  The only hypothesis I feel confident about is using the cloud to set up virtual machines with specific specs. That way, you’d have full control over the process, but I’m not sure if that business model makes sense for standard consumers.
  No matter what paid service you use, there’s no guarantee it’ll still exist in three months. So, all the know-how you build optimizing for “that platform and that model” might partially go to waste. If you have an active business, that could be a problem.

Conclusions

Embrace this historic technological moment we’re living through—it’s rare and unrepeatable. Always do what you love, listen to others’ opinions, but don’t hate anyone for thinking differently. I wish you all the best in your career, your art, your families (if you have them), and always stay curious. I’m sure that out of this chaos, new opportunities will always emerge.

A.S. – Translation by my dear imaginary friend Grokky