Built My Own AI. It Sucks. 😜

So, I’ll admit up front here. This issue is a little different than normal. A little bit nerdier than usual, perhaps? 🤓

Because, I traveled a bit down a rabbit hole this last weekend. Into the world of self-hosted AI. AI that sits on your desk and you don’t have to rent from a big corporation. And it was a learning experience, for sure.

So, the issue is going to dive directly into what I learned. How I managed to make it work, but the AI was pretty stupid. 😜 And why buying an expensive computer these days is like the modern gold bar.

This issue is a little different. But, AI is quite literally the biggest thing to happen to solo business owners since the invention of the internet. And, I’m the big digital sovereignty guy. So, being able to self-host AI is naturally something I would look into.

So, let’s do this, shall we? 😎

I Built My Own AI This Weekend. It’s Terrible.

I spent a good chunk of my weekend down a rabbit hole. Let me tell you about it, because I think where it’s heading matters more than where I actually got.

I’ve got a spare mini PC sitting in the office. It is a powerful little PC, given its size. 32GB of RAM in it. And I decided I wanted to see what it would take to run an AI model entirely in-house. No cloud. No API. Just… a computer in my office, doing AI things.

First, I installed Linux to it. My research said that Linux was better than Windows for local AI because it was more streamlined. So, I did it. But, it didn’t take long for me to realize why I don’t prefer Linux. 🤪 Too many dependencies. Too much manual work. Couldn’t get things to work. So, I went back to Windows.

Even Windows was more fiddly than I’d like, but I got there. Installed Ollama on the mini PC, which is the thing that actually runs the model. Then on my Mac I installed AnythingLLM, which gives you a nice chat interface. Pointed it across my home network at the little Windows box… and it worked.

I was chatting with an AI model running on a computer 3 feet away from me. Nothing left my office. No subscription. No terms of service. Very cool!

And also… not very good. 😂

The model I ran was Qwen 2.5, a 7B model. (“7B” means 7 billion parameters. Roughly: smaller number, dumber model, but it fits on smaller hardware.) It works. It answers you. But it’s like talking to a considerably stupider version of Claude. Every question I asked, I could hear the fan on that little PC spin up like it was trying to take off.

And that’s the thing about these downloadable AI models. They have to load into memory… and not just any memory. It needs to be memory your graphics chip (the GPU) can reach directly. As a rough rule of thumb, that 7B model wants somewhere in the neighborhood of 5GB to work with, and a 14B model roughly doubles it. The exact number moves around depending on how the model has been packaged up, but the shape of it holds: these things get needy about memory and GPU power really fast. So while this little PC had plenty of muscle for normal daily work, it was puny by AI standards.

So that’s the honest report: I built it, it runs, and it’s not remotely good enough to use for anything I actually do. And it is a nerdy experience to get it working at all.

So why am I bothering with local AI? Why even geek around with this when we can also use ChatGPT, Claude, or any of the other AI tools people talk about?

Because of the direction, not the destination.

My long-time readers know that digital sovereignty is one of my core principles. Own your own platform. Own your data. Don’t be completely dependent on an outside service for the things your business depends on.

And right now, my AI usage is completely dependent on the cloud. These big corporations have the big data centers and I don’t. They make it easy to use. And so, we use it. If pricing changes, if terms change, if they decide we’re not allowed to use certain features, then there’s no fallback. Not to mention… aren’t those big companies USING our data, logging it, and training their AI with it? Of course they are. I have all my privacy settings set correctly with Claude, but obviously a lot of data is heading up to their servers. It is just how this works.

I gladly pay for Claude. It is like having an employee on my desk. And I’m not about to go rip powerful AI like that out of my business and replace it with a tiny little 7B model sitting on my desk that can barely tie its shoes. 😜

But there’s a difference between “the cloud is fine” and “I have no other option.” I don’t like having no other option.

That’s really what this weekend was about. Not replacing anything. Just proving to myself that a path exists. And ultimately, deciding what I’m going to do about it. Truth is, you CAN have powerful AI that is fully private to you. But, you do need the hardware to pull it off.

Now here’s the part I find genuinely fascinating... (continued below…)

WordPress News & Updates

WordPress finally learned responsive text. WordPress 7.1 Beta 1 shipped with responsive styling built into core, landing for real on August 19, along with client-side media processing and the admin toolbar showing up everywhere. Mark McNeece of 365i Hosting put it perfectly: “WordPress couldn’t do responsive text for 23 years, and that one gap sold a decade of page builder licenses”. → the 7.1 Beta 1 announcement

Update your site right now. WordPress 7.0.2 is a security release patching a SQL injection issue and a REST API problem that could lead to remote code execution. Not the kind of thing to sit on. (All Concierge clients were already updated, so if you’re with me, you’re covered and there’s nothing to do. 😎) → the 7.0.2 release post

The Classic block got a stay of execution. The plan to hide it by default in 7.1 got pulled after community pushback, along with the migration tools that were coming with it. If you’ve still got old classic content sitting around, you’ve got more time… but I’d still stop pretending that day isn’t coming. → the reversal

Here’s what everybody is actually running. Somebody pulled plugin data from 80,000 agency-managed sites: Yoast SEO on 46%, Elementor on 40%, Wordfence on 30%, ACF Pro on 28%. Useful mostly as a sanity check… if your stack looks nothing like that list, fine, but it’s worth knowing what “normal” looks like. → the numbers

78% of WordPress users are on AI daily… and 41% can’t get it to work. SiteGround surveyed over 7,000 users and found near-universal AI adoption paired with people saying it takes several tries to get anything useful out of it. Which tracks with everything I’ve been telling you. Everybody’s using it, almost nobody feels good at it, and you’re not behind. → SiteGround’s report

The page builder era keeps quietly ending. Fresh install data shows the classic page builders sliding while Gutenberg-native options like Kadence pick up ground. On the SEO side it’s even starker… Rank Math is adding roughly 60,000 installs a week while All in One SEO is losing about 34,000. Worth watching if you’re sitting on either one. → the install-trend data

Selling with Stripe? Go check your version. Woo patched a flaw in the Stripe for WooCommerce extension (versions 10.6.0 through 10.8.3) where a customer with Adaptive Pricing enabled could change the amount they actually got charged. Yes, really. Patched versions are 10.6.2, 10.7.1, and 10.8.4… check Dashboard → Updates, and if you can’t update right this second, switch off Adaptive Pricing in the meantime. → the Stripe for WooCommerce plugin page

WooCommerce 11.0 arrives July 28. Product object caching switched on by default for new stores, checkout recovery emails, and reporting improvements. Nothing dramatic, but the performance work matters if your store has gotten big enough to feel slow. → the 11.0 pre-release notes

WooCommerce still owns the web… but not the big end of town. GravityKit ran the HTTP Archive numbers and found WooCommerce leading overall ecommerce share at 6.64% against Shopify’s 4.76%. Flip to the top million sites and it reverses, with Shopify ahead by nearly two to one… and Shopify is the only large platform actually growing at scale right now. Shopify is definitely giving WooCommerce a run for its money. → the full breakdown

FluentCart 1.5.4 brings native Bricks blocks. A full set of them… product grids, checkout, mini cart, product search, the customer dashboard… so Bricks users can build a whole store without leaving the builder. PayPal subscription handling also got a lot more reliable, which matters more than it sounds if you’ve ever had a first payment fail to activate the right subscription. → the 1.5.4 release notes

Your plugins are starting to talk to AI directly. WP Rocket 3.23 shipped an MCP server so you can point Claude at your sites and have it read real performance data and adjust settings. Fluent Support 2.3.0 did the same thing for your helpdesk. This is the quiet shift I keep flagging… it’s no longer about AI writing your copy, it’s about AI being handed the keys to the tools you already run. → WP Rocket’s announcement and Fluent Support’s

And finally, WordPress TV is now an Apple TV app. Thousands of free WordCamp talks and tutorials, on the big screen, for free. I make no promises about what this does to your evening. 😂 → 9to5Mac has the details

Computers Aren’t Supposed To Do This.

If I actually want to run fully private and local AI models, I need a real machine. Something like a maxed-out Mac Mini or a decked-out Mac Studio. And when I say decked-out, I mean literally maximum amount of memory you can get. This is a real business investment.

If you watch any YouTube videos about local AI, you’ll see people with juiced up PCs and VERY large (and super expensive) video cards. It is the kind of thing you usually associate with gamers. But, they’re not gaming. Those video cards are where the local AI runs – using that GPU and all of the memory on the card.

I looked at the NVIDIA DGX Spark too. Interesting box. Very expensive. And every story I read says the software side is finicky and difficult. Which kind of defeats the whole point for me. If I’m spending my evenings hunting down terminal commands, I’m not running a business, I’m running a hobby. That’s what pulls me back toward the Mac… local AI just seems to work better there, and Macs are easy.

But here’s the thing that stopped me cold.

If I order a tricked-out Mac from Apple right now, it is about a 3-month waiting period before I will receive that machine! 🤯 I have never seen that before. And this is after a big price increase recently from Apple. And, if I try to buy one somewhere else (eBay, resellers, whoever), they’ve slapped ridiculous premiums on them.

So, what’s happening here? Well, this rapid increase in AI has caused so much demand on these high-end machines with a lot of memory that it has pushed prices HIGHER. Video cards have gone UP in value. Maxed out Macs we use to laugh at because of the crazy Apple prices… are now actually worth that because so many people are buying these things.

Computers with a lot of RAM that people bought six months ago have gone up in value. That doesn’t happen. Computers don’t go UP in value?! Usually, PCs are used for a few years and become worth less than a paper weight over time. They depreciate fast. But…

What’s happening here is that AI has quietly turned raw computing power into a RESOURCE. Something scarce that people compete over and stockpile like a gold bar. People and companies are treating computing power and memory like the world thinks of OIL. Everybody building AI needs compute…. there isn’t enough of it…. and that shortage has worked its way all the way down to you and me trying to buy a desktop computer.

So what would I actually tell you to do? Nothing, mostly. 😊

I’m not telling you to go build a local AI setup. I did it and I’m telling you plainly that what I ended up with can’t do my job. If you’re running a business on AI today, pay for the good cloud tool, mind your privacy settings, and get on with your work.

But I’d pay attention to the direction. Local models get better every few months, and the hardware to run them is becoming a thing worth thinking about instead of an afterthought.

Me? I’m definitely considering the investment into an AI machine. To me, it is an investment into digital sovereignty. Truth is, the big AI companies are offering these AI plans to us far cheaper than they should be. Over time, I think prices will increase. AI is literally a geopolitical issue now. Things will change. I think cloud-based AI will get more expensive and more “out of reach” as time goes on. Being able to own and control my own AI means I am not dependent on any of that.

But I’ve stopped thinking of it as a computer purchase. It’s starting to feel more like buying a small piece of a resource. Like buying a plot of land… or an office building for my business to work in. A tricked-out Mac on my server rack here in the office would be able to run some advanced AI models that could replace the paid models for me.

So, we shall see. 😎 Ask me again in a month. 😎